Slides - Politecnico di Milano
Transcription
Slides - Politecnico di Milano
Heart-Pacing Simulation and Control via Multiagent Systems Alessandro Beda1 and Nicola Gatti and Francesco Amigoni2 1 ISVR, U NIVERSITY OF S OUTHAMPTON , S OUTHAMPTON , UK 2 DEI, P OLITECNICO DI M ILANO, M ILANO, I TALY Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 1/17 Outline • We introduce some problems related to the modeling of physiological processes Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 2/17 Outline • We introduce some problems related to the modeling of physiological processes • We present a cooperative negotiation protocol to combine partial models of physiological processes embedded in agents Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 2/17 Outline • We introduce some problems related to the modeling of physiological processes • We present a cooperative negotiation protocol to combine partial models of physiological processes embedded in agents • We applied and evaluated the proposed protocol in heart-pacing modeling Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 2/17 The Global Approach • Physiological processes are complex to study and model [Goldberger, West, 1987] ◦ Each process emerges from the interaction of several elements belonging to an intricate network of relationships ◦ Each element is involved in more processes Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 3/17 The Global Approach • Physiological processes are complex to study and model [Goldberger, West, 1987] ◦ Each process emerges from the interaction of several elements belonging to an intricate network of relationships ◦ Each element is involved in more processes • Physiological processes are currently described using a set of partial models [Amigoni, Gatti, 2003] ◦ Open problem: how combine the partial models? ◦ Current solutions: weighