Technical Reference Guide The Johns Hopkins ACG System

Transcription

The Johns Hopkins
ACG System
®
Technical Reference Guide
Version 9.0
December 2009
Important Warranty Limitation and Copyright Notices
Copyright 2009, The Johns Hopkins University. All rights reserved.
This document is produced by the Health Services Research & Development Center at
The Johns Hopkins University, Bloomberg School of Public Health.
The terms The Johns Hopkins ACG® System, ACG® System, ACG®, ADG®, Adjusted
Clinical Groups®, Ambulatory Care GroupsTM, Aggregated Diagnostic GroupsTM,
Ambulatory Diagnostic GroupsTM, Johns Hopkins Expanded Diagnosis ClustersTM,
EDCsTM, ACG Predictive Model , Rx-Defined Morbidity Groups , Rx-MGs , ACG
Rx GapsTM,, ACG Coordination Markers™, ACG-PM , Dx-PM , Rx-PM and DxRxPM are trademarks of The Johns Hopkins University. All materials in this document
are copyrighted by The Johns Hopkins University. It is an infringement of copyright law
to develop any derivative product based on the grouping algorithm or other information
presented in this document.
This document is provided as an information resource on measuring population morbidity
for those with expertise in risk-adjustment models. The documentation should be used
for informational purposes only. Information contained herein does not constitute
recommendation for or advice about medical treatment or business practices.
No permission is granted to redistribute this documentation. No permission is granted to
modify or otherwise create derivative works of this documentation.
Copies may be made only by the individual who requested the documentation initially
from Johns Hopkins or their agents and only for that person's use and those of his/her coworkers at the same place of employment. All such copies must include the copyright
notice above, this grant of permission and the disclaimer below must appear in all copies
made; and so long as the name of The Johns Hopkins University is not used in any
advertising or publicity pertaining to the use or distribution of this software without
specific, written prior authorization.
Disclaimer: This documentation is provided AS IS, without representation as to its
fitness for any purpose, and without warranty of any kind, either express or implied,
including without limitation the implied warranties of merchantability and fitness for a
particular purpose. The Johns Hopkins University and the Johns Hopkins Health System
shall not be liable for any damages, including special, indirect, incidental, or
consequential damages, with respect to any claim arising out of or in connection with the
use of the documentation, even if it has been or is hereafter advised of the possibility of
such damages.
Documentation Production Staff
Senior Editor: Jonathan P. Weiner, Dr. P.H.
Managing Editor: Chad Abrams, M.A.
Production assistance provided by: David Bodycombe Sc.D., Klaus Lemke, Ph.D.,
Patricio Muniz, M.D., MPH, MBA, Thomas M. Richards, Barbara Starfield, M.D., MPH
and Erica Wernery.
Special thanks to Lorne Verhulst M.D., MPA, of the British Columbia Ministry of Health
in Vancouver, Canada, for his contribution to the chapter titled Practitioner Profiling:
Assessing Individual Physician Performance Provider Performance Assessment.
Additional production assistance and original content provided by Rosina DeGiulio, Lisa
Kabasakalian, Meg McGinn, and Amy Salls of DST Health Solutions, LLC. The ACG
Team gratefully acknowledges the support provided by our corporate partner in helping
to move this publication forward.
If users have questions regarding the software and its application, they are advised to
contact the organization from which they obtained the ACG software. Questions about
grants of rights or comments, criticisms, or corrections related to this document should be
directed to the Johns Hopkins ACG team (see below). Such communication is
encouraged.
ACG Project Coordinator
624 N. Broadway - Room 607
Baltimore, MD 21205-1901 USA
Telephone (410) 955-5660
Fax: (410) 955-0470
E-mail: [email protected]
Website: http://acg.jhsph.edu
Third Party Library Acknowledgements
This product includes software developed by the following companies:
Health Plus Technologies (http://www.healthplustech.com)
Karsten Lentzsch (http://www.jgoodies.com)
Sentintel Technologies, Inc. (http://www.healthplustech.com)
This product includes software developed by The Apache Software Foundation
(http://www.apache.org)
This product includes the Java Runtime Environment developed by Sun Microsystems
(http://java.sun.com)
This product includes the following open source:
JDOM library (http://www.jdom.org)
iText library (http://www.lowagie.com/iText)
JasperReports library (http://www.jasperforge.org)
i
Table of Contents
1 Introduction ...................................................................................................... 1-i
Introduction to The Johns Hopkins ACG® System .................................... 1-1
Technical Reference Guide Objective ......................................................... 1-1
Technical Reference Guide Navigation ....................................................... 1-1
Technical Reference Guide Content ............................................................ 1-2
Installation and Usage Guide Content ........................................................ 1-3
Applications Guide Content ......................................................................... 1-4
Customer Commitment and Contact Information .................................... 1-5
2 Diagnosis and Code Sets .................................................................................. 2-i
Introduction ................................................................................................... 2-1
Coding Issues Using the International Classification of Diseases (ICD) .. 2-1
Using National Drug Codes (NDC).............................................................. 2-4
Using Anatomical Therapeutic Chemical (ATC) Codes............................ 2-6
3 Adjusted Clinical Groups (ACGs) .................................................................. 3-i
A Brief Overview ........................................................................................... 3-1
Overview of the ACG Assignment Process ................................................. 3-3
Major ADGs .................................................................................................. 3-6
Clinical Aspects of ACGs ........................................................................... 3-26
Clinically Oriented Examples of ACGs..................................................... 3-30
4 Expanded Diagnosis Clusters (EDCs) ............................................................ 4-i
Overview ........................................................................................................ 4-1
The Development of EDCs ........................................................................... 4-1
Understanding How EDCs Work ................................................................ 4-3
The Johns Hopkins ACG System, Version 9.0
ii
5 Medication Defined Morbidity ....................................................................... 5-i
Objectives ....................................................................................................... 5-1
Conclusion.................................................................................................... 5-23
6 Special Population Markers ............................................................................ 6-i
Introduction ................................................................................................... 6-1
Hospital Dominant Conditions .................................................................... 6-1
Frailty Conditions ......................................................................................... 6-3
Chronic Condition Count ............................................................................. 6-5
Condition Markers ........................................................................................ 6-9
7 Predicting Future Resource Use ..................................................................... 7-i
Objectives ....................................................................................................... 7-1
Conceptual Basis of Predictive Modeling ................................................... 7-1
Development of ACG Predictive Models .................................................. 7-13
Resource Bands ........................................................................................... 7-16
High, Moderate and Low Impact Conditions ........................................... 7-17
High Pharmacy Utilization Model ............................................................. 7-17
8 Predictive Modeling Statistical Performance ................................................ 8-i
Overview of ACG Predictive Models Statistical Performance ................. 8-1
Validation Data ............................................................................................. 8-1
Adjusted R-Square Values ........................................................................... 8-1
Sensitivity and Positive Predictive Value .................................................... 8-8
ROC Curve .................................................................................................. 8-10
Clinical Information Improves on Prior Cost .......................................... 8-11
9 Predicting Hospitalization ............................................................................... 9-i
Introduction ................................................................................................... 9-1
iii
Framework for Predicting Hospitalization Using ACG ............................ 9-1
Empiric Validation of the Likelihood of Hospitalization Model .............. 9-5
10 Coordination................................................................................................. 10-i
Introduction ................................................................................................. 10-1
Assessing Care Coordination ..................................................................... 10-1
Majority Source of Care (MSOC) ............................................................. 10-7
Unique Provider Count (UPC): ................................................................. 10-9
Specialty Count (SC) ................................................................................. 10-10
No Generalist Seen (NGS) ........................................................................ 10-11
Conclusion.................................................................................................. 10-13
11 Gaps in Pharmacy Utilization ..................................................................... 11-i
Objectives ..................................................................................................... 11-1
Significance of Pharmacy Adherence for Effective Care ........................ 11-1
Development of Possession/Adherence Markers ...................................... 11-2
Validation and Testing.............................................................................. 11-13
The Future ................................................................................................. 11-15
Appendix A: ACG Publication List ................................................................ A-1
Appendix B: Variables Necessary to Locally Calibrate the ACG
Predictive
Models ................................................................................................................. B-i
Introduction ...................................................................................................B-1
Index ............................................................................................................. IN-1
iv
This page was left blank intentionally.
Introduction
1-i
1 Introduction
Introduction to The Johns Hopkins ACG® System .................................... 1-1
Technical Reference Guide Objective ......................................................... 1-1
Technical Reference Guide Navigation ....................................................... 1-1
Technical Reference Guide Content ............................................................ 1-2
Installation and Usage Guide Content ........................................................ 1-3
Applications Guide Content ......................................................................... 1-4
Customer Commitment and Contact Information .................................... 1-5
1-ii
Introduction
Introduction
1-1
Introduction to The Johns Hopkins ACG® System
The ACG (Adjusted Clinical Groups) System was developed by faculty at the Johns
Hopkins Bloomberg School of Public Health to help make health care delivery more
efficient and more equitable. Because the ACG System can be used for numerous
management, finance, and analytical applications related to health and health care, it has
become the most widely used, population-based, case-mix/risk adjustment methodology.
Precisely because of the diversity of ACG applications, one size does not fit all in terms
of methodology. Like health management and analysis itself, using case-mix or risk
adjustment methods involves art as well as science, and these applications are particularly
context and objective driven. We hope this documentation will provide you with much of
the guidance you will need in order to apply the ACG System to most effectively meet
the risk adjustment and case-mix needs of your organization.
Technical Reference Guide Objective
The Technical Reference Guide was designed to augment the Installation and Usage
Guide that accompanies the ACG Software. The objective is to provide you with the
clinical basis and empirical performance of the ACG System.
Technical Reference Guide Navigation
Locating information in the Technical Reference Guide is facilitated by the following
search methods:
Master Table of Contents. The master table of contents contains the chapter names
and principal headings for each chapter.
Chapter Table of Contents. Each chapter has a table of contents, which lists the
principal headings and subheadings and figures and tables.
Index. Each chapter is indexed and organized alphabetically.
1-2
Introduction
Technical Reference Guide Content
For your convenience, a list of the Technical Reference Guide chapters is provided.
Chapter 1: Introduction. Provides a general overview of the physical organization
of the manual as well as content.
Chapter 2: Diagnosis and Code Sets. This chapter discusses the applicability of
diagnosis data to the field of risk assessment and the challenges of managing multiple
standards for diagnosis coding.
Chapter 3: Adjusted Clinical Groups (ACGs). This chapter provides a brief
overview of the history of the clinical origin of the ACG System and describes the
details of the ACG assignment algorithm.
Chapter 4: Expanded Diagnosis Clusters (EDCs). This chapter explains the
development and evolution of the EDC methodology.
Chapter 5: Medication Defined Morbidity. This chapter discusses the
applicability of pharmacy claims in risk assessment and defines how drug codes are
assigned to morbidity groups.
Chapter 6: Special Population Markers. This chapter discusses the definitions and
clinical criteria for the HOSDOM, Frailty, Chronic Condition Count, and Chronic
Condition Markers.
Chapter 7: Predicting Resource Use. This chapter discusses the methods and
variations of ACG Predictive Models developed to predict resource use.
Chapter 8: Predictive Modeling Statistical Performance. This chapter
demonstrates the ACG predictive models statistical performance while describing the
various ways in which they can be applied in health care applications.
Chapter 9: Predicting Hospitalization. This chapter discusses the methods and
variations of ACG Predictive Models developed to predict hospitalization risk.
Chapter 10: Coordination. This chapter discusses the methods for evaluating
coordination of care.
Chapter 11: Gaps in Pharmacy Utilization. This chapter discusses the methods
and metrics associated with medication possession and gaps in pharmacy adherence.
Appendix A: ACG Publication List
Appendix B: Variables Necessary to Locally Calibrate the ACG Predictive
Models
Introduction
1-3
Index
Installation and Usage Guide Content
For your convenience, a list of the Installation and Usage Guide chapters is provided.
Chapter 1: Getting Started. Provides a general overview of the physical
organization of the manual as well as content.
Chapter 2: Overview of the ACG Toolkit. Intended for all users, this chapter
provides a brief overview of the ACG toolkit to provide a baseline introduction to the
nomenclature of the ACG System components.
Chapter 3: Installing the ACG Software. Intended for the programmer/analyst,
this chapter discusses the technical aspects of installing the software.
Chapter 4: Basic Data Requirements. Intended for the programmer/analyst, this
chapter discusses at a high level the minimum data input requirements and other
necessary data requirements for performing ACG-based risk adjusted analyses.
Included are discussions of augmenting or supplementing diagnosis information with
optional user supplied flags as well as consideration of the use of pharmacy
information.
Chapter 5: Using the ACG Sample Data. Intended for the programmer/analyst, this
chapter walks through the example of processing the sample data provided with the
installation. This sample data is provided to allow users to understand the outputs of
the ACG System and demonstrate system functionality prior to the availability of the
user’s own input data.
Chapter 6: Using Client Data. Intended for the programmer/analyst, the purpose of
this chapter is to describe the process of importing user-supplied data files into the
system.
Chapter 7: Operating the ACG Software. Intended for the programmer/analyst,
this chapter discusses the technical aspects of using the software, and importing and
exporting data and reports.
Chapter 8: Validating Results. Intended for the programmer/analyst, the purpose of
this chapter is to provide examples of the ACG System functionality that was
designed to assist in the validation of user-supplied data.
Chapter 9: Troubleshooting. Intended for the programmer/analyst, the purpose of
this chapter is to leverage prior user experience with the software to describe the
symptom and solution to common user issues.
Appendix A: Output Data Dictionary.
1-4
Introduction
Appendix B: Report Detail.
Appendix C: Batch Mode Processing.
Appendix D: Java API.
Index
Applications Guide Content
For your convenience, a list of the Applications Guide chapters is provided.
Chapter 1: Introduction.
Chapter 2: Health Status Monitoring. This chapter demonstrates the application
of the ACG System markers and analyses to measures disease prevalence and support
health status monitoring.
Chapter 3: Performance Assessment. This chapter outlines the basic steps to
taking a population-based approach to profiling.
Chapter 4: Clinical Screening by Care and Disease Managers. This chapter
demonstrates the application of the ACG System markers to high risk case selection,
risk stratification and amenability.
Chapter 5: Managing Financial Risk for Pharmacy Benefits. This chapter
describes use of the ACG System markers to the application of managing pharmacy
risk.
Chapter 6: Capitation and Rate Setting. This chapter describes various methods of
applying the ACG System to capitation and rate setting.
Chapter 7: Final Considerations.
Index
Introduction
1-5
Customer Commitment and Contact Information
As part of our ongoing commitment to furthering the international state-of-the-art of riskadjustment methodology and supporting users of the ACG System worldwide, we will
continue to perform evaluation, research, and development. We will look forward to
sharing the results of this work with our user-base via white papers, our web site, peerreviewed articles, and in-person presentations. After you have carefully reviewed the
documentation supplied with this software release, we would welcome your inquiries on
any topic of relevance to your use of the ACG System within your organization.
Technical support is available during standard business hours by contacting your
designated account representative directly. If you do not know how to contact your
account representative, please call 866-287-9243 or e-mail [email protected].
We thank you for using the ACG System and for helping us to work toward meeting the
Johns Hopkins University’s ultimate goal of improving the quality, efficiency, and equity
of health care across the United States and around the globe.
1-6
Introduction
Diagnosis and Code Sets
2-i
2 Diagnosis and Code Sets
Introduction ................................................................................................... 2-1
Coding Issues Using the International Classification of Diseases (ICD) .. 2-1
Diagnosis Codes with Three and Four Digits ............................................ 2-2
Rule-Out, Suspected, and Provisional Diagnoses ...................................... 2-3
Special Note for ICD-10 Users .................................................................. 2-4
Using ICD-9 and ICD-10 Simultaneously ................................................. 2-4
Using National Drug Codes (NDC) .............................................................. 2-4
Table 1: Classification of Metformin ........................................................ 2-5
Using Anatomical Therapeutic Chemical (ATC) Codes............................ 2-6
Table 2: Classification of Metformin ........................................................ 2-6
2-ii
2-1
Introduction
The Johns Hopkins ACG System characterizes population health status based upon
readily available medical and pharmacy claims, patient registries or other administrative
sources. These data sources often feed financial or payment systems and were not
originally intended as a source of clinical information. When proper care is taken in
preparing and using the data in the Johns Hopkins ACG System, these administrative
sources provide a very robust input for clinical classification purposes. This chapter
discusses the challenges related to extending the use of administrative coding systems
and subsequent considerations in preparing data for use by the ACG System.
Coding Issues Using the International Classification of
Diseases (ICD)
Diagnosis codes are the primary data requirement of the Johns Hopkins ACG System.
The user must ensure, to the extent possible, the diagnosis codes recorded on the claims
encounter records and the resulting machine-readable data records are comprehensive and
consistent with the source medical records. For the purpose of assessing the quality of
diagnosis code data, a rudimentary understanding of the structure and limitations of the
International Classification of Diseases (ICD-9, ICD-9-CM, and/or ICD-10) is needed.
The two current editions of the International Classification of Diseases (ICD-9 and ICD10) are developed and maintained by the World Health Organization. In the United
States, a clinical modification of ICD-9 was prepared by the National Institutes of Health
(NIH). Known as ICD-9-CM, this system has been in use since the early 1980s and is
expected to be replaced by ICD-10-CM in 2013. ICD-10 was adopted by the WHO in
1993 and it and its various adaptations are in use by several other countries.
The ICD system was designed to serve primarily as an epidemiologic tool for tabulating
causes of mortality throughout the world. As accountability and reporting requirements in
the health care delivery and financing system have multiplied, so has the integration of
ICD diagnosis coding into claims management, medical management, and managed care
system oversight.
ICD-9-CM employs a five-digit coding scheme whereas ICD-9 uses only four digits. In
both systems, codes with as few as three digits are sometimes valid. The system is almost
entirely numeric with the exception of selected codes that begin with the letter V (Factors
Influencing Health Status) or the letter E (External Causes of Injury and Poisoning).
There are roughly 15,000 ICD-9-CM codes, but the lack of specification or agreement as
to what constitutes an invalid code renders this number an estimate.
The most obvious difference between ICD-9 and ICD-10 is the format of the codes to
include alphanumeric categories. Some chapters and conditions are organized differently
and ICD-10 has almost twice as many categories as ICD-9.
2-2
Since the ICD was originally developed to code causes of death, its underlying
assumptions lack an appreciation for the problem-oriented nature of differential diagnosis
in clinical medicine, particularly for conditions seen in primary care and other
ambulatory care settings. Many clinical problems have uncertain, or at best, tentative
diagnoses in these settings. As a result, rule-out diagnoses may be coded as definitive
diagnoses when claim forms are submitted (see the Rule-Out, Suspected, and Provisional
Diagnosis section below).
Furthermore, the use of ICD diagnosis codes by providers is inconsistent and often
confusing. Nonetheless, it is our belief (supported by evaluation of many health plan
databases) that the overwhelming majority of providers strive to report codes that
adequately characterize the condition of their members. The JHU team and other
researchers have repeatedly assessed the integrity of diagnosis codes assigned by care
providers and have found that they convey a sufficiently accurate picture of patients’
health status and resource requirements. The next sections describe some ICD coding
issues of which ACG Software users should be aware.
Diagnosis Codes with Three and Four Digits
The ICD coding scheme is structured hierarchically, with the fourth or fifth digits used to
further define or subdivide diseases or conditions that are described in general terms with
the first three digits. With the majority serving as headers for the more specific four- and
five-digit codes that follow, only a minority of three-digit ICD-9 or ICD-10 codes are
clinically valid as separately defined conditions. Therefore, these three-digit codes often
will not be accepted by payers on insurance claims.
The difficulty for the analyst is that there is no official list of valid three-digit codes.
While the Center for Medicare and Medicaid Services’ Diagnosis Related Groups (e.g.,
the CMS DRG) grouper does contain a list of valid ICD-9-CM codes, these are geared to
the inpatient setting. For ambulatory care services, the only source of information lies
with the various ICD-9-CM publications produced by the general publishing houses and
software vendors, and these differ on the specific codes they consider valid. Many of
these entities produce color-coded ICD-9 books that indicate whether a code is valid for
billing or if it requires a fourth or fifth digit. JHU encourages you to obtain one of these
books and use it to compare the results from the Non-Matched ICD-9 List produced by
the ACG Software.
Given the common use of three-digit codes, the ACG system does accept many threedigit codes and other invalid codes when their meaning is clear and their categorization is
precise enough for assignment into a single category. However, very few three digit
codes are considered for highly specialized markers such hospital dominant conditions
and frailty conditions.
2-3
Rule-Out, Suspected, and Provisional Diagnoses
One of the most frequent criticisms of the ICD system is the lack of codes that allow a
provider to stipulate that a particular diagnosis be designated as rule-out (R/O),
suspected, or provisional. Providers may record diagnoses as R/O on medical records
even though they do not strongly suspect them because certain tests, procedures or trials
of therapy are used to make a more definitive diagnosis. However, because ICD has no
rule-out code or modifier, diagnoses such as coronary artery disease, subarachnoid
hemorrhage, and hiatal hernia, just to name a few, may remain in the patient’s claim
database because they were recorded on one or more of the claim forms in the course of
the patient’s work-up.
With the exception of excluding diagnoses from lab and x-ray claims (which frequently
are rule-out or provisional in nature), the Johns Hopkins ACG Development Team does
not believe that R/O or suspected diagnoses have a dramatic effect on ACG assignment.
One reason is that in a retrospective application, R/O diagnoses still affect the
consumption of healthcare resources. For example, a patient who has R/O coronary artery
disease or R/O hiatal hernia still consumes the resources associated with the differential
diagnosis of these disorders. Although the extent of their impact is not well understood in
applications designed to predict resource consumption in the next time period, the
presence of rule-out or suspected diagnosis codes may have an effect if they appear in
large numbers or if certain providers or groups use these more than other providers or
groups. This impact is especially relevant if the ruled-out diagnoses resolve to ADGs that
the patient would not be otherwise assigned to, based upon the array of his/her other
confirmed diagnoses. For patients with multiple co-morbidities, the probability of this is
lower than for patients who are relatively healthy. While it is certainly possible for ruleout diagnoses to make healthy individuals appear sicker than they really are, this
distortion should occur for only a small subgroup of patients. To some extent, the user
can assess this by linking a count of ADGs assigned to a broad measure of resource
consumption, such as total charges, and a narrow one, such as office visits, and then
comparing the correlation between ADG counts and the two resource consumption
measures. Persons with many ADGs, low total charges, and many visits, may suggest that
rule-out diagnoses play a role in the assignment of the ADGs. When a particular health
plan or physician consistently appears to have a high morbidity mix but relatively low
resource use, it may be useful to ascertain, using medical records, if the use of R/O
diagnoses is higher in these instances. For example, this situation could occur if certain
experienced diagnosticians are referred a disproportionate share of difficult patients with
unclear symptoms.
While the only way to validate the impact of R/O diagnoses is by undertaking a complex
and expensive review of medical records, our experience suggests that ACG applications
will not be adversely impacted by a random distribution of rule-out diagnosis codes.
2-4
Special Note for ICD-10 Users
The WHO version of the ICD-10 was first incorporated into the ACG grouper in August
of 2003. Users of ICD-10 are encouraged to pay special attention to the discussion on
augmenting their pregnancy, delivery, and low birth weight information as the usefulness
of ICD-10 data for these purposes is not well established in the United States.
The ACG System uses the standard WHO version of 4 digit ICD-10 codes. This
excludes country-specific adaptations, specialty-specific adaptations and supplementary
divisions that utilize a 5th digit.
 Tip: The ACG System supports the WHO version of ICD-10. If you have a need
for a country-specific adaptation, please contact your ACG software distributor to
discuss the potential for local customization.
Using ICD-9 and ICD-10 Simultaneously
It is possible to simultaneously use both ICD-9 and ICD-10 data collected on the same
population. These codes can be processed as one data stream; however, ICD-9 data must
be stored in separate fields (or columns on the input data) from the ICD-10 data (see the
“Basic Data Requirements” chapter in the Installation and Usage Guide for more detail).
The ACG System will normalize the codes from each coding system into ACG System
markers.
Using National Drug Codes (NDC)
The National Drug Code (NDC) is a drug product classification system. First compiled
and organized as part of a Medicare outpatient drug reimbursement plan, it has grown
and spread to numerous sectors within the health care industry among which include
managed care organizations, pharmaceutical manufacturers, wholesalers, hospitals, and
Medicaid. Its usages span from clinical patient profile screening, to inventory control
and drug claims processes. Recorded within a database headed by the Food and Drug
Administration, it is used specifically by the government for product tracking,
evaluations, research, and drug approval within the United States.
2-5
The code itself is comprised of three segments. Two forms exist – a ten and an eleven
digit configuration. The ten digit code, referred to as a regulation NDC, is used mainly
by the FDA. However, the majority of government agencies and health care
organizations employ the 11 digit code format, including the Johns Hopkins ACG
System. It follows the form 5-4-2 (referring to the digit lengths of each individual subcode segment). The first segment, issued by the FDA, identifies the
labeler/manufacturer code. The next four digits – called the product code - impart
information regarding drug strength, dosage form, and formulation. The last two digits,
the package code, refer to package size and type. Together, these three number
sequences form the NDC number. With these pieces of information one can ascertain:
generic name/active ingredient; manufacturer; strength; route of administration; package
size; and, trade name, for any medication. Note that because the NDC code itself
contains specific packaging information, a large number of NDC codes are generated and
retired each month. Package sizes are not standardized across drugs or manufacturers.
This structure limits the value of monitoring of individual NDC codes. We suggest users
process all NDC codes over the period of interest.
Table 1 provides examples of NDC codes related to the ingredient Metformin. There are
54 manufacturers of Metformin with nearly 300 individual NDC codes for singleingredient doses of Metformin and an additional 350 NDC codes for combination drugs
containing Metformin.
Table 1: Classification of Metformin
NDC Code
Manufacturer
Drug Strength Package Size
00087-6060-05
BRISTOL MYERS
SQUIBB
500 mg
100
00087-6060-10
BRISTOL MYERS
SQUIBB
500 mg
500
00093-1048-05
TEVA
500 mg
PHARMACEUTICALS
500
00172-4330-80
IVAX
850 mg
PHARMACEUTICALS
1000
00185-0213-05
EON LABS
500
49884-*741-05
PAR
1000 mg
PHARMACEUTICALS
500
55289-*919-30
PDRX
1000 mg
PHARMACEUTICALS
270
500 mg
2-6
Using Anatomical Therapeutic Chemical (ATC) Codes
The World Health Organization’s (WHO) Anatomical Therapeutic Chemical (ATC)
codes may also be processed with the ACG System. In the ATC classification system, the
drugs are divided into different groups according to the organ or system on which they
act and their chemical, pharmacological and therapeutic properties. Drugs are classified
in groups at five different levels. The drugs are divided into fourteen main groups (first
level), with one pharmacological/therapeutic subgroup (second level). The third and
fourth levels are chemical/pharmacological/therapeutic subgroups and the fifth level is
the chemical substance. The second, third and fourth levels are often used to identify
pharmacological subgroups when that is considered more appropriate than therapeutic or
chemical subgroups. Table 2 provides the complete classification of metformin and code
structure.
Table 2: Classification of Metformin
The complete classification of metformin illustrates the structure of the code.
Code
Description
A
Alimentary tract and metabolism (first level, anatomical main group)
A10
Drugs used in diabetes (second level, therapeutic subgroup)
A10B
Blood glucose lowering drugs, excludes insulins (third level, pharmacological
subgroup)
A10BA
Biguanides (fourth level, chemical subgroup)
A10BA02
Metformin (fifth level, chemical substance)
The ATC system was created to serve as a tool for drug utilization research. Because the
ATC system has been specifically designed to capture the therapeutic use of the main
active ingredient, there is much more relevant information imbedded in an ATC code for
making Rx-MG assignments (reference the chapter entitled, “Medication Defined
Morbidity” in the Technical Reference Guide for a more detailed description of the RxMG assignment methodology.)
Adjusted Clinical Groups (ACGs)
3-i
3 Adjusted Clinical Groups (ACGs)
A Brief Overview ........................................................................................... 3-1
The Development of the ACGs .................................................................. 3-1
How ACGs Work ....................................................................................... 3-2
Overview of the ACG Assignment Process ................................................. 3-3
Step 1: Mapping ICD Codes to a Parsimonious Set of Aggregated
Diagnosis Groups (ADGs) ......................................................................... 3-3
Table 1: ADGs and Common ICD-9-CM Codes Assigned to Them ........ 3-4
Major ADGs .................................................................................................. 3-6
Table 2: Major ADGs for Adult and Pediatric Populations ...................... 3-6
Step 2: Creating a Manageable Number of ADG Subgroups (CADGs) ... 3-6
Table 3: The Collapsed ADG Clusters and the ADGs That They Comprise3-7
Step 3: Frequently Occurring Combinations of CADGs (MACs) ............ 3-8
Table 4: MACs and the Collapsed ADGs Assigned to Them ................... 3-8
Step 4: Forming the Terminal Groups (ACGs) ......................................... 3-9
Concurrent ACG-Weights .......................................................................... 3-9
Resource Utilization Bands (RUBs) ........................................................ 3-10
Table 5: The Final ACG Categories, Reference Concurrent Weights and
RUBs ........................................................................................................ 3-11
Figure 1: ACG Decision Tree ................................................................. 3-19
Figure 2: Decision Tree for MAC-12—Pregnant Women ...................... 3-21
Figure 3: Decision Tree for MAC-26—Infants....................................... 3-23
Figure 4: Decision Tree for MAC-24—Multiple ADG Categories ........ 3-25
Clinical Aspects of ACGs ........................................................................... 3-26
Duration .................................................................................................... 3-26
Severity..................................................................................................... 3-27
Diagnostic Certainty ................................................................................. 3-27
Etiology .................................................................................................... 3-27
Expected Need for Specialty Care ........................................................... 3-27
Table 6: Duration, Severity, Etiology, and Certainty of the Aggregated
Diagnosis Groups (ADGs) ....................................................................... 3-28
Clinically Oriented Examples of ACGs..................................................... 3-30
Chronic Illnesses ...................................................................................... 3-30
Example 1: Hypertension ........................................................................ 3-31
3-ii
Example 2: Diabetes Mellitus ................................................................. 3-32
Pregnancy ................................................................................................. 3-34
Example 3: Pregnancy/Delivery with Complications ............................. 3-34
Table 7: Clinical Classification of Pregnancy/Delivery ACGs ............... 3-35
Infants ....................................................................................................... 3-35
Example 4: Infants .................................................................................. 3-36
Table 8: Clinical Classification of Infant ACGs ..................................... 3-37
3-1
A Brief Overview
The Johns Hopkins ACG Case-Mix System (―
the ACG System‖) is a statistically valid,
diagnosis-based, case-mix methodology that allows healthcare providers, health plans,
and public-sector agencies to describe or predict a population‘s past or future healthcare
utilization and costs. The ACG System is also widely used by researchers and analysts to
compare various patient populations‘ prior health resource use, while taking into account
morbidity or illness burden.
Adjusted Clinical Group actuarial cells, or ACGs, are the building blocks of the Johns
Hopkins ACG Case-Mix System (―
the ACG System‖) methodology. ACGs are a series
of mutually exclusive, health status categories defined by morbidity, age, and sex. They
are based on the premise that the level of resources necessary for delivering appropriate
healthcare to a population is correlated with the illness burden of that population.
ACGs are used to determine the morbidity profile of patient populations to more fairly
assess provider performance, to reimburse providers based on the health needs of their
patients, and to allow for more equitable comparisons of utilization or outcomes across
two or more patient or enrollee aggregations.
The Development of the ACGs
The concept of ACG actuarial cells grew out of research by Dr. Barbara Starfield and her
colleagues in the late 1970s when they examined the relationship between morbidity or
illness burden and healthcare services utilization among children in managed care
settings. The research team theorized that the children using the most healthcare
resources were not those with a single chronic illness, but rather were those with
multiple, seemingly unrelated conditions. To test their original hypothesis, illnesses
found within pediatric HMO populations were grouped into five discrete categories:
1. Minor illnesses that are self-limited if treated appropriately, e.g., the flu, or chicken
pox.
2. Illnesses that are more severe but also time-limited if treated appropriately, e.g., a
broken leg or pneumonia.
3. Medical illnesses that are generally chronic and which remain incurable even with
medical therapy, e.g., diabetes or cystic fibrosis.
4. Illnesses resulting from structural problems that are generally not curable even with
adequate and appropriate intervention, e.g., cerebral palsy or scoliosis.
5. Psychosocial conditions, e.g., behavior problems or depression.
The Johns Hopkins team‘s findings supported the hypothesis that clustering of morbidity
is a better predictor of health services resource use than the presence of specific diseases.
3-2
This finding forms the basis of the current ACG System and remains the fundamental
concept that differentiates ACGs from other case-mix adjustment methodologies.
Since this early research more than two decades ago, the Johns Hopkins ACG Case-Mix
System has been refined and expanded based on years of extensive research and
development in collaboration with over 25 private and public health plans. The peer
reviewed research literature on the ACG System by the Johns Hopkins team and others is
quite extensive and we encourage you to study it. (See Bibliography at the end of this
guide for a complete listing of ACG-related publications.)
How ACGs Work
ACGs are a person-focused method of categorizing patients‘ illnesses. Over time, each
person develops numerous conditions. Based on the pattern of these morbidities, the
ACG approach assigns each individual to a single ACG category. Thus, an ACG captures
the specific clustering of morbidities experienced by a person over a given period of time,
such as a year.
The ACG System assigns all ICD (-9,-9-CM,-10) codes to one of 32 diagnosis clusters
known as Aggregated Diagnosis Groups, or ADGs. Individual diseases or conditions are
placed into a single ADG cluster based on five clinical dimensions:
Duration of the condition (acute, recurrent, or chronic): How long will healthcare
resources be required for the management of this condition?
Severity of the condition (e.g., minor and stable versus major and unstable): How
intensely must healthcare resources be applied to manage the condition?
Diagnostic certainty (symptoms versus documented disease): Will a diagnostic
evaluation be needed or will services for treatment be the primary focus?
Etiology of the condition (infectious, injury, or other): What types of healthcare
services will likely be used?
Specialty care involvement (e.g., medical, surgical, obstetric, hematology): To what
degree will specialty care services be required?
All diseases--even those yet to be discovered--can be classified along these dimensions
and categorized into one of these 32 ADG clusters1.
Because most management applications for population-based case-mix adjustment
systems require that patients be grouped into single, mutually exclusive categories, the
1
Please note that originally when the ACG System was termed the Ambulatory Care Group System, ADG
stood for Ambulatory Diagnostic Groups. Today, we retain the ADG acronym and call them Aggregated
Diagnosis Groups.
3-3
ACG methodology uses a branching algorithm to place people into one of 932 discrete
categories based on their assigned ADGs, their age and their sex. The result is that
individuals within a given ACG have experienced a similar pattern of morbidity and
resource consumption over the course of a given year.
ACGs can be assigned to individuals using readily available diagnostic information
derived from outpatient or ambulatory physician visit claims records, encounter records,
inpatient hospital claims, and computerized discharge abstracts. A patient/enrollee is
assigned to a single ACG based on the diagnoses assigned by all clinicians seeing them
during all contacts, regardless of setting. Thus ACGs are truly person-oriented and are
not based on visits or episodes.
Typically, ACGs perform up to ten times better than age and sex adjustment, the
traditional risk-adjustment mechanism used within the health insurance industry.
Overview of the ACG Assignment Process
The ACG System relies on automated claims or encounter data derived from healthcare
settings to characterize the degree of overall morbidity in patients and populations. ACGs
are designed to be conceptually simple, accessible, and practical to both clinicians and
managers. A significant advantage of the ACG System is its open architecture. Unlike
other case-mix, severity, and profiling systems‘ decision rules, those used for ACGs are
not housed in a ―
black box.‖ The ACG grouping logic is easily accessible to the user.
This chapter provides a detailed description of the ACG grouping logic. There is a
section in this chapter entitled, ―
Clinical Aspects of ACGs‖ that expands this discussion
by providing a clinical discussion of the ACG System.
Step 1: Mapping ICD Codes to a Parsimonious Set of Aggregated
Diagnosis Groups (ADGs)
There are roughly 25,000 ICD (-9 or -10) diagnosis codes that clinicians can use to
describe patients‘ health conditions. The first step of the ACG grouping logic is to assign
each of these codes to one of 32 diagnosis groups referred to as Aggregated Diagnosis
Groups, or ADGs. The ICD-to-ADG mapping embedded in the ACG Software includes
an ADG assignment for all3 ICD codes in the Center for Medicare and Medicaid Services
2
There are 82 default categories and 93 if all optional branchings are turned on.
3
Because they indicate the cause of injury rather than an underlying morbidity, E-codes have generally
been excluded from the ICD-9 to ADG mapping. Some E-codes representing iatrogenic conditions and
adverse effects are included in the mapping.
3-4
list of ICD-9 codes4. ICD-10 codes are sourced from the Official ICD-10 Updates
published by the World Health Organization.
Each ADG is a grouping of diagnosis codes that are similar in terms of severity and
likelihood of persistence of the health condition treated over a relevant period of time
(such as a year of HMO enrollment). ICD codes within the same ADG are similar in
terms of both clinical criteria and expected need for healthcare resources. Just as
individuals may have multiple ICD diagnosis codes, they may have multiple ADGs (up to
32). Table 1 lists the 32 ADGs and exemplary diagnosis codes.
ADGs are distinguished by several clinical characteristics (time limited or not,
medical/specialty/pregnancy, physical health/psycho-social), and degree of refinement of
the problem (diagnosis or symptom/sign). They are not categorized by organ system or
disease. Instead, they are based on clinical dimensions that help explain or predict the
need for healthcare resources over time. The need for healthcare resources is primarily
determined by the likelihood of persistence of problems and their level of severity rather
than organ system involvement. See the section entitled, ―
Clinical Aspects of ACGs,‖ in
this manual for further discussion of these clinical criteria.
Table 1: ADGs and Common ICD-9-CM Codes Assigned to Them
ADG
1.
Time Limited: Minor
2.
Time Limited: Minor-Primary
Infections
Time Limited: Major
3.
4.
5.
Time Limited: Major-Primary
Infections
Allergies
6.
Asthma
7.
Likely to Recur: Discrete
8.
Likely to Recur: Discrete-Infections
9.
Likely to Recur: Progressive
10. Chronic Medical: Stable
11. Chronic Medical: Unstable
12. Chronic Specialty: Stable-Orthopedic
4
ICD-9-CM Diagnosis Code
558.9
691.0
079.9
464.4
451.2
560.3
573.3
711.0
477.9
708.9
493.0
493.1
274.9
724.5
474.0
599.0
250.10
434.0
250.00
401.9
282.6
277.0
721.0
Noninfectious Gastroententris
Diaper or Napkin Rash
Unspecified Viral Infection
Croup
Phlebitis of Lower Extremities
Impaction of Intestine
Hepatitis, Unspecified
Pyogenic Arthritis
Allergic Rhinitis, Cause Unspecified
Unspecified Urticaria
Extrinsic Asthma
Intrinsic Asthma
Gout, Unspecified
Backache, Unspecified
Chronic Tonsillitus
Urinary Tract Infection
Adult Onset Type II Diabetes w/ Ketoacidosis
Cerebral Thrombosis
Adult-Onset Type 1 Diabetes
Essential Hypertension
Sickle-Cell Anemia
Cystic Fibrosis
Cervical Spondylosis Without Myelopathy
Available for download at http://www.cms.gov.
3-5
ADG
13. Chronic Specialty: Stable-Ear, Nose,
Throat
14. Chronic Specialty: Stable-Eye
15. No Longer in Use*
16. Chronic Specialty: UnstableOrthopedic
17. Chronic Specialty: Unstable-Ear,
Nose, Throat
18. Chronic Specialty: Unstable-Eye
19. No Longer in Use*
20. Dermatologic
21. Injuries/Adverse Effects: Minor
22. Injuries/Adverse Effects: Major
23. Psychosocial: Time Limited, Minor
24. Psychosocial: Recurrent or Persistent,
Stable
25. Psychosocial: Recurrent or Persistent,
Unstable
26. Signs/Symptoms: Minor
27. Signs/Symptoms: Uncertain
28. Signs/Symptoms: Major
29. Discretionary
30. See and Reassure
31. Prevention/Administrative
32. Malignancy
33. Pregnancy
34. Dental
ICD-9-CM Diagnosis Code
718.8
389.14
385.3
367.1
372.9
Other Joint Derangement
Central Hearing Loss
Cholesteatoma
Myopia
Unspecified Disorder of Conjunctiva
724.02
732.7
386.0
383.1
365.9
379.0
Spinal Stenosis of Lumbar Region
Osteochondritis Dissecans
Meniere's Disease
Chronic Mastoiditis
Unspecified Glaucoma
Scleritis/Episcleritis
078.1
448.1
847.0
959.1
854.0
972.1
Viral Warts
Nevus, Non-Neoplastic
Neck Sprain
Injury to Trunk
Intracranial Injury
Poisoning by Cardiotonic Glycosides and
Similar Drugs
Cannabis Abuse, Unspecified
Brief Depressive Reaction
Panic Disorder
Bulimia
Catatonic Schizophrenia
Alcohol Withdrawal Delirium Tremens
Headache
Pain in Limb
Effusion of Lower Leg Joint
Malaise and Fatigue
Cardiomegaly
Syncope and Collaspe
Inguinal Hernia (NOS)
Sebaceous Cyst
Hypertrophy of Breast
Localized Adiposity
Routine Infant or Child Health Check
Gynecological Examination
Malignant Neoplasm of Breast (NOS)
Hodgkin's Disease, Unspecified Type
Pregnant State
Delivery in a Completely Normal Case
Dental Caries
Chronic Gingivitis
305.2
309.0
300.01
307.51
295.2
291.0
784.0
729.5
719.06
780.7
429.3
780.2
550.9
706.2
611.1
278.1
V20.2
V72.3
174.9
201.9
V22.2
650.0
521.0
523.1
*Note: Only 32 of the 34 markers are currently in use.
When the user applies the lenient diagnostic certainty option, any single diagnosis
qualifying for an ADG marker will turn the marker on. However, the user may also
apply a stringent diagnostic certainty option. For a subset of diagnoses, there must be
more than one diagnosis qualifying for the marker in order for the ADG to be assigned.
3-6
This was designed to provide greater confidence in the ADGs assigned to a patient. For
more information, refer to the Installation and Usage Guide, Chapter 4.
Major ADGs
Some ADGs have very high expected resource use and are labeled ―
Major ADGs.‖ Table
2 presents major ADGs for adult and pediatric populations.
Table 2: Major ADGs for Adult and Pediatric Populations
Pediatric Major ADGs (ages 0-17 years)
3 Time Limited: Major
9 Likely to Recur: Progressive
11 Chronic Medical: Unstable
12 Chronic Specialty: Stable-Orthopedic
13 Chronic Specialty: Stable-Ear, Nose, Throat
18 Chronic Specialty: Unstable-Eye
25 Psychosocial: Recurrent or Persistent, Unstable
32 Malignancy
Adult Major ADGs (ages 18 and up)
4 Time Limited: Major-Primary Infections
16 Chronic Specialty: Unstable-Orthopedic
22 Injuries/Adverse Effects: Major
25 Psychosocial: Recurrent or Persistent,
Unstable
32 Malignancy
Step 2: Creating a Manageable Number of ADG Subgroups
(CADGs)
The ultimate goal of the ACG algorithm is to assign each person to a single morbidity
group (i.e., an ACG). There are 4.3 billion possible combinations of ADGs, so to create a
more manageable number of unique combinations of morbidity groupings, the 32 ADGs
are collapsed into 12 CADGs or Collapsed ADGs (Table 3). Like ADGs, CADGs are not
mutually exclusive in that an individual can be assigned to as few as none or as many as
12.
Although numerous analytic techniques could be used to form CADGs from ADGs, the
ACG Case-Mix System has placed the emphasis on clinical cogency. Three clinical
criteria are used:
the similarity of likelihood of persistence or recurrence of diagnoses within the
ADG, i.e., time-limited, likely to recur, or chronic groupings;
the severity of the condition, i.e., minor versus major and stable versus unstable; and
the types of healthcare services required for patient management--medical versus
specialty, eye/dental, psychosocial, prevention/administrative, and pregnancy.
ADGs and CADGs can be used for various analytic and research applications that do not
require mutually exclusive categories such as multivariate predictive or explanatory
models. See the ACG publication list in Appendix A for additional examples of these
approaches.
3-7
Table 3: The Collapsed ADG Clusters and the ADGs That They
Comprise
Collapsed ADG (CADG)
1.
Acute Minor
2.
Acute Major
3.
Likely to Recur
4.
5.
Asthma
Chronic Medical: Unstable
6.
Chronic Medical: Stable
7.
Chronic Specialty: Stable
8.
Eye/Dental
9.
Chronic Specialty: Unstable
10. Psychosocial
11. Preventive/Administrative
12. Pregnancy
ADGs in Each
1 Time Limited: Minor
2 Time Limited: Minor-Primary Infections
21 Injuries/Adverse Events: Minor
26 Signs/Symptoms: Minor
4 Time Limited: Major-Primary Infections
22 Injuries/Adverse Events: Major
27 Signs/Symptoms: Uncertain
28 Signs/Symptoms: Major
5 Allergies
7 Likely to Recur: Discrete
8 Likely to Recur: Discrete-Infections
20 Dermatologic
29 Discretionary
6 Asthma
32 Malignancy
10 Chronic Medical: Stable
30 See and Reassure
12 Chronic Specialty: Stable-Orthopedic
13 Chronic Specialty: Stable-Ear, Nose, Throat
14 Chronic Specialty: Stable-Eye
34 Dental
16 Chronic Specialty: Unstable-Orthopedic
17 Chronic Specialty: Unstable-Ear, Nose, Throat
18 Chronic Specialty: Unstable-Eye
23 Psycho-social: Time Limited, Minor
24 Psycho-social: Recurrent or Persistent, Stable
25 Psycho-social: Recurrent or Persistent, Unstable
31 Prevention/Administrative
33 Pregnancy
3-8
Step 3: Frequently Occurring Combinations of CADGs (MACs)
The third step in the ACG categorization methodology assigns individuals into a single,
mutually exclusive category, called a MAC. This grouping algorithm is based primarily
on the pattern of CADGs. Table 4, below, shows the MACs and the Collapsed ADGs
which comprise them.
There are twenty-three commonly occurring combinations of CADGs which form MACs
1 through 23:
The first 11 MACs correspond to presence of a single CADG.
MAC-12 includes all pregnant women, regardless of their pattern of CADGs.
MACs 13 through 23 are commonly occurring combinations of CADGs.
MAC-24 includes all other combinations of CADGs.
MAC-25 is used for enrollees with no service use or invalid diagnosis input data.
MAC-26 includes all infants (age<12 months), regardless of the pattern of CADGs.
Table 4: MACs and the Collapsed ADGs Assigned to Them
MACs
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
Acute: Minor
Acute: Major
Likely to Recur
Asthma
Chronic Medical: Unstable
Eye/Dental
Chronic Specialty: Unstable
Psychosocial
Prevention/Administrative
Pregnancy
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
Acute: Minor and Acute: Major
Acute: Minor and Likely to Recur
Acute: Minor and Chronic Medical: Stable
Acute: Minor and Eye/Dental
Acute: Minor and Psychosocial
Acute: Major and Likely to Recur
Acute: Minor and Acute: Major and Likely to Recur
Acute: Minor and Likely to Recur and Eye and Dental
Acute: Minor and Likely to Recur and Psychosocial
Acute: Minor and Major and Likely to Recur
and Chronic Medical: Stable
CADGs
1
2
3
4
5
6
7
8
9
10
11
All CADG combinations that include
CADG 12
1 and 2
1 and 3
1 and 6
1 and 8
1 and 10
2 and 3
1, 2 and 3
1, 3 and 8
1, 3, and 10
1, 2, 3, and 6
3-9
MACs
23. Acute: Minor and Major and Likely to Recur
and Psychosocial
24. All Other Combinations Not Listed Above
25. No Diagnosis or Only Unclassified Diagnosis
26. Infants (age less than 1 year)
CADGs
1, 2, 3, and 10
All Other Combinations
No CADGs
Any CADGs combinations and less
than 1 year old
Step 4: Forming the Terminal Groups (ACGs)
MACs form the major branches of the ACG decision tree. The final step in the grouping
algorithm divides the MAC branches into terminal groups, the actuarial cells known as
ACGs. The logic used to split MACs into ACGs includes a combination of statistical
considerations and clinical insight. During the ACG development process, the
overarching goal for ACG assignment was to identify groups of individuals with similar
needs for healthcare resources who also share similar clinical characteristics. Yale
University‘s AUTOGRP Software (which performs recursive partitioning) was used to
identify subdivisions of patients within a MAC who had similar needs for healthcare
resources based on their overall expenditures. The variables taken into consideration
included: age, sex, presence of specific ADGs, number of major ADGs, and total number
of ADGs. (Note: Because prevention/administrative needs do not reflect morbidity, we
do not include ADG31 in the count of total ADGs5).
See Table 5, below, for a complete listing and description of all ACGs.
Concurrent ACG-Weights
A concurrent ACG weight is an assessment of the relative resource use for individuals in
the ACG. The concurrent weight is calculated as the mean cost of all patients in an ACG
divided by the mean cost of all patients in the population. A fixed set of concurrent
ACG-weights derived from external reference data is available as part of the software
output file. Separate sets of weights exist for under age 65 working age populations and
for over 65 Medicare eligible populations and determined by the Risk Assessment
Variables selected during processing. The weights produced by the software are relative
weights, i.e., relative to a population mean, and are standardized to a mean of 1.0. An
individual weight is associated with each ACG. In the case of an elderly reference set,
the weights of some ACGs (e.g., those associated with pediatrics, pregnancy or
newborns) may be extrapolated from an under 65 population. The software-supplied
weights may be considered a national reference or benchmark for comparisons with
5
Readers are referred to Weiner (91) and Starfield (91) for more detail on the historical origins of the
current system including the original Version 1.0 development process.
3-10
locally calibrated ACG-weights. In some instances (e.g., for those with limited or no cost
data), these weights may also be used as a reasonable proxy for local cost data. Table 5
provides a complete listing of ACGs and their corresponding nationally representative
concurrent ACG-weight from the US Non-elderly and US Elderly Risk Assessment
Variables. (See the following discussion regarding the importance of rescaling so that
dollars are not over predicted or under predicted.)
The software-supplied reference ACG-weights are supplied in two forms: unscaled and
rescaled. Unscaled ACG-weights are simply the values of the reference ACG-weights
applied to a population of interest. The mean value of the unscaled ACG-weights
provides a rudimentary profiling statistic. If the mean of the unscaled ACG-weight is
greater than 1.0 it indicates the rating population (the population to which the weights are
being applied) is sicker than the reference population (the national reference database). If
the mean is less than 1.0, it indicates the rating population is healthier. To ensure that
dollars in the system are not over or under-estimated, we have also made available a
rescaled or standardized ACG-weight that mathematically manipulates the unscaled
ACG-weight to have a mean of 1.0 in the local population. The steps for performing this
manually are discussed in Applications Guide, Chapter 3.
Our experience indicates that concurrent (also referred to as retrospective) ACG-weights,
especially when expressed as relative values, have remarkable stability. Where
differences in ACG-weights across plans are present, it is almost universally attributable
to differences in covered services reflected by different benefit levels. The software
provided concurrent weights associated with the US Non-elderly Risk Assessment
Variables were developed from a nationally representative database comprising
approximately 4.7 million lives with comprehensive benefit coverage.
If local cost data are available, the ACG Software also calculates local ACG-weights.
These local weights more accurately reflect local benefit levels and area practice patterns.
In general it is recommended that the reference population (on which the weights are
developed) should be as similar as possible to the assessment population to which the
weights are applied. However in the absence of local cost data, the reference weights may
prove useful for calculating reasonably representative profiling statistics (reference
Applications Guide, Chapter 3).
Resource Utilization Bands (RUBs)
ACGs were designed to represent clinically logical categories for persons expected to
require similar levels of healthcare resources. However, enrollees with similar predicted
(or expected) overall utilization may be assigned different ACGs because they have
different epidemiological patterns of morbidity. For example, a pregnant woman with
significant morbidity, an individual with a serious psychological condition, or someone
with two chronic medical conditions may all be expected to use approximately the same
level of resources even though they each fall into different ACG categories. In many
instances users may find it useful to collapse the full set of ACGs into fewer categories,
particularly where resource use similarity and not clinical cogency is a desired objective.
3-11
Often a fewer number of combined categories will be easier to handle from an
administrative perspective. ACGs can be combined into what we term Resource
Utilization Bands (RUBs). The software automatically assigns 6 RUB classes:
• 0 - No or Only Invalid Dx
• 1 - Healthy Users
• 2 - Low
• 3 - Moderate
• 4 - High
• 5 - Very High
The relationship between ACG categories and RUBs is defined in Table 5 below.
Table 5: The Final ACG Categories, Reference Concurrent Weights
and RUBs
Reference Concurrent Weight
ACG
DESCRIPTION
0100
0200
0300
0400
Acute Minor, Age 1
Acute Minor, Age 2 to 5
Acute Minor, Age 6+
Acute Major
Likely to Recur, w/o
Allergies
Likely to Recur, w/
Allergies
Asthma
Chronic Medical:
Unstable
Eye & Dental
Chronic Specialty:
Unstable
Psychosocial, w/o
Psychosocial Unstable
Psychosocial, w/
Psychosocial Unstable,
w/o Psychosocial Stable
Psychosocial, w/
Psychosocial Unstable, w/
Psychosocial Stable
Preventive/Administrative
Pregnancy, 0-1 ADGs
Pregnancy, 0-1 ADGs,
Delivered
Pregnancy, 0-1 ADGs,
0500
0600
0700
0800
0900
1000
1100
1200
1300
1400
1500
1600
1710*
1711
1712
Commercially
Insured
(0 to 64 Years)
Medicare
Beneficiaries
(65 Years
and Older)
0.31
0.14
0.16
0.33
0.09
0.04
0.10
0.18
2
1
1
2
0.21
0.12
2
0.23
0.25
0.12
0.16
2
2
1.38
0.35
0.25
0.13
0.47
0.15
0.25
0.12
3
2
2
1
0.23
0.13
2
0.32
0.14
2
0.73
0.36
3
1.24
0.13
1.89
0.43
0.09
0.56
3
1
3
2.34
1.65
0.70
0.49
4
3
RUB
3-12
ACG
1720*
1721
1722
1730*
1731
1732
1740*
1741
1742
1750*
1751
1752
1760*
1761
1762
1770*
1771
1772
1800
1900
DESCRIPTION
Not Delivered
Pregnancy, 2-3 ADGs, no
Major ADGs
Major ADGs, Delivered
Major ADGs, Not
Delivered
Pregnancy, 2-3 ADGs, 1+
Major ADGs
Major ADGs, Not
Delivered
Major ADGs
Major ADGs, Not
Delivered
Major ADGs
Major ADGs, Not
Delivered
Pregnancy, 6+ ADGs, no
Major ADGs
Major ADGs, Not
Delivered
Pregnancy, 6+ ADGs, 1+
Major ADGs
Major ADGs, Not
Delivered
Acute Minor/Acute Major
Acute Minor/Likely to
Recur, Age 1
Commercially
Insured
(0 to 64 Years)
Medicare
Beneficiaries
(65 Years
and Older)
2.17
0.65
3
2.81
0.84
4
1.84
0.55
3
2.64
0.79
4
3.07
0.92
4
2.26
0.67
3
2.44
0.73
4
3.25
0.97
4
2.10
0.63
3
3.06
0.91
4
3.70
1.10
4
2.60
0.78
4
2.84
0.85
4
3.82
1.14
4
2.48
0.74
4
4.45
1.33
4
5.45
1.63
4
3.91
0.49
1.17
0.23
4
2
0.50
0.15
2
RUB
3-13
ACG
2000
2100
2200
2300
2400
2500
2600
2700
2800
2900
3000
3100
3200
3300
3400
3500
DESCRIPTION
Recur, Age 2 to 5
Recur, Age 6+, w/o
Allergy
Recur, Age 6+, w/
Allergy
Acute Minor/Chronic
Medical: Stable
Acute Minor/Eye &
Dental
Acute
Minor/Psychosocial, w/o
Psychosocial Unstable
Acute
Minor/Psychosocial, w/
Psychosocial Unstable,
w/o Psychosocial Stable
Acute
Minor/Psychosocial, w/
Psychosocial Unstable, w/
Psychosocial Stable
Acute Major/Likely to
Recur
Acute Minor/Acute
Major/Likely to Recur,
Age 1
Acute Minor/Acute
Age 2 to 5
Acute Minor/Acute
Age 6 to 11
Acute Minor/Acute
Age 12+, w/o Allergy
Acute Minor/Acute
Age 12+, w/ Allergy
Recur/Eye & Dental
Recur/Psychosocial
Commercially
Insured
(0 to 64 Years)
Medicare
Beneficiaries
(65 Years
and Older)
0.27
0.08
2
0.29
0.18
2
0.34
0.11
2
0.42
0.17
2
0.24
0.12
2
0.40
0.14
2
0.88
0.28
3
1.40
0.49
3
0.53
0.24
3
0.94
0.28
3
0.57
0.17
3
0.51
0.15
3
0.79
0.30
3
0.82
0.21
3
0.42
0.19
2
0.65
0.18
3
RUB
3-14
ACG
3600
3700
3800
3900
4000
4100
4210
4220
4310
4320
4330
4410
4420
4430
4510
DESCRIPTION
Acute Minor/Acute
Major/Likely to
Recur/Chronic Medical:
Stable
Acute Minor/Acute
Major/Likely to
Recur/Psychosocial
2-3 Other ADG
Combinations, Age 1 to
17
2-3 Other ADG
Combinations, Males Age
18 to 34
2-3 Other ADG
Combinations, Females
Age 18 to 34
2-3 Other ADG
Combinations, Age 35+
4-5 Other ADG
17, no Major ADGs
4-5 Other ADG
17, 1+ Major ADGs
4-5 Other ADG
44, no Major ADGs
4-5 Other ADG
44, 1 Major ADGs
4-5 Other ADG
44, 2+ Major ADGs
4-5 Other ADG
Combinations, Age 45+,
no Major ADGs
4-5 Other ADG
Combinations, Age 45+, 1
Major ADGs
4-5 Other ADG
2+ Major ADGs
6-9 Other ADG
Combinations, Age 1 to 5,
no Major ADGs
Commercially
Insured
(0 to 64 Years)
Medicare
Beneficiaries
(65 Years
and Older)
1.41
0.46
3
1.28
0.55
3
0.50
0.15
2
0.59
0.18
3
0.55
0.16
3
0.76
0.31
3
0.66
0.20
3
1.29
0.39
3
0.72
0.22
3
1.39
0.41
3
2.36
0.70
4
0.97
0.32
3
1.71
0.56
3
2.95
0.88
4
1.20
0.36
3
RUB
3-15
ACG
4520
4610
4620
4710
4720
4730
4810
4820
4830
4910
4920
4930
4940
DESCRIPTION
6-9 Other ADG
Combinations, Age 1 to 5,
1+ Major ADGs
6-9 Other ADG
17, no Major ADGs
6-9 Other ADG
17, 1+ Major ADGs
6-9 Other ADG
Combinations, Males,
Age 18 to 34, no Major
ADGs
6-9 Other ADG
Age 18 to 34, 1 Major
ADGs
6-9 Other ADG
Age 18 to 34, 2+ Major
ADGs
6-9 Other ADG
Combinations, Females,
Age 18 to 34, no Major
ADGs
6-9 Other ADG
Age 18 to 34, 1 Major
ADGs
6-9 Other ADG
Age 18 to 34, 2+ Major
ADGs
6-9 Other ADG
0-1 Major ADGs
6-9 Other ADG
Major ADGs
6-9 Other ADG
Major ADGs
6-9 Other ADG
4+ Major ADGs
Commercially
Insured
(0 to 64 Years)
Medicare
Beneficiaries
(65 Years
and Older)
2.27
0.68
4
1.11
0.33
3
2.50
0.75
4
1.13
0.34
3
1.99
0.59
3
3.93
1.17
4
1.22
0.36
3
1.93
0.58
3
3.48
1.04
4
2.09
0.70
3
4.05
1.21
4
6.89
1.87
5
12.59
2.89
5
RUB
3-16
ACG
5010
5020
5030
5040
5050
5060
5070
5110
5200
5310*
5311
5312
5320*
5321
5322
5330*
5331
DESCRIPTION
10+ Other ADG
17, no Major ADGs
10+ Other ADG
17, 1 Major ADGs
10+ Other ADG
17, 2+ Major ADGs
10+ Other ADG
0-1 Major ADGs
10+ Other ADG
Major ADGs
10+ Other ADG
Major ADGs
10+ Other ADG
4+ Major ADGs
No Diagnosis or Only
Unclassified Diagnosis (2
input files)
Non-Users (2 input files)
Infants: 0-5 ADGs, no
Major ADGs
Major ADGs, Low Birth
Weight
Major ADGs, Normal
Birth Weight
Infants: 0-5 ADGs, 1+
Major ADGs
Weight
Major ADGs, Normal
Birth Weight
Infants: 6+ ADGs, no
Major ADGs
Weight
Commercially
Insured
(0 to 64 Years)
Medicare
Beneficiaries
(65 Years
and Older)
2.25
0.67
3
3.95
1.18
4
12.72
3.80
5
3.32
1.06
4
5.28
1.60
4
8.28
2.40
5
18.85
4.60
5
0.00
0.90
0.00
0.27
1
0
7.45
2.22
3
0.85
0.25
4
3.66
1.09
3
15.66
4.67
4
2.39
0.71
5
1.69
0.50
4
5.34
1.59
3
1.63
0.49
4
RUB
3-17
ACG
5332
5340*
5341
5342
9900
DESCRIPTION
Major ADGs, Normal
Birth Weight
Infants: 6+ ADGs, 1+
Major ADGs
Weight
Major ADGs, Normal
Birth Weight
Invalid Age or Date of
Birth
Commercially
Insured
(0 to 64 Years)
Medicare
Beneficiaries
(65 Years
and Older)
10.15
3.03
3
28.35
8.46
5
6.14
1.83
5
0.86
0.26
4
0.31
0.09
0
RUB
Source: PharMetrics, Inc., a unit of IMS, Watertown, MA; national cross-section of managed care plans;
population of 4,740,000 commercially insured lives (less than 65 years old) and population of 257,404
Medicare beneficiaries (65 years and older), 2007.
*Note: The default is to subdivide these groups on delivery or low birth weight status.
Grouping the ACGs without these divisions is optional and must be turned on by the user in
order to be used.
3-18
Figure 1, on the following page, illustrates the main branches of the ACG decision tree.
Some MACs are not subdivided by the characteristics listed above because doing so did
not increase the explanatory power of the ACG model. Some include only a single
CADG: for instance, MAC-02 is composed of individuals with only acute major
conditions. Others, such as MAC-01, acute conditions only, are subdivided into three age
groups: ACG 0100 (Age=1 year), ACG 0200 (Age=2-5 years), and ACG 0300 (6 or more
years) because resource use differs by age for individuals with this pattern of morbidity.
MAC-10, including individuals with psychosocial morbidity only and MAC-17,
including individuals with psychosocial and acute minor conditions, are further split by
the presence of ADG-24 (recurrent or chronic stable psychosocial conditions) and ADG25 (recurrent or chronic unstable psychosocial conditions).
3-19
Figure 1: ACG Decision Tree
ACG 9900
Age <1
w/ diagnosis
Entire Population
Missing Age
To
MAC-26
tree
MAC-26
Age>= 1
Split into MACs
Based on CADGs
MAC-1
MAC-2
ACG
0400
Age
MAC-4
ACG
0700
Yes
No
ACG 0200
MAC-9
ACG
1200
MAC-8
ACG
1100
MAC-11
ACG
1600
MAC-10
ACG 0600
ACG 0500
Yes
ADG
24?
MAC-13
ACG
1800
MAC-12
Key
Collapsed ADG
Aggregated Diagnosis Group
No
MAC-14
MAC-17
MAC-16
ACG
2400
MAC-19
MAC-18
ACG
2800
No
ACG 1300
Yes
1 ACG 1900
ADG
24?
2-5 ACG 2000
ADG
05?
Yes
ACG 2200
Yes
No
ACG 2100
MAC-23
ACG
3700
MAC-22
ACG
3600
MAC-20
ACG
3400
MAC-25
MAC-24
To
MAC-24
tree
1 ACG 0100
No
ACG 2500
1 or 2
input
files?
2-5 ACG 0200
6+ ACG 0300
1
ACG 5100
12+
ACG 1500
ACG 1400
MAC-21
ACG
3500
Age
ADG
25?
Age
6+
Yes
MAC-15
ACG
2300
To
MAC-12
tree
ADG
25?
6+ ACG 0300
CADG
ADG
MAC-7
ACG
1000
MAC-6
ACG
0900
ADG
05?
1 ACG 0100
2-5
MAC-5
ACG
0800
MAC-3
ADG
05?
ACG 2700
Yes
No
ACG 2600
ACG 2200
No
ACG 2100
2
Claims
Info?
Yes
ACG 5110
No
ACG 5200
3-20
Figure 2, on the following page, illustrates the grouping logic for pregnant women. All
women with at least one diagnosis code indicating pregnancy are assigned to MAC-12.
The ACGs for pregnant women are formed with subdivisions first on total number of
ADGs (0-1, 2-3, 4-5, 6+) and second, for individuals with two or more ADGs, a split on
none versus one or more major ADGs. These two splits yield seven ACGs for pregnant
women.
Users may want to further subdivide the standard seven ACGs for pregnant women
according to whether delivery has occurred during the time period of interest, yielding a
total of 14 ACGs for women with a diagnosis of pregnancy. Either diagnosis codes for
delivery or user-supplied information on delivery (e.g., CPT-4 codes for delivery) can be
used to separate pregnant women according to delivery status. Because of the marked
differences in resource consumption for women with and without delivery and generally
adequate validity of diagnoses associated with delivery, most users will find this option
desirable to use. By default, the software will use diagnosis codes to subdivide based on
delivery status
Refer to the chapter entitled, ―
Basic Data Requirements‖ in the Installation and Usage
Guide for a more detailed discussion of appropriate means of identifying delivery status
using user-defined flags.
3-21
Figure 2: Decision Tree for MAC-12—Pregnant Women
MAC-12
# of
ADGs
0-1
# of
Major
ADGs
0
ACG 1710
ACG 1720
2-3
# of
Major
ADGs
1+
0
ACG 1730
ACG 1740
6+
4-5
# of
Major
ADGs
1+
0
ACG 1750
ACG 1760
1+
ACG 1770
Note: This level of
branching is optional.
Delivered?
Yes
No
ACG 1711
ACG 1712
Delivered?
Yes
No
ACG 1721
ACG 1722
Delivered?
Yes
No
Delivered?
ACG 1731
Yes
ACG 1732
No ACG 1742
ACG 1741
Delivered?
Yes
No
Delivered?
ACG 1751
Yes
ACG 1752
No
ACG 1761
ACG 1762
Delivered?
Yes
ACG 1771
No
ACG 1772
Key
CADG
ADG
Collapsed ADG
3-22
Figure 3, on the following page, illustrates the branching algorithm for MAC-26, which
includes all infants, regardless of their pattern of CADGs. The first bifurcation is made on
the total number of ADGs. Each group is further subdivided by presence of the number of
major ADGs. These two splits yield four ACG groups.
For the infant ACGs, there is an optional additional split on birth weight. If a user has
accurate birth weight information that can be linked with claims and enrollment files, the
four standard infant ACGs can be further split into low birth weight (<2,500 grams) and
normal birth weight (>2,500 grams). Our developmental work suggests that this
additional split improves the explanatory power of the ACG System. However, two
caveats are important to consider before using this ACG option. First, our research
indicates poor validity for existing ICD-9-CM birth weight codes in some administrative
data sets. Second, some populations may have such low rates of low birth weight infants
that the number of infants grouped into an ACG may be too small for accurate estimates.
In general, we recommend that at least 30 individuals per ACG are needed to obtain
stable estimates of average resource use for that ACG. By default, the ACG System will
divide infants based upon the presence or absence of a diagnosis code indicating low
birth weight.
Refer to the chapter entitled, ―
Basic Data Requirements‖ in the Installation and Usage
Guide for a more detailed discussion of appropriate means of identifying low birth weight
status using user-defined flags.
3-23
Figure 3: Decision Tree for MAC-26—Infants
MAC-26
# of
ADGs
0-5
# of
Major
ADGs
0
ACG 5310
Note: This level of
branching is optional.
Low
Birth weight?*
*Low birth weight
refers to infants
weighing less than
2500 grams.
Key
CADG
ADG
Yes
No
# of
Major
ADGs
1+
0
ACG 5320
Low
Birth weight?*
ACG 5311
ACG 5312
Yes
No
1+
ACG 5340
ACG 5330
Low
Birth weight?*
ACG 5321
Yes
ACG 5322
No
6+
Low
Birth weight?*
ACG 5331
ACG 5332
Yes
No
ACG 5341
ACG 5342
Collapsed ADG
3-24
Figure 4, on the following page, illustrates the last branch of the ACG tree, MAC-24,
which includes less frequently occurring combinations of CADGs. There are 33 ACGs
within MAC-24. With MAC-24, the first two splits are total number of ADGs (2-3, 4-5,
6-9, and 10+) and then, within each of these four groups, by age. The age splits separate
children (1-17 years) from adults (18+), and in some cases further subdivides within these
groups. Additional divisions are made on sex and number of major ADGs.
3-25
Figure 4: Decision Tree for MAC-24—Multiple ADG Categories
MAC-24
# of
ADGs
2-3
4-5
Age
1-17
18-34
Age
Age
ACG 3800
1-17
# of
Major
ADGs
18-44
Male
ACG 3900
# of
Major
ADGs
Female
ACG 4000
35+
ACG 4100
Key
CADG
ADG
0
ACG 4210
1+
ACG 4220
0
ACG 4310
Sex
Collapsed ADG
10+
6-9
1-5
6-17
1
ACG 4320
2+
ACG 4330
18-34
Age
# of
Major
ADGs
# of
Major
ADGs
# of
Major
ADGs
1
ACG 4420
2+
ACG 4430
1+
ACG 4520
0
ACG 4610
1+
ACG 4620
0-1
ACG 4910
35+
0
ACG 5010
M
# of
Major
ADGs
0
ACG 4710
1-17
1
ACG 4720
# of
Major
ADGs
1
ACG 5020
2+
ACG 5030
2+
ACG 4730
Sex
0
ACG 4410
45+
0
ACG 4510
# of
Major
ADGs
2
ACG 4920
3
ACG 4930
4+
ACG 4940
F
# of
Major
ADGs
0
ACG 4810
0-1
ACG 5040
1
ACG 4820
2+
ACG 4830
18+
# of
Major
ADGs
2
ACG 5050
3
ACG 5060
4+
ACG 5070
3-26
Clinical Aspects of ACGs
An ACG, the ‗building block‘ of the ACG System, is assigned based on all diagnosis
codes assigned to a person by providers during a predetermined period of time. This
makes ACGs different from most other case-mix measures (e.g., Diagnosis-Related
Groups--DRGs, Ambulatory Patient Categories--APCs, or Episode Treatment Groups-ETGs). In these other systems, the case-mix unit of analysis is based on a designated
service period and usually a single distinct clinical condition. For example, the service
period may be defined based on a single procedure or an episode of care. In contrast,
ACGs are based on all morbidities for which a person receives services over a defined
period of time.
The first step in the ACG assignment process is to categorize every ICD diagnosis code
given to a patient into a unique morbidity grouping known as an ―
ADG.‖ Each ADG is a
group of ICD diagnosis codes that are homogenous with respect to specific clinical
criteria and their demand on healthcare services. Patients with only one diagnosis over a
time period are assigned only one ADG, while a patient with multiple diagnoses can be
assigned to one or more ADGs:
 Example: A patient with both Obstructive Chronic Bronchitis (ICD-9-CM code
491.2) and Congestive Heart Failure (ICD-9-CM code 428.0) will fall into only one
ADG, Chronic Medical: Unstable (ADG-11), while a patient with Candidiasis of
Unspecified Site (ICD-9-CM code 112.9) and Acute Upper Respiratory Infections of
Unspecified Site (ICD-9-CM code 465.9) will have two ADGs, Likely to Recur:
Discrete Infections (ADG-8) and Time Limited: Minor-Primary Infections (ADG-2),
respectively.
ADGs represent a shift from standard methods for categorizing diagnosis codes. A class
of diagnoses, such as all those for diabetes or bacterial pneumonia, may indicate a
process that affects a specific organ system or pathophysiologic process. The criteria for
ADG assignment depends on those features of a condition that are most helpful in
understanding and predicting the duration and intensity of healthcare. Five clinical
criteria guide the assignment of each diagnosis code into an ADG: duration, severity,
diagnostic certainty, type of etiology, and expected need for specialty care. Table 6
shows how each of these five clinical criteria is applied to the 32 ADGs.
Duration
What is the expected length of time the health condition will last? Acute conditions are
time limited and expected to resolve completely. Recurrent conditions occur episodically
with inter-current disease-free intervals. Chronic conditions persist and are expected to
require long-term management generally longer than one year.
3-27
Severity
What is the expected prognosis? How likely is the condition to worsen or lead to
impairment, death, or an altered physiologic state? The ADG taxonomy divides acute
conditions into minor and major categories corresponding to low and high severity,
respectively. The system divides chronic conditions into stable and unstable based on the
expected severity over time. Unstable conditions are more likely to have complications
(related co-morbidities) than stable conditions and are expected to require more resources
on an ongoing basis (e.g., more likely to need specialty care).
Diagnostic Certainty
Some diagnosis codes are given for signs/symptoms and are associated with diagnostic
uncertainty. As such, they may require watchful waiting only or substantial work-up. The
three ADGs for signs/symptoms are arrayed by expected intensity of diagnostic work-up,
from low to intermediate to high.
Etiology
What is the cause of the health condition? Specific causes suggest the likelihood of
different treatments. Infectious diseases usually require anti-microbial therapy; injuries
may need emergency medical services, surgical management, or rehabilitation; anatomic
problems may require surgical intervention; neoplastic diseases could involve surgical
care, radiotherapy, chemotherapy; psychosocial problems require mental health services;
pregnancy involves obstetric services; and, medical problems may require
pharmacologic, rehabilitative, or supportive management.
Expected Need for Specialty Care
Would the majority of patients with this condition be expected to require specialty care
management from one of the following types of specialized providers: orthopedic
surgeon, otolaryngologist, ophthalmologist, dermatologist? In addition to these subspecialties, certain other ADG categories imply that specialty care is more likely.
3-28
Table 6: Duration, Severity, Etiology, and Certainty of the Aggregated Diagnosis Groups (ADGs)
Note: ADGs 15 and 19 are no longer used.
ADG
Expected Need for
Specialty Care
Duration
Severity
Etiology
Medical, noninfectious
Medical,
infectious
Medical,
infectious
Allergy
Mixed
High
Unlikely
High
Unlikely
High
Likely
High
Likely
High
High
Possibly
Possibly
Medical,
infectious
Anatomic/
Musculoskeletal
Anatomic/Ears,
Nose, Throat
Anatomic/Eye
Anatomic/
Musculoskeletal
Anatomic/Ears,
Nose, Throat
High
Unlikely
High
Unlikely
High
Likely
High
Unlikely
High
Likely
High
Likely: orthopedics
High
Likely: ENT
High
High
Likely: ophthalmology
Likely: orthopedics
High
Likely: ENT
1.
Time Limited: Minor
Acute
Low
2.
Time Limited: Minor-Primary Infections
Acute
Low
3.
Time Limited: Major
Acute
High
4.
Time Limited: Major-Primary Infections
Acute
High
5.
6.
Allergies
Asthma
Low
Low
7.
Likely to Recur: Discrete
Recurrent
Recurrent or
Chronic
Recurrent
8.
Likely to Recur: Discrete-Infections
Recurrent
Low
9.
Likely to Recur: Progressive
Recurrent
High
10. Chronic Medical: Stable
Chronic
Low
11. Chronic Medical: Unstable
Chronic
High
12. Chronic Specialty: Stable-Orthopedic
Chronic
Low
13. Chronic Specialty: Stable-Ear, Nose, Throat
Chronic
Low
14. Chronic Specialty: Stable-Ophthalmology
16. Chronic Specialty: Unstable-Orthopedics
Chronic
Chronic
Low
High
15. Chronic Specialty: Unstable-Ear, Nose, Throat
Chronic
High
Diagnostic
Certainty
Low
ADG
16. Chronic Specialty: Unstable-Ophthalmology
20. Dermatologic
17.
18.
19.
20.
Injuries/Adverse Effects: Minor
Injuries/Adverse Effects: Major
Psychosocial: Time Limited, Minor
Psychosocial: Recurrent or Chronic, Stable
21. Psychosocial: Recurrent or Persistent, Unstable
22.
23.
24.
25.
Signs/Symptoms: Minor
Signs/Symptoms: Uncertain
Signs/Symptoms: Major
Discretionary
26.
27.
28.
29.
30.
See and Reassure
Malignancy
Pregnancy
Dental
Technical Reference Manual
3-29
Duration
Severity
Etiology
Diagnostic
Certainty
Chronic
Acute,
recurrent
Acute
Acute
Acute
Recurrent or
Chronic
Recurrent or
Chronic
Uncertain
Uncertain
Uncertain
Acute
High
Low to High
Anatomic/Eye
Mixed
High
High
Likely: ophthalmology
Likely: dermatology
Low
High
Low
Low
Injury
Injury
Psychosocial
Psychosocial
High
High
High
High
Unlikely
Likely
Unlikely
Likely: mental health
High
Psychosocial
High
Likely: mental health
Low
Uncertain
High
Low or High
Mixed
Mixed
Mixed
Anatomic
Low
Low
Low
High
Low
N/A
High
Low
Low to High
Anatomic
N/A
Neoplastic
Pregnancy
Mixed
High
N/A
High
High
High
Unlikely
Uncertain
Likely
Likely: surgical
specialties
Unlikely
Unlikely
Likely: oncology
Likely: obstetrics
Likely: dental
Acute
N/A
Chronic
Acute
Acute,
recurrent,
chronic
Expected Need for
Specialty Care
3-30
Clinically Oriented Examples of ACGs
Patients are categorized into an ACG based on the pattern of ADGs experienced over a
predetermined interval and, in some cases, their age and sex. This approach focuses on
the totality of diseases experienced by a person rather than any specific disease. Because
this method diverges from the traditional biomedical, categorical method of examining
morbidity, we show how ACGs classify patients with specific types of diseases.
In the examples that follow, we categorize patients by choosing a specific clinical feature
that they have, such as a disease, pregnancy, or by their age. These examples show how
the presence of other diseases or total number of ADGs changes ACG assignment.
Chronic Illnesses
On the following pages, Example 1 presents three patients with hypertension and
Example 2 presents three patients with diabetes. These individuals were actual patients
selected from a private health plan database. The input data used by the ACG grouping
software, the output produced by the software, and the associated resource consumption
variables are presented. As these patients demonstrate, there is substantial variability in
patterns of morbidity and need for healthcare for different patients classified by a specific
condition such as hypertension or diabetes. Thus, knowing only that a patient has a
particular medical problem, even if it is a chronic condition, provides little information
about the need for medical services. In general, as the number of different types of
morbidities increases, the total number of ambulatory visits increases as does total
expenditures. However, the total burden of morbidity as represented by the ACG – that
is, the constellation of ADGs and presence of major ADGs is the most important
determinant of resource consumption.
In Example 2, for diabetes, during the assessment period Patient 1 had diagnosis codes
given only for uncomplicated diabetes mellitus and a routine medical exam and is
therefore classified into the ACG for patients with stable, chronic medical conditions
(ACG-0900). In contrast, all three of the patients with hypertension as well as Patients 2
and 3 with diabetes are in ACGs that branch from MAC-24 (combinations of ADGs not
otherwise classified). This occurs because their combinations of ADGs occur too
infrequently to merit a separate ACG. Patients in MAC-24 have both high levels of
morbidity and high levels of health need. There is a strong link between the total number
of ADGs/major ADGs and resource consumption.
Although not shown in the examples, there are two additional ACGs that describe
commonly occurring combinations of morbidity for individuals with stable, chronic
medical conditions. ACG-2300 (Chronic Medical--Stable and Acute Minor) is assigned
to patients with uncomplicated diabetes, hypertension, or other stable chronic conditions
and a minor illness, injury, or symptom. Individuals in ACG-3600 have four types of
morbidities: stable chronic medical conditions (which includes the diagnosis of
3-31
hypertension), acute minor conditions, conditions of low severity likely to reoccur, and
acute major conditions.
Example 1: Hypertension
The following patient types demonstrate the levels of hypertension, ADGs, and
associated costs.
Patient 1: Low Cost Patient with Hypertension
Input Data/Patient
Characteristics
ACG Output
Age/Sex: 56/Male
ACG-4100: 2-3
Other ADG Combinations, age > 35
Conditions: Hypertension, Disorders
of lipid metabolism,
Glaucoma, and Bursitis/synovitis
ADGs: 07, 10, and 18.
Likely to Recur: Discrete, Chronic Medical:
Stable, and Chronic Specialty: Unstable Eye
Resource
Consumption
Variables
Total Cost: $318
Ambulatory visits: 2
>1 Hospitalization: N
Patient 2: High Cost Patient with Hypertension
Input Data/Patient
Characteristics
Age/Sex: 53/Male
Conditions: Hypertension, General
medical exam, Cardiovascular
symptoms; Ischemic heart disease,
Disorders of lipid metabolism,
Debility/fatigue, Cerobrovascular
disease, Arthralgia, and
Bursitis/synovitis
ACG Output
ACG-4430: 4-5
Other ADG Combinations, Age >45, 2
+Majors
ADGS: 01, 09*, 10, 11*, 27, and 31.
Time Limited: Minor,
Likely to Recur: Progressive, Chronic
Medical: Stable, Chronic Medical:
Unstable, Signs/Symptoms: Uncertain and
Resource
Consumption
Variables
Total Cost: $1,968
*Major ADG, all ages
3-32
Patient 3: Very High Cost Patient with Hypertension
Input Data/Patient
Characteristics
Age/Sex: 47/Male
Conditions: Hypertension,
General medical exam,
Ischemic heart disease,
Congenital heart disease,
Cardiac valve disorders,
Gastrointestinal signs/
symptoms, Diverticular disease
of colon, Chest pain, and Lower
back pain
ACG Output
ACG- 4920: 6-9
Other ADGs Combination, Age
>35, 2 Majors
ADGs: 07, 09*, 11*, 27, 28, and
31.
Likely to Recur: Discrete,
Likely to Recur: Progressive,
Chronic Medical: Stable, Chronic
Medical: Unstable,
Signs/Symptoms: Uncertain,
Signs/Symptoms: Major, and
Resource Consumption
Variables
Total Cost: $16,960
>1 Hospitalization: Y
Example 2: Diabetes Mellitus
The following patient types demonstrate the levels of diabetes mellitus, ADGs, and
associated costs.
Patient 1: Low Cost Patient with Diabetes
Input Data/Patient
Characteristics
ACG Output
Age/Sex: 57/Female
ACG-0900: Chronic Medical, Stable
Conditions: Diabetes mellitus and
General medical exam
ADGs: 10 and 32.
Chronic Medical: Stable and
Resource
Consumption
Variables
Total Cost: $418
3-33
Patient 2: High Cost Patient with Diabetes
Input Data/Patient
Characteristics
Age/Sex: 54/Female
Conditions: Diabetes mellitus,
Congestive heart failure,
Thrombophlebitis, contusions and
abrasions, Non-fungal infections of
skin, disease of nail, Chest pain,
Vertiginous syndromes,
Fibrositismyalgia, Respiratory
signs/symptoms, and cough
ACG Output
ACG-4930: 6-9
Other ADG Combinations, age >35, 3
Major
ADGs: 01, 04*, 09*, 10, 11*, 21, 27,
28, and 31.
Time Limited: Major Primary
Infection,
Likely to Recur: Progressive, Chronic
Medical: Stable, Chronic Medical:
Unstable, Injuries/Adverse
Resource
Consumption
Variables
Total Cost: $2,122
Patient 3: Very High Cost Patient with Diabetes
Input Data/Patient
Characteristics
ACG Output
Age/Sex: 38/Female
ACG-5060: 10 + Other ADG
Combination, age >17, 3 Majors
Conditions: Diabetes mellitus,
Cardiovascular symptoms,
Cardiac arrhythmia, Sinusitis,
Abdominal pain, Anorectal
conditions, Benign/unspecified
neoplasm, Otitis media,
Cholelithiasis, Cholecystitis, Acute
lower respiratory
ADGs: 01, 02, 03*, 06, 07, 08, 09*,
10, 11*, 27, 28, 29, 30, and 31.
Time Limited: Minor- Primary
Infections,
Time Limited: Major, Asthma,
Likely to Recur: Discrete,
Likely to Recur: Discrete-Infections,
Likely to Recur: Progressive,
Chronic Medical: Stable, Chronic
Medical: Unstable, Signs/Symptoms:
Uncertain, Signs/ Symptoms: Major,
Discretionary, See/ Reassure, and
Resource
Consumption
Variables
Total Cost: $12,944
3-34
Pregnancy
Using diagnosis codes for pregnancy, the ACG system identifies all women who were
pregnant during the assessment period and places them into the pregnancy MAC. ACGs
are formed based on (1) total number of ADGs, (2) presence of ―
complications‖ (i.e.,
presence of a major ADG), and (3) whether the woman delivered (the user may override
this default level of assignment).
Example 3 shows how the ACG System groups women with a complicated
pregnancy/delivery. Both women in the example had ICD-9-CM codes that map to ADG03 (an acute major morbidity). The salient difference between the two that explains the
difference in resource consumption is that Patient #2 had a greater number of ADGs and
more major ADGs and thus fits into a more resource intensive ACG. That is, Patient #2
had a higher level of morbidity than #1, even though both women experienced a
complicated pregnancy/delivery.
Table 7 presents an alternative clinical categorization of the pregnancy/delivery ACGs.
Three dimensions are used to classify the ACGs – number of ADGs, presence of major
ADGs, and whether the women delivered during the assessment period. Resource
consumption increases along each of the three axes: presence of delivery, presence of a
major ADG, and number of ADGs. Using various combinations of these ACGs, a
clinician or manager can determine the proportion of women with complicated
pregnancies and deliveries overall, and with different levels of morbidity. The need for
specialty services will be greatest for those women with higher levels of morbidity and
complications as defined by presence of a major ADG.
Example 3: Pregnancy/Delivery with Complications
The following patient types demonstrate the levels of pregnancy and delivery with
complications, ADGs, and associated costs.
Patient 1: Pregnancy/Delivery with Complications, Low Morbidity
Input Data/Patient
Characteristics
Age/Sex: 32/Female
Conditions: General medical
exam, Pregnancy and delivery uncomplicated and Pregnancy
and delivery - with
complications.
ACG Output
ACG-1731: 2-3 ADGs, 1+ Major
ADGs, Delivered
ADGs: 03*, 28, 31, and 33.
Time Limited: Major,
Signs/Symptoms: Major, Prevention/
Administrative, and Pregnancy
Resource
Consumption
Variables
Total Cost: $8,109
Ambulatory visits: N/A
3-35
Patient 2: Pregnancy/Delivery with Complications, High Morbidity
Input Data/Patient
Characteristics
ACG Output
Age/Sex: 36/Female
exam,
Dermatitis and eczema,
Benign and unspecified
neoplasm,
Urinary tract infection, Multiple
sclerosis,
Pregnancy & deliveryuncomplicated, and Pregancy &
delivery-with complications.
ACG-1771: 6+ ADGs, 1+ Major
ADGs, Delivered
ADGs: 01, 03*, 08, 10, 11, 26, 28,
31, and 33.
Likely to recur: discrete-infections,
Chronic Medical: Stable,
Chronic Medical: Unstable,
Signs/Symptoms: Major, Prevention/
Administrative, and Pregnancy
Resource
Consumption
Variables
Total Cost: $10,859
Ambulatory visits: N/A
Table 7: Clinical Classification of Pregnancy/Delivery ACGs
ACG Levels
Morbidity
Level
Low
(1-3 ADGs)
Mid
(4-5 ADGs)
High
(6+ ADGs)
Pregnancy Only
Uncomplicated Complicated
(No Major
(1+ Major
ADGs)
ADGs)
Delivered
Uncomplicated Complicated
(No Major
(1+ Major
ADGs)
ADGs)
1712, 1722
1732
1711, 1721
1731
1742
1752
1741
1751
1762
1772
1761
1771
Infants
The ACG System places all infants into an infant MAC. By definition, all had at least one
hospitalization (at time of delivery). ACG groups are formed based on total number of
ADGs and the presence of a complication or major ADG. Table 8 provides a clinical
classification of the four infant ACGs. Example 4 compares an infant in the low
morbidity/no complications ACG (5310, the most frequent infant ACG) with an infant in
the higher morbidity/with complications ACG (5340, the most resource intensive infant
ACG). Infant #1 had a typical course: a hospitalization at birth, routine check-ups, and
illness management for upper respiratory tract infection and otitis media. Infant 2
presents a completely different picture in terms of pattern of morbidity and resource
consumption, both of which are substantially greater in comparison with Infant 1.
3-36
Example 4: Infants
The following patient types demonstrate the levels of infants with complications, ADGs,
and associated costs.
Patient 1: Infant With Low Morbidity, No Complications
Input Data/Patient
Characteristics
Age/Sex: 0/Female
exam,
Otitis media and Acute upper
respiratory tract infection.
ACG Output
ACG 5310: 0-5 ADGs, No Major
ADGs
ADGs: 02, 08 and 31
Likely to Recur: Discrete-Infections,
And Prevention/Administration
Resource
Consumption
Variables
Total Cost: $1,842
Patient 2: Infant With High Morbidity, No Complications
Input Data/Patient
Characteristics
Age/Sex: 0/Male
exam, Cardiovascular symptoms,
Dermatitis and eczema, Otitis
media, Fluid/ electrolyte
disturbance, Diarrhea, Abdominal
pain, Nausea, vomiting, Acute
upper respiratory tract infection,
and Acute lower respiratory tract
infection
ACG Output
ACG 5340: 6+ ADGs, 1+ Major
ADGs
ADGs: 01, 02, 03*, 08, 11*, 13, 26,
27, 28, 29, and 31
Time Limited: Minor - Primary
Infections,
Likely to Recur: Discrete- Infections,
Chronic Medical: Unstable, Chronic
Specialty: Stable - ENT,
Signs/Symptoms: Minor,
Signs/Symptoms: Uncertain,
Signs/Symptoms: Major,
Discretionary, and
Resource
Consumption
Variables
Total Cost: $18,703
3-37
Table 8: Clinical Classification of Infant ACGs
Morbidity
Level
Low
(0-5 ADGs)
Mid
(6+ ADGs)
No Complications
(No Major ADGs)
Complication
(1+ Major ADGs)
5310
5320
5330
5340
3-38
Expanded Diagnosis Clusters (EDCs)
4-i
4 Expanded Diagnosis Clusters (EDCs)
Overview ..................................................................................................................... 4-1
The Development of EDCs ........................................................................................ 4-1
Understanding How EDCs Work ............................................................................. 4-3
Table 1: ICD-9-CM Codes Assigned to Otitis Media EDC (EAR01) .................... 4-3
Table 2: Examples of Diabetes Complications ....................................................... 4-5
Table 3: Annual Prevalence Ratios for Expanded Diagnosis Clusters (EDCs) for
Commercially Insured and Medicare Beneficiaries ................................................. 4-6
Diagnostic Certainty .............................................................................................. 4-13
MEDC Types ......................................................................................................... 4-14
Table 4: MEDC Types .......................................................................................... 4-14
Some Important Differences Between EDCs and ADGs ...................................... 4-15
Important Differences Between EDCs and Episode-of-Care Systems .................. 4-15
Applications of EDCs ............................................................................................ 4-16
4-ii
4-1
Overview
The Johns Hopkins Expanded Diagnosis Clusters (EDCs) complement the unique personoriented approach that underpins the ACG System.
EDCs are a tool for easily identifying people with specific diseases or symptoms,
without having to create your own algorithms. The EDC methodology assigns ICD codes
found in claims or encounter data to one of 267 EDCs, which are further organized into
27 categories called Major Expanded Diagnosis Clusters (MEDCs). As broad groupings
of diagnosis codes, EDCs help to remove differences in coding behavior between
practitioners.
 Example: There are 67 ICD-9-CM codes that practitioners can record as a diagnosis
for otitis media. The EDC for otitis media combines these codes into a single rubric,
allowing the healthcare analyst to easily summarize information on this condition.
As a stand-alone tool, EDCs can be used to select patients with a specific condition or
combination of conditions and can also be used to compare the distribution of
conditions in one population with another. When combined with ACGs, the result is a
powerful combination tool for demonstrating variability of cost within disease categories.
This is useful for many profiling applications and can help to target individuals for casemanagement purposes.
The Development of EDCs
EDCs build on the methodology developed by Ronald Schneeweiss and colleagues in
1983.1 In Schneeweiss’s original work, diagnosis codes were classified according to
clinical criteria. Schneeweiss’s original 92 categories, so-called Diagnosis Clusters,
underwent considerable updating and extensive expansions based on significant
additional development at Johns Hopkins University, led by Dr. Christopher Forrest. In
2000, Dr. Forrest’s team of generalist and specialist physicians used their clinical
judgment to assign the most commonly used ICD-9 codes to EDCs. In all, the team
assigned approximately 9,400 diagnosis codes to 190 EDCs.
The development team reviewed the Clinical Classification for Health Policy Research
produced by the Agency for Health Care Policy and Research,2 another disease-oriented
grouping methodology, to identify additional conditions that were excluded from
Schneeweiss’s original taxonomy. Other modifications made to the original Schneeweiss
clusters included deletion of some clusters, expansion of the content of other clusters by
1
Schneeweiss R, Rosenblatt RA, Cherkin DC, Kirkwood R, Hart G. Diagnosis Clusters: Tool for
Analyzing the Content of Ambulatory Medical Care. Med Care 1983; 21:105-122.
2
Elixhauser A, Andrews RM, Fox S. (1993) Clinical classifications for health policy research: Discharge
statistics by principal diagnosis and procedure (AHCPR Publication No. 93-0043).
4-2
updating the range of ICD codes, improving the clinical homogeneity of certain clusters,
and splitting existing clusters into new categories.
Additionally, each of the Expanded Diagnosis Clusters was classified into one of 27
broad clinical categories, termed a Major EDC (MEDC). For example, three EDCs-allergic reactions, asthma, and allergic rhinitis--all fall within the Allergy MEDC
classification.
The original assignments of EDCs to MEDCs were made based on the assessment of our
clinician panel as to the physician specialty most likely to provide care for the class of
conditions, when other than primary care services were received.
 Example: Prostatic hypertrophy (GUR04) was mapped to the genito-urinary MEDC
and gout (RHU02) to the rheumatologic MEDC.
Further review of data and actual specialist use within each EDC category led to changes
in the final assignment of EDCs to their respective MEDC. It should be noted that for
many EDCs, the most common service provider is the primary care clinician, and not a
specialist.
Since the initial development of the EDC typology in 2000, the EDC categories have
undergone several revisions and expansions so that now all ICD (-9, -9-CM and -10
version) codes are assigned. There are 267 EDCs in the current version of the ACG
System.
Recently, the ACG developers added several new common pediatric conditions (many of
which also occur during adulthood) and added new categories to broaden the spectrum of
signs and symptoms covered (e.g., gastrointestinal signs and symptoms, ophthalmic signs
and symptoms, musculoskeletal signs and symptoms). For example, cancers were divided
into 17 groups: 15 site specific cancers, and low impact and high impact cancers (based
on the expected costs of treatment). Several EDC conditions were refined to support the
pharmacy adherence methodology and identify conditions requiring chronic medication
therapy.
4-3
Understanding How EDCs Work
Each ICD code maps to a single EDC. ICD codes within an EDC share similar clinical
characteristics and are expected to evoke similar types of diagnostic and therapeutic
responses.
There are 67 ICD-9-CM codes (shown in Table 1 below) that practitioners can record as
a diagnosis for otitis media. The EDC for otitis media combines these codes into a single
rubric, which lessens the impact of physicians’ coding styles on your analyses.
Table 1: ICD-9-CM Codes Assigned to Otitis Media EDC (EAR01)
ICD-9
Description
0552
381
3810
38100
38101
38102
38103
38104
38105
38106
3811
38110
38119
3812
38120
38129
3813
3814
3815
38150
38151
38152
3816
38160
38161
38162
38163
3817
3818
38181
38189
3819
382
Postmeasles otitis media
Nonsuppurative otitis media and Eustachian tube disorders
Acute nonsuppurative otitis media
Unspecified acute nonsuppurative otitis media
Acute serous otitis media
Acute mucoid otitis media
Acute sanguinous otitis media
Acute allergic serous otitis media
Acute allergic mucoid otitis media
Acute allergic sanguinous otitis media
Chronic serous otitis media
Simple or unspecified chronic serous otitis media
Other chronic serous otitis media
Chronic mucoid otitis media
Simple or unspecified chronic mucoid otitis media
Other chronic mucoid otitis media
Other and unspecified chronic nonsuppurative otitis media
Nonsuppurative otitis media, not specified as acute or chronic
Eustachian salpingitis
Unspecified Eustachian salpingitis
Acute Eustachian salpingitis
Chronic Eustachian salpingitis
Obstruction of Eustachian tube
Unspecified obstruction of Eustachian tube
Osseous obstruction of Eustachian tube
Intrinsic cartilagenous obstruction of Eustachian tube
Extrinsic cartilagenous obstruction of Eustachian tube
Patulous Eustachian tube
Other disorders of Eustachian tube
Dysfunction of Eustachian tube
Other disorders of Eustachian tube
Unspecified Eustachian tube disorder
Suppurative and unspecified otitis media
4-4
ICD-9
Description
3820
38200
38201
38202
3821
3822
3823
3824
3829
384
3840
38400
38401
38409
3841
3842
38420
38421
38422
38423
38424
38425
3848
38481
38482
3849
3886
38860
38869
3887
38870
38871
38872
3889
Acute suppurative otitis media
Acute suppurative otitis media without spontaneous rupture of eardrum
Acute suppurative otitis media with spontaneous rupture of eardrum
Acute suppurative otitis media in diseases classified elsewhere
Chronic tubotympanic suppurative otitis media
Chronic atticoantral suppurative otitis media
Unspecified chronic suppurative otitis media
Unspecified suppurative otitis media
Unspecified otitis media
Other disorders of tympanic membrane
Acute myringitis without mention of otitis media
Unspecified acute myringitis
Bullous myringitis
Other acute myringitis without mention of otitis media
Chronic myringitis without mention of otitis media
Perforation of tympanic membrane
Unspecified perforation of tympanic membrane
Central perforation of tympanic membrane
Attic perforation of tympanic membrane
Other marginal perforation of tympanic membrane
Multiple perforations of tympanic membrane
Total perforation of tympanic membrane
Other specified disorders of tympanic membrane
Atrophic flaccid tympanic membrane
Atrophic nonflaccid tympanic membrane
Unspecified disorder of tympanic membrane
Otorrhea
Unspecified otorrhea
Other otorrhea
Otalgia
Unspecified otalgia
Otogenic pain
Referred otogenic pain
Unspecified disorder of ear
The EDCs related to Diabetes receive special treatment in the EDC assignment process.
Each Diabetes diagnosis is associated with either EDC END06 Type 2 Diabetes without
complication or EDC END08 Type 1 Diabetes without complication. The EDC
assignment process then looks for the presence of a potential complication diagnosis
code. If a complication is found (see Table 2 below), a patient with END06 will be
changed to END07 while a patient with END08 will be changed to END09. The
implication of this method is that a patient cannot be assigned to a diabetes condition
with and without complications simultaneously.
4-5
Table 2: Examples of Diabetes Complications
ICD-9
2501
2502
2503
2504
410
411
412
413
414
581
582
583
584
585
586
V56
Description
Diabetes with ketoacidosis
Diabetes with hyperosmolar coma
Diabetes with coma NEC
Diabetes with renal manifestation
Acute myocardial infarction
Other acute ischemic heart disease
Old myocardial infarction
Angina pectoris
Other chronic ischemic heart disease
Nephrotic syndrome
Chronic nephritis
Nephritis NOS
Acute renal failure
Chronic renal failure
Renal failure NOS
Dialysis encounter
Also, patients will not be assigned to a type 1 Diabetes EDC and a type 2 Diabetes EDC
concurrently. If the patient has diagnoses indicating both type 1 Diabetes and type 2
Diabetes, the EDC assignment will reflect type 1 Diabetes only.
4-6
Table 3: Annual Prevalence Ratios for Expanded Diagnosis Clusters
(EDCs) for Commercially Insured and Medicare Beneficiaries
Expanded Diagnosis Cluster (EDC)
Administrative
Surgical aftercare
Transplant status
Administrative concerns and non-specific laboratory
abnormalities
Preventive care
Allergy
Allergic reactions
Allergic rhinitis
Asthma, w/o status asthmaticus
Asthma, with status asthmaticus
Disorders of the immune system
Cardiovascular
Cardiovascular signs and symptoms
Ischemic heart disease (excluding acute myocardial
infarction)
Congenital heart disease
Congestive heart failure
Cardiac valve disorders
Cardiomyopathy
Heart murmur
Cardiac arrhythmia
Generalized atherosclerosis
Disorders of lipid metabolism
Acute myocardial infarction
Cardiac arrest, shock
Hypertension, w/o major complications
Hypertension, with major complications
Cardiovascular disorders, other
Dental
Disorders of mouth
Disorders of teeth
Gingivitis
Stomatitis
Ear, Nose, Throat
Otitis media
Tinnitus
Number of Persons with EDC per 1,000
Enrollees
Commercially Insured Medicare Beneficiaries
(0 to 64 Years)
(65 Years and Older)
41.3
0.8
198.1
2.1
176.9
419.1
403.7
547.6
16.0
81.0
36.0
3.2
2.6
13.9
70.3
43.8
6.2
11.5
40.4
156.3
17.6
3.3
3.3
11.7
2.7
5.5
14.1
5.1
117.8
1.3
0.5
114.7
10.3
6.4
196.1
10.3
68.1
94.6
26.9
18.6
142.8
70.7
480.3
14.8
5.0
559.3
85.7
41.9
7.2
8.9
0.7
2.3
10.3
2.6
0.8
2.2
74.7
3.4
31.7
9.3
Temporomandibular joint disease
Foreign body in ears, nose, or throat
Deviated nasal septum
Otitis externa
Wax in ear
Deafness, hearing loss
Chronic pharyngitis and tonsillitis
Epistaxis
Acute upper respiratory tract infection
ENT disorders, other
Endocrine
Osteoporosis
Short stature
Hypothyroidism
Other endocrine disorders
Type 2 diabetes, w/o complication
Type 2 diabetes, w/ complication
Type 1 diabetes, w/o complication
Type 1 diabetes, w/ complication
Eye
Ophthalmic signs and symptoms
Blindness
Retinal disorders (excluding diabetic retinopathy)
Disorders of the eyelid and lacrimal duct
Refractive errors
Cataract, aphakia
Conjunctivitis, keratitis
Glaucoma
Infections of eyelid
Foreign body in eye
Strabismus, amblyopia
Traumatic injuries of eye
Diabetic retinopathy
Eye, other disorders
Age-related macular degeneration
Female Reproductive
Pregnancy and delivery, uncomplicated
Female genital symptoms
Endometriosis
Pregnancy and delivery with complications
Female infertility
4-7
Enrollees
(0 to 64 Years)
2.6
1.5
5.8
13.0
13.2
10.5
7.3
3.7
183.9
15.1
2.2
1.6
6.8
12.7
49.3
45.5
1.6
9.2
83.5
20.1
7.8
0.7
39.1
21.3
28.3
5.3
4.7
1.7
82.0
0.0
119.4
38.9
115.5
70.8
11.2
15.5
38.8
0.9
6.7
13.1
69.4
14.2
41.3
13.3
8.2
2.5
6.1
5.2
2.5
16.2
1.1
83.0
3.6
54.2
61.8
112.8
234.4
41.2
111.6
22.5
2.7
7.0
5.3
19.7
69.6
47.4
21.7
26.2
2.8
18.2
3.2
0.3
10.5
0.4
0.7
0.1
4-8
Abnormal pap smear
Ovarian cyst
Vaginitis, vulvitis, cervicitis
Menstrual disorders
Contraception
Menopausal symptoms
Utero-vaginal prolapse
Female gynecologic conditions, other
Gastrointestinal/Hepatic
Gastrointestinal signs and symptoms
Inflammatory bowel disease
Constipation
Acute hepatitis
Chronic liver disease
Peptic ulcer disease
Gastroenteritis
Gastroesophageal reflux
Irritable bowel syndrome
Diverticular disease of colon
Acute pancreatitis
Chronic pancreatitis
Lactose Intolerance
Gastrointestinal/Hepatic disorders, other
General Signs and Symptoms
Nonspecific signs and symptoms
Chest pain
Fever
Syncope
Nausea, vomiting
Debility and undue fatigue
Lymphadenopathy
Edema
General Surgery
Anorectal conditions
Appendicitis
Benign and unspecified neoplasm
Cholelithiasis, cholecystitis
External abdominal hernias, hydroceles
Chronic cystic disease of the breast
Other breast disorders
Varicose veins of lower extremities
Enrollees
(0 to 64 Years)
13.7
9.3
28.5
33.7
22.8
17.9
2.4
5.8
3.5
3.2
21.2
5.9
0.1
25.6
11.0
5.2
36.1
3.4
14.0
2.3
6.2
19.0
38.7
45.4
8.6
11.5
1.0
0.6
1.3
10.3
99.2
6.6
33.4
3.2
12.5
47.5
50.8
118.3
14.8
67.8
3.3
3.1
3.6
28.8
42.4
50.0
30.1
7.8
27.7
54.6
7.6
10.1
96.6
141.1
23.0
35.1
36.2
107.8
8.9
55.8
20.5
1.7
55.7
6.3
6.7
10.0
23.5
3.6
49.0
1.1
147.3
17.9
18.2
12.3
39.8
13.5
Nonfungal infections of skin and subcutaneous
tissue
Abdominal pain
Peripheral vascular disease
Burns--1st degree
Aortic aneurysm
Gastrointestinal obstruction/perforation
Genetic
Chromosomal anomalies
Inherited metabolic disorders
Genito-urinary
Vesicoureteral reflux
Undescended testes
Hypospadias, other penile anomalies
Prostatic hypertrophy
Stricture of urethra
Urinary symptoms
Other male genital disease
Urinary tract infections
Renal calculi
Prostatitis
Incontinence
Genito-urinary disorders, other
Hematologic
Hemolytic anemia
Iron deficiency, other deficiency anemias
Thrombophlebitis
Neonatal jaundice
Aplastic anemia
Deep vein thrombosis
Hemophilia, coagulation disorder
Hematologic disorders, other
Infections
Tuberculosis infection
Fungal infections
Infectious mononucleosis
HIV, AIDS
Sexually transmitted diseases
Viral syndromes
Lyme disease
Septicemia
4-9
Enrollees
(0 to 64 Years)
36.1
70.6
1.5
0.4
1.1
4.7
62.5
111.1
32.0
0.3
20.3
20.5
0.5
6.4
0.1
19.1
0.8
0.3
0.5
10.3
0.8
42.1
13.2
41.2
9.3
3.8
6.7
8.3
0.7
0.0
0.1
91.9
4.7
121.1
31.6
95.3
18.5
10.1
28.8
55.8
0.8
21.1
1.9
2.8
0.5
2.5
1.6
3.6
2.0
105.5
9.6
0.1
3.6
16.3
10.4
16.4
0.3
4.9
1.8
0.9
7.3
33.1
0.7
1.5
0.5
8.8
0.1
0.3
4.6
10.8
1.3
12.5
4-10
Infections, other
Malignancies
Malignant neoplasms of the skin
Low impact malignant neoplasms
High impact malignant neoplasms
Malignant neoplasms, breast
Malignant neoplasms, cervix, uterus
Malignant neoplasms, ovary
Malignant neoplasms, esophagus
Malignant neoplasms, kidney
Malignant neoplasms, liver and biliary tract
Malignant neoplasms, lung
Malignant neoplasms, lymphomas
Malignant neoplasms, colorectal
Malignant neoplasms, pancreas
Malignant neoplasms, prostate
Malignant neoplasms, stomach
Acute leukemia
Malignant neoplasms, bladder
Musculoskeletal
Musculoskeletal signs and symptoms
Acute sprains and strains
Degenerative joint disease
Fractures (excluding digits)
Torticollis
Kyphoscoliosis
Congenital hip dislocation
Fractures and dislocations/digits only
Joint disorders, trauma related
Fracture of neck of femur (hip)
Congenital anomalies of limbs, hands, and feet
Acquired foot deformities
Cervical pain syndromes
Low back pain
Bursitis, synovitis, tenosynovitis
Amputation status
Musculoskeletal disorders, other
Neonatal
Newborn Status, Uncomplicated
Newborn Status, Complicated
Low Birth Weight
Enrollees
(0 to 64 Years)
5.6
11.7
6.1
3.3
2.7
4.3
1.3
0.5
0.1
0.5
0.2
0.7
1.3
1.1
0.1
1.8
0.1
0.3
0.4
46.8
16.8
19.0
25.0
3.0
1.9
0.9
4.1
1.4
9.7
7.5
12.0
1.6
39.0
1.2
1.0
8.5
141.9
76.7
28.7
20.1
1.5
4.8
0.2
5.1
33.1
0.3
2.9
10.5
69.0
120.5
49.3
0.4
45.3
296.0
78.0
174.1
40.3
1.1
8.0
0.0
4.6
40.8
6.8
2.7
36.9
79.1
197.5
100.4
2.0
125.0
7.3
0.7
0.5
0.1
0.2
0.0
Prematurity
Disorders of Newborn Period
Neurologic
Neurologic signs and symptoms
Headaches
Peripheral neuropathy, neuritis
Vertiginous syndromes
Cerebrovascular disease
Parkinson's disease
Seizure disorder
Multiple sclerosis
Muscular dystrophy
Sleep problems
Dementia and delirium
Quadriplegia and paraplegia
Head injury
Spinal cord injury/disorders
Paralytic syndromes, other
Cerebral palsy
Developmental disorder
Central nervous system infections
Neurologic disorders, other
Migraine Headaches
Nutrition
Failure to thrive
Nutritional deficiencies
Obesity
Nutritional disorders, other
Psychosocial
Anxiety, neuroses
Substance use
Tobacco use
Behavior problems
Attention deficit disorder
Family and social problems
Schizophrenia and affective psychosis
Personality disorders
Depression
Psychologic signs and symptoms
Psychosocial disorders, other
Bipolar disorder
4-11
Enrollees
(0 to 64 Years)
0.9
1.5
0.0
0.5
14.6
44.3
33.7
20.3
6.9
1.9
6.1
1.8
0.4
21.9
1.7
0.3
7.5
1.6
0.9
0.4
2.9
0.8
6.0
18.4
67.3
41.4
82.3
68.8
95.8
14.3
13.1
2.0
1.4
30.3
44.3
0.8
14.8
4.7
6.9
0.2
0.4
2.4
18.6
7.9
2.7
5.5
20.2
0.1
3.5
24.4
27.3
0.2
55.6
8.0
21.6
4.1
15.0
1.9
4.7
1.1
47.3
2.2
2.5
5.5
47.7
6.2
20.7
4.1
0.7
0.9
9.1
2.2
53.4
1.4
4.8
3.6
4-12
Reconstructive
Cleft lip and palate
Lacerations
Chronic ulcer of the skin
Burns--2nd and 3rd degree
Renal
Fluid/electrolyte disturbances
Acute renal failure
Nephritis, nephrosis
Renal disorders, other
Respiratory
Respiratory signs and symptoms
Acute lower respiratory tract infection
Cystic fibrosis
Emphysema, chronic bronchitis, COPD
Cough
Sleep apnea
Sinusitis
Pulmonary embolism
Tracheostomy
Respiratory failure
Respiratory disorders, other
Rheumatologic
Autoimmune and connective tissue diseases
Gout
Arthropathy
Raynaud's syndrome
Rheumatoid arthritis
Skin
Contusions and abrasions
Dermatitis and eczema
Keloid
Acne
Disorders of sebaceous glands
Sebaceous cyst
Viral warts and molluscum contagiosum
Other inflammatory conditions of skin
Exanthems
Skin keratoses
Dermatophytoses
Enrollees
(0 to 64 Years)
0.2
25.2
2.1
2.2
0.0
33.1
22.2
1.9
3.4
15.6
1.3
1.1
3.8
53.9
76.5
21.2
5.0
13.8
45.2
74.2
0.2
10.1
53.6
14.5
109.0
0.9
0.2
6.7
10.1
151.1
111.9
0.2
95.7
89.2
27.4
71.0
6.4
1.2
46.2
50.8
5.7
4.8
11.7
0.8
4.7
17.3
23.2
52.0
1.6
17.3
35.4
55.7
1.4
28.3
4.8
10.8
22.2
7.5
28.4
26.0
18.1
50.0
79.9
1.9
12.5
11.1
20.0
12.7
18.0
33.6
142.1
76.3
4-13
Enrollees
(0 to 64 Years)
Psoriasis
4.5
8.6
Disease of hair and hair follicles
9.9
7.0
Pigmented nevus
1.6
2.0
Scabies and pediculosis
1.2
0.7
Diseases of nail
7.7
36.2
Other skin disorders
29.3
62.3
Benign neoplasm of skin and subcutaneous tissues
46.4
94.0
Impetigo
3.9
1.3
Dermatologic signs and symptoms
16.4
31.4
Toxic Effects and Adverse Events
Toxic effects of nonmedicinal agents
2.8
2.8
Adverse effects of medicinal agents
4.5
10.0
Adverse events from medical/surgical procedures
3.4
11.4
Complications of mechanical devices
3.4
18.7
Diagnostic Certainty
The recording of provisional diagnosis codes can improperly attribute chronic conditions
to a patient subsequently misidentifying individuals for care management programs and
elevating cost predictions. The goal of stringent diagnostic certainty is to provide greater
certainty of a given diagnosis for subset of chronic conditions.
When the user applies the lenient diagnostic certainty option, any single diagnosis
qualifying for an EDC marker will turn the marker on. However, the user may also apply
a stringent diagnostic certainty option. For a subset of diagnoses, there must be more
than one diagnosis qualifying for the marker in order for the EDC to be assigned. This
was designed to provide greater confidence in the EDC conditions assigned to a patient.
For several conditions, namely Hypertension, Asthma, Diabetes and Pregnancy, there are
multiple EDCs to describe the condition. When applying stringent diagnostic certainty, if
multiple diagnoses are required, they must qualify for the same EDC. For example, a
single diagnosis for Type 1 diabetes and a single diagnosis for Type 2 diabetes will not
turn on any diabetes-related EDC, but 2 diagnoses for Type 2 diabetes will turn on EDC
END06. For more information about the inclusion and exclusion of diagnoses, refer to
the Installation and Usage Guide, Chapter 4.
4-14
MEDC Types
To assist analysts, the 27 Major EDCs may be further aggregated into five MEDC Types.
Specifically, the categorization is presented in Table 2.
Table 4: MEDC Types
MEDC-Type
1.
2.
Administrative
Medical
3.
Surgical
4.
5.
Obstetric/Gynecologic
Psychosocial
MEDCs
Administrative
Allergy, Cardiovascular, Endocrine, Gastrointestinal/Hepatic, General
Signs and Symptoms, Genetic, Hematologic, Infections, Malignancies,
Neonatal, Neurologic, Nutrition, Renal, Respiratory, Rheumatologic,
Skin, Toxic Effects
Dental, ENT, Eye, General Surgery, Genito-urinary, Musculoskeletal,
Reconstructive
Female reproductive
Psychosocial
In summary, EDCs can be examined at three levels. The broadest level is the MEDC-type
and includes five categories. MEDC-Types are based on 27 clinically oriented groupings
of MEDCs. The full EDC taxonomy is composed of 267 groups.
4-15
Some Important Differences Between EDCs and ADGs
Both EDCs and Aggregated Diagnosis Groups (ADGs) are aggregations of ICD codes.
However, there is a significant difference in the methodology underlying the grouping of
ICD codes: ADGs, the building block of the ACG System, are groups of diagnoses with
similar expected healthcare need, while EDCs are clinically similar clusters. When
applying the EDCs, it is important to recognize that the criteria used to assign ICD codes
are quite different from the criteria used to determine the appropriate ADG. The main
criterion used for the ICD-to-EDC assignment is diagnostic similarity, whereas the ICD
to ADG assignments (and the overall ACG System in general) are based on a unique set
of clinical criteria that captures various dimensions of the range and severity of an
individual’s co-morbidity. This subject is discussed more detail in the chapter entitled,
“Adjusted Clinical Groups.”
In brief, the key dimensions of the ICD to ADG assignment process are:
Expected duration of illness (e.g., acute, chronic, or recurrent)
Severity (e.g., expected prognosis with respect to disability or longevity)
Diagnostic certainty (i.e., signs/symptoms versus diagnoses)
Etiology (e.g., infectious, neoplastic, psychosocial conditions)
Expected need for specialist care or hospitalization
In the ADG assignment process, ICD codes representing multiple diseases and conditions
may be assigned to the same ADG. This occurs when the diseases are expected to have a
similar impact on the need for healthcare resources. In contrast, the EDC assignment does
not account for differences in disease severity, chronicity, or the expected need for
resources.
Because of the marked differences between the conceptual underpinnings of EDC and the
ADG/ACG framework, each ADG will usually contain ICD codes that map into more
than one EDC. (The notable exception is the asthma ADG and asthma EDC, which
contain the same ICD codes). Likewise, a single EDC may map into multiple ADGs.
Important Differences Between EDCs and Episode-of-Care Systems
It is important to recognize that while the EDC tool is effective for identifying persons
with specific conditions, the current configuration of the EDC tool is not an episode-ofcare grouping methodology (i.e., while EDCs can be used to identify persons under
treatment for otitis media (or any select condition), there is no count of the number of
occurrences otitis media infections during the analysis period). While episodes-of-care
tools also consider ICD codes, their main purpose is to group together services,
procedures, drugs, and tests that are provided to a patient in the course of management of
a specific disease or condition. Currently the EDC assignment is based only on diagnosis
codes. In the future, as the software’s input stream becomes more sophisticated and date-
4-16
of-service and/or other data elements become available, the functionality of the EDC
module may be increased.
Sophisticated analysts wishing to process data of shorter durations than the typical 12
month time period could, for example, process data monthly to create monthly-EDCindicators. Summing these indicators for each patient by each EDC could provide rough
approximations of the number of times a patient received services related to specific
EDCs and proxy clinical treatment groups could be envisioned.
Applications of EDCs
EDCs have many applications, particularly in areas of profiling and disease/case
management. EDCs can be used to:
1. Describe the prevalence of specific diseases within a single population;
2. Compare disease distributions across two or more populations; and,
3. Aid disease management/case management processes by identifying individual
patients by condition and displaying a patient condition profile.
Each of these applications is highlighted within the Applications Guide chapters for
Health Status Monitoring and Clinical Screening by Care and Disease Managers.
Medication Defined Morbidity
5-i
5 Medication Defined Morbidity
Objectives ....................................................................................................... 5-1
National Drug Code System ....................................................................... 5-1
Anatomical Therapeutic Chemical (ATC) System .................................... 5-2
Table 1: ATC Drug Classification Levels ................................................. 5-2
Rx-Defined Morbidity Groups ................................................................... 5-3
Table 2: Counts of Associated Therapeutic Classes per Rx-Defined
Morbidity Groups ....................................................................................... 5-4
Table 3: Counts of Associated Second Level ATCs per Rx-Defined
Table 4: Counts of Associated Third Level ATCs per Rx-Defined
Table 5: Major Rx-MG Categories ........................................................... 5-7
NDC to Rx-MG Assignment Methodology ............................................... 5-9
ATC to Rx-MG Assignment Methodology ................................................ 5-9
Table 6: Number of Pharmacy Codes for Each Rx-MG ......................... 5-10
Table 7: Rx-MG Morbidity Taxonomy and Clinical Characteristics of
Medications Assigned to Each Rx-MG ................................................... 5-14
Table 8: Annual Prevalence Ratios for Rx-MGs for Commercially Insured
and Medicare Beneficiaries ...................................................................... 5-20
Generic Drug Count Variable .................................................................. 5-22
Conclusion.................................................................................................... 5-23
5-ii
5-1
Objectives
This chapter is intended to describe the pharmacy based morbidity markers, Rx-MGs and
the clinical criteria used to assign medications to morbidity groups. The Rx-MGs provide
further methods to describe the unique morbidity profile of a population and form the
basis of the pharmacy based predictive model, Rx-PM. Specifics of the predictive model
are discussed in greater detail in later chapters discussing Predicting Resource Use and
Predicting Hospitalization.
National Drug Code System
The National Drug Code (NDC) system serves as a universal product identifier for
human drugs. It is maintained by the Food and Drug Administration (FDA) of the US
government. An NDC identifies the following attributes of a medication: generic
name/active ingredient; manufacturer; strength; route of administration; package size;
and, trade name. The same generic drug generally has many NDC codes, because
different manufacturers may produce it, it may come in several strengths, package sizes
can vary, etc. Not all the NDC attributes are useful for defining the type of morbidity
that the drug is intended to influence. Manufacturer and trade name provide no clinical
information, whereas the generic name of the drug gives essential information on its
chemical structure and known mechanism(s) of action. The same generic drug can be
given via different routes (e.g., by mouth, inhaled, topical). In many cases, the route of
administration is critical for identifying the specific morbidity managed. A good
example of this is the drug class of corticosteroids, which are powerful anti-inflammatory
medications. Oral corticosteroids can be used to manage pulmonary disease that result
from airway inflammation, such as asthma, whereas topical forms of the medication are
used to manage inflammatory skin conditions, such as eczema. Although medication
strength may provide some indication of disease stability, it may also be suggestive of
patient non-adherence to medication regimens, inappropriate over-treatment, patient
idiosyncratic response, or size of the patient. For these reasons, we elected to ignore
strength as a classification criterion, and restricted the assignment process to unique
generic drug-route of administration combinations. In so doing, we reduced the nearly
90,000 NDCs to approximately 2,700 units. Each generic drug-route of administration
has a set of NDCs associated with it.
5-2
Anatomical Therapeutic Chemical (ATC) System
The Anatomical Therapeutic Chemical (ATC) System serves as the World Health
Organization (WHO) standard for drug consumption studies. It is maintained by the
WHO Collaborating Centre for Drug Statistics Methodology.
In the Anatomical Therapeutic Chemical (ATC) classification system, the drugs are
divided into different groups according to the organ or system on which they act in
addition to their chemical, pharmacological, and therapeutic properties. Drugs are
classified into five different group levels. The drugs are divided into fourteen main
groups (first level), with one pharmacological/therapeutic subgroup (second level). The
third and fourth levels are chemical/pharmacological/therapeutic subgroups and the 5th
level is the chemical substance. The second, third and fourth levels are often used to
identify pharmacological subgroups when that is considered more appropriate than
therapeutic or chemical subgroups.
The complete classification of metformin illustrates the structure of the code as show in
Table 1:
Table 1: ATC Drug Classification Levels
Level
Description
A
Alimentary tract and metabolism
(1st level, anatomical main group)
A10
Drugs used in diabetes
(2nd level, therapeutic subgroup)
A10B
Blood glucose lowering drugs, excluding insulins
(3rd level, pharmacological subgroup)
A10BA
Biguanides
(4th level, chemical subgroup)
A10BA02
Metformin
(5th level, chemical substance)
Therefore, in the ATC system, all plain metformin preparations are given the code
A10BA02.
5-3
ATCs are classified based upon the therapeutic use of the main active ingredient. Unlike
NDCs, the ATC classification has been specifically designed to capture therapeutic use.
In fact, the same generic substance can be given more than one ATC code if it is
available in two or more strengths or formulations with clearly different therapeutic uses.
However, an international classification system has the challenges of capturing countryspecific, main therapeutic use of a drug that often results in several different classification
alternatives. As a general guideline, the ATC system has attempted to assign such drugs
to one code, the main indication being decided on the basis of the available literature.
Rx-Defined Morbidity Groups
Rx-defined morbidity groups (Rx-MGs) represent the building blocks of a pharmacybased predictive model, Rx-PM. Each generic drug/route of administration combination
is assigned to a single Rx-MG. We found that generic drug/route of administration
combinations within therapeutic classes were sometimes logically assigned to different
Rx-MGs, which further reinforced the need make assignments at the individual drug
level.
Table 2 shows the number of therapeutic classes represented within Rx-MGs. Just six
Rx-MGs included a single therapeutic class, and about a third of all Rx-MGs included
medications from five or more therapeutic classes.
5-4
Table 2: Counts of Associated Therapeutic Classes per Rx-Defined
Morbidity Groups
Number of Therapeutic
Classes Represented in an
Rx-MG
Number of
Rx-MGs
1
2
3
4
5
6
7
8
9
11
12
16
26
6
12
12
11
6
5
3
1
3
1
2
1
1
Percent
9.38
18.75
18.75
17.19
9.38
7.81
4.69
1.56
4.69
1.56
3.13
1.56
1.56
Cumulative
Frequency
Cumulative
Percent
6
18
30
41
47
52
55
56
59
60
62
63
64
9.38
28.13
46.88
64.06
73.44
81.25
85.94
87.50
92.19
93.75
96.88
98.44
100.00
Although the Rx-MGs generally included medications from more than one therapeutic
class, the majority of therapeutic classes were assigned to a single Rx-MG or 2 Rx-MGs.
Specifically, 50% of therapeutic classes were assigned to a single Rx-MG, 28% had
medications grouped to 2 Rx-MGs, and 22% had medications assigned to 3 or more RxMGs.
5-5
Table 3 shows the number of secondlevel (therapeutic subgroup) ATC codes represented
within Rx-MGs. A little over half of the all Rx-MGs included medications from five or
more second level codes.
Table 3: Counts of Associated Second Level ATCs per Rx-Defined
Morbidity Groups
Number of 2nd Level
ATCs Represented in an
Rx-MG
1
2
3
4
5
6
7
8
9
11
12
33
Number of
Rx-MGs
15
10
11
7
6
1
3
1
3
1
1
1
Percent
25.00
16.67
18.33
11.67
10.00
1.67
5.00
1.67
5.00
1.67
1.67
1.67
Cumulative
Frequency
Cumulative
Percent
15
25
36
43
49
50
53
54
57
58
59
60
25.00
41.67
60.00
71.67
81.67
83.33
88.33
90.00
95.00
96.67
98.33
100.00
5-6
Table 4 shows the number of third level (pharmacological subgroup) ATC codes
represented within Rx-MGs. Half of the all Rx-MGs included medications from nine or
more third level codes.
Table 4: Counts of Associated Third Level ATCs per Rx-Defined
Morbidity Groups
Number of 3rd Level
ATCs Represented in an
Rx-MG
Number of
Rx-MGs
1
2
3
4
5
6
7
8
9
12
14
15
19
22
26
67
10
8
5
9
10
3
1
1
3
1
2
3
1
1
1
1
Percent
16.67
13.33
8.33
15.00
16.67
5.00
1.67
1.67
5.00
1.67
3.33
5.00
1.67
1.67
1.67
1.67
Cumulative
Frequency
Cumulative
Percent
10
18
23
32
42
45
46
47
50
51
53
56
57
58
59
60
16.67
30.00
38.33
53.33
70.00
75.00
76.67
78.33
83.33
85.00
88.33
93.33
95.00
96.67
98.33
100.00
The specific clinical criteria that we used to assign medications to an Rx-MG category
were the primary anatomico-physiological system, morbidity differentiation, expected
duration, and severity of the morbidity type being targeted by the medication. These four
clinical dimensions not only characterize medications by morbidity type, they also have
major consequences for predictive modeling. Higher levels of differentiation, chronicity,
and greater severity would all be expected to increase resource use. Each of these criteria
is discussed below.
Primary Anatomico-Physiological System. The Rx-MG classification system is
organized into 19 Major Rx-MG categories: 16 anatomico-physiological groupings; 1
general signs and symptoms category; 1 toxic effects/adverse events group; and 1 other
and non-specific medications category. Table 5 lists these categories. Medications
assigned to the non-specific category are not reflective of any underlying morbidity type;
examples of these medication types are diagnostic agents and multi-vitamins.
5-7
Table 5: Major Rx-MG Categories
Allergy/Immunology
Cardiovascular
Ears-Nose-Throat
Endocrine
Eye
Female Reproductive
Genito-Urinary
Hematologic
Infections
Malignancies
Musculoskeletal
Neurologic
Psychosocial
Respiratory
Skin
Toxic effects/adverse events
Other and non-specific medications
5-8
Morbidity Differentiation. Very few medications are used for a single disease. Even
insulin treatment, which is employed in the management of diabetes, does not separate
patients into Type 1 and Type 2 diabetes, because it can be used to manage either form.
A single medication is often administered for multiple clinical indications; thus, a
medication classification system that is solely disease-based will not validly reflect
patient morbidity. Rx-MGs assign medications to a specific disease when there is a
logical 1-to-1 assignment (e.g., digoxin  congestive heart failure). However, most
medications are assigned to broader morbidity groups, which are reflective of the
patient’s underlying patho-physiology. Furthermore, a large amount of medication
treatment is for physically (e.g., pain medications) and emotionally (e.g., anxiolytics)
experienced body sensations. These symptoms are undifferentiated morbidities that
generally require palliative therapy. In contrast, specific medical diseases are more
differentiated morbidity types. Because we sought to create a comprehensive
classification system of medications, the Rx-MGs include categories for symptoms
(assigned to an acute minor category within an organ system grouping or to the general
signs and symptoms category), fully developed diseases that have a 1-to-1
correspondence with medication, and more general morbidity-types.
Expected Duration. This dimension refers to the expected time period that the
morbidity will require treatment and is used to characterize conditions as acute, recurrent,
or chronic. An acute condition is a time-limited morbidity that is expected to last less
than 12 months. The classic example of acute conditions is infections. Recurrent
conditions are episodic health problems. They tend to occur repeatedly, and over time.
Migraine headaches and gout are two examples of recurrent conditions. Each episode is
time-limited and, in isolation, could be considered an acute health problem. However,
the repetitive occurrences may span several years with asymptomatic inter-current
periods; this episodic relapsing nature of the morbidity type is the hallmark feature of
recurrent morbidities. Chronic conditions are persistent health states that generally last
longer than 12 months.
Severity. The severity of morbidities refers to somewhat different concepts for acute and
recurrent/chronic conditions. For acute problems, morbidity severity is related to the
expected impact of the condition on the physiological stability of the patient or the
patient’s functional status. Low impact conditions have minimal effects on functioning
or physiological stability, whereas high impact problems can have significant effects on
the ability of an individual to perform daily activities. For recurrent and chronic
conditions, morbidity severity is related to the stability of the problem over several years.
Without adequate treatment, unstable chronic conditions are expected to worsen over
time, whereas stable conditions are expected to change less rapidly.
5-9
NDC to Rx-MG Assignment Methodology
The Rx-MG classification system was developed with an a priori assumption that
morbidity categories needed to be broad enough to encompass multiple FDA-approved
and common off-label uses of a drug. Each unique generic medication/route of
administration combination was assigned to a single Rx-MG according to the clinical
criteria described above and the most common evidence-based use of the drug. Both
FDA approval and empirical data from the literature were accepted as evidence for these
assignments.
A team of Johns Hopkins physicians and pharmacist health professionals made the initial
NDC-to-Rx-MG assignment. Disagreements were reconciled during consensus meetings.
The face validity of the assignments was tested by having another team of physicians and
pharmacist health professionals examine the preliminary NDC-to-Rx-MG look-up table.
This second review led to a few modifications, but overall confirmed the validity of the
initial assignments.
All NDC codes representing drugs are assigned to a single Rx-MG, and just 9.8% are
mapped to the Non-Specific category. There are NDC codes representing medical
supplies, exclusively over-the counter medications and other non-pharmaceutical items
that are not assigned to a Rx-MG. The ACG team updates the assignment of new NDC
codes to Rx-MG regularly. Table 6 shows the distribution of the NDC codes assigned to
each Rx-MG. The pain and acute minor infections Rx-MGs have the most NDC codes
assigned to them. This is because these categories include some of the generic drugs with
the most NDC codes: ibuprofen ranked #1 with 1,336 NDCs; acetaminophen #2 with
1,253 NDCs; amoxicillin #6 with 689 NDCs; and, erythromycin #8 with 668 NDCs.
ATC to Rx-MG Assignment Methodology
The methodology for assigning ATC-to-Rx-MGs is largely the same as that used in
making NDC-to Rx-MGs assignments. However, there were four important differences
as summarized below:
1. Larger number of codes assigned. In the NDC system, the 100,000 codes can be
collapsed to roughly 2,700 unique drug route combinations which then get assigned
to a Rx-MG. The ATC assignment involved assigning close to 5,100 ATC codes (900
4th Level and nearly 4,200 5th Level ATC codes) to an Rx-MG.
2. International variability. Incorporating international variability and off label use
was also considered. Literature was extensively reviewed for evidence of alternate
and prevalent international use of medications. Whenever evidence showed that the
predominant use of the medication was different from the FDA approved use, the RxMG assignment was modified to reflect this.
3. Route of administration was not available for some ATC codes. Route is not
provided in the ATC dictionary for nearly 41% of the ATC codes. However, this was
not an impediment in establishing Rx-MG assignments due to the fact that the ATC
hierarchy is therapeutic centric; and, in almost every case, it provides much more
information than the drug/route combinations.
5-10
4. Use of original Drug/Route-to-Rx-MG Assignments. In order to preserve
assignment consistency from the NDC based Rx-MGs, Drug/Route-to-Rx-MG
assignments were systematically reviewed and were considered as a guide when
making ATC-to-Rx-MG assignments.
Once again, a two-step process for Rx-MG assignments was undertaken. A team of
international clinicians made the initial NDC-to-Rx-MG assignments. Consensus
meetings were held to reconcile disagreements. Afterwards a separate team of six
clinicians composed of Johns Hopkins physicians and pharmacists and an international
group of physicians made independent assignment to a randomly selected sample of 300
codes. The result of the review led to very few modifications. In fact, 99% of the codes
were assigned to the same Rx-MG by the majority of the clinician panel (83% of the
codes were assigned unanimously by the six clinicians to a single Rx-MG, 10% received
the same Rx-MG assignment by five of the six clinicians, 6% were assigned to the same
Rx-MG by four of the six clinicians, and the remaining 1% was assigned to the same
assignment by three of the six clinicians).
Table 6: Number of Pharmacy Codes for Each Rx-MG
Rx-Defined Morbidity Group
% of All
NDC
Codes
% of 4th
and 5th
Level ATC
Codes.
Allergy/Immunology
Acute Minor
2.36
1.54
Chronic Inflammatory
1.26
0.40
Immune Disorders
0.15
0.18
Transplant
0.08
0.32
Chronic Medical
3.15
2.61
Congestive Heart Failure
0.56
1.03
High Blood Pressure
7.18
6.08
Disorders of Lipid Metabolism
1.19
1.01
Cardiovascular / Vascular Disorders
0.80
2.40
0.35
1.01
Cardiovascular
Ears, Nose, Throat
Acute Minor
Endocrine
5-11
% of All
NDC
Codes
% of 4th
and 5th
Level ATC
Codes.
Bone Disorders
0.11
0.83
Chronic Medical
0.95
2.26
Diabetes With Insulin
0.10
0.67
Diabetes Without Insulin
1.52
0.79
Thyroid Disorders
1.00
0.46
Growth Problems
0.06
0.00
Weight Control
0.63
0.00
Acute Minor: Curative
0.61
1.86
Acute Minor: Palliative
0.62
1.48
Glaucoma
0.66
0.91
Hormone Regulation
0.62
1.11
Infertility
0.08
0.30
Pregnancy and Delivery
0.55
0.67
Acute Minor
4.18
4.47
Chronic Liver Disease
0.18
0.73
Chronic Stable
0.02
0.44
Inflammatory Bowel Disease
0.12
0.30
Pancreatic Disorder
0.16
0.10
Peptic Disease
1.65
1.03
Nausea and Vomiting
1.64
0.59
Pain
8.48
3.66
Pain and Inflammation
6.46
3.58
Eye
Female Reproductive
5-12
Severe Pain
% of All
NDC
Codes
% of 4th
and 5th
Level ATC
Codes.
1.45
0.00
Acute Minor
1.32
0.97
Chronic Renal Failure
0.09
0.16
0.24
0.38
Acute Major
0.27
2.77
Acute Minor
9.23
7.44
HIV/AIDS
0.34
0.69
Tuberculosis
0.27
0.63
1.24%
0.00
1.00
4.26
Gout
0.32
0.30
Inflammatory Conditions
0.62
0.77
Alzheimer's Disease
0.16
0.30
Chronic Medical
0.09
1.01
Migraine Headache
0.25
0.48
Parkinson's Disease
0.62
0.87
Seizure Disorder
2.26
0.97
Genito-Urinary
Hematologic
Coagulation Disorders
Infections
Severe Acute Major
Malignancies
Musculoskeletal
Neurologic
Psychosocial
5-13
% of All
NDC
Codes
% of 4th
and 5th
Level ATC
Codes.
Acute Minor
1.13
0.16
Addiction
0.12
0.34
Anxiety
1.80
0.89
Attention Deficit Hyperactivity Disorder
0.37
1.33
Chronic Unstable
2.07
1.33
Depression
4.73
1.41
Acute Minor
5.49
2.77
Airway Hyperreactivity
1.50
0.48
Chronic Medical
0.10
0.04
Cystic Fibrosis
0.08
2.06
Acne
0.60
0.83
Acute and Recurrent
6.18
6.93
Chronic Medical
0.10
0.65
0.32
0.89
8.50
15.05
140,335
4,949
Respiratory
Skin
Toxic Effects/Adverse Effects
Acute Major
Other and Non-Specific Medications
Total Number of Pharmacy Codes
The final set of 64 Rx-MG categories, along with their distinguishing clinical
characteristics, is shown in Table 7.
5-14
Table 7: Rx-MG Morbidity Taxonomy and Clinical Characteristics of Medications Assigned to Each
Rx-MG
Severity
Differentiation
Duration
Acute Conditions:
Impact
Symptoms/General
Acute
Low Impact
General
Chronic
Unstable
Immune Disorders
General
Chronic
Unstable
Transplant
General
Chronic
Unstable
Chronic Medical
General
Chronic
Stable
Disease
Chronic
Unstable
High Blood Pressure
Disease
Chronic
Stable
Disease
Chronic
Stable
Vascular Disorders
General
Chronic
Stable
Symptoms/General
Acute
Chronic and Recurrent
Conditions: Stability
Allergy/Immunology
Acute Minor
Cardiovascular
Ears-Nose-Throat
Acute Minor
Low Impact
5-15
Severity
Acute Conditions:
Impact
Differentiation
Duration
Bone Disorders
General
Chronic
Stable
Chronic Medical
General
Chronic
Stable
Disease
Chronic
Unstable
Disease
Chronic
Stable
Thyroid Disorders
General
Chronic
Stable
Growth Problems
General
Chronic
Stable
Weight Control
General
Chronic
Stable
General
Acute
Low Impact
Symptoms
Acute
Low Impact
Disease
Chronic
Stable
Hormone Regulation
General
Chronic
Stable
Infertility
General
Acute
Endocrine
Eye
Glaucoma
Female Reproductive
High Impact
5-16
Severity
Differentiation
Duration
Acute Conditions:
Impact
General
Acute
High Impact
Symptoms/General
Acute
Low Impact
General
Chronic
Unstable
Chronic Stable
General
Chronic
Stable
Disease
Chronic
Stable
Pancreatic Disorder
General
Chronic
Unstable
Peptic Disease
Disease
Recurrent
Stable
Nausea and Vomiting
Symptoms
Acute
Low Impact
Pain
Symptoms
Acute
Low Impact
Symptoms
Acute
Low Impact
Severe Pain
Symptoms
Acute
High Impact
Acute Minor
5-17
Severity
Differentiation
Duration
Acute Conditions:
Impact
Symptoms/General
Acute
Low Impact
Disease
Chronic
Unstable
General
Chronic
Unstable
Acute Major
General
Acute
High Impact
Acute Minor
General
Acute
Low Impact
HIV/AIDS
Disease
Chronic
Tuberculosis
Disease
Acute
Low Impact
Severe Acute Major
General
Acute
High Impact
General
Chronic
Genito-Urinary
Acute Minor
Hematologic
Infections
Unstable
Malignancies
Malignancies
Unstable
5-18
Severity
Acute Conditions:
Impact
Differentiation
Duration
Gout
Disease
Recurrent
Stable
General
Chronic
Unstable
Alzheimers Disease
Disease
Chronic
Unstable
Chronic Medical
General
Chronic
Stable
Migraine Headache
Disease
Recurrent
Stable
Parkinsons Disease
Disease
Chronic
Unstable
Seizure Disorder
General
Chronic
Stable
Attention Deficit Disorder
Disease
Chronic
Stable
Addiction
General
Chronic
Unstable
Symptoms/General
Recurrent
Stable
Disease
Chronic
Stable
Symptoms/General
Acute
Musculoskeletal
Neurologic
Psychosocial
Anxiety
Depression
Acute Minor
Low Impact
5-19
Severity
Acute Conditions:
Impact
Differentiation
Duration
General
Chronic
Symptoms/General
Acute
Airway Hyper-reactivity
Disease
Chronic
Stable
Chronic Medical
General
Chronic
Stable
Cystic Fibrosis
Disease
Chronic
Unstable
Disease
Chronic
Stable
Symptoms/General
Acute and
Recurrent
General
Chronic
General
Chronic
High Impact
N/A
N/A
N/A
Chronic Unstable
Unstable
Respiratory
Acute Minor
Low Impact
Skin
Acne
Acute and Recurrent
Chronic Medical
Low Impact
Stable
Acute Major
Other and Non-Specific
Medications
N/A
5-20
The frequency of occurrence of each Rx-MG in a commercially insured and Medicare
population is shown in Table 8.
Table 8: Annual Prevalence Ratios for Rx-MGs for Commercially
Insured and Medicare Beneficiaries
Number of Persons with Rx-MG per
1,000 Enrollees
Rx-Defined Morbidity Group (Rx-MG)
Allergy/Immunology
Acute Minor
Immune Disorders
Transplant
Commercially
Insured
(0 to 64 Years)
Medicare
Beneficiaries
(65 Years and
Older)
111.7
76.9
0.3
0.9
136.1
101.6
1.2
1.7
15.9
8.0
129.3
82.9
12.2
164.1
71.0
589.4
433.2
141.1
23.2
16.6
11.9
33.1
9.0
30.1
37.2
2.2
0.2
116.0
58.2
39.7
142.3
134.6
0.9
0.0
42.7
19.1
4.6
70.9
55.8
67.7
70.8
0.8
4.1
0.0
Cardiovascular
Chronic Medical
High Blood Pressure
Vascular Disorders
Ears, Nose, Throat
Acute Minor
Endocrine
Bone Disorders
Chronic Medical
Thyroid Disorders
Weight Control
Growth Problems
Eye
Glaucoma
Female Reproductive
Hormone Regulation
Infertility
5-21
1,000 Enrollees
Commercially
Insured
(0 to 64 Years)
Medicare
Beneficiaries
(65 Years and
Older)
11.0
0.4
41.2
0.7
0.5
3.0
0.5
69.1
111.2
1.7
1.4
6.7
2.0
214.1
44.3
194.4
104.3
16.3
62.3
275.9
159.2
28.8
29.4
0.4
155.4
5.6
0.2
0.9
3.5
411.1
1.0
1.1
0.3
11.7
446.5
0.5
1.2
1.3
4.3
26.5
5.8
6.0
36.9
16.2
0.3
2.9
30.3
3.4
Acute Minor
Chronic Stable
Pancreatic Disorder
Peptic Disease
Nausea and Vomiting
Pain
Severe Pain
Genito-Urinary
Acute Minor
Hematologic
Infections
Acute Major
Acute Minor
HIV/AIDS
Tuberculosis
Severe Acute Major
Malignancies
Musculoskeletal
Gout
Neurologic
Alzheimer's Disease
Chronic Medical
5-22
1,000 Enrollees
Migraine Headache
Parkinson's Disease
Seizure Disorder
Commercially
Insured
(0 to 64 Years)
Medicare
Beneficiaries
(65 Years and
Older)
15.3
4.0
30.8
5.0
19.7
64.1
44.7
2.3
51.5
19.3
8.3
107.2
73.8
1.5
85.5
2.6
17.7
154.9
114.2
78.2
4.9
0.2
97.5
108.6
44.0
2.3
25.2
112.2
3.3
6.1
193.8
3.0
0.1
0.4
37.7
138.4
Psychosocial
Acute Minor
Addiction
Anxiety
Attention Deficit Hyperactivity Disorder
Chronic Unstable
Depression
Respiratory
Acute Minor
Airway Hyperactivity
Chronic Medical
Cystic Fibrosis
Skin
Acne
Acute and Recurrent
Chronic Medical
Acute Major
Generic Drug Count Variable
A generic drug count is calculated as the count of unique generic drug (active
ingredient)/route of administration combinations encountered in the patient’s drug claims.
This marker is a proxy for identifying poly-pharmacy members and contributes
independently and significantly to the prediction of cost. This marker is incorporated into
all pharmacy-based predictive models (Rx-PM and DxRx-PM). Members with a generic
drug count of 13 or greater will get additional weight in the predictive models.
5-23
Conclusion
The Rx-MGs were created using a clinical framework that makes sense to health
professionals and medical managers. As we will discuss in the Chapters on Predicting
Resource Use and Predicting Hospitalizations, pharmacy derived morbidity markers add
a great deal of clinical context and improved statistical performance when applied to
predictive modeling applications. This page was left blank intentionally.
5-24
Special Population Markers
6-i
6 Special Population Markers
Introduction ................................................................................................... 6-1
Hospital Dominant Conditions .................................................................... 6-1
Table 1: Examples of Diagnostic Codes Included in the Hospital
Dominant Marker ....................................................................................... 6-1
Table 2: Effects of Hospital Dominant Conditions During a Baseline
Year on Next Year’s Hospitalization Risk, Total Healthcare Costs,
and Pharmacy Costs ................................................................................... 6-2
Frailty Conditions ......................................................................................... 6-3
Table 3: Medically Frail Condition Marker – Frailty Concepts and
Diagnoses ................................................................................................... 6-3
Table 4: Effects of Frailty Conditions During a Baseline Year on Next
Year’s Hospitalization Risk, Total Healthcare Costs, and Pharmacy
Costs ........................................................................................................... 6-4
Chronic Condition Count ............................................................................. 6-5
Table 5: EDCs considered in the Chronic Condition Count Marker ........ 6-5
Table 6: Distribution of the Chronic Condition Count Marker ................. 6-8
Condition Markers ........................................................................................ 6-9
Table 7: Definitions of Condition Markers ............................................. 6-10
6-ii
6-1
Introduction
The ACG System includes a number of special population markers to identify patient
populations requiring specialized care. These markers enhance the clinical screening
process by providing meaningful filtering criteria and clinical context. Several of the
markers are also independent variables in the predictive models.
Hospital Dominant Conditions
Hospital dominant conditions are based on diagnoses that, when present, are associated
with a greater than 50 percent probability among affected patients of hospitalization in
the next year. All these diagnoses are setting-neutral, i.e., they can be given in any
inpatient or outpatient face-to-face encounter with a health professional. The variable is a
count of the number of morbidity types (i.e., ADGs) with at least one hospital dominant
diagnosis. When the stringent diagnostic certainty option is selected, more than one
diagnosis from the hospital dominant condition list may be required for the marker to be
turned on. Table 1 provides examples of diagnostic codes that have been identified as
hospital dominant.
Table 1: Examples of Diagnostic Codes Included in the Hospital
Dominant Marker
Individuals with any one of these diagnostic codes have a greater than 50% chance of
hospitalization in the subsequent year.
ICD-9-CM Code
Description
162.3
Malignant Neoplasm, Upper Lobe, Bronchus or Lung
261
Nutritional Marasmus
2638
Other Protein Calorie Malnutrition Nec
284.8
Other specified Aplastic Anemia
2894
Hypersplenism
29181
Alcohol Withdrawal
29643
Bipolar affective disorder, manic, severe, without mention of psychotic
behavior
0380
Streptococcal Septicemia
4150
Acute cor pulmonale
4821
Pseudomonal Pneumonia
491.21
Obstructive Chronic Bronchitis with acute exacerbation
518.81
Acute Respiratory Failure
5722
Hepatic coma
584.9
Acute Renal Failure, unspecified
6-2
ICD-9-CM Code
Description
785.4
Gangrene
7895
Ascites
The risk of hospitalization, total healthcare costs, and medication costs rise dramatically
in association with the number of hospital dominant conditions (See Table 2). The count
of hospital dominant conditions is a powerful predictor of greater resource use next year.
Commercially insured individuals who have a single hospital dominant condition have a
5-fold greater chance of having at least 1 hospitalization next year, compared with
individuals without such a condition. There is a 10-fold increase for individuals in the
group with 2 or more hospital dominant conditions. For commercially insured
individuals and Medicare beneficiaries alike, significant increases in healthcare costs are
associated with the presence of one or more hospital dominant conditions.
Table 2: Effects of Hospital Dominant Conditions During a Baseline
Year on Next Year’s Hospitalization Risk, Total Healthcare Costs, and
Pharmacy Costs
HOSPITAL DOMINANT CONDITIONS
Next Year’s Outcomes
Baseline Year Risk Factors
% with 1+
hospitalization
Mean total
healthcare costs
Mean pharmacy
costs
Commercially Insured Individuals (<65 years-old)
Number of hospital dominant
conditions
None
1
2+
4.2
$1,875
$445
20.4
$12,652
$2,342
45.8
$35,802
$4,098
Medicare Beneficiaries (65 years and older)
Number of hospital dominant
conditions
None
1
2+
14.4
$5,189
$1,007
35.3
$14,810
$2,112
55.0
$28,407
$2,361
Medicare beneficiaries (65 years and older), 2001-2002.
6-3
Frailty Conditions
The Medically frail condition marker is a dichotomous (on/off) variable that indicates
whether an enrollee has a diagnosis falling within any 1 of 12 clusters that represent
medical problems associated with frailty. When the stringent diagnostic certainty option
is selected, more than one diagnosis from the frailty condition list may be required for the
marker to be turned on. Examples of these problems are shown in Table 3. In
collaboration with a team of Johns Hopkins Geriatricians, we identified diagnostic codes
that are highly associated with marked functional limitations among older individuals
(i.e., frailty). Presence of any one of these diagnoses turns on the FRAILTY indicator
(yes/no) variable. Among commercially insured populations, about 2% of individuals
have at least one of these frailty diagnoses, whereas among Medicare beneficiaries the
proportion is 4%.
Table 3: Medically Frail Condition Marker – Frailty Concepts and
Diagnoses
Frailty Concept
Malnutrition
Dementia
Impaired Vision
Decubitus Ulcer
Incontinence of Urine
Loss of Weight
Incontinence of Feces
Obesity (morbid)
Poverty
Barriers to Access of Care
Difficulty in Walking
Fall
Diagnoses (Examples)
Nutritional Marasmus
Other severe protein-calorie malnutrition
Senile dementia with delusional or depressive features
Senile dementia with delirium
Profound impairment, both eyes
Moderate or severe impairment, better eye/lesser eye: profound
Decubitus Ulcer
Incontinence without sensory awareness
Continuous leakage
Abnormal loss of weight and underweight
Feeding difficulties and mismanagement
Incontinence of feces
Morbid obesity
Lack Of Housing
Inadequate Housing
Inadequate material resources
No Med Facility For Care
No Med Facilities Nec
Difficulty in walking
Abnormality of gait
Fall On Stairs Or Steps
Fall From Wheelchair
The risk of hospitalization, total healthcare costs, and medication costs rise in association
with a frailty condition (See Table 4).
6-4
Table 4: Effects of Frailty Conditions During a Baseline Year on Next
Year’s Hospitalization Risk, Total Healthcare Costs, and Pharmacy
Costs
FRAILTY CONDITIONS
Next Year’s Outcomes
Baseline Year Risk Factors
% with 1+
hospitalization
Mean total
healthcare costs
Mean pharmacy
costs
Number of frailty conditions:
None
15.3
$5,819
$1,065
1+
30.4
$9,208
$1,349
6-5
Chronic Condition Count
The ACG System includes a chronic condition count as an aggregate marker of case
complexity. A chronic condition is an alteration in the structures or functions of the body
that is likely to last longer than twelve months and is likely to have a negative impact on
health or functional status.
The ACG System defines a limited set of Expanded Diagnosis Clusters (EDCs) that
represent high impact and chronic conditions likely to last more than 12 months with or
without medical treatment (see Table 5). From this list of EDCs, individual diagnosis
codes were tested against the criteria for chronic conditions stated above. The diagnoses
identified by these EDCs were compared against a chronic condition list provided by the
Center for Child and Adolescent Health Policy, MassGeneral Hospital for Children, in
Boston, Massachusetts. Differences between the lists were identified as psychological or
medical/surgical conditions. The psychological codes were reviewed by a practicing
psychologist and health services researcher for congruence with the operational
definition. Several anxiety and substance abuse codes were deleted from the chronic
condition list as a result of this review. Several acute conditions of the retina were
deleted from the chronic condition list after a similar review of medical/surgical
conditions. The Center for Child and Adolescent Health Policy list and the ACG System
differ in definitions related to infectious diseases such as tuberculosis, peptic ulcer
disease, congenital heart disease (which is generally resolved through surgical
interventions at birth), gastrointestinal obstructions and perforations (likely to be acute
and treatable conditions), osteomyelitis and prematurity. These conditions are not
considered chronic conditions in the ACG System chronic condition marker.
The chronic diagnosis codes that were identified were further classified and aggregated
into Expanded Diagnosis Cluster (EDC) categories. The chronic diagnosis codes
triggered a chronic EDC flag (see Table 5).
Table 5: EDCs considered in the Chronic Condition Count Marker
EDC Description
Acute hepatitis
Acute leukemia
Adverse events from medical/surgical procedures
Age-related macular degeneration
Anxiety, neuroses
Aplastic anemia
Asthma, with status asthmaticus
Behavior problems
Benign and unspecified neoplasm
Bipolar disorder
Blindness
Cardiac arrhythmia
GAS04
MAL16
TOX03
EYE15
PSY01
HEM05
ASTH
ASTH
PSY05
RHU01
PSY04
GSU03
PSY12
EYE02
CAR09
6-6
EDC Description
Cardiomyopathy
Cataract, aphakia
Cerebral palsy
Bipolar disorder
Blindness
Cardiac arrhythmia
Cardiomyopathy
Cataract, aphakia
Cerebral palsy
Chronic cystic disease of the breast
Congenital anomalies of limbs, hands, and feet
Cystic fibrosis
Deafness, hearing loss
Depression
Diabetic retinopathy
Disorders of lipid metabolism
Disorders of Newborn Period
Disorders of the immune system
Endometriosis
Eye, other disorders
Failure to thrive
Gastrointestinal signs and symptoms
Gastrointestinal/Hepatic disorders, other
Genito-urinary disorders, other
Glaucoma
Gout
Headaches
Hematologic disorders, other
Hemolytic anemia
CAR06
CAR07
CAR16
EYE06
NUR20
NUR18
NUR05
GTC01
PSY12
EYE02
CAR09
CAR06
CAR07
CAR16
EYE06
NUR20
NUR18
NUR05
GTC01
GSU06
GAS05
GAS12
REN01
REC03
REC01
MUS11
CAR04
CAR05
RES03
EAR08
MUS03
NUR11
PSY09
NUR19
EYE13
CAR11
NEW05
ALL06
RES04
FRE03
EYE14
NUT01
GAS01
GAS14
CAR10
GUR12
EYE08
RHU02
NUR02
HEM08
HEM01
6-7
EDC Description
Hemophilia, coagulation disorder
HIV, AIDS
Hypertension, with major complications
Hypothyroidism
Inherited metabolic disorders
Irritable bowel syndrome
Ischemic heart disease (excluding acute myocardial infarction)
Kyphoscoliosis
Low back pain
Malignant neoplasms of the skin
Malignant neoplasms, cervix, uterus
Multiple sclerosis
Muscular dystrophy
Musculoskeletal disorders, other
Nephritis, nephrosis
Neurologic disorders, other
Obesity
Osteoporosis
Other endocrine disorders
Other skin disorders
Paralytic syndromes, other
Parkinson's disease
Psychosocial disorders, other
Quadriplegia and paraplegia
Renal disorders, other
Respiratory disorders, other
Seizure disorder
Short stature
HEM07
MAL03
INF04
HYPT
HYPT
END04
GAS02
GTC02
GAS09
CAR03
MUS06
MUS14
MAL02
MAL01
MAL18
MAL04
MAL05
MAL12
MAL07
MAL08
MAL09
MAL10
MAL11
MAL06
MAL13
MAL14
MAL15
NUR08
NUR09
MUS17
REN04
NUR21
NUT03
END02
END05
SKN17
NUR17
NUR06
NUR03
GSU11
PSY08
GUR04
PSY11
NUR12
REN05
RES11
EYE03
RHU05
PSY07
NUR07
END03
6-8
EDC Description
Sleep apnea
Spinal cord injury/disorders
RES06
NUR16
The chronic condition count represents the number of unique chronic EDC Flags present
in the individual’s diagnosis history. When the stringent diagnostic certainty option is
selected, more than one diagnosis from the chronic condition list may be required for the
marker to be turned on. Table 6 provides a representative distribution of the chronic
condition count for elderly and non-elderly populations.
Table 6: Distribution of the Chronic Condition Count Marker
Chronic Condition Count Non-Elderly
Elderly
0
65.9%
16.1%
1
16.4%
11.2%
2
8.2%
14.3%
3
4.6%
15.1%
4
2.4%
12.9%
5
1.2%
9.7%
6
0.6%
6.9%
7
0.3%
4.7%
8
0.2%
3.3%
9
0.1%
2.2%
10+
0.1%
3.6%
6-9
Condition Markers
The ACG System uses condition markers to highlight specific conditions that are high
prevalence chronic conditions, commonly selected for disease management or warranting
ongoing medication therapy. Conditions that are used to measure pharmacy adherence
are discussed in more detail in Chapter 11: Gaps in Pharmacy Utilization.
The specific criteria for each condition are listed in Table 7. The values within these
condition markers indicate the evidence used to identify members with the condition.
The following values are possible:
NP – The condition is not present; there is no evidence of the condition from any data
source.
TRT – The condition was identified according to specific treatment criteria (listed in
Table 7) which include a minimum of two prescriptions (on different dates of
service) from an appropriate chronic medication drug class. Regardless of the
diagnostic certainty option selected, minimum visit counts will apply as listed in
Table 7. It is possible to select the stringent diagnostic certainty option requiring two
diagnoses for a related EDC, yet assign a condition marker to the value of TRT based
on a single inpatient or emergency department diagnosis.
ICD – The condition was identified only from diagnosis information. Each condition
is identified by one or more EDCs. When the stringent diagnostic certainty option is
selected, more than one diagnosis may be required for the condition to be identified.
Rx – The condition was identified only from pharmacy information and minimum
treatment criteria were not met.
BTH – The condition was identified by both diagnosis and pharmacy criteria but
minimum treatment criteria were not met.
For each condition, a separate column indicates if the condition is untreated. The
specific criteria for an untreated condition are listed in Table 7. Regardless of the
specific visit criteria specified, the patient must also have the EDC assigned under the
diagnostic criteria column to be identified as untreated. It is possible to meet the
untreated criteria based on a single inpatient or emergency department diagnosis. If
the stringent diagnostic certainty option is applied requiring two diagnoses for a
related EDC, the patient will not be designated as untreated. The untreated column
will have the following values:
Y – The patient has no prescriptions in an appropriate chronic medication drug
class.
P - The patient is classified as potentiallyuntreated based upon one of the
following criteria:
1. The patient has only one prescription in an appropriate chronic medication
drug class.
2. The patient has more than one prescription in an appropriate chronic
medication drug class but all prescriptions occurred on only one fill date.
6-10
3. The patient has more than one prescription from multiple chronic medication
drug classes but does not have two or more from a single chronic medication
drug class.
N – The patient meets treatment criteria and has two or more prescriptions in an
appropriate chronic medication drug class.
If none of the values apply to a particular member, the untreated column will be blank
indicating that there was insufficient evidence to confirm the presence of the specific
condition.
Table 7: Definitions of Condition Markers
Condition
Diagnostic Pharmacy
Criteria
Criteria
Treatment
Criteria
Untreated
Criteria
Bipolar disorder
PSY12
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
prescription fills in
at least one of the
drug classes:
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
fills in a single drug
class:
Anti-convulsants,
Anti-psychotics
Anti-convulsants,
Anti-psychotics
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
at least one of the
drug classes:
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
class:
ACEI/ARB
ACEI/ARB
Aldosterone
receptor blockers
Aldosterone
receptor blockers
Beta-blockers
Beta-blockers
Diuretics
Diuretics
Inotropic agents
Inotropic agents
Vasodilators
Vasodilators
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
the drug class:
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
fills in the drug
class:
Depression
CAR05
PSY09
N/A
CARx020
PSYx040
Anti-depressants
Anti-Depressants
6-11
Condition
Diagnostic Pharmacy
Criteria
Criteria
Treatment
Criteria
Untreated
Criteria
Diabetes
END06,
END07,
END08,
END09
2 prescription fills
in at least one of
the drug classes:
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
class:
Meglitinides
Non-sulfonylureas
Sulfonylureas
Thiazolidinediones
Other
antihyperglycemic
agents
Long and shortacting Insulins
Meglitinides
Non-sulfonylureas
Sulfonylureas
Thiazolidinediones
Other
antihyperglycemic
agents
Long acting
Insulins
Meglitinides
Non-sulfonylureas
Sulfonylureas
Thiazolidinediones
Other
antihyperglycemic
agents
Long acting
Insulins
CAR11
Bile acid
sequestrants
Cholesterol
absorption
inhibitors
Fibric acid
derivatives
HMG-CoA
reductase
inhibitors
Miscellaneous
antihyperlipidemic
agents
in at least one of
the drug classes:
Bile acid
sequestrants
Cholesterol
absorption
inhibitors
Fibric acid
derivatives
HMG-CoA
reductase inhibitors
Miscellaneous
antihyperlipidemic
agents
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
class:
Bile acid
sequestrants
Cholesterol
absorption
inhibitors
Fibric acid
derivatives
HMG-CoA
reductase inhibitors
Miscellaneous
antihyperlipidemic
agents
Glaucoma
EYE08
EYEx030
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
the drug class:
Ophthalmic
glaucoma agents
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
fills in the drug
class:
Ophthalmic
6-12
Condition
Diagnostic Pharmacy
Criteria
Criteria
Treatment
Criteria
Untreated
Criteria
glaucoma agents
Human Immunodeficiency
Virus
INF04
HAART*
in the drug class:
HAART*
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
fills in the drug
class:
HAART*
Hypertension
Hypothyroidism
CAR14,
CAR15
END04
CARx030
Thyroid
replacement drugs
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
at least one of the
drug classes:
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
class
ACEI/ARB
ACEI/ARB
Aldosterone
receptor blockers
Aldosterone
receptor blockers
Anti-adrenergic
agents
Anti-adrenergic
agents
Beta-blockers
Beta-blockers
Calcium channel
blockers
Calcium channel
blockers
Diuretics
Diuretics
Vasodilators
Vasodilators
in the drug class:
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
fills in the drug
class:
Thyroid drugs
Thyroid drugs
Immunosuppression/Transplant
ADM03
ALLx050
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
the drug class:
Immunologic
agents
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
fills in the drug
class:
Immunologic
Condition
6-13
Diagnostic Pharmacy
Criteria
Criteria
Treatment
Criteria
Untreated
Criteria
agents
Ischemic heart disease
Osteoporosis
CAR03
END02
N/A
ENDx010
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
at least one of the
drug classes:
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
class:
Antianginal agents
Antianginal agents
Beta-blockers
Beta-blockers
Calcium channel
blockers
Calcium channel
blockers
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
the drug class:
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
fills in the drug
class:
Osteoporosis
Hormones
Parkinson’s disease
NUR06
Anticholinergic
antiparkinson
agents
in at least one of
the drug classes:
Dopaminergic
antiparkinsonism
agents
Anticholinergic
antiparkinson
agents
Dopaminergic
antiparkinsonism
agents
Osteoporosis
Hormones
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
class:
Anticholinergic
antiparkinson
agents
Dopaminergic
antiparkinsonism
agents
Persistent asthma
ALL04,
ALL05
RESx040
1 inpatient or ER
principal diagnosis
or 4 outpatient
diagnoses plus 2
prescription fills at
least one of the
drug classes or 4
the same drug
class:
1 inpatient or ER
principal diagnosis
or 4 outpatient
diagnoses with less
than 2 prescription
class:
Adrenergic
bronchodilators
Immunosuppressive
6-14
Condition
Diagnostic Pharmacy
Criteria
Criteria
Treatment
Criteria
Untreated
Criteria
Adrenergic
bronchodilators
monoclonal
antibodies
Immunosuppressive
monoclonal
antibodies
Inhaled
corticosteroids
Inhaled
corticosteroids
Leukotriene
modifiers
Leukotriene
modifiers
Mast cell stabilizers
Methylxanthines
Methylxanthines
Schizophrenia
RHU05
PSY07
MUSx020
N/A
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
at least one of the
drug classes:
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
class:
Disease-modifying
anti-rheumatic
drugs (DMARDs)
Disease-modifying
anti-rheumatic
drugs (DMARDs)
Immunologic
agents
Immunologic
agents
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
the drug class:
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
fills in the drug
class:
Anti-psychotics
Seizure disorders
NUR07
NURx050
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses plus 2
the drug class:
Anti-convulsants
Age-related macular
degeneration
EYE15
Anti-angiogenic
ophthalmic agents
N/A
Anti-psychotics
1 inpatient or ER
diagnosis or 2
outpatient
diagnoses with less
than 2 prescription
fills in the drug
class:
Anti-convulsants
N/A
6-15
Condition
Diagnostic Pharmacy
Criteria
Criteria
Treatment
Criteria
Untreated
Criteria
COPD
RES04
N/A
N/A
N/A
REN01
GURx020
N/A
N/A
Low back pain
MUS14
N/A
N/A
N/A
*HAART represents a multi-drug cocktail that can be delivered in a number of
configurations.
1. HAART (a three drug combination filled with a single prescription)
2. Any NRTI combination drug including either zidovudine , abacavir or tenofovir with
an additional NNRTI, entry inhibitor, integrase inhibitor or protease inhibitor
3. A NNRTI plus a ritonavir-boosted PI or nelfinavir
4. Three of more drugs from two or more different drug classes (Entry Inhibitors,
Integrase Inhibitors, NNRTIs, NRTIs, Protease Inhibitors)
6-16
Predicting Future Resource Use
7-i
7 Predicting Future Resource Use
Objectives ....................................................................................................... 7-1
Conceptual Basis of Predictive Modeling ................................................... 7-1
What is Predictive Modeling? .................................................................... 7-1
Text Box 1: Definition of Population-Based Predictive Modeling ........... 7-2
Forecasting Future Costs with the Components of the ACG Case-Mix
System ........................................................................................................ 7-3
Table 1: Selected ACG Variables versus Hospitalization as Predictors
of High Cost Patients.................................................................................. 7-3
Table 2: Proportion of Total Healthcare Costs Consumed by the
Highest Cost 10% and Lowest Cost 50% of Enrollees .............................. 7-4
Multi-Morbidity as the Clinical Basis of PM ............................................. 7-5
Figure 1: Percentage of Patients with Selected Chronic Condition and
Their Associated Number of Co-Morbidities............................................. 7-6
Figure 2: Total Healthcare Costs per Beneficiary by Count of Chronic
Conditions .................................................................................................. 7-8
Elements of a Predictive Model: Five Key Dimensions ........................... 7-9
Conceptual Basis ................................................................................... 7-9
Target Population ................................................................................ 7-10
Statistical Modeling Approach ............................................................ 7-11
Data Inputs........................................................................................... 7-12
Outcomes Assessed ............................................................................. 7-13
Development of ACG Predictive Models .................................................. 7-13
Dx-PM: A Diagnosis-Defined Predictive Model ..................................... 7-13
Figure 3: Dx-PM Risk Factors ................................................................ 7-14
Development of Rx-PM ........................................................................... 7-15
Rx-PM Score ............................................................................................ 7-16
DxRx-PM: Combined ACG Predictive Model ........................................ 7-16
Resource Bands ........................................................................................... 7-16
High, Moderate and Low Impact Conditions ........................................... 7-17
High Pharmacy Utilization Model ............................................................. 7-17
High Pharmacy Utilization Model Outputs .............................................. 7-18
Probability of Unexpected Pharmacy Cost .......................................... 7-18
High Risk for Unexpected Pharmacy Cost .......................................... 7-18
Relevant to Clinicians and Case-Managers .............................................. 7-18
Table 3: Overlap of the High Pharmacy Utilization Model and Dx-PM
(Top 1.5% of Scores) ............................................................................... 7-19
7-ii
This page left was blank intentionally.
7-1
Objectives
This chapter is intended to describe the basics of ACG Predictive Modeling. The chapter
provides background information on the conceptual and clinical basis underlying
predictive modeling and provides the history of the development of the suite of ACG
Predictive Models™ (referred to collectively as ACG-PM™).
Conceptual Basis of Predictive Modeling
What is Predictive Modeling?
Predictive modeling (PM) is used in a variety of manufacturing and service sectors to
forecast, for example, costs, profitability, financial risk, purchasing behavior, and
likelihood of defaulting on a loan. A predictive model comprises a set of variables (often
called risk factors) that are selected because they are likely to influence the occurrence of
the event or trend of interest. The predictions generated by the PM are used to improve
decision-making in the face of uncertainty.
In healthcare, PM has two purposes—one focuses on the individual patient to improve
clinical decision-making (i.e., clinical prediction tools) and the other (i.e., populationbased predictive models) draws on data from groups of patients to forecast healthcare
cost trends and to identify candidates for healthcare interventions such as care or disease
management programs.
In clinical medicine, PM is used to improve decision-making regarding prognosis, need
for diagnostic testing, or likelihood of treatment responsiveness for individual patients.
For example, data from the Framingham Heart Study are used to determine an
individual’s risk of coronary heart disease according to age, sex, cigarette smoking, LDL
and HDL cholesterol levels, blood pressure, and diabetes history.1 Each of these risk
factors is given an integer weight, and an individual’s probability of future heart disease
is related to the sum of the risk factor weights. Patients and health professionals use the
Framingham Heart Disease Prediction Score to guide cardiovascular disease prevention
behavior.
Population-based predictive models, such as the ACG models, are used to forecast trends
in healthcare costs for groups of patients or to identify candidates for intensive healthcare
interventions (Text Box 1). (Note: Throughout this manual, we will refer to them simply
as predictive models, rather than population-based PMs) Currently, ACG-PM uses health
and healthcare risk factors from administrative claims and encounter data to produce risk
scores. Future versions of ACG-PM will use clinical information from patients’ health
records once these data are commonly collected and stored in digital formats.
1
Wilson PWF et al. Prediction of Coronary Heart Disease using risk factor categories. Circulation
1998;97:1837-1847.
7-2
Text Box 1: Definition of Population-Based Predictive Modeling
Definition of Population-Based
Predictive Modeling
Population-based predictive models use
health and healthcare information derived
from all members of a population to
predict future health events or forecast
healthcare costs and service use for
populations or individual patients.
7-3
Forecasting Future Costs with the Components of the ACG Case-Mix
System
Before ACG-PM was developed, users of the ACG System successfully employed ACGs
themselves and their building blocks, ADGs, to identify patients at high risk for future
costs. A high burden of morbidity, defined by (1) the number of morbidity-types (ADGs)
overall, 2) the number of major ADGs (i.e., those ADGs that have a high impact on
future resources), or selected ACG categories perform well in predicting the future risk of
using healthcare resource. The performance of these components compared to prior
hospitalization in predicting high risk users is shown in Table 1.
Table 1: Selected ACG Variables versus Hospitalization as Predictors
of High Cost Patients
Baseline (Year 1) Predictor Variables
Used to Define Risk Groups
Percentage
of
Members
Percentage Risk
Group Who are High
Cost (top 5% of the
population) in Next
Year (Year 2)
Prior Use
2+ Hospitalizations
Commercial population
Medicare population
0.6%
4.1%
39.6%
19.5%
2.5%
11.3%
34.2%
12.6%
0.4%
4.2%
54.1%
19.5%
2.3%
11.2%
35.1%
12.7%
Morbidity Measures
10+ Morbidity types (ADGs)
Medicare population
4+ Major Morbidity types (Major ADGs)
Medicare population
Selected ACGs
Medicare population
7-4
For example, commercially insured patients who were hospitalized two or more times in
the baseline year comprised 0.6% of the total population and had a risk of 39.6% for
being in the high cost group (top 5% of total healthcare costs) in the next year. On the
other hand, patients with the highest morbidity burden ACG categories, those with four
or more major ADGs comprised 0.4% of the total population and had a risk of 54.1% for
being in the high cost group. Observations such as these led our team to believe that the
clinical logic that underpinned the ACG system could be successfully modified and
expanded to form more comprehensive predictive modeling applications.
Retrospectively, we know with certainty which persons were high-cost consumers of
health care resources and which were lower cost. One of the most consistent findings
about resource utilization is that 10% of people consume about 60% of resources,
whereas the lowest cost 50% of consumers account for only about 4% of costs (Table 2).
This concentration of health care resources in a small segment of the population is a
universal finding across all populations, and it is the key empirical observation that
underpins the conceptual rationale of risk adjustment. Predictive modeling reduces the
uncertainty of who will be high-cost in the future. In terms of statistical performance, the
best predictive models are those that are most accurate at either forecasting future costs
for groups of patients or classifying patients as potentially high cost.
Table 2: Proportion of Total Healthcare Costs Consumed by the
Highest Cost 10% and Lowest Cost 50% of Enrollees
% Enrolled Population
% Total Healthcare Costs
Highest cost 10%
Commercial
Medicare
63%
55%
Lowest cost 50%
Commercial
Medicare
4%
9%
The unique insight provided by the ACG System is that morbidity can be categorized in a
way that is both clinically logical and informative of future healthcare resources. The
clinical risk factors in ACG-PM are likely to persist over time and are indicative of
sustained health need (i.e., capacity to benefit from health services) and demand (i.e.,
patient or provider desire for services). The predictions from ACG-PM can be used to
identify high-cost persons prospectively to better forecast their future costs and to proactively identify individuals whose high intensity healthcare needs may be amenable to
organized care management or expanded primary care interventions.
7-5
Multi-Morbidity as the Clinical Basis of PM
In the 21st century, people are living longer and experiencing healthier later lives than
their parents and grandparents. These improvements in health are a reflection of
advances in public health, medicine, and general improvements in the economy that
resulted in greater affluence for all people. With the development of vaccines and
antimicrobial therapy, the morbidity profile of patients presenting to medical care has
shifted from the common acute diseases of the 1900s to longer term, chronic diseases,
such as hypertension, diabetes, or ischemic heart disease. This change in disease patterns
has been called the ―
epidemiological shift.‖ Along with this greater burden of chronic
illness, it became increasingly obvious that individuals are likely to have not just a single
chronic condition, but multiple co-occurring chronic diseases, a phenomenon known as
multi-morbidity. In fact, the ACG System was developed to classify multi-morbidity, the
new clinical reality faced by populations in developed nations.
Among commercially insured populations, just 14.5% of patients with diabetes and
24.7% of those with hypertension had that condition only and no other chronic conditions
(Figure 1). On the other hand, among Medicare beneficiaries, 51.3% of those with
diabetes, 59.8% of those with ischemic heart disease, 47.1% of those with arthritis, and
38.2% of those with hypertension had four or more other chronic conditions. In short,
co-morbidity is the norm among adults with chronic illness.
7-6
Figure 1: Percentage of Patients with Selected Chronic Condition and
Their Associated Number of Co-Morbidities
Commercially Insured Population (less than 65 years old)
100%
90%
25.1
80%
70%
15.7
60%
50%
20%
10%
0%
14.9
12.3
12
18.6
20.1
18.2
22
40%
30%
35.2
16.5
21.3
24.7
28.3
28
24.7
Arthritis
Hypertension
22.6
17.4
14.5
Diabetes
4+
3
2
1
0
8
Ischemic Heart
Disease
100%
90%
80%
70%
51.3
47.1
38.2
59.8
60%
17.2
50%
40%
30%
20%
16.5
14.1
11.2
0%
4.4
Diabetes
19.6
2
1
17.1
0
17.2
15.9
10%
3
18.1
17.1
7.7
1.9
Ischemic Heart
Disease
12.5
4+
5.1
7.9
Arthritis
Hypertension
7-7
Source: PharMetrics, Inc., a unit of IMS, Watertown, MA; national cross-section of managed care
plans; population of 2,259,584 commercially insured lives (less than 65 years old) and population of
93,620 Medicare beneficiaries (65 years and older), 2001-2002.
An extensive amount of research produced by our team and others in the academic and
commercial sectors has repeatedly shown that multi-morbidity is a common clinical
feature of individuals who were high-cost in the past and those likely to be high cost in
the future. With each additional chronic condition, health spending per beneficiary rises
dramatically (Figure 2). Compared with individuals with no chronic conditions, those
with five or more are 25 times more costly among the commercially insured and 20 times
more costly among Medicare beneficiaries. In sum, multi-morbidity provides critically
important clinical insight into why a small percentage of the population consumes such a
large share of resources, and is therefore a logical clinical basis for predictive modeling.
7-8
Figure 2: Total Healthcare Costs per Beneficiary by Count of Chronic
Conditions
Commercially Insured Population (less than 65 years old)
20000
$18,118
total healthcare costs/person ($)
18000
16000
14000
12000
10000
$7,950
8000
$5,537
6000
$3,601
4000
2000
$2,180
$725
0
0
1
2
3
4
5+
# chronic conditions
14000
$12,429
total healthcare costs/person ($)
12000
10000
8000
6000
$5,198
$3,600
4000
2000
$1,706
$2,482
$637
0
0
1
2
3
4
5+
# chronic conditions
7-9
Elements of a Predictive Model: Five Key Dimensions
The structure of a predictive model can be characterized in five dimensions: conceptual
basis; target population; data inputs; statistical modeling approach; and, outcomes
predicted. Each of these dimensions is discussed in more detail below. For each
dimension, we provide a specific description of that attribute for ACG-PM.
Conceptual Basis
Dozens of predictive models have been developed and are commercially available to
medical managers, actuaries, and other healthcare professionals to predict future costs
and identify the highest opportunity patients for care management and other intensive
population-based interventions. Most of these PMs, however, are (unfortunately)
atheoretical—that is, they are not based on any conceptual model related to the patient,
provider, and health system factors that determine health need, demand, and supply.
Because PMs are increasingly being used as supplemental decision tools in the care of
high cost patients, to be meaningful to medical managers and health professionals, every
PM should be grounded in a strong conceptual framework. For example, ACG-PM is
based on a rich and comprehensive set of multi-morbidity clinical frameworks (the logic
that classifies diagnostic codes to ADGs and EDCs, the logic that forms ACG categories,
and the logic that classifies pharmacy codes to Rx-MGs) that distinguish ACGs from all
other risk adjustment methodologies.
Two important problems result from PMs that are purely data-driven. First, without a
conceptual basis for the PM, the meaning of the PM output, commonly referred to as a
risk score, is unclear. Does the risk refer to future resource use, future health, future
specific utilization events such as hospitalization or something else? A second problem
with the atheoretical approach to PM development is that these models do not bridge the
financial and clinical enterprises in health care organizations. Although a risk score that
is predictive of future costs may be useful to an actuary interested in small group
underwriting, the same score has no meaning to health professionals who are more
concerned with clinical factors that are responsible for those costs.
An empirically based argument can be developed for using the current year’s costs to
predict future costs. In fact, this has been the strategy for many years. In terms of
statistical performance, prior cost is a fairly good predictor of future cost, but it has
important limitations. First, prior cost has no inherent clinical meaning, and is therefore
of no relevance to clinicians who wish to intervene. It is not tied to morbidity and,
thereby, cannot be translated into clinical action. Second, prior cost is subject to the
phenomenon of regression to the mean; very high costs in one year tend to drift towards
average costs in the subsequent year. This occurs because uncommon and high-cost
medical events (e.g., a hospitalization) tend not to repeat themselves. Third, prior-use
measures are also not appropriate as risk factors for risk-adjusted rate setting or profiling
as they potentially could provide incentives to use excessive and inappropriate resources.
7-10
The ACG predictive models have comparable, and in many applications, superior,
statistical performance compared with prior cost measures. Clinically based models,
such as ACG-PM, produce risk scores that can be adapted for financial applications such
as underwriting and actuarial forecasting. Importantly, these risk scores can be
decomposed into the clinical cost drivers. Prior cost measures are useful only for
financial forecasting and have little to no relevance for the clinical mission of health care
organizations.
ACG-PM is grounded in clinical logic. Each of the risk factors in the models is a clinical
variable that has meaning to health professionals and when combined produce a risk
score that is a robust predictor of future costs, utilization, and health status. Because of
this clinical grounding, ACG-PM can provide a common language between financial and
clinical managers in health care organizations. This is an important innovation for risk
adjustment and health care management more generally.
Target Population
PMs can be developed for a general population (generic approach) or those distinguished
by age, sex, disease category, or some other personal characteristic (categorical
approach). A significant advantage of the generic approach is that the same set of
predictors is used for all modeling applications, which allows one to compare cost drivers
across groups. Categorical model development, however, is optimized for sub-groups
within a population defined by disease class or age group, for example. Greater
predictive accuracy may be achieved for sub-groups if model risk factors are specifically
selected for patients within the sub-group. Whether this holds true statistically is unclear
and requires further investigation.
ACG-PM applies to a general population, which is consistent with the whole-person
orientation of the entire ACG system. However, ACG-PM can be optimized for disease
or age groups by fitting the model (i.e., establishing new weights for the risk factor set)
for a sub-population of interest. With the availability of laboratory and patient-reported
outcomes it seems likely that optimal performance of models will entail more
customization to the target population.
Our research has shown that the primary difference between age groups in predictive
modeling applications was the distribution of morbidity. Nonetheless, there is a need for
a small number of age-specific conditions (e.g., Alzheimer’s disease occurs almost
exclusively in older patients, while pregnancy and delivery is a condition of younger
women) in PMs that use a generic approach. To address this need, ACG-PM includes all
conditions that are important predictors of future health and resource use for any age
group. However, for age-specific applications, different weights are attached to the
predictors to reflect that the conditions or other clinical variables have different levels of
influence on future health and healthcare by age group. (For example, pregnancy as a
risk factor would get a weight of zero for an over 65 population.) In effect, ACG-PM is
calibrated differently by age with age-specific weights attached to the predictors. The
same type of calibration can be done for patients with specific diseases.
7-11
Statistical Modeling Approach
Several types of statistical approaches can be used to develop predictive models. Here
we summarize the most common.
Algebraic. The simplest approach is to use current cost data or a count of conditions (or
some other algebraically computed clinical marker) as a predictor of future cost. The
model might add additional factors such as the rate of medical inflation and, perhaps,
demographics such as age and gender. Such models are not well suited to individual
predictions and have no inherent clinical meaning. They are also subject to regression to
the mean (i.e., the natural tendency of groups of individuals who are high cost one year to
move towards mean costs in the following years). Furthermore, the incorporation of
prior use potentially conflates demand with the actual need for health care services,
which makes such models particularly unsuited for profiling applications in which the
goal is to assess the match between services delivered and services needed.
Linear Regression. Linear regression using ordinary least squares parameter estimation
is widely used and provides fairly stable estimates. The individual predictions have an
intuitive appeal, because they can readily be translated into dollar values. However, cost
data do not fulfill the assumptions for these models (which are best suited for data
following a normal distribution with common variance across the variable range) but the
models are sufficiently robust that some departures from the assumptions can be
tolerated. Because it is important to be cautious about extreme values, or outliers, that
could dramatically alter the fitting of the regression line, linear models frequently employ
truncation or top coding to minimize outlier effects.
More sophisticated generalized linear models that use transformations of the dependent
variable and assume a non-normal distribution of error terms have been used in some
predictive modeling applications. The superiority of these alternative models compared
with standard ordinary least squares estimation has not been demonstrated as yet. Our
work suggests, for example, that generalized linear models using a log link function and a
gamma distribution of error terms provide PM estimates that are no better, and in many
cases inferior to, standard linear regression methods. Thus, ACG-PM was developed
using a standard linear regression approach with ordinary least squares estimation in
order to provide risk factor weights (the beta coefficients from the regression equations)
with intuitive, real-dollar meaning.
Logistic Regression. This regression approach is applied to dichotomous outcomes
events such as whether someone is hospitalized or is in a high-cost group (e.g., in the top
5% of the highest cost persons in the population). Logistic regression assumes a specific
functional form between predictors and the outcome (a logit relationship) that may not be
observed in health care resource use applications. The approach also attenuates
differences between cases because the risk output is a predicted probability that ranges
between 0 and 1.
7-12
Neural Networks. These modeling strategies (also referred to as artificial intelligence
approaches) represent collections of mathematical models that attempt to emulate the
processes observed in biological nervous systems, including the capacity to adapt and
learn from prior experiences or observations. They are fundamentally atheoretical and
lack a logical clinical or conceptual model. Neural networks require no specification of
the nature of the relationship between predictors and the outcome in advance. An
optimally fitted model is created as part of an iterative statistical fitting or learning
process. Neural networks are computationally intense, can be difficult to understand, and
every solution is unique (which can be a problem for models used in rate setting or
comparing data across populations). From the perspective of either empirical accuracy or
performance, neural network models have yet to demonstrate any advantage over those
developed using linear regression methods.
Data Inputs
One of the main challenges facing PM development is the limited risk factor data on
which the models can be based. Currently, most PMs depend on standard medical claims
(age, sex, diagnoses, procedures, outpatient medications, service dates, and costs)
available from administrative billing records. ACG-PM uses age, sex, diagnostic codes,
medication codes, and optional summary cost variables.
For a given patient, all demographic, diagnostic, and pharmaceutical information
available in administrative claims records is used by ACG-PM assignment algorithms.
Alternatively, there are PM approaches that organize care into a number of discrete timedelimited episodes of care. While episode approaches have some appeal to those who
envision the care process in these terms, commonly available claims data used for
predictive modeling are encouraged as encounters rather than as episodes. This requires
episode-based approaches to use complex decision rules to assign services, medications,
and diagnostic information to a unique time-dependent cluster, called the episode-of-care.
These algorithms reject large numbers of claims during the assignment process,
effectively leaving a substantial amount of clinical data on ―
the cutting room floor.‖
Thus, the clinical picture created by episode of care PMs is incomplete, which contrasts
with the use-every-clinical-datum approach of the ACG system. Further, episode-based
approaches encourage a disease by disease approach to the care process which, as we
have noted above, is best characterized by the simultaneous presence of multiple
morbidities.
Several data sources that offer exciting new avenues for model development will soon be
available: clinical information from electronic health records; physiological outcome data
from laboratory assessments; patient-reported psychosocial outcomes; and, background
socio-demographic information. These additional data sources will stimulate
development of new generations of PMs that are more individualized to each patient’s
context and needs. With the limited source of risk factor data available to the current
generation of PMs, model predictive capabilities are limited as well. All predictive
models are decision tools that must be used with good clinical and managerial judgment
and augmented with other sources of information.
7-13
Outcomes Assessed
ACG-PM can predict a large range of financial and clinical outcomes, including:
Total healthcare costs
Pharmacy costs
Outpatient costs (and sub-components, e.g., imaging, physician, laboratory, specialty,
ED)
Inpatient hospitalization
ICU admissions
Physician visits
Specialty referrals
Morbidity events (such as acute complications of chronic disease)
Mortality
Optimal performance of ACG-PM occurs when weights for each of the risk factors are
obtained from models that employ the same outcome as that which is being predicted. If
the outcome is risk of mortality, using risk factor weights obtained from a test model that
regress the occurrence of death on the full set of risk factors will give the most accurate
prediction results, and would be superior to utilization-based weights. In this way ACGPM can be customized to the outcome of interest.
Development of ACG Predictive Models
The ACG system includes two types of predictive models, Dx-PM and Rx-PM, which
differ according to the data inputs. The first is based on diagnostic codes (i.e., ICD-9CM, ICD-10) and the second is built with medication codes (i.e., NDC codes). Both are
excellent predictive models for forecasting healthcare costs and identifying high-risk
patients. In this section we describe the clinical basis and development of both model
types. The section concludes with a discussion of a combined model, DxRx-PM, that
includes both diagnostic and medication code inputs.
Dx-PM: A Diagnosis-Defined Predictive Model
The ACG System’s Dx-PM is a predictive model constructed from diagnoses given to
patients in all inpatient and outpatient clinical settings, age, and sex. The Dx-PM
modeling strategy makes use of the comprehensive array of metrics available within the
entire ACG risk adjustment system. The sole data inputs to the model are age, sex, and
diagnostic codes. The set of risk factors that constitute Dx-PM is shown pictorially in
Figure 3, and is discussed in detail below.
7-14
Figure 3: Dx-PM Risk Factors
Age and gender are included to assess age-related and gender-based health needs.
Overall morbidity burden is measured using the ACG categorization of morbidity
burden. The variable in the model includes 3 groupings of low resource intensity ACG
categories and 24 individual ACGs that are the most resource intensive morbidity groups.
High-impact chronic conditions were selected for inclusion in the predictive model
because when they are present they have a high impact on resource consumption. They
are common chronic conditions that are associated with greater than average resource
consumption, uncommon diseases with high impact on both cost and health,
complications of chronic disease that signify high disease severity (e.g., diabetic
retinopathy), or conditions that are a major biological influences on health status (e.g.,
transplant status, malignancy). Only conditions for which the evidence linking health
care to outcomes is strong were included in the model. A sub-set of the ACG system’s
Expanded Diagnostic Clusters (EDCs) is used to identify the high impact conditions in
the model.
Hospital dominant condition markers are based on diagnoses that, when present, are
associated with a greater than 50 percent probability among affected patients of
hospitalization in the next year. All these diagnoses are setting-neutral, i.e., they can be
given in any inpatient or outpatient face-to-face encounter with a health professional.
The variable is a count of the number of morbidity types (i.e., ADGs) with at least one
hospital dominant diagnosis.
7-15
Development of Rx-PM
The first PM that the ACG team developed was based entirely on diagnostic codes, age,
and sex (Dx-PM). Although Dx-PM was successfully used in many health care
organizations, it was noted that in many cases the ACG system user had access to
NDC/ATC codes, but no ICD information. Some plans had medication data only and
others wanted to identify high risk patients (and enroll them proactively in care
management programs) before the 6 months needed to assign Dx-PM had elapsed. RxPM was developed to address these needs.
Incorporating medication data into ACG-PM presented several new opportunities for our
development efforts. First, medication-based predictive modeling can be done before the
full set of ICD codes are obtained, which is the case for new enrollees in a health plan,
for example. We believe that a thorough clinical history and recording of medications
during a single encounter may be sufficient to assign an Rx-PM risk score. Second,
medication data captures a unique constellation of clinical information that overlaps,
although not entirely, with diagnostic codes. The combination of the two data sources
complement one another, particularly when the outcome of interest is total health care
costs, and offer the promise of superior PM performance. Third, we sought to create a
clinically-based, medication-defined PM that comprised a set of risk factors that group
medications according to the types of morbidity that they are commonly used to treat.
We were unaware of any Rx-based PM that used this clinically oriented approach, which
we feel is one of the more important innovations presented by the Rx-PM. Some
medication-defined PMs group a sub-set of medication codes into disease entities,
whereas other medication-defined PMs are no more than medication therapeutic classes.
Both approaches fail to fully capture the unique attributes of morbidity conveyed by
medication data, which is the clinical innovation offered by Rx-PM.
Several challenges are posed by medication-only PMs. The sheer volume of drug codes
(there are now over 100,000 NDCs, and over 5,000 fourth and fifth Level ATC codes)
makes creating a streamlined set of risk factors highly challenging. Moreover,
medication use is not synonymous with presence of specific disease. One medication can
be used to treat many diseases, and one disease often has many medication options for its
management. This relationship between medication and disease entities convinced us
that an NDC/ATC-based PM should be based on a new way of classifying morbidity-one that captures the unique clinical information embedded in medication use data--rather
than an approach that attempts to attach disease labels to medications. Because risk
scores based entirely on NDC/ATC codes are subject to artificial inflation associated with
excess and inappropriate medication usage, any NDC/ATC grouping algorithm
underpinning a PM should be developed in a way to blunt this potential circularity
problem.
7-16
Rx-PM Score
Rx-PM produces a score that is the sum of the Rx-MG, generic drug count, age, and sex
risk factor weights. Each Rx-MG is entered into the model as an indicator (yes/no)
variable. An Rx-MG is either turned on or off, regardless of the number of medications
that may be prescribed within that category or the number of times each medication may
have been prescribed. This all-or-none attribute of the Rx-MGs minimizes the potential
bias introduced by excessive prescribing or medication changing within therapeutic
classes, leading to spuriously high risk scores. A prior cost variable can be added as a risk
factor and is considered a measure of demand for services not captured by NDC codes.
This strategy maximizes model parsimony and clinical utility of the predictor variables
while optimizing statistical performance.
DxRx-PM: Combined ACG Predictive Model
When both diagnostic and medication codes are available for a patient population, a
model that combines Dx-PM and Rx-PM can be used. We call this combined model
DxRx-PM. This model includes all the Dx-PM and Rx-PM predictors along with the
appropriate prior cost variable. Combining these risk factors is clinically logical because
the morbidity information obtained from diagnostic codes complements that which is
obtained from medication codes. In effect, the combined model provides the fullest
clinical picture of the patient’s health needs and demand for care that is possible using
administrative data.
The innovation that the Dx-PM, Rx-PM, and DxRx-PM present is their unique clinical
logic coupled with their excellent statistical performance. Each employs risk factors that
were created using clinical frameworks that make sense to health professionals and
medical managers. As we will show in the next section, each has superlative statistical
performance. The combination of clinical logic with excellent financial predictions
provides a common language that can link medical and financial managers in health care
organizations.
Resource Bands
The software incorporates both prior total cost and prior pharmacy cost bands into the
ACG predictive models. They are a useful adjunct to analysts wishing to stratify their
populations.
Possible values include:
0 – 0 or no pharmacy costs
1 – 1-10 percentile
7-17
High, Moderate and Low Impact Conditions
Within the Patient Clinical Profile report, the EDCs and Rx-MGs associated with an
individual are categorized as High, Moderate or Low impact. The definition for these
categories is based on the individual contribution of the condition to the predictive
modeling score. The specific criteria are as follows:
High Impact EDC: Coefficient >1 in the non-elderly Dx-PM without prior cost
predicting total cost model
Moderate Impact EDC: Coefficient between 0.1 and 1.0 in the non-elderly Dx-PM
without prior cost predicting total cost model
Low Impact EDC: Coefficient less than 0.1 or not included in the non-elderly Dx-PM
High Impact Rx-MG: Coefficient >1 in the non-elderly Rx-PM without prior cost
predicting total cost model
Moderate Impact Rx-MG: Coefficient between 0.1 and 1.0 in the non-elderly Rx-PM
Low Impact Rx-MG: Coefficient less than 0.1 in the non-elderly Rx-PM without
prior cost predicting total cost model
High Pharmacy Utilization Model
Traditional predictive modeling focuses on using diagnoses and/or pharmacy information
to predict future expenditures. Current models work well and have provided a means of
identifying individuals at risk for future high pharmacy expenditures. However, a
limitation of the current approach is that these models, by definition, have tended to
identify a multi-morbid population whose high use of pharmacy services may well be
justified; or, more specifically, their disease profile warrants the high use of medications.
The High Pharmacy Utilization model is specifically designed to address this
shortcoming and is focused on predicting the subset of the multi-morbid population who
are consuming drugs above and beyond what might be anticipated based on their
morbidity burden.
7-18
High Pharmacy Utilization Model Outputs
Probability of Unexpected Pharmacy Cost
The probability of unexpected pharmacy cost is a numerical probability score
representing the result of applying weights from a logistic regression model on the
development data set predicting individuals with moderate or high morbidity with
unusually large pharmacy expenditures. The model independent variables include age,
gender, ACG categories, and select high impact conditions defined by EDCs, Rx-MGs,
hospital dominant conditions, and frailty. The markers in the model are the same as the
DxRx-PM model markers without prior cost. In the development data set individuals
were deemed to have moderate or multi-morbidity if they were assigned to a moderate or
higher resource utilization band and they were considered to have unusually large
pharmacy expenditures if their standardized pharmacy expenditures were more than 1.75
standard deviations from their predicted amount based on diagnoses information2.
High Risk for Unexpected Pharmacy Cost
High Risk for Unexpected Pharmacy Cost is a binary flag that indicates individuals with
a Probability of Unexpected Pharmacy Cost greater than 0.4.
Relevant to Clinicians and Case-Managers
Objectives for success for the High Pharmacy Utilization Model included PPV/SENS and
ROC statistics similar to traditional predictive modeling strategies; however, to warrant
inclusion of an additional model in the ACG suite of predictive models the High
Pharmacy Utilization Model needed to identify a different group of individuals than
traditional predictive modeling methods. Table 3 summarizes the overlap of the High
Pharmacy Utilization Model and Dx-PM to identify high pharmacy users with moderate
morbidity burden or higher and indicates that the overlap between the two models was
only 16%. In short, the High Pharmacy Utilization Model is identifying a new group of
patients not identified by our prior methods.
2
This is a laymen’s interpretation of applying methodologies outlined in Gray Woodal’s 1993 article ―
The
Maximum Size of Standardized and Internally Studentized Residuals in Regression Analyses‖.
7-19
Table 3: Overlap of the High Pharmacy Utilization Model and Dx-PM
(Top 1.5% of Scores)
Total
Identified by
Both High
Pharmacy
Utilization and
Dx-PM
Total
20,709
7,822
49,253
42.05%
15.88%
100.00%
Identified by
High
Pharmacy
Utilization
Model only
Identified by
Dx-PM only
20,722
42.07%
There are a variety of reasons why this population subgroup may be of interest and
explanations as to why pharmacy utilization might be higher than expected. Sometimes
pharmacy costs are higher because of the use of an expensive medication. Other times
unanticipated high drug costs may be related to quality issues (i.e., poor or uncoordinated
care). Unexpected high pharmacy utilization can also be a result of both cost and quality
issues. These can range from poor data (i.e., missing ICD codes), to poor or inadequate
care and/or gaps in care. Other times it is simply a matter of polypharmacy and/or patient
abuse. The hope is that this model can help clinical or case managers to:
Better identify those at risk for future high pharmacy utilization
Provide information about what you can do to improve outcomes and/or reduce cost
for those at risk including:
Providing better coordination
Refining the drug regimen
Other interventions yet to be identified.
The ACG Team looks forward to feedback on the success of this model and will continue
to explore the methodologies introduced here in anticipation of additional predictive
models for target populations.
7-20
Predictive Modeling Statistical Performance
8-i
8 Predictive Modeling Statistical Performance
Overview of ACG Predictive Models Statistical Performance ................. 8-1
H
H
Validation Data ............................................................................................. 8-1
H
H
Adjusted R-Square Values ........................................................................... 8-1
H
H
Table 1: Predictive Models Evaluated ...................................................... 8-2
Table 2: Proportion of Variance Explained (Adjusted R-Square
Values) Using Year 1 Risk Factors and Year 2 Resource Use
Measures for Alternative Predictive Models.............................................. 8-3
Values) Using Year 1 Risk Factors and Year 2 Resource Use
Measures for Prior Cost and ICD & NDC Based Combined Models ........ 8-5
Interpreting Adjusted R-Square Values ..................................................... 8-6
Predictive Ratios ........................................................................................ 8-6
Table 4: Predictive Ratio by Year-1 Cost Quintile for Commercial
and Medicare Populations .......................................................................... 8-7
H
H
H
H
H
H
H
H
H
H
H
H
Sensitivity and Positive Predictive Value.................................................... 8-8
H
H
Table 5: Sensitivity and PPV for Top Five Percent .................................. 8-9
H
H
ROC Curve.................................................................................................. 8-10
H
H
Table 6: C-Statistics ................................................................................ 8-10
H
H
Clinical Information Improves on Prior Cost.......................................... 8-11
H
H
Figure 1: Venn Diagram Comparing Overlap of Prior, Predicted,
and Actual High Cost Users ..................................................................... 8-12
H
H
8-ii
8-1
Overview of ACG Predictive Models Statistical Performance
This section demonstrates the ACG predictive models statistical performance while
describing the various ways in which they can be applied in health care applications.
In developing the ACG Predictive Modeling suite, the goal was to attain the highest level
of statistical performance as measured by commonly used benchmarks such as the
percent of variation in cost explained by the model (R-Squared statistic) while retaining
the underlying clinical sense and ease of use that are hallmarks of the ACG System.
Validation Data
A very large longitudinal claims database, licensed from PharMetrics, a unit of IMS,
Watertown, MA, was used for our modeling efforts. The database is generally
representative of the U.S. commercially insured population. Two populations were used
for modeling: (1) commercial (all ages to 65 years old) and (2) Medicare Advantage
(Medicare managed care). Two years of data for the period 2006-2007 were used. Only
persons with at least six months of enrollment in year one and one month of enrollment in
year two were included. For the pharmacy model, the population was restricted to
include enrollees that had both pharmacy and medical eligibility. For the commercial
population, the database was split into two randomly assigned halves for development
and validation, with roughly 2.5 million persons in each of these halves. Given that the
Medicare population was much smaller, we chose to base development on the entire
population of roughly 250,000.
Adjusted R-Square Values
The conventional measure of model performance is the R-Squared statistic. This statistic
measures how well the model fits the data and has become a standard measure of
performance, especially among underwriters and actuaries who must price products
across a range of populations. Comparative performance statistics have been calculated
for the models outlined in Table 1.
8-2
Table 1: Predictive Models Evaluated
Risk Model
Description
Age and Sex
Simple demographic model
Age, Sex, and ADG
Simple demographic model augmented with ADGs,
the ACG building blocks
Prior Cost
Total Cost or Pharmacy Cost as appropriate
ACG DxRx-PM
ICD- and NDC-based predictive model
Charlson Index
17 ICD-based high impact conditions
ACG Dx-PM
ICD-based predictive model
Chronic Disease Score
29 NDC-based disease-oriented markers
ACG Rx-PM
NDC-based predictive model
Generally where the data are available it is recommended that a combined model, the
DxRx-PM, be implemented as it takes advantage of all available data streams. Whether
or not to include prior cost as a predictor in any model is a more difficult decision.
Including prior cost has the impact of increasing explained variance. Therefore, for
financial models the choice is easy, incorporate prior cost. If all you care about is
explained variance than including prior cost increases explanatory power. However,
having the highest R-square value does not necessarily always equate to the optimal
methodology. For example, for care management applications it is less clear that using
prior cost offers clear advantages. Prior high cost users may or may not be amenable to
disease or case-management intervention programs. In contrast, individuals identified as
high risk using only diagnosis or pharmacy data (and excluding prior cost information),
are oftentimes better candidates for such intervention programs. Therefore, for care
management purposes, diagnosis or pharmacy based (or combination models) may
provide the best performance. More information on model performance under particular
circumstances is presented below.
The results of this comparative assessment in both commercial and Medicare populations
are shown in Table 2 for diagnosis-based and pharmacy-based models.
8-3
Table 2: Proportion of Variance Explained (Adjusted R-Square Values) Using Year 1 Risk Factors
and Year 2 Resource Use Measures for Alternative Predictive Models
Commercially Insured Enrollees
Predictive Total
Model
Costs
Age, Sex
Medicare Beneficiaries
Pharmacy Physician
Costs
Visits
5%
7%
5%
ICD/Diagnostic Code Based Models
Predictive Total
Model
Costs
Age, Sex
Pharmacy Physician
Costs
Visits
1%
0%
0%
ICD/Diagnostic Code Based Models
Age, Sex,
ADG
17%
20%
25%
Age, Sex,
ADG
11%
7%
25%
ACG
16%
17%
22%
ACG
10%
5%
22%
12%
7%
11%
16%
10%
26%
Age, Sex,
Charlson
Disease
Indicators
13%
17%
10%
Age, Sex,
Charlson
Disease
Indicators
Dx-PM
21%
29%
23%
Dx-PM
8-4
Predictive Total
Model
Costs
Pharmacy Physician
Costs
Visits
Predictive Total
Model
Costs
Pharmacy Physician
Costs
Visits
NDC/Medication Code Based Models
NDC/Medication Code Based Models
Age, Sex,
Chronic
Disease
Score
16%
35%
12%
Age, Sex,
Chronic
Disease
Score
10%
23%
13%
Rx-PM
19%
44%
17%
Rx-PM
12%
27%
16%
Notes: Total costs were
truncated at $50,000 and
pharmacy costs at $20,000.
For each resource use
measure, the highest R2
value is highlighted. Data
are from IMS Health
Incorporated.
Notes: Total costs were
truncated at $50,000 and
pharmacy costs at
$20,000. For each
resource use measure, the
highest R2 value is
highlighted. Data are
from IMS Health
Incorporated.
8-5
Dx-PM is the best performer across populations in terms of predicting total costs. Not
surprisingly, Rx-PM is a much better predictor of pharmacy costs.
Where the models are combined we see the best performance of all. For these models we
chose to present results from models that both include and exclude prior cost information
as it is a powerful predictor of future cost. As can be seen by the results summarized in
Table 3, prior cost gives a boost to explaining pharmacy costs but appears to have little
impact on estimates of total cost. This option is available for Dx-PM alone and Rx-PM
alone as well.
Values) Using Year 1 Risk Factors and Year 2 Resource Use Measures
for Prior Cost and ICD & NDC Based Combined Models
Predictive Model
Total
Costs
Pharmacy
Costs
Combined Models
Predictive Model
Total
Costs
Pharmacy
Costs
Combined Models
DxRx-PM w/out Prior
Costs
24%
46%
DxRx-PM w/out Prior
Costs
19%
28%
DxRx-PM with Prior
Costs
26%
57%
DxRx-PM with Prior
Costs
19%
38%
Notes: Total costs were truncated at
$50,000 and pharmacy costs at
$20,000. For each resource use
measure, the highest R2 value is
highlighted. Data are from IMS
Health Incorporated.
Notes: Total costs were truncated at
$50,000 and pharmacy costs at
$20,000. For each resource use
measure, the highest R2 value is
highlighted. Data are from IMS
Health Incorporated.
8-6
Interpreting Adjusted R-Square Values
For predictive modeling of future high resource use, we are focused on the small segment
of highly co-morbid patients. This represents a very small portion of the range of values
being fitted to the regression. A model could perform well for the bulk of the population,
but perform poorly for these especially risky individuals (typically a small group) and
still provide credible R-Square performance statistics. It is also possible to precisely fit
models to any dataset by adding more variables and manipulating these variables
mathematically to assume a perfect form which could be logarithmic, exponential,
quadratic, or some other transformation. This, in essence, is what happens when
advanced neural networking approaches are used. The problem with such approaches is
two fold. First, the more convoluted the model, the less sense it will convey to clinicians
and other decision makers who are supposed to take action. Second, highly fitted models
will change with the dataset, often from year to year, which can introduce considerable
instability. Instability in terms of payment applications can adversely impact health plans
and care providers. Instability in terms of performance assessment will undermine
credibility among those being assessed.
Predictive Ratios
The credibility of models does not solely rest on regression fit. Regression provides
scores related to individuals while most applications of predictive modeling relate to
populations of individuals. To determine how well predictions perform in estimating cost
across the full range of risk, we also recommend looking at persons in different prior cost
cohorts (we use quintiles). Prior to the advent of diagnosis-based predictive models,
actuaries often relied on the age and gender distribution of the population as the basis for
making financial predictions. The next analysis examines predictive ratios for an
age/gender model compared to the suite of ACG predictive models and is summarized in
Table 4.
Predictive ratios represent the ratio of the expected cost predicted by particular modeling
strategies to the observed year two costs. The goal is to approach 1.0 as closely as
possible, especially in the high cost quintile. Ratios less than 1.0 mean that the model
under-predicts future costs while ratios over 1.0 suggest that the model is over-predicting
costs. The following table shows these ratios for predicting total cost for Dx-PM, RxPM, DxRx-PM. These results are contrasted to a model that that relies solely on an age
and gender as inputs.
8-7
Table 4: Predictive Ratio by Year-1 Cost Quintile for Commercial and
Medicare Populations
Model
Predictive Ratio,
Predictive Ratio,
Commercial
Medicare
Quintile (quintile=180,800) (quintile=18,700)
Age and Gender
1
2.72
3.35
2
1.66
2.09
3
1.17
1.50
4
0.89
1.05
5
0.49
0.48
1
1.45
1.34
2
1.31
1.15
3
1.20
0.99
4
1.06
0.97
5
0.84
0.92
1
1.56
1.56
2
1.33
1.29
3
1.19
1.11
4
1.09
1.01
5
0.82
0.77
Dx-PM w/o Prior
Cost
Rx-PM w/o Prior
Cost
8-8
Model
Predictive Ratio,
Predictive Ratio,
Commercial
Medicare
Quintile (quintile=180,800) (quintile=18,700)
DxRx-PM w/o
Prior Cost
1
1.14
1.13
2
1.12
1.10
3
1.11
1.00
4
1.07
1.00
5
0.92
0.95
There is a tendency for all of the models to somewhat overestimate costs among the
lowest cost quintile while somewhat underestimate costs in the highest cost quintile. This
effect is most pronounced in the age and gender model. The diagnosis- and pharmacybased models all substantially outperform age and gender. The performance of the
combined model is exemplary and especially so for Medicare, yielding predictions close
to 1.0 in the three highest cost groups. The diagnostic model performs a little better than
the pharmacy-based model, especially for Medicare. If the predicted cost is pharmacy
cost, the effectiveness of the Rx-PM model dramatically improves. We chose not to use
any model that includes prior cost in these comparisons since using prior cost as the basis
for forming the quintiles will favorably bias the results.
Sensitivity and Positive Predictive Value
An increasingly important use of the ACG Suite of Predictive Models has been for high
risk case identification. For this application, models are used more like a diagnostic test
and performance is measured at how well true cases are identified and false positives
ignored. The focus tends to be on two key indicators, sensitivity and positive predictive
value.
The computational approach for these indicators is as follows:
•
Sensitivity = True Cases Identified/All True Cases In Population
•
Positive Predictive Value = True Cases Identified/(True Cases Identified + False
Positives)
Sensitivity is likely to be of greatest interest to epidemiologists and others who are
focused in the health of the population since they are considering how well the “test”
captures all of the high risk individuals in a population.
Positive predictive value will likely be of greatest interest to clinicians/care managers
who want to know the likelihood that a particular patient is actually high risk.
8-9
In evaluating model performance in terms of PPV and sensitivity, the choice of cutpoints
can be very important. The cut point defines the percentage threshold one sets for
defining someone as a high cost case (i.e., defining some with the disease of interest). If
the cut point for high cost was set to 1%, which bounds the population of interest (true
positives) to a relatively small group of very costly individuals, misidentification of cases
is likely to be much higher than if the cut point is set to define a larger high risk pool.
Setting a cut point of 5%, for example, will improve sensitivity and PPV by roughly 10
percentage points over an approach using the 1% cut point. Decisions about cut points
can be as much economic as they are statistical. If the costs of identifying cases are high
(e.g., identifying persons who will receive one-on-one case management) then a more
restrictive cut point is probably the best approach. When screening for large populationbased disease management initiatives, threshold of 5% or even larger may be appropriate.
In the largest populations, distinctions between persons at the 1% and 5% thresholds are
likely to blur, they will all tend to be very co-morbid.
Selected model performance for commercial and Medicare populations in terms of their
diagnostic capabilities is presented in Table 5.
Table 5: Sensitivity and PPV for Top Five Percent
Sensitivity/PPV
Predictive Model (Predicting Total Cost)
Population Percent
Commercial Medicare
Prior Cost
35
26
Age and Gender
20
12
Dx-PM w/o Prior Cost
33
28
Rx-PM w/o Prior Cost
34
23
DxRx-PM w/o Prior Cost
37
24
Notes: The true positive rate is the same as the sensitivity and positive predictive value,
because the risk score cut-point is set equal to the outcome cut-point (in this example, this
is the top 5%).
In this example, the model to beat is prior cost which is a strong predictor of future cost.
None of the ACG models presented here used prior cost as an independent variable
although the use of prior cost in our models tends to improve predictive performance.
The ACG models are all strong performers, exceeding the performance of prior cost
models under some circumstances. It is important to note that the cases identified by the
ACG models are not always the same as those identified by the prior cost model. They
were selected using clinical criteria and are usually much more actionable than those
selected with prior cost alone.
8-10
Prior cost is a good predictor of future cost and for high risk case identification has been
one of the anchors for case selection in the absence of other information. However, prior
cost has a number of notable limitations. First, it provides no information on the clinical
condition of the patients. Further, some prior cost cases will revert to lower costs in the
subsequent year because the period of instability and medical crisis has passed. This
phenomenon is referred to as “regression to the mean.” Selection based upon diagnostic
and pharmacy codes, while still susceptible to this phenomenon, is more likely to yield
appropriate cases for care management.
ROC Curve
In addition to sensitivity and specificity in assessing the value of a predictive model as a
diagnostic test, there is a measure of model fit for case identification, the Receiver
Operating Characteristics Curve or ROC. The ROC was originally developed in World
War II to evaluate the performance of radar units, addressing the fact that as the
sensitivity or gain of a test is raised, there is a greater likelihood to introduce false
positives or noise. The C-Statistic is a measure of the fit where models produce the
greatest signal in relation to noise. A C-Statistic of 0.5 indicates that true cases are
indistinguishable from false positives (or a model no better than chance). A C-Statistic of
at least 0.8 is widely accepted as a threshold for good test performance. For models
predicting total cost with a 5% cut point, (i.e., top 5% of actual year two costs defines
high risk), C-statistics are summarized in Table 6.
Table 6: C-Statistics
C-Statistic
Model
Commercial
Medicare
Dx Alone
.820
.755
Dx + Total Cost
.833
.760
Rx Alone
.797
.718
Rx + Rx Cost
.802
.719
Combined Alone
.830
.764
Combined + Total Cost
.835
.765
Model performance overall for the commercial models generally exceeds the 0.8
threshold. We would expect somewhat better performance from the commercial models
because the commercial validation dataset is 10x larger than the Medicare dataset. The
inclusion of prior cost data in the models adds a slight performance boost. Pharmacy data
perform well in predicting future high cost users even where the predicted metric is total
cost and not pharmacy cost.
8-11
Clinical Information Improves on Prior Cost
Some of the challenges of using predictive models for case identification are expressed in
the Venn diagram (Figure 1) on the next page. The three ellipses cover those who are
high cost in the prior year, those who the model predicts as being high cost, and those
who turn out to be high cost. The G represents a particularly intriguing group since it
includes those who are actually high risk and who have been solely identified by the
predictive model and, therefore, would have been missed if only prior cost was used. If
all the available data is being effectively used, then the area covered by H should be
minimized.
8-12
Figure 1: Venn Diagram Comparing Overlap of Prior, Predicted, and
Actual High Cost Users
Actual High Cost Year-1
(Prior Use)
Predicted High Risk in
Year-2 (Using Year-1 Data)
B
C
E
D
F
High Risk,
Current Costs
G
Low, Future
H
Costs High
Actual High
Cost Year-2
A
Not High Risk
Some summary statistics can be applied that can help to capture the essence of this
diagram. Two perspectives on gauging model performance were taken with respect to
this diagram. First, how well does using both ACG predictive models and prior cost do
in identifying true high cost cases, the areas covered by D, E, and G in the diagram? Our
analyses indicate that by using both approaches one can identify about 45% of true cases
where high risk cases are defined as those in the highest 5% of total costs in year two.
Second, we focused on just the G portion of the chart, those cases that were uniquely
identified by ACG predictive models. If you used both prior cost and an ACG predictive
model as the basis for identifying potentially high risk cases, then the ACG System
uniquely identifies around 20% of the patients. In other words, ACG predictive modeling
identifies individuals who would NOT have been identified if only a prior cost model had
been used. This is an important population for case managers as they represent a highly
actionable group of patients, those who are currently not in the highest cost category.
The needs of these case managers might be best met by continuing to use prior cost as
one level of screening but to also apply the pure diagnostic (or pharmaceutical) predictive
models, where ACG predictive models can provide clinical context for all those cases
that they identify. This will leave Group D, those cases only identified by prior cost, for
whom there will be no additional clinical detail.
Predicting Hospitalization
9-i
9 Predicting Hospitalization
Introduction ................................................................................................... 9-1
Framework for Predicting Hospitalization Using ACG ............................ 9-1
Why Predict Hospitalization ...................................................................... 9-1
What Hospitalizations Are We Predicting ................................................. 9-1
Figure 1: Typology of Acute Care Inpatient Hospitalization ..................... 9-2
How is ACG Used to Predict Hospitalization ............................................ 9-2
Figure 2: Schematic Framework for Predicting Hospitalization ................ 9-3
Why Include Utilization in Prediction Models for Hospitalization ........... 9-3
Utilization Markers .................................................................................... 9-3
Likelihood of Hospitalization..................................................................... 9-4
Empiric Validation of the Likelihood of Hospitalization Model .............. 9-5
Sensitivity and Positive Predictive Value .................................................. 9-6
Table 1: PPV and Sensitivity for Predicting Hospitalization ..................... 9-6
C-Statistic ................................................................................................... 9-6
Table 2: C-Statistics for Predicting Hospitalization .................................. 9-7
9-ii
9-1
Introduction
This chapter introduces the ACG model for predicting hospitalization. It describes the
framework for predicting hospitalization and concludes with an empiric validation of the
model.
Framework for Predicting Hospitalization Using ACG
Why Predict Hospitalization
The ability to predict hospitalizations with some level of accuracy will help improve the
allocation of healthcare resources, in budgetary systems and in systems that link financial
incentives to efficiency and quality of care.
Inpatient care is part of the healthcare system in every country. A model-based screening
strategy can be of value in general for identifying high risk individuals for interventions
that are designed to improve population health.
Inpatient care is more resource intensive compared to other types of healthcare. Hospital
financing in the U.S. takes the efficiency and quality of inpatient care into consideration.
One particular area of interest related to quality is the rate at which patients are readmitted within a short period after a prior hospitalization. A tool for identifying
individuals at risk for re-admission is valuable in this context.
What Hospitalizations Are We Predicting
Hospitalizations for childbirth can be anticipated with near certainty. Other
hospitalizations are predictable based on prior medical history and a variety of other
factors. Our efforts are directed at inpatient hospitalizations that are predictable with a
degree of uncertainty. We hypothesize that unanticipated hospitalizations have a potential
for being predictable from individual morbidity data. For example, hospitalizations that
are clinically attributable to congestive heart failure can have a medical history leading
up to the event.
Unanticipated hospitalizations may also have an external cause such as poisonings and
injuries suffered in accidents. These hospitalizations are an additional focus of our
predictive modeling. Figure 1 gives an overview over a guiding typology of
unanticipated hospitalizations.
The hospitalization prediction model differs from the Hospital Dominant Condition
marker. Hospital Dominant Conditions represent a very small subset of diagnoses
associated with very high rates of admission in the following 12 months (See Chapter 6,
Special Population Markers for more information). The Hospital Dominant Conditions
contribute significantly ACG-PM cost predictions. Hospital Dominant Conditions have
high Positive Predictive Value (PPV) but lack sensitivity in identifying the full pool of
patients at risk for hospitalization. The Hospitalization Prediction models use ACGdefined comorbidity and previous utilization (See Figure 2 below) to identify risk of
hospitalization across the entire population.
9-2
Figure 1: Typology of Acute Care Inpatient Hospitalization
UNANTICIPATED HOSPITALIZATION
NO EXTERNAL CAUSE
Examples:
Elective joint replacement surgery (Arthritis)
WITH EXTERNAL CAUSE
Examples:
Injuries (due to accidents)
Poisonings
ANTICIPATED HOSPITALIZATION
Example:
Childbirth
How is ACG Used to Predict Hospitalization
There is a conceptual link and an empiric correlation between predicting resource use and
predicting inpatient hospitalization. The ACG predictive model framework, first
introduced in the chapter on predicting resource use, also applies to the prediction of
hospitalization. As mentioned in the earlier chapter, optimal performance of an ACG
predictive model occurs when the weighting of risk factors employs the outcome that is
being predicted. This fact motivates the formulation of a predictive model that is aimed at
hospitalization events. Figure 2 presents a schematic of the ACG framework for
predicting hospitalization.
A tiered approach to making predictions about hospitalizations has been adopted. We
make predictions about future hospitalization based on whether a person had a
hospitalization in the prior time period. Persons who have had a prior hospitalization are
thought to have an increased likelihood of hospitalization in a subsequent time period. In
the absence of a prior hospitalization, a secondary criterion for making predictions about
hospitalization relates to the age of a person. Older persons are thought to have an
increased likelihood of hospitalization compared to younger individuals. Our approach
uses age 55 as the threshold for separating the two age groups.
9-3
Figure 2: Schematic Framework for Predicting Hospitalization
Why Include Utilization in Prediction Models for Hospitalization
The ACG predictive model framework has been amended to improve the prediction of
hospitalization events. The framework for predicting hospitalization includes certain
utilization markers which represent a plausible set of valid predictors. For example,
emergency care and inpatient hospitalization can be precursors to future inpatient
hospitalization. This chapter introduces the utilization markers that are included in the
predictive model framework for hospitalization events.
Utilization Markers
Several utilization markers support the calculation of hospitalization models. These
markers require detailed dates, place of service, procedures, and revenue codes in the
Medical Services Input File. If the required fields are absent, the markers will not be
calculated. Alternatively, these markers can be drawn from the patient file to support the
hospitalization models.
Dialysis Service
Dialysis service is a binary flag to indicate that a patient with chronic renal failure (EDC
REN01) has received at least one dialysis service during the observation period.
Nursing Service
Nursing service is a binary flag to indicate that a patient has used services in a nursing
home, domiciliary, rest home, assisted living, or custodial care facility during the
observation period.
9-4
Major Procedure
Major procedure is a binary flag attached to a patient to indicate that a major procedure
was performed in an inpatient setting during the observation period. This marker
considers several CPT1 and HCPCS codes as major procedures, according to the
Berenson-Eggers Type of Service (BETOS) classification system.
The BETOS classification system consists of readily understood clinical categories and is
widely used by the CMS, the Medicare Payment Advisory Commission, and many health
services researchers and organizations as the standard classification system for types of
services.
Emergency Visit Count
This count is an integer count of emergency room visits that did not lead to a subsequent
acute care inpatient hospitalization during the observation period. This marker considers
place of service, procedure code, and revenue code to identify emergency room visits.
Inpatient Hospitalization Count
The inpatient hospitalization count is an integer count of inpatient confinements during
the observation period. The count of inpatient hospitalizations excludes admissions with
a primary diagnosis for pregnancy, delivery, newborns, and injuries. Transfers made
within and between providers count as a single hospitalization event. This marker is
dependent upon revenue codes and dates of service.
Outpatient Visit Count
The outpatient visit count is an integer count of outpatient encounters with place of
service of physician office (11), outpatient hospital (22), and other (24, 25, 26, 50, 53, 60,
62, 65, 71, 72) place of service codes. Visits are counted a maximum of once per date of
service and provider.
Likelihood of Hospitalization
Hospitalization prediction is a refinement to the ACG predictive models specifically
calibrated to identifying patients with risk of future hospitalization. The hospital
prediction models focus on unanticipated hospitalizations and are complementary to the
ACG cost prediction models.
There are five predictive model outputs related to the likelihood of hospitalization. These
models are intended to be used for the indicated outcome. A value of providing multiple
model outputs is greater sensitivity of each model calibrated to a particular outcome, as
compared to using a single model.
1
CPT codes copyright 2009 American Medical Association. All rights reserved. CPT is a trademark of the
AMA.
9-5
These are logistic models that generate a probability score indicating the likelihood of a
future hospitalization event. These models incorporate the utilization markers of inpatient
hospitalizations, emergency department visits, outpatient visits, dialysis services, nursing
services, and major procedures, in addition to traditional ACG predictive modeling
variables. Hospitalization models require the Medical Services Input File and optionally
include the Pharmacy Input File when available. The model outputs are discussed below.
Probability IP Hospitalization Score
Probability IP hospitalization score is the probability score for an acute care inpatient
hospital admission within the 12 months subsequent to the observation period.
Probability IP Hospitalization Six Months Score
The probability IP hospitalization six months score is the probability score for an acute
care inpatient hospital admission within the six months subsequent to the observation
period.
Probability ICU Hospitalization Score
The probability ICU hospitalization score is the probability score for an Intensive Care
Unit or Critical Care Unit admission within the 12 months subsequent to the observation
period.
Probability Injury Hospitalization Score
The probability injury hospitalization score is the probability score for an injury-related
admission within the 12 months subsequent to the observation period.
Probability Extended Hospitalization Score
The probability extended hospitalization score is the probability score for being admitted
to an acute care hospital for 12 or more days (across one or more admissions) within the
12 months subsequent to the observation period.
Empiric Validation of the Likelihood of Hospitalization Model
A longitudinal healthcare claims database, licensed from Phar-Metrics, a unit of IMS,
Watertown, MA, was used for our model development and validation. The database is
generally representative of the U.S. insured population in commercial and Medicare
managed care plans. Data for the period 2005-2006 were used for model development,
and data for the period 2006-2007 were used for model validation. Persons included in
the analysis had at least six months of enrollment in the first year and one or more
months of enrollment in the second year. The validation database included 4.5 million
persons under age 65 and 200,000 Medicare eligible beneficiaries.
9-6
Sensitivity and Positive Predictive Value
An important use of the ACG Predictive Models is for high risk case identification. For
this application, model performance is measured by how well true cases are identified
and false positives are avoided. The focus is on two key indicators, sensitivity and
positive predictive value:
Sensitivity = True Cases Identified/All True Cases In Population
Positive Predictive Value = True Cases Identified/(True Cases Identified + False
Positives)
Sensitivity measures how well the model captures all hospitalized individuals in a
population. Positive predictive value measures the likelihood that a particular patient is
being hospitalized in the next time period.
Table 1 illustrates, using sensitivity and positive predictive value measures, that the ACG
models are strong performers compared to using prior cost alone. The top 5% of
hospitalization probability scores define cases that are identified.
Table 1: PPV and Sensitivity for Predicting Hospitalization
Predictive Model
Positive Predictive Value
Sensitivity
IP Hospitalization with prior cost
and diagnosis and pharmacy data
input
21.2%
33.3%
IP Hospitalization with prior cost
and diagnosis data input
20.8%
32.6%
Prior Cost (as a benchmark to the IP
Hospitalization model)
14.2%
22.4%
C-Statistic
The C-Statistic is a measure of model fit. Chapter 8 on Predictive Modeling Statistical
Performance describes the relationship between this summary measure and the area under
the ROC curve. A C-Statistic of 0.5 indicates that true cases are indistinguishable from
false positives, or in other words, the test model is no better than tossing a fair coin. A CStatistic of 0.8 is widely accepted as a threshold for good model performance.
C-statistics for models that predict hospitalization are summarized in Table 2. The table
presents results for each of the models and population groups included in the tiered
approach to making predictions about hospitalization. The predictive model is IP
Hospitalization with predictors that are based on prior cost and diagnosis and pharmacy
data.
9-7
Table 2: C-Statistics for Predicting Hospitalization
Predictive Model
Persons with Prior
Hospitalization
Persons Age
55 or older
Persons Age less
than 55
IP Hospitalization
.751
.718
.741
IP Hospitalization Six Months
.754
.728
.747
ICU Hospitalization
.805
.754
.757
Injury Hospitalization
.808
.748
.668
Extended Hospitalization
.842
.793
.721
The C-Statistics for predicting hospitalization in Table 2 are above 0.7 for most
outcomes and groups. Several plausible patterns emerge from the empiric validation.
First, ACG better predicts hospitalization events for persons who had a hospitalization
within the previous 12 months, compared to persons who did not. Second, ACG better
predicts specialized hospitalization events, compared to a more general hospitalization
event. This finding validates the fact that optimal performance of an ACG predictive
model occurs when the weighting of risk factors employs the outcome that is being
predicted.
9-8
Coordination
10-i
10 Coordination
Introduction................................................................................................. 10-1
H
H
Assessing Care Coordination ..................................................................... 10-1
H
H
Table 1: Face to Face Physician Visits .................................................... 10-3
Table 2: Eligible Specialties..................................................................... 10-5
Table 3: Specialties not Eligible ............................................................. 10-6
Table 4: Generalist Specialties................................................................ 10-7
H
H
H
H
H
H
H
H
Majority Source of Care (MSOC) ............................................................. 10-7
H
H
Figure 1: Majority Source of Care ........................................................... 10-8
H
H
Unique Provider Count (UPC): ................................................................. 10-9
H
H
Figure 2: Unique Provider Count............................................................. 10-9
H
H
Specialty Count (SC) ................................................................................ 10-10
H
H
Figure 3: Specialty Count ..................................................................... 10-10
H
H
No Generalist Seen (NGS)........................................................................ 10-11
H
H
Figure 4: No Generalist Seen................................................................ 10-11
H
H
Conclusion ................................................................................................. 10-13
H
H
10-ii
Coordination
Coordination
10-1
Introduction
The ACG System 9 introduces ACG Coordination Markers™ in order to identify
populations that are at risk for poorly coordinated care. Used by themselves or in
conjunction with other special population markers (See Technical Reference Guide,
Chapter 6), this set of markers adds another dimension so as to enhance the clinical
screening process. The basic premise behind the creation of ACG Coordination Markers
is that individuals receiving poorly coordinated care have worse clinical outcomes and
have higher medical expenses than individuals who are being provided coordinated care.
Assessing Care Coordination
Up to now, care coordination has been assessed by instruments that have not been able to
use administrative claims information. Traditional coordination assessment approaches
have been survey based targeting the patient, family member’s or providers perceptions
of care collaboration. Chart reviews have also been used to assess the information
exchanged between physicians, teamwork processes and performance. Finally, resources
and structures that support care coordination have also been used to indirectly evaluate
coordination. Surveys, chart reviews and resource information are not readily available
for coordination assessment, making it challenging to incorporate this dimension of care
into the clinical screening process.
Using only administrative claims information, ACG Coordination Markers are able to
assess whether an individual is at risk for receiving poorly coordinated care. Four patient
markers make up ACG Coordination Markers:
•
Majority Source of Care (MSOC): An assessment of the level of participation of
those providers that provided care to each patient.
•
Unique Provider Count (UPC): A count of the number of unique providers that
provided care to the patient.
•
Specialty Count (SC): A count of the number of specialty types (not the same as
number of specialists seen) that provided care to the patient.
•
No Generalist Seen (NGS): A marker for the absence of a generalist’s participation in
an individual’s care.
10-2
Coordination
The ACG Coordination Markers introduce the use of two new data sources:
•
CPT 1 (“Procedure Code” field of the Medical Services Input File) provides standard
Evaluation and Management procedures.
•
NPI or NUCC taxonomy codes 2 (“ProviderSpecialty_NPI” field of the Medical
Services Input File) provide standardized definitions for provider specialties.
F
F
Each of the ACG Coordination Markers limit their evaluation to outpatient face-to-face
physician visits as defined by CPT Evaluation and Management Codes.
1
CPT codes copyright 2009 American Medical Association. All rights reserved. CPT is a trademark of the
AMA.
2
Provider Taxonomy copyright 2009 American Medical Association on behalf of the National Uniform
Claim Code Committee (NUCC). All rights reserved. Current code lists are available at
http://www.nucc.org/
Coordination
10-3
Table 1: Face to Face Physician Visits
CPT Codes defining Face to Face
Physician Visit
92002
92004
92012
92014
99201
99202
99203
99204
99205
99211
99212
99213
99214
99215
99217
99218
99219
99220
99241
99242
99243
99244
99245
99341
99342
99343
99344
99345
99347
99348
99349
99350
99354
99355
99374
99375
99377
99378
99379
99380
CPT Codes defining Face to Face
Physician Visit
99381
99382
99383
99384
99385
99387
99391
99392
99393
99394
99395
99396
99397
99401
99402
99403
99404
99411
99412
99420
99429
99455
99456
99499
10-4
Coordination
The Majority Source of Care (MSOC), Unique Provider Count (UPC) and Specialty
Count (SC) further consider only those visits provided by eligible specialties (i.e.,
specialties that could reasonably manage the overall care for a patient). Examples of
eligible specialties are provided in Table 2.
Coordination
10-5
Table 2: Eligible Specialties
Eligible Specialties
Allergy & Immunology
Colon & Rectal Surgery
Family Medicine
Internal Medicine
Neurological Surgery
Neuro-musculoskeletal Medicine
Nuclear Medicine
Obstetrics & Gynecology
Ophthalmology
Oral & Maxillofacial Surgery
Orthopedic Surgery
Otolaryngology
Pain Medicine
Pediatrics
Physical Medicine & Rehabilitation
Plastic Surgery
Preventive Medicine
Psychiatry and Neurology
Radiology
Surgery
Thoracic Surgery (Cardiothoracic Vascular Surgery)
10-6
Coordination
Eligible Specialties
Transplant Surgery
Urology
Advanced Practice Midwife
Clinical Nurse Specialist
Nurse Practitioner
Physician Assistant
Face-to-face visits with provider specialties that are not eligible are not considered as part
of the Majority Source of Care (MSOC), Unique Provider Count (UPC) and Specialty
Count (SC) markers. Examples of specialties that are not considered eligible are
provided in Table 3.
Table 3: Specialties not Eligible
Specialties not Eligible:
Ambulance Services
Agencies
Dental Providers
Dermatologists
Dietary & Nutritional Service Providers
Facilities (Clinics, Ambulatory Health Centers, Emergency Departments, Hospitals,
Assisted living facilities)
Laboratories
Managed Care Organizations
Pharmacy Service Providers
Podiatrists
Psychologists
Social Service Providers
Speech, Language and Hearing Service Providers
Suppliers of Durable Medical Equipment & Supplies
Therapists (Physical – Respiratory)
Technicians
Coordination
10-7
A subset of provider specialties defines generalists. Patients with at least one outpatient
face-to-face visit to a generalist will have the No Generalist Seen marker set to “N”.
Patients with no outpatient face-to-face visits to a generalist will have the No Generalist
Seen marker set to “Y”. Table 4 provides the list of specialties considered generalists for
this marker.
Table 4: Generalist Specialties
Generalist Specialties
Family Medicine
Internal Medicine
Geriatricians
Pediatricians
Preventive Medicine
Nurse Practitioners
Obstetrics and Gynecology
Majority Source of Care (MSOC)
The Majority Source of Care marker will determine the percent of the outpatient visits
provided by eligible physicians that saw the member most over the measurement period.
For each patient, the application determines the following:
•
The percentage of care provided from a majority source.
•
The provider ID(s) that is responsible for the majority source of care. In the event that
more than one provider is assigned the same MSOC percentage, the application will
identify all the providers as being the majority source of care.
Figure 1 shows two patients. The patient on the left saw three eligible providers and
received eight of his ten outpatient services from the geriatrician. In this case, the
geriatrician is identified as the MSOC provider with a score of 0.80 (80%.) The patient
on the right saw four eligible providers and received four of her outpatient services from
the endocrinologist. The endocrinologist provider is identified as the MSOC with a score
of 0.40 (40%.)
10-8
Coordination
Figure 1: Majority Source of Care
Coordination
10-9
Unique Provider Count (UPC):
The Unique Provider Count (UPC) marker determines the number of unique eligible
providers that imparted outpatient care over the measurement period for any condition.
In Figure 2 the patient on the left saw three eligible providers giving him a UPC of 3.
The patient on the right saw four eligible providers and received a UPC of 4.
Figure 2: Unique Provider Count
10-10
Coordination
Specialty Count (SC)
The Specialty Count (SC) determines the number of eligible specialty types that provided
outpatient care over the measurement period for any condition.
In Figure 3 the patient on the left saw three distinct specialty types. The patient on the
right saw four eligible providers. This patient received a specialty count of 4.
Figure 3: Specialty Count
Coordination
10-11
No Generalist Seen (NGS)
The No Generalist Seen (NGS) marker determines whether or not a patient has been
provided outpatient care by a generalist in the measurement period.
In Figure 4 the patient on the left saw a generalist for outpatient care. The patient on the
right did not see a generalist and was flagged as No Generalist Seen.
Figure 4: No Generalist Seen
Risk of Poor Coordination
The ACG Coordination Markers can be used together to provide a comprehensive picture
of coordination of care. Figure 5 considers only members with high morbidity levels (in
Resource Utilization Bands 4 and 5) to control for illness burden. Patients were defined
as at high risk for poor coordination when the MSOC Percent was less than 30%, Unique
Provider Count was greater than 5, and No Generalist Seen was Y. In contrast, patients
at low risk for poor coordination typically had less than 5 providers involved in their care
including at least one generalist. There is a dramatic impact on year one and year two
costs for members at high risk for poor coordination.
Figure 5: Impact of Poor Coordination on High Morbidity Users
10-12
Coordination
Figure 6 stratifies the low morbidity population using the same definitions for risk of
poor coordination. In this low morbidity population, members at high risk for poor
coordination represent a population of emerging risk typically not identified based on
predictive models.
Figure 6: Impact of Poor Coordination on Low Morbidity Users
Coordination
10-13
Conclusion
The combination of the ACG Coordination Markers provides a meaningful way to
identify members at risk for poor coordination. Members with poorly coordinated care
are more likely to have excess utilization as a result of redundant testing, potentially
harmful drug-disease interactions and overall lower quality care. These markers
complement the ACG-PM risk scores and additional system markers in further defining
the “at risk” population.
10-14
Coordination
Gaps in Pharmacy Utilization
11-i
11 Gaps in Pharmacy Utilization
Objectives..................................................................................................... 11-1
H
H
Significance of Pharmacy Adherence for Effective Care........................ 11-1
H
H
Development of Possession/Adherence Markers...................................... 11-2
H
H
Figure 1: Gap in Medication Possession.................................................. 11-2
Figure 2: Gap in Medication Possession Following Oversupply............. 11-3
Figure 3: Gap in Medication Possession Following Hospitalization ....... 11-3
Table 1: Markers of Pharmacy Adherence/Possession ........................... 11-4
Table 2: MPR Example 1........................................................................ 11-6
Table 5: CSA Example 1 ........................................................................ 11-7
How Medication Gaps are Defined And Constructed ............................. 11-8
Table 8: Condition-Drug Class Pairings ............................................... 11-10
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
H
Validation and Testing ............................................................................. 11-13
H
H
Table 9: Possible Types of Analyses .................................................... 11-14
H
H
The Future ................................................................................................. 11-15
H
H
11-ii
11-1
Objectives
This chapter is intended to introduce the concept of pharmacy possession or adherence
and describe the methods used by the ACG System to measure pharmacy adherence.
Pharmacy adherence represents an additional method for identifying candidates for care
management and provides new information that can be used to manage and improve
patient care.
Significance of Pharmacy Adherence for Effective Care
There is a considerable literature on pharmacy adherence and a substantial body of
evidence that high levels of medication adherence yield improved therapeutic and cost
effective outcomes. Medication adherence represents an important dimension of
effective treatment.
Medication adherence is one dimension of patient compliance with therapy. Optimally,
medication adherence means taking the correct drugs, for the correct indications, at the
correct times, at the correct dose, and under the proper conditions for safe and effective
use (storage, shelf life, avoiding substances that could affect efficacy).
Probably the only way to accurately assess medication adherence on all of these
dimensions is through direct patient observation. Such a measurement strategy would be
highly impractical. Researchers have developed a number of less intrusive strategies for
assessing medication adherence including patient logs, periodic assessments of the
patient’s medication supply, and electronic sensors in the medication bottles that record
when the bottle was opened. These approaches constitute of indicators of pharmacy
adherence but none fully address all of the dimensions of medication adherence that were
defined earlier.
Much research on medication adherence relies completely on what is reported within
pharmacy claims. With claims, what we are really talking about is an indicator of
medication possession. You know when the medication prescription was filled and how
many days supply were given. There is also a fairly mature literature on pharmacy
possession and a number of widely used metrics. In constructing our new markers, we
chose to make good use of this prior work.
There is a substantial literature on metrics for capturing prescribing gaps. For an
excellent review see Hess et al. (Annals of Pharmacotherapy, 40:1280-1288, 2006).
11-2
Development of Possession/Adherence Markers
Generally, measurement strategies have targeted specific possession events (i.e., gaps) or
average possession of time expressed as a ratio (supply over prescribing period).
We chose to employ markers using both of these strategies as they address different
dimensions of adherence. Gaps capture acute occurrences and may represent a
therapeutically significant event that could be overlooked if only averages were
considered. Gaps if captured on a timely basis are also amenable to care management
intervention. Averages represent how well medication is supplied over a span of time
(i.e, one year). They are a good summary indicator of possession in assessing the overall
status of a patient but are less amenable to direct clinical action.
Both approaches follow the same conceptual model of prescribing gaps. Figure 1 depicts
a gap event:
Figure 1: Gap in Medication Possession
A gap begins with a prescription for a particular medication in the time period of interest.
The prescription covers a span of time that is defined by the associated days supply. The
interval between the end of the days supply and the onset of a new prescription for this
medication represents a gap. A common approach to gap measurement is to include a
grace period at the end of the days supply to account for such factors as surplus supply
from an earlier prescription and to ensure that the reported gap is clinically significant.
For this example we have chosen to use 15 days as the grace period.
11-3
We have also chosen to address events which could complicate the measurement of gaps
and potentially introduce false positives. The first event is when a patient renews a
prescription before their current supply has been completely exhausted. This scenario is
presented in Figure 2:
Figure 2: Gap in Medication Possession Following Oversupply
Under these circumstances, there is a surplus supply between the start of the renewal and
the end of the supply from the original prescription. We have chosen to add this surplus
to the end of the days supply of the renewal. So the computation of a gap considers both
this surplus and the 15 day grace period.
Another less common event that we have chosen to address is if the patient is
hospitalized at some point. Customarily, the hospital pharmacy is responsible for all
prescriptions during the period of hospitalization. We have chosen to address this
scenario as presented in Figure 3:
Figure 3: Gap in Medication Possession Following Hospitalization
If a hospitalization occurs during the prescribing period, whatever supply the patient had
on hand when they were hospitalized is presumed to continue after discharge. This
remaining supply plus the 15-day grace period must be completed before we begin to
measure a gap.
11-4
Table 1: Markers of Pharmacy Adherence/Possession
There are four markers that have been adopted for pharmacy adherence/possession. Each
marker represents a slightly different perspective on adherence.
Note that gaps identify significant intervals where there is no medication supply over the
observation period and consider individual problematic events while averages provide an
overall perspective on adherence over time. The averaging and gaps approaches are
intended to complement each other. It is possible to have an acceptable average and still
experience a major gap in prescribing. Averages are also appropriate measures to
measure adherence at a population level.
Marker
Definition
Application/Interpretation
Number of Gaps
Count of occurrences where time
interval between end of supply of
one prescription and onset of next
prescription within the same drug
class is more than the grace
period.
To be significant, the gap must
extend beyond a grace period
(currently 15 or 30 days
depending on the condition and
drug class pair).
Medication Possession Ratio
(MPR)
Total number of days for which
medication is dispensed
(excluding final prescription)
divided by the total number of
days between the first and last
prescription. If a person is on
multiple drug classes for a single
condition, the days supply and
prescribing period are evaluated
separately and weighted across
drug classes.
The Medication Possession Ratio
(MPR) represents an average
medication possession over the
prescribing period. In common
usage, an MPR of .80 or higher is
considered good adherence. The
MPR can conceivably exceed 1.0
if there is excess supply. In
general, this measure is more
sensitive to large gaps than to
frequent gaps.
Continuous Single-Interval
Measure of Medication
Availability (CSA)
Ratio of days supply to days until
the next prescription averaged
across all prescriptions
The Continuous Single Interval
Measure of Medication
Acquisition (CSA), also an
average, considers adherence
over discrete prescribing events.
This approach equally weights
each prescribing event. This
metric is also allowed to exceed
1.0. In general, this measure is
more sensitive to frequent gaps
than to large gaps.
Untreated
No evidence of treatment with a
designated class of
pharmaceuticals. (Refer to the
“Special Population Markers”
chapter in the Technical
Reference Guide for Definitions
of the Untreated Condition
Marker.)
The untreated marker indicates
instances where although chronic
drug administration is warranted
for a condition, there is not
evidence of such. Non-treatment
represents another distinct type of
care management issue. It should
be noted that although this metric
Marker
11-5
Definition
Application/Interpretation
indicates an absence of
prescribing, prescribing could be
occurring or samples could be
used. Pharmacy benefits could be
outsourced and there is no claims
data stream available. There also
may be no pharmacy benefits.
Medications may be administered
by injection in the physician’s
office and recorded in medical
claims (J-codes) rather than
pharmacy claims (NDC Codes).
Patients may have newly begun
medication with only a single
prescription. Due to side effects
or performance of the drug,
patients may need to switch
between therapeutic alternatives
in different drug classes. In these
instances, we have insufficient
information to evaluate
possession and the patient may
erroneously be identified as
untreated. Issues such as these
should be researched before
trying to draw conclusions from
this metric.
11-6
Medication Possession Ratio (MPR)
Medication possession ratio is calculated for each of the condition-drug class pairings in
Table 8. For each condition, there is a field with the naming convention of the specific
condition appended with _MPR containing the ratio of days supply to days per
prescribing period. If a person is on multiple drug classes for a single condition, the days
supply and prescribing period will be evaluated separately and weighted across drug
classes. Tables 2, 3, and 4 provide examples of different MPRs.
Table 2: MPR Example 1
Supply Available
Upon Refill
Days Exceeding
Grace Period
Rx_fill_date
Days_supply
1/15/2009
30
2/10/2009
30
4/2/2009
30
2
5/19/2009
30
2
4
MPR=90 days supply / (5/19/2009 – 1/15/2009) = 0.73
Rx_fill_date
Days_supply
1/15/2009
30
2/10/2009
30
3/18/2009
30
4/19/2009
30
Supply Available
Upon Refill
Days Exceeding
Grace Period
4
MPR=90 days supply / (4/19/2009 – 1/15/2009) = 0.96
Supply Available
Upon Refill
Rx_fill_date
Days_supply
1/15/2009
30
2/10/2009
30
4
3/10/2009
30
6
Days Exceeding
Grace Period
4/06/2009
11-7
30
9
MPR=90 days supply / (4/06/2009 – 1/15/2009) = 1.11
In general, this measure is more sensitive to large gaps than to frequent gaps. MPR can
be greater than 1.0 if the member consistently refills prior to exhausting days supply.
Additional detail of drug classes and refill dates is available in the Pharmacy Spans
Export File.
Continuous, Single-interval Measure of Medication Availability (CSA)
The Continuous, Single-interval Measure of Medication Availability is calculated for
each of the condition-drug class pairings in Table 8. For each condition, there is a field
with the naming convention of the specific condition appended with _CSA. Tables 5, 6,
and 7 provide examples of different CSAs.
Table 5: CSA Example 1
Supply Available
Upon Refill
Days Exceeding
Grace Period
Rx_fill_date
Days_supply
1/15/2009
30
2/10/2009
30
4/2/2009
30
2
5/19/2009
30
2
4
CSA = ((30 / (2/10/2009 – 1/15/2009)) + (30 / (4/2/2009 – 2/10/2009)) + (30 / (5/19/2009
– 4/2/2009)))/3 = 0.79
Rx_fill_date
Days_supply
1/15/2009
30
2/10/2009
30
3/18/2009
30
4/19/2009
30
Supply Available
Upon Refill
Days Exceeding
Grace Period
4
CSA = ((30 / (2/10/2009 – 1/15/2009)) + (30 / (3/18/2009 – 2/10/2009)) + (30 /
(4/19/2009 – 3/18/2009)))/3 = 0.97
11-8
Supply Available
Upon Refill
Rx_fill_date
Days_supply
1/15/2009
30
2/10/2009
30
4
3/10/2009
30
6
4/06/2009
30
9
Days Exceeding
Grace Period
CSA = ((30 / (2/10/2009 – 1/15/2009)) + (30 / (3/10/2009 – 2/10/2009)) + (30 / (4/6/2009
– 3/10/2009)))/3 = 1.11
In general, this measure is more sensitive to frequent gaps than to large gaps. CSA can
be greater than 1.0 if the member consistently refills prior to the exhausting days supply.
Additional detail of drug classes and refill dates is available in the Pharmacy Spans
Export File.
How Medication Gaps are Defined And Constructed
Medication possession has a different meaning in the context of individual diseases and
drugs. The majority of drugs are given acutely and gaps have little relevance in this case.
What distinguishes ACG Rx Gaps™ from other approaches that are commonly used in
the literature or provided by pharmacy benefits managers is the effort dedicated to
identifying diseases and drugs that are intended to be administered chronically. We have
identified seventeen chronic diseases where the chronic administration of medication is,
in most instances, appropriate. The condition definitions are described more thoroughly
in the “Special Population Markers” chapter in the Technical Reference Guide.
•
Bipolar Disorder
•
•
•
Ischemic Heart Disease/CAD
•
Depression
•
Osteoporosis
•
Diabetes ◄
•
Parkinson’s Disease◄
•
Glaucoma
•
Persistent Asthma
•
Human Immunodeficiency Virus ◄
•
Rheumatoid Arthritis
•
Disorders of Lipid Metabolism◄
•
Schizophrenia
•
Hypertension
•
Seizure Disorders
•
Hypothyroidism◄
11-9
To reduce the numbers of false positives and to avoid issues associated with drugs used
for more than one condition, the majority of the conditions identified will require an
inpatient diagnosis, an emergency department diagnosis or a minimum of two outpatient
diagnoses for the condition. Alternately, a minimum of two prescriptions in an
applicable drug class can be used to identify a condition when the usage pattern is clear.
Prescribing alone can be used to identify those conditions listed above that are marked
with an arrow.
Each targeted condition is associated with one or more target drug classes identified by
our clinician advisors as a subset of drugs that should be given continuously. The
resultant disease-drug class pairings are presented in Table 8. Note: The measurement of
gaps is confined to only these pairs. It is possible for patients to take multiple ingredients
within the same drug class and/or temporarily substitute one ingredient for another within
the drug class. To prevent the appearance of oversupply when multiple ingredients are
prescribed, Gap, MPR and CSA calculations are performed by ingredient.
11-10
Table 8: Condition-Drug Class Pairings
Condition
Drug Category
Condition
Drug Category
Bipolar Disorder
Bipolar Disorder
Congestive Heart
Failure
Congestive Heart
Failure
Congestive Heart
Failure
Congestive Heart
Failure
Congestive Heart
Failure
Congestive Heart
Failure
Depression
Diabetes
Diabetes
Anti-convulsants
Anti-psychotics
Hypertension
Hypertension
Aldosterone receptor blockers
Anti-adrenergic agents
ACEI/ARB
Hypertension
Beta-blockers
Aldosterone receptor blockers
Hypertension
Calcium channel blockers
Beta-blockers
Hypertension
Diuretics
Diuretics
Hypertension
Vasodilators
Inotropic agents
Hypothyroidism
Thyroid drugs
Vasodilators
Anti-depressants
Insulins
Meglitinides
Miscellaneous antidiabetic
agents
Ischemic Heart Disease
Osteoporosis
Antianginal agents
Beta-blockers
Calcium channel blockers
Hormones
Anticholinergic antiparkinson
agents
Dopaminergic antiparkinsonism
agents
Diabetes
Diabetes
Diabetes
Diabetes
Diabetes
Glaucoma
Human
Immunodeficiency
Virus
Disorders of Lipid
Parkinson's Disease
Non-Sulfonylureas
Other Anti-Hyperglycemic
Agents
Persistent Asthma
Sulfonylureas
Thiazolidinediones
Ophthalmic glaucoma agents
Persistent Asthma
Persistent Asthma
Persistent Asthma
Adrenergic bronchodilators
Immunosuppressive monoclonal
antibodies
Inhaled corticosteroids
Leukotriene modifiers
HAART (see below)
Bile acid sequestrants
Persistent Asthma
Persistent Asthma
Methylxanthines
Parkinson's Disease
Condition
11-11
Drug Category
Condition
Drug Category
Cholesterol absorption inhibitors
Disease-modifying antirheumatic drugs (DMARDs)
Fibric acid derivatives
Immunologic agents
HMG-CoA reductase inhibitors
Miscellaneous
antihyperlipidemic agents
ACEI/ARB
Schizophrenia
Anti-psychotics
Seizure Disorder
Anti-convulsants
Immunologic agents
metabolism
Disorders of Lipid
metabolism
Disorders of Lipid
metabolism
Disorders of Lipid
metabolism
Disorders of Lipid
metabolism
Hypertension
11-12
Our definition of a gap includes a grace period between the end of one prescription and
the start of another where a gap would not be counted. Thus a patient could be reported
as having no gaps in possession but actually have a number of small gaps that fall within
the grace period. There was a trade off made in the choice of grace periods. A longer
grace period improves the positive predictive value but reduces sensitivity. In most
instances, a 15-day grace period has been employed. We have used a 30-day grace
period when our clinical advisors agreed that the impact of an extended gap would be
minimal or is essential to accommodate attributes of dispensing or use. A 30-day grace
period has been adopted for the following conditon-drug class pairs:
•
Glaucoma-Ophthalmic Glaucoma Agents
•
Hyperlipidemia-All Drug Classes
•
Hypertension-All Drug Classes
•
Hypothyroidism-Thyroid Drugs
•
Osteoporosis-Hormones
•
Persistent Asthma-Adrenergic Bronchodilators
Reporting of gaps is summarized at the disease level. A number of conditions only have
one disease-drug class pair (e.g., osteoporosis). However, many others have multiple
pairs (e.g., persistent asthma). If a disease is associated with multiple drug classes, all
gaps associated with these drugs are reported. If a patient experiences eight gaps for five
different classes of tracked medications then eight gaps are reported by the software.
Further if a gap occurs for a combination drug, the gap is reported for each component
that falls into a separate drug class. Specific drugs and prescribing periods are stored and
made available in the Pharmacy Spans export file which is described in more detail in
Appendix A of the Installation and Usage Guide.
Human Immunodeficiency Virus presents some unique issues in computing gaps since a
combination of medications (HAART) is required for effective treatment. We organized
HIV/AIDS medications into the following classes:
•
Entry Inhibitors
•
HAART
•
Integrase Inhibitors
•
NNRTIs
•
NRTIs
•
Protease Inhibitors
11-13
For HIV/AIDS we tallied gaps according to whether or not a patient was taking some
form of HAART, a combination of drugs that represents best evidence-based practice.
Our operational definition for HAART uses the following scenarios to identify the
administration of HAART:
1. HAART (a three drug combination filled with a single prescription)
2. Any NRTI combination drug including either zidovudine , abacavir or tenofovir with
an additional NNRTI, entry inhibitor, integrase inhibitor or protease inhibitor
3. A NNRTI plus a ritonavir-boosted PI or nelfinavir
4. Three of more drugs from two or more different drug classes (Entry Inhibitors,
Integrase Inhibitors, NNRTIs, NRTIs, Protease Inhibitors)
Gaps differ greatly in terms of duration and the duration has potential implications for
care management. The phenomenon of someone with an extended gap in prescribing is
likely a very different phenomenon from someone experiencing a shorter gap. The
former problem appears to be one of treatment while the latter is one of possession. In
addition, substitution of an alternate ingredient within the same drug class may
occasionally be made in between prescribing events for various reasons; potentially
negating what may appear to be a gap. In both of these situations, extended gaps and
substitutions, we have chosen to exclude the span from the gap count, so as not to
potentially overstate the number of actual gaps. “Excluded” spans are flagged in the Rx
Eligible for Adherence column within the Pharmacy Spans export file, which is described
in more detail in Appendix A of the Installation and Usage Guide.
Validation and Testing
Pharmacy adherence/possession has strong face validity among clinicians. If patients fail
to comply with therapy then there are likely to be consequences downstream. For some
conditions these will be felt immediately while for others years may pass before serious
adverse events are experienced.
As part of our validation efforts for pharmacy adherence markers, we looked at the
relationship between gap counts and several important outcomes. The purpose of this
validation exercise was to determine whether or not there was an association between
numbers of gaps and either concurrent or prospective total cost. We were also looking at
the internal consistency of our markers with higher numbers of gaps being related to a
lower MPR.
Table 9 suggests the types of analyses that are possible. The example evaluates patients
with Congestive Heart Failure prescribed beta blockers, and the data are only shown for
persons in the highest two RUB categories (RUBs 4 and 5).
11-14
Table 9: Possible Types of Analyses
Number Of Gaps
Dependent Effects
Number of Cases
Median Total Cost Year 1
Median Pharmacy Cost Year 1
Median Total Cost Year 2
Median Pharmacy Cost Year 2
Median MPR
No Gap
One Gap
Two Gaps
3+ Gaps
4,990
2,314
882
692
$18,027
$23,617
$23,105
$17,377
$4,094
$3,642
$3,270
$2,908
$11,621
$13,190
$12,306
$12,685
$4,248
$3,754
$3,389
$3,069
0.99
0.84
0.71
0.57
This particular extract used patients who had a minimum of 120 prescribing days to
improve the homogeneity of the patient population. Setting such a threshold will also
winnow out persons who are receiving prophylactic treatment. For example, short
courses of HAART are routinely prescribed to persons who have been exposed to the
HIV virus.
These data exemplify a pattern that is commonly observed in medication possession data.
While one or two gaps may identify persons with higher concurrent and/or prospective
costs, the predicted dollars diminish with more gaps. It is possible that persons who
experience a higher number of gaps are chronic under-utilizers of medical services.
11-15
Medication possession is a complex issue that will require a close partnership between
analysts and clinicians. There are subtle distinctions in the significance of gaps
depending upon the condition. We have tried to capture some of these differences by
varying the number of days used in the grace period. Some gaps may occur because of
uncertainty in treatment where clinicians are trying different combinations of
medications. One form of treatment may not be achieving the desired results, a new
treatment is begun in a different class, and then the old treatment is resumed because of
complications associated with the alternative regimen. This sequence of events would
produce a large gap over the period where therapies are swapped out. We have tried to
avoid measurement problems associated with switching therapies by defining medication
classes where such behavior would be unlikely. But where evidence of such still exists,
as per the discussion above, we have chosen to exclude the span from the gap count, so as
not to potentially overstate the number of actual gaps.
Gaps may also potentially be indicative of good management as well. For persons with
persistent asthma, gaps in adrenergic bronchodilators could mean that the asthma is under
good control and that over-supply could mean the converse, poor control.
Because of the subtleties of meaning in interpreting gaps, it is highly recommended that
ACG results be shared with one or more clinical reviewers before identifying which
disease-drug class pairs are targeted for gap assessment and potential intervention.
The Future
Pharmacy adherence/possession represents one part of an ongoing effort by the ACG
Team to extract as much actionable information as possible from administrative claims
data. While the pharmacy adherence/possession markers are offered as a “complete”
solution, we anticipate that a number of improvements will be made. First, we will
incorporate these markers into our predictive models where they contribute to the
explanation of future costs. We will also expand the types of markers to permit more
nuanced analyses. Likely improvements will include the addition of both maximum gaps
and the total gap duration for the period. Finally we will continue to refine and
modification the specific chronic conditions that are tracked for medication/possession
adherence, enhancing definitions and including new conditions when warranted.
11-16
A-1
December 2009
Corrections or updates should be forwarded to [email protected]
“Adjust Utilization for Case Mix and Make Physicians Responsible for Remaining
Variation.” (1998). Data Strategy Benchmarks, 2(2), 25-28.
http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=PubMed&list_
uids=10345360&dopt=Abstract
Adams, E. K., Bronstein, J. M., Raskind-Hood, C. (2002) “Adjusted Clinical Groups:
Predictive Accuracy for Medical Enrollees in Three States.” Health Care
Financing Review. 24(1), 43-61.
http://www.cms.hhs.gov/review/02fall/02fallpg43.pdf
Adams, E. K., Bronstein, J. M., Becker, E. (2000). “Medi-Cal and Managed Care: Risk,
Costs, and Regional Variation.” Public Policy Institute of California.
http://www.ppic.org/content/pubs/R_1200KAR.pdf
AdvanceMed Corp., a DynCorp Company (2003). “Applicability of Predictive Risk
Groupers to MHS Data Model.”
Aguado, A. (2008). “Variability in prescription drug expenditures explained by adjusted
clinical groups (ACG) case-mix: A cross-sectional study of patient electronic
records in primary care.” BMC Health Services Research.
http://www.biomedcentral.com/1472-6963/8/53/abstract.
Anderson, G., Weller, W. (1999). “Methods of Reducing the Financial Risk of
Physicians Under Capitation.” AMA, Archives of Family Medicine, 8(2), 149-155.
http://archfami.ama-assn.org/cgi/content/short/8/2/149
Ash, A, McCall, N. “Risk Assessment of Military Populations to Predict Health Care
Cost and Utilization.” Final report prepared for Center for Health Care
Management Studies, HPA&E, TRICARE Management Activity, RTI Project
Number 08490.006. November, 2005.
Baldwin L.M., Klabunde C.N., Green P., Barlow W., Wright G. (2006). “In Search of
the Perfect Comorbidity Measure for Use With Administrative Claims Data.
Does it Exist?” Med Care, 44(8), 745-753.
Barnard D.K., Hu W. (2005). “The Population Health Approach: Health GIS as a
Bridge from Theory to Practice.” International Journal of Health Geographics,
4(23).
http://www.ij-healthgeographics.com/content/4/1/23
Bolibar-Ribas, B., Juncosa-Font, S. (2000). “Los Grupos De Atencion Ambulatoria
Como Sistema De Clasificacion De Pacientes para La Atencion Primaria.”
Cuadernos De Destion, 6(1), 19-24.
A-2
Blustein, J., Hanson, K., Shea, S. (1998). “Preventable Hospitalizations and
Socioeconomic Status.” Health Affairs, 17(2), 177-189.
http://content.healthaffairs.org/cgi/reprint/17/2/177
Bono, C., Shenkman, E., Hope-Wegener, D. (2000). “The Actual Versus Expected
Health Care Use Among Healthy Kids Enrollees.” Institute for Child Health
Policy.
Bonomi, A.E., Anderson, M.L., Reid, R.J., (2009) “Medical and Psychosocial Diagnoses
in Women with a History of Inmate Partner Violence.” Arch Internal Medicine
169(18): 1692-7.
Briggs. L. W., Rohrer. J. E., Ludke. R. L., Hilsenrath, P. E., Phillips, K. T. (1995).
“Geographic Variation in Primary Care Visits in Iowa.” Health Services Research
30(5), 657-671.
Broemeling A. (2005) “Chronic Care Management: Understanding Chronic Health
Conditions & Co-morbidity.” Presentation. Centre for Health Services and
Policy Research.
Broemeling, A. (2008) “Chronic conditions and co-morbidity among residents of British
Columbia.” Center for Health Services Policy and Research.
http://www.chspr.ubc.ca/files/publications/2005/chspr05-08.pdf
Broemeling, A., Watson, D., and Black, C. (2005). “Chronic conditions and comorbidity among residents of British Columbia.” Centre of for Health Services
and Policy Research, University of British Columbia. www.chspr.ubc.ca
Buntin, M.J.B., Newhouse, J. P. (1998). “Employer Purchasing Coalitions and
Medicaid: Experiences with Risk Adjustment.” The Commonwealth Fund.
http://www.cmwf.org/publications/publications_show.htm?doc_id=221418
Bureau of Primary Health Care by MDS Associates, I. (1998). “Differences in Case Mix
Between CHC Users and Non Users: Washington and Missouri, 1992 Final
Report.”
Carlin, C. (2006). “Optimal Pricing of Employment-Based Health Plans” Dissertation,
Graduate School of Minnesota.
Carlsson, L. (2004). “Burden of Illness in Defined Populations.” Department of
Clinical Services, Center of Family Medicine Karolinska Institutet.
Carlsson, L., Borjesson, U., Edgren, L. (2002). “Patient-based „Burden-of-Illness‟ in
Swedish Primary Health Care. Applying the Johns Hopkins ACG Case-mix
System in a Retrospective Study of Electronic Patient Records.” International
Journal of Health Planning and Management, 17(3), 269-282.
http://www3.interscience.wiley.com/cgi-bin/fulltext/97518189/PDFSTART
A-3
Carlsson, L., Strender, Lars-Erik, Fridh, Gerd and Nilsson, Gunnar. (2004). “Types of
Morbidity and Categories of Patients in a Swedish Country: Apply the Johns
Hopkins Adjusted Clinical Groups System to Encounter Data in Primary Health
Care.” Scandinavian Journal of Primary Health Care, 22, 174-179.
http://taylorandfrancis.metapress.com/app/home/contribution.asp?wasp=42f3515c
6f774b308efb5dccf3a34cfe&referrer=parent&backto=issue,11,14;journal,4,31;lin
kingpublicationresults,1:102104,1
Center for Health Quality Outcomes & Economic Research. “Health Quality Section
Update, ACGs: An Ambulatory Care Case-mix Measure.” CHQOER Quarterly,
ENRM VA Hospital, Bedford, MA. Fall, 1997.
Clark, D. O., Von Korff, M., Saunders, K., Baluch, W. M., Simon, G. E. (1995). “A
Chronic Disease Score with Empirically Derived Weights.” Medical Care, 33(8),
783-795.
Conrad, D.A., Fishman, P., Grembowski, D. et al. (2008). “Access Intervention in an
Integrated, Prepaid Group Practice: Effects on Primary Care Physician
Productivity.” Health Services Resarch, Volume 43(5p2) 1888-1905.
Conrad, D. A., Maynard, C., Cheadle, A., Ramsey. S., Marcus-Smith, M., Kirz, H., et. al.
(1998). “Primary Care Physician Compensation Method in Medical Groups:
Does It Influence The Use and Cost of Health Services For Enrollees in Managed
Care Organizations?” JAMA, 279(11), 853-858.
http://jama.ama-assn.org/cgi/content/full/279/11/853
Crampton, C., Davis, P., Lay-Yee, R., Raymont, A., Forrest, C., Starfield, B. (2004).
“Comparison of Private For-Profit with Private Community-Governed Not-ForProfit Primary Care Services in New Zealand.” Journal of Health Services
Research and Policy, 9(2), 17-22.
Crampton, C., Davis, P., Lay-Yee, R., Raymont, A., Forrest, C., Starfield, B. (2005).
“Does Community-Governed Nonprofit Primary Care Improve Access to
Services? Cross-Sectional Survey of Practice Characteristics.” International
Journal of Health Services, 35(3), 465-478.
Crosson, F.J., (2004). “The Changing Shape of The Physician Workforce In Prepaid
Group Practice.” Health Affairs – Web Exclusive, W4, 60-63.
Diamond, L. H. (2000). “Profiling: Doing It Right.” Healthplan, 41(3), 74-75.
Duckett, S. J., and Agius, P. A. (2002). “Performance of Diagnosis-based Risk
Adjustment Measures In a Population of Sick Australians.” Australian and New
Zealand Journal of Public Health, 26(6), 500-507.
Dunn, D., Rosenblatt A, Taira, D., Latimer, E., Bertko, J., Stoiber, T., Braun, P.,
Busch, S. (1995). “A Comparative Analysis of Methods of Health Risk
Assessment: Final Report.” Harvard University School of Public Health.
A-4
Ellis, R. P. (2001). “Formal Risk Adjustment by Private Employers.” Inquiry, 38(3),
299-309.
http://www.inquiryjournalonline.org/inqronline/?request=getabstract&issn=0046-9580&volume=038&issue=03&page=0299
Ettner, S. L., Frank, R. G., McGuire, T.G., Hermann, R. C. (2001). “Risk Adjustment
Alternatives in Paying for Behavioral Health Care Under Medicaid.” Health
Services Research, 36(4), 793-811.
Ettner, S. L., and Johnson, S. (2003). “Do Adjusted Clinical Groups Eliminate
Incentives for HMOs to Avoid Substance Abusers? Evidence from the Maryland
Medicaid HealthChoice Program.” Journal of Behavioral Health Services and
Research, 30(1), 63-77.
Fakhraei, S. H., Kaelin, J., Conviser, R. (2001). “Comorbidity-Based Payment
Methodology for Medicaid Enrollees with HIV/AIDS.” Health Care Financing
Review, 23(2), 53-68.
http://72.14.207.104/search?q=cache:LmlPaXNKzVQJ:www.umbc.edu/chpdm/ex
pertise/managed-care/medicaid/comorbidity.pdf+Comorbiditybased+payment+methodology+for+medicaid+enrollees+with+HIV/AIDS&hl=en
Falik, M., Needleman, J., Wells, B., Korb, J. (2001). “Ambulatory Care Sensitive
Hospitalizations and Emergency Visits: Experiences of Medicaid Patients Using
Federally Qualified Health Centers.” Medical Care, 39(6) 551-561.
http://www.lww-medicalcare.com/pt/re/medcare/abstract.00005650-20010600000004.htm;jsessionid=CCbIYY9msBm2JEWd4SgnnyBwS1sahmvkx1cAhvhOQt
FpjbGP9y1B!-1738921856!-949856031!9001!-1
Fick, D.M., Maclean, J.R., Rodriguez, N.A., Short, L., Heuvel, R.V., Waller, J.L.,
Rogers, R.L. (2004). “A Randomized Study to Decrease the Use of Potentially
Inappropriate Medications Among Community-dwelling Older Adults in a
Southeastern Managed Care Organization.” The American Journal of Managed
Care, 10(11), 761-768.
http://www.ajmc.com/files/articlefiles/AJMC04novPrt1_Fick761to68.pdf
Fishman, P., Shay, D. K. (1999). “Development and Estimation of a Pediatric Chronic
Disease Score Using Automated Pharmacy Data.” Medical Care, 37(9), 874-883.
http://www.lww-medicalcare.com/pt/re/medcare/fulltext.00005650-19990900000004.htm;jsessionid=CCaV0cmAKYfzTqwX47k8mc1cj2phpZV2hBN3VGUTk
fz7iwOn1yod!1362100327!-949856032!9001!-1
FitzHenry, F. and Shultz, E. K. (2000). “Health-Risk-Assessment Tools Used to Predict
Costs in Defined Populations.” Journal of Healthcare Information Management,
14(2), 31-57.
A-5
Foldes, S. (1998). “Managed Care Holds Down Costs, Doesn‟t Hurt Auality.” January
3, 1998. Minneapolis Star Tribune, A13.
Forrest, C. B, Kinder, K., Klaus, K. W., Reid, R. J. (2004). “Adjusted Clinical Group –
ein Instrument zur Prognose des Ressourcenverbrauchs.” Gesimdheits und
Sozialpolitik, 1-2/2004, 8-15.
Forrest, C. B, Majeed, A., Weiner, J., Carroll, K., Bindman, A. B. (2003). “Referral of
Children to Specialists in the United States and the United Kingdom.” Archives of
Pediatric Adolescent Medicine, 157(3), 279-285.
http://archpedi.ama-assn.org/cgi/content/full/157/3/279
Forrest, C. B, Majeed, A., Weiner, J., Carroll, K., Bindman, A. (2002). “Comparison of
Specialty Referral Rates in the United Kingdom and United States: Retrospective
Cohort Analysis.” British Medical Journal, 325(7360), 370-371.
http://bmj.bmjjournals.com/cgi/content/full/325/7360/370/
Forrest, C. B., Weiner, J. P., Fowles, J., Vogeli, C., Frick, K. D., Lemke, K. W., et. al.
(2001). “Self-Referral in Point-of-Service Health Plans.” JAMA, 285(17), 22232231.
http://jama.ama-assn.org/cgi/content/short/285/17/2223
Forrest, C. B., Reid, R. J. (2001). “Prevalence of Health Problems and Primary Care
Physicians‟ Specialty Referral Decisions.” Journal of Family Practice, 50(5),
427-432.
http://www.jfampract.com/content/2001/05/jfp_0501_04270.asp
Forrest, C. B., Whelan, E. M. (2000). “Primary Care Safety-Net Delivery Sites in the
United States: A Comparison of Community Health Centers, Hospital Outpatient
Departments, and Physicians‟ Offices.” JAMA, 284(16), 2077-2083.
http://jama.ama-assn.org/cgi/content/short/284/16/2077
Fowler, L., Anderson G. (1996). “Capitation Adjustment For Pediatric Populations.”
Pediatrics, 98(1), 10-17.
http://pediatrics.aappublications.org/cgi/content/abstract/98/1/10
Fowles, J., Weiner, J., Knutson, D., Fowler, E., Tucker, A., Ireland, M., “Taking Health
Status into Account When Setting Capitation Rates: A Comparison of Risk
Adjustment Methods.” JAMA, 276(16), 1316-1321.
http://jama.ama-assn.org/cgi/content/abstract/276/16/1316
Franks, P., and Fiscella, K. (2002) “Effect of Patient Socioeconomic Status on Physician
Profiles for Prevention, Disease Management, and Diagnostic Testing Costs.”
Medical Care, 40(8), 717-724.
http://www.lww-medicalcare.com/pt/re/medcare/fulltext.00005650-20020800000011.htm
A-6
Franks, P., Fiscella, K., Beckett, L., Zwanziger, J., Mooney, C., and Gorthy, S. (2003).
“Effects of Patient and Physician Practice Socioeconomic Status on the Health
Care of Privately Insured Managed Care Patients.” Medical Care, 41(7), 842852.
Fye, W.B., (2004). “Cardiology Workforce: A Shortage, Not a Surplus.” Health Affairs –
Web Exclusive, W4, 64-66.
Gifford, G. A.,, Edwards, K. R., Knutson, D. (2004). “Health-Based Capitation Risk
Adjustment in Minnesota Public Health Care Programs.” Health Care Financing
Review, 26(2), 21-41.
http://www.cms.hhs.gov/review/04winter/04winterpg21.pdf
Glazier, R. (2008). “The Impact of Not Having a Primary Care Physician Among People
with Chronic Conditions.” ICES Investigative Report July 2008.
Goodman, D.C., (2004). “Do We Need More Physicians?” Health Affairs – Web
Exclusive, W4, 67-69.
Green, B., Barlow, J., Newman, C. (1996). “Ambulatory Care Groups and The Profiling
of Primary Care Physician Resource Use: Examining the Application of Case-Mix
Adjustments.” Journal of Ambulatory Care Management, 19(1), 86-89.
http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve&db=pubmed&dopt
=Abstract&list_uids=10154372
Gifford, G. A., Edwards, K. R., and Knutson, D. J. (2004). “Health-Based Capitation
Risk Adjustment in Minnesota, Public Health Care Programs.” Health Care
Financing Review, Vol. 26(2), 21-41.
http://www.cms.hhs.gov/review/04winter/04winterpg21.pdf
Hall, J., Moore, J., “Does High Risk Poll Coverage Meet the Needs of People at Risk for
Disability.” Inquiry Journal. 340-353.
Halling, A., Fridh, G., Ovhed, I. (2006). “Validating the Johns Hopkins ACG Case-Mix
System of the Elderly in Swedish Primary Health Care.” BMC Public Health,
6(171).
Harris, L. (2009). “Diabetes Quality of Care and Outpatient Utilization Associated with
Electronic Patient-Provider Messaging: A Cross-Sectional Analysis.” Diabetes
Care.
Harry, N. (1997). “Hopkins Health Pricing Tool a Managed Care Hit.” Baltimore
Business Journal, 15(9), 20.
http://www.bizjournals.com/baltimore/stories/1997/07/21/focus2.html
Holahan, J., Rangarajan, S., Schirmer, M. (1999). “Medicaid Managed Care Payment
Rates in 1998.” Health Affairs (Millwood), 18(3), 217-227.
http://content.healthaffairs.org/cgi/reprint/18/3/217
A-7
Hollander, M., Kadlec, H., Hamdi, R., “Increasing value for money in the Canadian
healthcare system: new findings on the contribution of primary care services.”
Healthcare Quarterly, Vol. 12(4), 30-43.
Jackson, J. E., Doescher, M. P., Saver, B. G., and Fishman, P. (2004). “Prescription
Drug Coverage, Health, and Medication Acquisition Among Seniors With One or
More Chronic Conditions.” Medical Care, 42(11), 1056-1065. http://www.lwwmedicalcare.com/pt/re/medcare/fulltext.00005650-200411000-00004.htm
Johnson M.L., El-Serag H.B., Tran T.T., Hartman C., Richardson P., et. al. (2006).
“Adapting the Rx-Risk-V for Mortality Prediction in Outpatient Populations.”
Med Care, 44(8), 793-797.
Juncosa, S., Bolibar, B. (1997). [A Patient Classification System for Our Primary Care:
The Ambulatory Care Groups (ACGs)]. Gac Sanit, 11(2), 83-94.
Juncosa, S., Carrillo, E., Bolibar, B., Prados, A., and Grevas, J. (1996). {Classification
Systems in Care-Mix Groups in Ambulatory Care. Perspectives for Our Primary
Health Care.] Aten Primaria, 17(1), 76-82.
Juncosa, S., Bonaventura, B., Rpset. M., Tomas. R. (1999). “Performance of an
Ambulatory Case-Mix Measurement in Primary Care in Spain: Ambulatory Care
Groups (ACGs).” European Journal of Public Health, 9, 27-35.
Juncosa, S., Bonaventura, B., Rpset. M., Tomas. R. (1999). “Descripcion,
Comportamiento, Usos Y Metodologia de Empleo Para Medir La Casuistica En
Nuestra Atencion Primaria: Los Ambulatory Care Groups (ACGs).” Fundacio
Salut, Empresa I Economia.
Klabunde, C. N., Warren, J. L., Legler, J. M. (2002). “Assessing Comorbidity Using
Claims Data: An Overview.” Medical Care, 40(8 Suppl), IV-26-35.
Knutson, D. (1998). “Case Study: The Minneapolis Buyers Health Care Action Group.”
Inquiry, 35, 171-177.
Kuhlthau K., Ferris, T. G., Beal, A. C., Gortmaker, S. L., and Perrin, J. M. (2001). “Who
Cares for Medicaid-Enrolled Children With Chronic Conditions?” Pediatrics,
108(4), 906-912.
http://pediatrics.aappublications.org/cgi/content/full/108/4/906
Kuhlthau K., Ferris T.G., Davis R.B., Perrin J.M., Iezzoni L.I. (2005). “Pharmacy and
Diagnosis Based Risk Adjustment for Children with Medicaid.” Med Care,
43(11), 1155-1159.
Lafferty, W.E., Tyree, P., Delvin, S., Anderson, M., Diehr, P., (2008) “Complementary
and Alternative Medicine Provider Use and Expenditures by Cancer Treatment
Phase.” The American Journal of Managed Care.” 14(5).
A-8
Lafferty W.E., Tyree P.T., Bellas A.S., Watts C.A., Lind B.K., et. al. (2006). “Insurance
Coverage and Subsequent Utilization of Complementary and Alternative
Medicine Providers.” Am J Manag Care, 12(7), 397-404.
http://www.ajmc.com/files/articlefiles/AJMC_06julLafferty397to404.pdf
Leibson, C. L., Katusic, S. K., Barbaresi, W. J., Ransom, J., and O‟Brien, P. C. (2001)
“Use and Costs of Medical Care for Children and Adolescents with and without
Attention-Deficit/Hyperactivity Disorder.” JAMA, 285(1), 60-66.
http://jama.ama-assn.org/cgi/content/full/285/1/60
Liu, C. F., Sales, A. E., Sharp, N. D., Fishman, P., Sloan, K. L., Todd-Stenberg, J., et. al.
(2003). “Case-mix Adjusting Performance Measures in a Veteran Population:
Pharmacy- and Diagnosis-based Approaches.” Health Services Research, 38(5),
1319-1337.
http://www.blackwell-synergy.com/doi/abs/10.1111/14756773.00179?cookieSet=1
Lind, B. (2006). “The Effect of Complementary and Alternative Medicine Claims on
Risk Adjustment.”Medical Care. Vol. 44 (12) 1078-1084.
Lind, B. (2007). “Use of Complementary and Alternative Medicine Providers by
Fibromyalgia Patients Under Insurance Coverage.” Arthritis & Rheumatism
(Arthritis Care & Research). 57, 71-76.
Maciejewski, M., Liu, C., Derleth, A., McDonell, M., Anderson, St., Fihn, S. (2005).
“The Performance of Administrative Self-Reported Measures for Risk
Adjustment of Veterans Affairs Expenditures.” HSR: Health Services Research,
40(3), 887-904.
Madden, C. W., Mackay, B. P., Skillman, S. M. (2001). “Measuring Health Status for
Risk Adjusting Capitation Payments.” Center for Health Care Strategies, Inc.
http://www.chcs.org/publications3960/publications_show.htm?doc_id=211914
Madden, C. W., Mackay, B. P., Skillman, S. M., Ciol, M., Diehr, P. (2000). “Risk
Adjusting Capitation: Applications in Employed and Disabled Populations.”
Health Care Management Science, 3(2), 101-109.
Madden, C. W., Stanley. M, Skillman, S. M., Blough, D. K., Mackay, B., Wilson V., et.
al. (1998). “Implementing a Model for Risk Distribution Among Competing
Health Plans.” Final Report for the Robert Wood Johnson Foundation.
http://www.rwjf.org/reports/grr/023352.htm
Madden, C. W., Skillman, S. M , Mackay, B. P. (1998). “Risk Distribution and Risk
Assessment Among Enrollees in Washington State‟s Medicaid SSI Population.
Center for Health Care Strategies, Inc.
Majeed, A., Bindman, A., Weiner, J. P. (2001). “Use of Risk Adjustment in Setting
Budgets and Measuring Performance in Primary Care I: How it Works.” British
Medical Journal, 323, 604-607.
www.bmj.com/cgi/reprint/323/7313/604.pdf
A-9
Majeed, A., Bindman, A., Weiner, J. P. (2001). “Use of Risk Adjustment in Setting
Budgets and Measuring Performance in Primary Care II: Advantages,
Disadvantages, and Practicalities.” British Medical Journal, 323, 607-610.
http://bmj.bmjjournals.com/cgi/reprint/323/7313/607.pdf
Mark. T. L., Ozminkowski, R.J., Kirk, A., Ettner, S. L. Drabek, J. (2003). “Risk
Adjustment for People With Chronic Conditions in Provide Sector Health Plans.”
Medical Decision Making, 23(5), 397-405.
=Abstract&list_uids=14570297&query_hl=5
Maryland Health Care Commission by the Project HOPE Center for Health Affairs
(2001). “Characteristics of Maryland Residents Who Obtain Health Insurance
from the Small and Large Group Markets.”
Massachusetts Rate Setting Commission, (1996). “Evaluation of Case-Mix Adjustments
Methods for the Massachusetts Division of Medical Assistance.” Final Report
and Recommendation.
McCall, D., Saruma, Angela, Hamman, Richard, Reusch, Barton, P. (2004). “Are LowIncome Elderly Patients At Risk for Poor Diabetes Care?” Diabetes Care, 27(5),
1060-1065.
McCracken, S. B. (1996). “Health Information Services Technologies.” Journal of
Ambulatory Care Management, 19(1), 90-97.
Medical College of VA Associated Physicians and Mathematica Policy Research, I.
(1999). “Minimal-Burden Risk Adjusters for the Medicare Risk Program.”
Report prepared for the Health Care Financing Review (HCFA Contract No. 17C90366/3/01), November.
Meenen, Richard T., Goodma, M. J., Fishman, P. A., Hornbrook, M. C., O‟KeeffeRosetti, M. C., Bachman, D. J. (2003). “Using Risk Adjustment Models to
Identify High-Cost Risks.” Medical Care, (41)11, 1301-1312.
Minnesota Department of Health (1998). “Risk Adjustment and Rate Setting Methods in
Public Programs.” A Report to Legislature Prepared by the Health Economics
Program for the Minnesota Department of Health.
Mitchell R.N. (2006). “Managing Costs with Software.” Editorial. Advance for Health
Information.
http://health-careit.advanceweb.com/Common/editorial/editorial.aspx?CTIID=339
Moore, H. W., K. J., Johnson, S., Mussman, M., O‟Brien, J. (2001). “Risk Adjustment
for Asthma: Variations by Methodology and Implications for Providers.” Working
Paper: Informed Purchasing Series, Center for Health Care Strategies, Inc.
A-10
Morgan, P., Shah, N., Kaufman, J., et al, (2008). “Impact of Physician Assistant Care on
Office Visit Resource Used in the United States.” Health Services Research. Vol.
43 (5p2) 1906-1922.
Moynihan, R. (2004). “Using Health Research in Policy and Practice: Case Studies from
Nine Countries.” Academy Health.
Mullan, F., (2004). “My Dad Was Not A Prepaid Group Practice Patient.” Health Affairs
– Web Exclusive, W4, 70-72.
National Health Information, LLC. (2004). “Use Predictive Modeling to Improve
Profitability under Cap.” Capitation Rates & Data, 9(6).
Naessens, J., Baird, M., Van Houten, H., Vanness, D., Campbell, C. (2005). “Predicting
Persistently High Primary Care Use.” Annals of Family Medicine. 3, 324-330.
http://www.annfammed.org/cgi/content/full/3/4/324
Omar, O.Z. (2008) “A model based on age, sex, and morbidity to explain variation in UK
general practice prescribing: cohort study.” BMJ, 337;a238.
Orueta, J. F., Lopez-De-Munain, J., Baez, K., Aiarzaguena, J. M., Aranguren, J. I., and
Pedrero, E. (1999). “Application of the Ambulatory Care Groups in the Primary
Care of a European National Health Care System: Does It Work?” Med Care,
37(3), 238-248.
Page, L. (1996). “Maryland to Test New Medicaid Managed Care Pay Formula.”
American Medical News.
Palsbo, S., P. R. (2001). “Implementing Risk Assessment and Risk Adjustment for
People with Disabilities in State Programs: Six Case Studies.” NRH Center for
Health and Disability Research.
Paramore, L. C., Halpern, M. T., Lapuerta, P., Hurley, J. S., Frost, F. J., Fairchild, D.G.,
et. al. (2001). “Impact of Poorly Controlled Hypertension on Healthcare
Resource Utilization and Cost.” Am J Manag Care, 7(4), 389-398.
http://www.ajmc.com/files/articlefiles/AJMC2001aprParamore389_98.pdf
Parente, S. T., Susan S. Kim, PharmD; Michael D. Finch, PhD; Lisa A. Schloff, M. T. S.
R., PharmD, PhD; and Raafat Siefeldin, P., PhD; and J. David Haddox, DDS,
MD. (2004). “Identifying Controlled Substance Patterns of Utilization Requiring
Evaluation Using Administrative Claims Data.” Am J Managed Care, 10, 783790.
http://www.ajmc.com/files/articlefiles/AJMC04nov_Parente783to90.pdfhttp://ww
w.ajmc.com/Article.cfm?Menu=1&ID=2752
Parente, S. T., Weiner, J. P., Garnick, D. W., Fowles, J., Lawthers, A. G., and Palmer, R.
H. (1996). “Profiling Medicare Beneficiary Resource Use by Primary-Care
Practices: Managed Medicare Implications.” Health Care Financ Rev. 17(4), 2342.
A-11
Parkerson, G. R., Jr., Harrell, F. E., Jr., Hammond, W. E., Wang, X. Q. (2001)
“Characteristics of Adult Primary Care Patients as Predictors of Future Health
Services Charges.” Med Care 39(11), 1170-1181.
Petersen, L. A., Peitz, Kenneth, Woodard, LeChauncy, D., and Byrne, Margaret. (2005).
“Comparison of the Predictive Validity of Diagnosis-Based Risk Adjusters for
Clinical Outcomes.” Medical Care, 43(1), 61-67.
Pietz, K. O., m. K., Byrne, M., Souchek, J., Petersen, N., Ashton, C. (2002).
“Physician-Level Variation in Practice Patterns in VA Healthcare System.”
Health Services & Outcomes Research Methodology, 2, 95-106.
Pietz, K., Ashton, C, McDonell, M, Wray, N. (2004). “Predicting Healthcare Costs in a
Population of Veterans Affairs Beneficiaries Using Diagnosis-Based Risk
Adjustment and Self-Reported Health Status.” Medical Care. 42(10):1027-1035.
Powe, N. R., Weiner, J. P., Starfield, B., Stuart, M., Baker, A., and Steinwachs, D. M.
(1996). “Systemwide Performance in a Medicaid Program: Profiling the Care of
Patients with Chronic Illnesses.” Med Car, 34(8), 798-810.
Powers C.A., Meyer C.M., Roebuck M.C., Vaziri B. (2005). “Predictive Modeling of
Total Healthcare Costs Using Pharmacy Claims Data.” Med Car, 43(11), 10651072.
Prosser, B., QUILTS Team. (2006). “The Utility of ACGs & ADGs in the Evaluation of
the BC Nurseline.” Presentation in British Columbia.
Quan H., Sundararajan V., Halfon P., Fong A., Burnand B., et. al. (2005). “Coding
Algorithms for Defining Comorbidities in ICD-9-CM and ICD-10 Administrative
Data.” Med Care, 43(11), 1130-1139).
Ralston, JD. (2007). “Patient Web Services Integrated with a Shared Medical Record:
Patient Use and Satisfaction.” JAMIA.
http://www.jamia.org/cgi/reprint/14/6/798.pdf
Ralston, JD. (2009). “Patient Use of Secure Electronic Messaging Within a Shared
Medical Record: a Cross-Sectional Analysis.” JGIM.
Ray, G., Mertens, J., Weisner, C., (2009) “Family members of people with alcohol or
drug dependence: health problems and medical cost compared to family members
of people with diabetes and asthma.” Society for the Study of Addiction.
104(203-214)
Ray, G., Croen, L., Habel, L., ”Mothers of children diagnosed with attentiondeficit/hyperactivity disorder: Health conditions and medical care utilization in
periods before and after the birth of a child.” Medical Care. 47(1)105-114.
A-12
Reid, R., B. B., Roos, N. P., Black, C., MacWilliam, L., Menec, V. “Do Some Physician
Groups See Sicker Patients than Others? Implications for Primary Care Policy in
Manitoba.” Manitoba Centre for Health Policy and Evaluation, Department of
Community Health Services Faculty of Medicine, University of Manitoba.
http://www.umanitoba.ca/centres/mchp/reports/pdfs/acg2001.pdf
Reid, R., M. L., Roos, N., Bogdanovic, B., Black, C. (1999). “Measuring Morbidity in
Populations: Performance of the Johns Hopkins Adjusted Clinical Group (ACG)
Case-Mix Adjustment System in Manitoba.” Manitoba Centre for Health Policy
and Evaluation, Department of Community Health Services Faculty of Medicine,
University of Manitoba.
http://www.umanitoba.ca/centres/mchp/reports/pdfs/acg.pdf
Reid, R. J., MacWilliam, L., Verhulst, L., Roos, N., and Atkinson, M. (2001).
“Performance of the ACG Case-Mix System in Two Canadian Provinces.”
Medical Care, 39(1), 86-99.
Reid, R. J., Roos, N. P., MacWilliam, L., Frohlich, N., and Black, C. (2002) “Assessing
Population Health Care Need Using a Claims-based ACG Morbidity Measure: A
Validation in the Province of Manitoba.” Health Services Research, 37(5): 13451364.
http://www.blackwell-synergy.com/doi/abs/10.1111/1475-6773.01029
Reid, R. J., Evans, R.., Barer, M., Sheps, S., Kerluke, K., McGrail, K. Hetzman, C. and
Pagliccia, N. (2003) “Conspicuous consumption: characterizing high users of
physician services in one Canadian province.” Journal of Health Services
Research & Policy, 8(4): 1345-1364.
http://www.blackwell-synergy.com/doi/abs/10.1111/1475-6773.01029
Robinson J.W., Zeger S.L., Forrest C.B. (2006). “A Hierarchical Multivariate Two-Part
Model for Profiling Providers‟ Effects on Health Care Changes.” Journal of the
American Statistical Association, 101(475), 911-923.
Rosen, A. K., Loveland, S. A., Rakovski, C. C., Christiansen, C., L., Berlowitz, D. R.
(2003). “Do Different Case-Mix Measures Affect Assessments of Provider
Efficiency: Lessons From the Department of Veteran Affairs.” Journal of
Ambulatory Care Management, 26(3), 229-242.
http://pt.wkhealth.com/pt/re/lwwgateway/abstract.00004479-20030700000006.htm
Rosen, A. K., Loveland, S., Anderson, J. J., Rothendler, J. A., Hankin, C. S.,
Rakovski, C., C., et. al. (2001). “Evaluating Diagnosis-Based Case-Mix
Measures: How Well Do They Apply to the VA Populations?” Medical Care,
39(7), 692-704. http://www.lwwmedicalcare.com/pt/re/medcare/fulltext.00005650-200107000-00006.htm
A-13
Rosen, A. K., Rakovski, C., Loveland, S. A., Anderson, J. J., Berlowitz, D. R. (2002)
“Profiling Resource Use: Do Different Outcomes Affect Assessments of Provider
Efficiency?” American Journal of Managed Care, 8(12, 1105-1115.
http://www.ajmc.com/files/articlefiles/AJMC2002decRosen1105-1115.pdf
Rosen, A. K., Reid, R., Broemeling, A., and Rakovski, C. C. (2003). “Applying a RiskAdjustment Framework to Primary Care: Can We Improve on Existing
Measures?” Annals of Family Medicine, 1(1), 44-51.
Rosenblatt, R. A., Baldwon, L. M., Chan, L., Fordyce, M. A., Hirsch, I. R., Palmer, J. P.,
et. al. (2001). “Improving the Quality of Outpatient Care for Older Patients With
Diabetes: Lesson From a Comparison of Rural and Urban Communities.”
Journal of Family Practice, 50(8), 676-680.
http://www.jfponline.com/content/2001/08/jfp_0801_06760.asp
Salem-Schatz, S., Moore, G., Rucker, M., and Pearson, S. (1994). “The Case For CaseMix Adjustment In Practice Profiling: When Good Apples Look Bad.” JAMA,
272(11), 871-874.
http://jama.ama-assn.org/cgi/content/abstract/272/11/871
Sales, A. E., Liu, C. F., Sloan, K. L., Malkin, J., Fishan, P. A., Rosen, A. K., et. al.
(2003). “Predictiing Costs of Care Using Pharmacy-Based Measure Risk
Adjustment in a Veteran Population.” Medical Care, 41(6), 753-760.
http://www.lww-medicalcare.com/pt/re/medcare/abstract.00005650-20030600000008.htm
Salsberg, E., Forte, G. (2004). “Benefits and Pitfalls in Applying the Experience of
Prepaid Group Practices to the U.S. Physician Supply.” Health Affairs – Web
Exclusive, W4, 73-75.
Schoenbaum, S.C., (2004). “Physicians and Prepaid Group Practices.” Health Affairs –
Web Exclusive, W4, 76-78.
Shen, Y. and Ellis, R. P. (2002). “Cost-Minimizing Risk Adjustment.” Journal of
Health Economic, 21(3), 515-530.
http://www.sciencedirect.com/science/article/B6V8K-459J6C71/2/750e4d9c422c203710b505981a0b9f0a
Shen, Y., Ellis, R. P. (2002). “How Profitable is Risk Selection? A Comparison of Four
Risk Adjustment Models.” Health Economics, 11(2), 165-174.
http://people.bu.edu/ellisrp/shenellis_profitability2002.pdf
Shenkman, E., Pendergast, J., Wegener, D. H., Hartzel, T., Naff, R., Freedman S., et. al.
(1997). “Children‟s Health Care Use in the Healthy Kids Program.” Pediatrics,
100(6), 947-953.
http://pediatrics.aappublications.org/cgi/content/full/100/6/947
Shenkman, E., Breiner, J.. (2001). “Characteristics of Risk Adjustment Systems.”
Working Paper Series, #2. Division of Child Health Services Research and
Evaluation, Institute for Child Health Policy, University of Florida.
A-14
Shields, A. E., Comstock, C., Finkelstein, J. A., and Weiss, K.B. (2003). “Comparing
Asthma Care Provided to Medicaid-Enrolled Children in a Primary Care Case
Manager Plan and a Staff Model HMO.” Ambulatory Pediatric, 3(5), 253-262.
http://ampe.allenpress.com/ampeonline/?request=getdocument&doi=10.1367%2F15394409(2003)003%3C0253:CACPTM%3E2.0.CO%3B2
Sheilds, A. E., Finkelstein, J. A., Comstock, C., Weiss, K. B. (2002). “Process of Care
for Medicaid-Enrolled Children with Asthma.” Medical Care, 40(4), 303-314
Smith, N. S., and Weiner, J. P. (1994). “Applying Population-Based Case-Mix
Adjustment in Managed Care: The Johns Hopkins Ambulatory Care Group
System.” Managed Care Quqrterly, 2(3), 21-34.
Starfield, B., Lemke, K. W., Herbert, R., Pavlovich, W. D., and Anderson, G. (2005).
“Comorbidity and the Use of Primary Care and Specialist Care in the Elderly.”
Annals of Family Medicine, 3(3), 215-222.
http://annalsfm.highwire.org/cgi/content/full/3/3/215
Starfield, B., Lemke, K. W., Bernhardt, T., Foldes, S. S., Forrest, C. B., and Weiner, J. P.
(2003). “Comorbidity: Implications for the Importance of Primary Care in Case
Management.” Annals of Family Medicine, 1(1), 8-14.
http://www.annfammed.org/cgi/content/full/1/1/8
Starfield, B., Weiner, J., Mumford, L., Steinwachs, D. (1991). “Ambulatory Care
Groups: A Categorization of Diagnoses For Research and Management.” Health
=Abstract&list_uids=1901841
Stuart, M. E., and Steinwachs, D. M. (1993). “Patient Mix Differences Among
Ambulatory Providers and Their Effects on Utilization and Payments For
Maryland Medicaid Users.” Medical Care, 31(12), 1119-1137.
Sullivan, C. Omar, R., Forrest, C., Majeed, A. (2004). “Adjusting for Case Mix and
Social Class in Examining Variation in Home Visits Between Practices.” Family
Practice, 21(4), 355-363.
http://fampra.oxfordjournals.org/cgi/content/full/21/4/355
Sylvia M.L., Shadmi E., Hsiao C.J., Boyd C.M., Schuster A.B., et. al. (2006). “Clinical
Features of High-Risk Older Persons Identified by Predictive Modeling.” Disease
Management, 9(1), 56-62.
“The Faces of Medicaid.” (2000). Publication for the Center of Health Care Strategies,
Inc., Finance, 71-75.
http://www.chcs.org/publications/pdf/cfm/Finance.pdf
A-15
Thomas, J. W., Grazier, K. L., and Ward, K. (2004). “Comparing Accuracy of RiskAdjustment Methodologies Used in Economic Profiling of Physicians.” Inquiry,
41(2), 218-231.
http://www.inquiryjournalonline.org/inqronline/?request=getabstract&issn=0046-9580&volume=041&issue=02&page=0218
Tucker, A., Weiner, J., Abrams, C. (2002). “Health-Based Risk Adjustment: Application
to Premium Development and Profiling.” Health Administration Press, Ann
Arbor.
Tucker, A. M., Weiner, J. P., Honigfeld, S., and Parton, R. A. (1996). “Profiling Primary
Care Physician Resource Use: Examining the Application of Case-Mix
Adjustment.” Journal of Ambulatory Care Management, 19(1), 60-80.
Verhulst, L., Reid, R. J., Forrest, C. B. (2001). “Hold It - My Patients are Sicker! The
Importance of Case Mix Adjustment to Practitioner Profiles in British Columbia.”
BC Medical Journal, 43(6), 328-333.
http://www.bcma.org/public/bc_medical_journal/BCMJ/2001/july_august_2001/P
remisePatientsSicker.asp
Volpel, A., O‟Brien, J. (2005). “Strategies for Assessing Health Plan Performance on
Chronic Diseases: Selecting Performance Indicators and Applying Health-Based
Risk Adjustment.” CHCS, Inc., Resource Paper.
Weiner, J. (1991). “Ambulatory Case-Mix Methodologies: Applications to Primary Care
Research.” In Grady, M. (ed) Primary Care Research: Theory and Methods.
Weiner, J. P., Dobson, A., Maxwell, S. L., Coleman, K., Starfield, B., Anderson, G. F.
(1996). “Risk-Adjusted Medicare Capitation Rates Using Ambulatory and
Inpatient Diagnoses.” Health Care Financing Review, 17(3), 77-99.
Weiner, J. P., Starfield, B. H., Lieberman, R. N. (1992). “Johns Hopkins Ambulatory
Care Groups (ACGs): A Case-Mix System For UR, QA, and Capitation
Adjustment.” HMO Practice, 6(1), 13-19.
Weiner, J. P., Starfield, B., Stuart, M., Powe, N. R., Steinwachs, D. M. (1996).
“Ambulatory Care Practice Variation Within a Medicaid Program.” Health
Weiner, J. P., Starfield, B. H., Steinwachs, D. M., and Mumford, L. M. (1991).
“Development and Application of a Population-Oriented Measure of Ambulatory
Care Case-Mix.” Medical Care, 29(5), 452-472.
A-16
Weiner, J. P., Tucker, A. M., Collins, A. M., Fakhraei, H., Liebermann, R., Abrams, C.,
et. al. (1998). “The Development of Risk-Adjusted Capitation Payment System:
The Maryland Medicaid Model.” Journal of Ambulatory Care Management,
21(4), 29-52.
Weir, Sharada, Phil, D., Aweh, Gideaon, Clark, Robin E. (2008). “Case Selection for a
Medicaid Chronic Care Management Program.” Health Care Financing Review,
Volume 30, Number 1, Fall 2008, pp 61-74.
http://www.cms.hhs.gov/HealthCareFinancingReview/downloads/08Fallpg61.pdf
Weyrauch, K. F., Boiko, P., and Feeny, D. (1995). “HMO Family Physicians: Men and
Women Differ in Their Work.” HMO Practice, 9(4), 155-161.
Willson, D. F., Horn, S.D., Hendley, J. O., Smout, R., and Gassaway, J. (2001). “Effect
of Practice Variation on Resource Utilization in Infants Hospitalized for Viral
Lower Respiratory Illness.” Pediatrics, 108(4), 851-855.
http://pediatrics.aappublications.org/cgi/reprint/108/4/851.pdf
Wolff, J. L., Starfield, B., and Anderson, G. (2002). “Prevalence, Expenditures, and
Complications of Multiple Chronic Conditions in the Elderly.” Arch of Internal
Medicine, 162, 2269-2276.
Woolf, S. H., Rothemich, S. F., Johnson, R. E., Marsland, D. W. (2000). “Selection Bias
from Requiring Patients to Give Consent to Examine Data for Health Services
Research.” AMA, Archives of Family Medicine, 9(10), 1111-1118.
http://archfami.ama-assn.org/cgi/reprint/9/10/1111.pdf
Zielinski, A., Kronogard, M., Lenhoff, H., & Halling, A. (2009). Validation of ACG
case-mix for equitable resource allocation in swedish primary health care. BMC
Public Health, 9(1), 347.
Appendix B: Variables Necessary to Locally Calibrate the ACG Predictive Models
B-i
Appendix B: Variables Necessary to Locally
Calibrate the ACG Predictive Models
Introduction ...................................................................................................B-1
Table 1: Individual ACG Categories Included in the Predictive Models .B-1
Table 2: ACGs Included in Three Prospectively Calibrated Resource
Utilization Bands ........................................................................................B-2
Table 3: Pregnancy Without Delivery .......................................................B-2
Table 4: EDCs Included in the Predictive Models ....................................B-3
Table 5: Rx-MGs Included in the Predictive Models ...............................B-6
Table 6: Special Population Markers ........................................................B-7
Table 7: Demographic Markers .................................................................B-7
Table 8: Optional Prior Cost Markers* .....................................................B-7
Table 9: Utilization Markers** .................................................................B-8
B-ii
B-1
Introduction
These tables indicate the variables used in the ACG Predictive Models. The ACG
Predictive Models have been developed using a reference population provided by
PharMetrics, a unit of IMS in Watertown, MA. Modest performance improvements may
be attained when predictive models are calibrated against a local population. The ACG
Software facilitates local calibration of predictive models with the export of the
independent variables as model markers. For reference, the independent variables used in
the ACG Predictive Models are presented in the following tables.
U.S. and international copyright law protect these assignment algorithms. It is an
infringement of copyright law to develop any derivative product based on the grouping
algorithm or other information presented in this document without the written permission
of The Johns Hopkins University.
Copyright © The Johns Hopkins University, 1995, 1997, 1998, 1999, 2000, 2001, 2002,
2003, 2004, 2005, 2007, 2008 and 2009. All Rights Reserved.
Table 1: Individual ACG Categories Included in the Predictive Models
ACG
Description
4220
4330
4420
4430
4510
4520
4610
4620
4730
4830
4910
4920
4930
4940
5010
5020
5030
5040
5050
5060
5070
5320
4-5 Other ADG Combinations, Age 1-17, 1+ Major ADGs
4-5 Other ADG Combinations, Age >44, 1 Major ADGs
4-5 Other ADG Combinations, Age >44, 2+ Major ADGs
6-9 Other ADG Combinations, Age 1-5, No Major ADGs
6-9 Other ADG Combinations, Age 6-17, No Major ADGs
6-9 Other ADG Combinations, Male, Age 18-34, 2+ Major ADGs
6-9 Other ADG Combinations, Female, Age 18-34, 2+ Major ADGs
6-9 Other ADG Combinations, Age >34, 0-1 Major ADGs
6-9 Other ADG Combinations, Age >34, 4+ Major ADGs
10+ Other ADG Combinations, Age 1-17 No Major ADGs
10+ Other ADG Combinations, Age 1-17, 1 Major ADGs
10+ Other ADG Combinations, Age 1-17, 2+ Major ADGs
10+ Other ADG Combinations, Age 18+, 0-1 Major ADGs
10+ Other ADG Combinations, Age 18+, 2 Major ADGs
10+ Other ADG Combinations, Age 18+, 3 Major ADGs
10+ Other ADG Combinations, Age 18+, 4+ Major ADGs
Infants: 0-5 ADGs, 1+ Major ADGs
B-2
ACG
Description
5321
5322
5330
5331
5332
5340
5341
5342
Infants: 0-5 ADGs, 1+ Major ADGs, low birth weight
Infants: 0-5 ADGs, 1+ Major ADGs, normal birth weight
Infants: 6+ ADGs, No Major ADGs
Infants: 6+ ADGs, No Major ADGs, low birth weight
Infants: 6+ ADGs, No Major ADGs, normal birth weight
Infants: 6+ ADGs, 1+ Major ADG
Infants: 6+ ADGs, 1+ Major ADG, low birth weight
Infants: 6+ ADGs, 1+ Major ADG, normal birth weight
Table 2: ACGs Included in Three Prospectively Calibrated Resource
Utilization Bands
ACG Prospective RUB Levels
ACG RUB Level 1 (included in Intercept)
ACG RUB Level 2
ACG RUB Level 3
Reference Group
ACG 0100 0200 0300 1100 1200 1600 5100
5110 5200 9900
ACG 0400 0500 0600 0900 1000 1300 1800
1900 2000 2100 2200 2300 2400 2500 2800
2900 3000 3100 3400 3900 4000 1711 1721
1731 1741
ACG 0700 0800 1400 1500 2600 2700 3200
3300 3500 3600 3700 3800 4100 4210 4310
4320 4410 4710 4720 4810 4820 5310 1751
1761 1771
Table 3: Pregnancy Without Delivery
ACG
Description
17x2
All non-delivered pregnancy related ACGs (1712, 1722, 1732, 1742, 1752, 1762 and 1772
are collapsed into 1 Boolean marker).
B-3
Table 4: EDCs Included in the Predictive Models
EDC
ADM03
ALL04
ALL05
ALL06
CAR03
CAR04
CAR05
CAR06
CAR07
CAR09
CAR10
CAR12
CAR13
CAR14
CAR15
END02
END06
END07
END08
END09
EYE03
EYE13
EYE15
GAS02
GAS04
GAS05
GAS06
GAS10
GAS11
GAS12
GSI08
GSU11
GSU13
GSU14
GTC01
GUR04
HEM01
HEM03
HEM05
HEM06
HEM07
INF04
INF08
Description
Transplant Status
Asthma, WITH status asthmaticus
Disorders of the Immune System
Ischemic heart disease (excluding acute myocardial infarction)
Cardiomyopathy
Cardiac arrhythmia
Acute Myocardial Infarction
Cardiac arrest, shock
Hypertension, WITH major complications
Osteoporosis
Type 2 Diabetes, w/o Complication
Type 2 Diabetes, w/ Complication
Type 1 Diabetes, w/o Complication
Type 1 Diabetes, w/ Complication
Diabetic Retinopathy
Age-related Macular Degeneration
Acute hepatitis
Peptic ulcer disease
Diverticular disease of colon
Acute pancreatitis
Edema
Aortic aneurysm
Gastrointestinal Obstruction/Perforation
Hemolytic anemia
Thrombophlebitis
Aplastic anemia
Deep vein thrombosis
Hemophilia, coagulation Disorder
HIV, AIDS
Septicemia
B-4
EDC
MAL02
MAL03
MAL04
MAL06
MAL07
MAL08
MAL09
MAL10
MAL11
MAL12
MAL13
MAL14
MAL15
MAL16
MAL18
MUS03
MUS10
MUS14
MUS16
NUR03
NUR05
NUR06
NUR07
NUR08
NUR09
NUR11
NUR12
NUR15
NUR16
NUR17
NUR18
NUR19
NUT02
PSY01
PSY02
PSY03
PSY05
PSY07
PSY08
PSY09
PSY12
REC01
REC03
REC04
REN01
Description
Acute Leukemia
Fracture of neck of femur (hip)
Low back pain
Amputation Status
Parkinson's disease
Seizure disorder
Multiple sclerosis
Muscular dystrophy
Quadriplegia and Paraplegia
Head Injury
Spinal Cord Injury/Disorders
Paralytic Syndromes, Other
Cerebral Palsy
Nutritional deficiencies
Anxiety, neuroses
Substance use
Tobacco abuse
Depression
Bipolar Disorder
Burns--2nd and 3rd degree
EDC
REN02
REN03
REN04
RES02
RES03
RES04
RES08
RES09
RES10
RHU01
RHU05
TOX02
TOX04
B-5
Description
Fluid/electrolyte disturbances
Acute renal failure
Nephritis/Nephrosis
Acute lower respiratory tract infection
Cystic fibrosis
Pulmonary embolism
Tracheostomy
Respiratory arrest
Adverse effects of medicinal agents
Complications of Mechanical Devices
B-6
Table 5: Rx-MGs Included in the Predictive Models
Rx-MG
ALLx010
ALLx030
ALLx040
ALLx050
CARx010
CARx020
CARx030
CARx040
CARx050
EARx010
ENDx010
ENDx020
ENDx030
ENDx040
ENDx050
ENDx060
ENDx070
EYEx010
EYEx020
EYEx030
FREx010
FREx020
FREx030
GASx010
GASx020
GASx030
GASx040
GASx050
GASx060
GSIx010
GSIx020
GSIx030
GSIx040
Description
Allergy/Immunology / Acute Minor
Allergy/Immunology / Chronic
Inflammatory
Allergy/Immunology / Immune
Disorders
Allergy/Immunology / Transplant
Cardiovascular / Chronic Medical
Cardiovascular / Congestive Heart
Failure
Cardiovascular / High Blood Pressure
Cardiovascular / Hyperlipidemia
Cardiovascular / Vascular Disorders
Ears, Nose, Throat / Acute Minor
Endocrine / Bone Disorders
Endocrine / Chronic Medical
Endocrine / Diabetes With Insulin
Endocrine / Diabetes Without Insulin
Endocrine / Thyroid Disorders
Endocrine / Growth Problems
Endocrine / Weight Control
Eye / Acute Minor: Curative
Eye / Acute Minor: Palliative
Eye / Glaucoma
Female Reproductive / Hormone
Regulation
Female Reproductive / Infertility
Female Reproductive / Pregnancy and
Delivery
Gastrointestinal/Hepatic / Acute Minor
Gastrointestinal/Hepatic / Chronic Liver
Disease
Gastrointestinal/Hepatic / Chronic Stable
Gastrointestinal/Hepatic / Inflammatory
Bowel Disease
Gastrointestinal/Hepatic / Pancreatic
Disorder
Gastrointestinal/Hepatic / Peptic Disease
General Signs and Symptoms / Nausea
and Vomiting
General Signs and Symptoms / Pain
General Signs and Symptoms / Pain and
Inflammation
General Signs and Symptons / Severe
Rx-MG
GURx010
GURx020
HEMx010
INFx010
INFx020
INFx030
INFx040
INFx050
MALx010
MUSx010
MUSx020
NURx010
NURx020
NURx030
NURx040
NURx050
PSYx010
PSYx020
PSYx030
PSYx040
PSYx050
PSYx060
RESx010
RESx020
RESx030
RESx040
SKNx010
SKNx020
SKNx030
TOXx010
ZZZx000
Description
Pain
Genito-Urinary / Acute Minor
Genito-Urinary / Chronic Renal Failure
Hematologic / Coagulation Disorders
Infections / Acute Major
Infections / Acute Minor
Infections / HIV/AIDS
Infections / Tuberculosis
Infections / Severe Acute Major
Malignancies
Musculoskeletal / Gout
Musculoskeletal / Inflammatory
Conditions
Neurologic / Alzheimer's Disease
Neurologic / Chronic Medical
Neurologic / Migraine Headache
Neurologic / Parkinsons Disease
Neurologic / Seizure Disorder
Psychosocial / Attention Deficit
Hyperactivity Disorder
Psychosocial / Addiction
Psychosocial / Anxiety
Psychosocial / Depression
Psychosocial / Acute Minor
Psychosocial / Chronic Unstable
Respiratory / Acute Minor
Respiratory / Chronic Medical
Respiratory / Cystic Fibrosis
Respiratory / Airway Hyperactivity
Skin / Acne
Skin / Acute and Recurrent
Skin / Chronic Medical
Toxic Effects/Adverse Effects / Acute
Major
B-7
Table 6: Special Population Markers
Special Population Marker
HOSDOM0 (included in Intercept)
HOSDOM1
HOSDOM2
HOSDOM3+
Frailty Marker
Generic Drug Count
Use in the Predictive Model
Boolean indicator for 0 HOSDOM
Boolean indicator for 1 HOSDOM indicator
Boolean indicator for 2 HOSDOM indicators
Boolean indicator for 3 or more HOSDOM
indicators
Boolean indicator
Boolean indicator for 13 or more Generic
Drugs
Table 7: Demographic Markers
Demographic Marker
Ten age groups
Female
Use in the Predictive Model
0-4; 5-11; 12-17; 18-34; 35-44; 45-54; 55-69
(Included in the Intercept); 70-74; 75-79; 8084; 85+
Boolean indicator
Table 8: Optional Prior Cost Markers*
Prior Cost Marker
Total Expense, Non-users (Included in the Intercept)
Total Expense, 1-10th Percentile
Total Expense, 76-91st Percentile
Total Expense, 91-93rd Percentile
Pharmacy Expense, Non-users (Included in the Intercept)
Pharmacy Expense, 1-10th Percentile
Pharmacy Expense, 76-91st Percentile
Pharmacy Expense, 91-93rd Percentile
*Either pharmacy expense markers or total expense markers will be used depending on
the model variant, but pharmacy expense and total expense will not be used
simultaneously. Prior pharmacy expense is preferred when the predicting pharmacy
expense and prior total expense is preferred when predicting total expense.
B-8
Table 9: Utilization Markers**
Utilization Markers
Dialysis Service
Nursing Service
Major Procedure
1 Inpatient Hospitalization
2 Inpatient Hospitalizations
5+ Inpatient Hospitalizations
1 Emergency Department Episode
2 Emergency Department Episodes
5+ Emergency Department Episodes
1-9 Outpatient Visits
60+ Outpatient Visits
**Utilization markers are used only in the hospitalization prediction models.
Index
IN-1
Index
A
ACG
assignment process, 3-3
clinical aspects, 3-26
clinically oriented examples, 3-30
decision tree, 3-19
development, 3-1
forecasting future costs with the components of the
ACG case mix system, 7-3
how ACGs work, 3-2
overview, 3-1
terminal groups, 3-9
ACG-PM
conceptual basis, 7-9
data inputs, 7-12
definition of population-based predictive modeling,
7-1
development of ACG predictive models, 7-13
elements of a predictive model – five key dimensions,
7-9
multi-morbidity as the clinical basis, 7-5
outcomes assessed, 7-13
predictive modeling, 7-1
statistical modeling approach, 7-11
target population, 7-10
ADG
diagnostic certainty, 3-27
duration, 3-26
etiology, 3-27
mapping ICD codes to a parsimonious set of aggregated
diagnosis groups (ADGs), 3-3
severity, 3-27
specialty care, 3-27
subgroups (CADG), 3-6
Adjusted R-square values
predictive modeling statistical performance, 8-1
Anatomical therapeutic chemical system, 5-2
Applications guide content, 1-4
introduction, 1-4
ATC
anatomical therapeutic chemical system, 5-2
ATC to RX-MG assignment methodology, 5-9
C
CADG
creating a manageable number of ADG subgroups
(CADGs), 3-6
occurring combinations, 3-8
Categorization methodology
EDC, 4-1
MEDC, 4-14
Chronic condition count, 6-5
Chronic illnesses, 3-30
Clinical aspects of ACGs, 3-26
Clinical information improves on prior cost
Clinically oriented examples
chronic illnesses, 3-30
diabetes mellitus, 3-32
hypertension, 3-31
infants, 3-36
pregnancy/delivery with complications, 3-34
Coding, 2-1
Combined ACG predictive model (DxRx-PM), 7-16
Condition markers, 6-9
Continuous single-interval measure of medication
availability (CSA), 11-7
Customer commitment and contact information, 1-5
introduction, 1-5
D
Decision tree
for MAC-12--pregnant women, 3-21
for MAC-26--infants, 3-23
Development of ACG-PM, 7-13
Development of possession/adherence markers, 11-2
Development of Rx-PM, 7-15
Diagnosis and code sets, 2-1
coding issues using the international classification of
diseases (ICD), 2-1
National Drug Codes (NDC), 2-4
provisional, 2-3
rule-out, 2-3
special note for ICD-10 users, 2-4
suspected, 2-3
three and four digits, 2-2
using ICD-9 and ICD-10 simultaneously, 2-4
Dialysis service, 9-3
Dx-PM
diagnosed-defined predictive model, 7-13
Dx-PM risks factors
age and gender, 7-14
high-impact chronic conditions, 7-14
overall morbidity burden, 7-14
DxRx-PM
combined ACG predictive model, 7-16
E
EDCs
applications of, 4-16
IN-2
clinical classification for health policy research
produced by the Agency for Health Care Policy and
Research, 4-1
development, 4-1
how they work, 4-3
important differences between EDCs and ADGs, 4-15
important differences between EDCs and episode-ofcare systems, 4-16
MEDC
types, 4-14
overview, 4-1
Schneeweiss diagnosis clusters, 4-1
Elements of a predictive model – five key dimensions, 7-9
Emergency visit count, 9-4
Empiric validation of the likelihood of hospitalization
model, 9-5
predicting hospitalization, 9-6
sensitivity and positive predictive value, 9-6
F
Figures
ACG decision tree, 3-19
decision tree for MAC-12--pregnant women, 3-21
decision tree for MAC-24--multiple ADG categories, 325
decision tree for MAC-26--infants, 3-23
Dx PM risk factors, 7-14
gap in medication possession, 11-2
gap in medication possession following hospitalization,
11-3
gap in medication possession following oversupply, 113
percentage of patients with selected chronic condition
and their associated number of co-morbidities, 7-6
schematic framework for predicting hospitalization,
9-3
total healthcare costs per beneficiary by count of
chronic conditions, 7-8
typology of acute care inpatient hospitalization, 9-2
Venn diagram comparing overlap of prior, predicted,
and actual high cost users, 8-12
Forming the terminal groups (ACGs), 3-9
Frailty conditions, 6-3
Framework for prediction hospitalization using ACG
how is ACG used to predict hospitalization, 9-2
likelihood of hospitalization, 9-4
utilization markers, 9-3
what hospitalizations are we predicting, 9-1
why include utilization in prediction models for
hospitalization, 9-3
why predict hospitalization, 9-1
Frequently occurring combinations of CADGS (MACs), 3-8
G
Gaps in pharmacy utilization
development of possession/adherence markers, 11-2
Index
how medication gaps are defined and constructed,
11-8
objectives, 11-1
significance of pharmacy adherence for effective care,
11-1
the future, 11-15
validation and testing, 11-13
H
High pharmacy utilization model, 7-17
High pharmacy utilization model outputs
high risk for unexpected pharmacy cost, 7-18
probability of unexpected pharmacy cost, 7-18
relevant to clinicians and case managers, 7-18
High risk for unexpected pharmacy cost, 7-18
High, moderate, and low impact conditions, 7-17
Hospital dominant conditions, 6-1
How medication gaps are defined and constructed, 11-8
I
ICD
10, 2-4
9, 2-4
coding issues, 2-1
mapping to ADGs, 3-3
special note for ICD-10 users, 2-4
Infants, 3-35
Inpatient hospitalization count, 9-4
Installation and usage guide content, 1-3
introduction, 1-3
Interpreting adjusted R-square values
Introduction
applications guide content, 1-4
customer commitment and contact information, 1-5
diagnosis and code sets, 2-1
installation and usage guide content, 1-3
objective of the technical reference guide, 1-1
the Johns Hopkins ACG system, 1-1
L
Likelihood of hospitalization, 9-4
probability extended hospitalization score, 9-5
probability ICU hospitalization score, 9-5
probability injury hospitalization score, 9-5
probability IP hospitalization score, 9-5
probability IP hospitalization six months score, 9-5
M
MAC
occurring combinations, 3-8
Major procedure, 9-4
Mapping ICD codes to a parsimonious set of aggregated
diagnosis groups (ADGs), 3-3
Medication define morbidity
Index
National Drug Code system, 5-1
Medication defined morbidity
conclusion, 5-23
objectives, 5-1
Medication possession ratio (MPR), 11-6
N
National Drug Code system, 5-1
Navigation
technical reference guide navigation, 1-1
NDC
coding issues, 2-4
NDC to Rx-MG assignment methodology, 5-9
New condition markers
continuous single-interval measure of medication
availability (CSA), 11-7
medication possession ratio (MRP), 11-6
Nursing service, 9-3
O
Objective of the technical reference guide, 1-1
Objectives
predicting future resource use, 7-1
Outpatient visit count, 9-4
Overview
ACG, 3-1
ACG assignment process, 3-3
P
Predicting, 9-1, 9-2, 9-3
Predicting future resource
resource bands, 7-16
Predicting future resource use, 7-1
high pharmacy utilization model, 7-17
high, moderate and low impact conditions, 7-17
Predicting hospitalization, 9-1
empiric validation of the likelihood of hospitalization
model, 9-5, 9-6
introduction, 9-1
Predictive modeling, 7-1
Predictive modeling statistical performance
adjusted R-square values, 8-1
clinical information improves on prior cost, 8-11
interpreting adjusted R-square values, 8-6
overview, 8-1
predictive ratios, 8-6
ROC curve, 8-10
sensitivity and positive predictive value, 8-8
validation data, 8-1
Predictive ratios
Pregnancy, 3-34
Probability extended hospitalization score, 9-5
Probability ICU hospitalization score, 9-5
Probability injury hospitalization score, 9-5
IN-3
Probability IP hospitalization score, 9-5
Probability IP hospitalization six months score, 9-5
Probability of unexpected pharmacy costs, 7-18
Provisional diagnosis, 2-3
R
Resource bands, 7-16
ROC curve
Rule-out
diagnosis, 2-3
suspected, and provisional diagnoses, 2-3
Rx-defined morbidity groups, 5-3
Rx-MG
ATC to Rx-MG assignment methodology, 5-9
expected duration, 5-8
morbidity differentiation, 5-8
NDC to Rx-MG assignment methodology, 5-9
primary anatomico-physiological system, 5-6
severity, 5-8
Rx-PM
development, 7-15
Rx defined morbidity groups, 5-3
score, 7-16
S
Sensitivity and positive predictive value
empiric validation of the likelihood of hospitalization
model, 9-6
Significance of pharmacy adherence for effective care,
11-1
Special note for ICD-10 users, 2-4
Special population markers
chronic condition count, 6-5
condition, 6-9
frailty conditions, 6-3
hospital dominant conditions, 6-1
introduction, 6-1
Statistical modeling approach, 7-11
algebraic, 7-11
linear regression, 7-11
logistic regression, 7-11
neural networks, 7-12
Subgroups (CADG)
ADG, 3-6
Suspected diagnosis, 2-3
T
Tables
ADGs and common ICD-9-CM codes assigned to them,
3-4
annual prevalence ratios for expanded diagnosis
clusters (EDCs) for commercially insured and
Medicare beneficiaries, 4-6
IN-4
Index
annual prevalence ratios for Rx-MGs for commercially
insured and Medicare beneficiaries, 5-20
ATC drug classification levels, 5-2
classification of metformin, 2-6
clinical classification of infant ACGs, 3-37
clinical classification of pregnancy/delivery ACGs, 3-35
counts of associated second level ATCs per Rx-Defined
morbidity groups, 5-5
counts of associated therapeutic classes per Rx-defined
counts of associated third level ATCs per Rx-Defined
CSA example 1, 11-7
CSA example 2, 11-7
C-statistics, 8-10
C-statistics for prediction hospitalization:, 9-7
definitions of condition markers, 6-10
disease-drug class pairings, 11-10
distribution of the chronic condition count marker,
6-8
duration, severity, etiology, and certainty of the
aggregated diagnosis groups (ADGs), 3-28
EDCs considered in the chronic condition count marker,
6-5
effects of frailty conditions during a baseline year on
next year’s hospitalization risk, total healthcare
costs, and pharmacy costs, 6-4
effects of hospital dominant conditions during a
baseline year on next year’s hospitalization risk,
total healthcare costs, and pharmacy costs, 6-2
examples of diabetes complications, 4-5
examples of diagnosic codes included in the hospital
dominant marker, 6-1
final ACG categories, 3-11
ICD-9-CM codes assigned to otitis media EDC (EAR01),
4-3
MACs and the collapsed ADGs assigned to them, 3-8
major Rx-MG categories, 5-7
markers of pharmacy adherence possession, 11-4
MEDC types, 4-14
medically frail condition marker-frailty concepts and
diagnoses, 6-3
MPR, 11-6
MPR example 1, 11-8
MPR example 2, 11-6
number of pharmacy codes for each Rx-MG, 5-10
overlap of High Pharamacy Utilization Model and DXPM (top 1.5% of scores, 7-19
possible types of analyses, 11-14
PPV and sensitivity for predicting hospitalization, 9-6
predictive models evaluated, 8-2
predictive ratio by year 1 cost quintile for commercial
and Medicare populations, 8-7
proportion of total healthcare costs consumed by the
highest cost 10% and lowest cost 50% of enrollees,
7-4
proportion of variance explained (adjusted R-square
values) using year 1 risk factors and year 2 resource
measures for prior cost and ICD & NDC based
combined models, 8-5
proportion of variance explained (adjusted r-square
values) using year 1 risk factors and year 2 resource
use measures for alternative predictive models, 8-3
Rx-MG morbidity taxonomy and clinical characteristics
of medication assigned to each Rx-MG, 5-14
selected ACG variables versus hospitalization as
predictors of high cost patients, 7-3
sensitivity and PPV for top five percent, 8-9
the collapsed ADG clusters and the ADGs that they
comprise, 3-7
Target population, 7-10
Technical reference guide
navigation, 1-1
Technical reference guide content, 1-2
introduction, 1-2
The Johns Hopkins ACG system
introduction, 1-1
U
Using ICD-9 and ICD-10 simultaneously, 2-4
Utilization markers, 9-3
dialysis service, 9-3
emergency visit count, 9-4
inpatient hospitalization count, 9-4
major procedure, 9-4
nursing service, 9-3
outpatient visit count, 9-4
V
Validation and testing, 11-13
Validation data

Technical Reference Guide The Johns Hopkins ACG System

Transcription

Similar documents

Kelly Order Form

investment bank prospectus - Texas ACG Capital Connection

Page 1 of 2 Wafer Biscuits, Waffles, Sugarcones, Wafers, Ice Cream

SPOnSORSHIP - SourceMedia

BMC Health Services Research

AHITU-Projection

Explanatory Ability of the ACG System Regarding

Hospital to Skilled Nursing Facilities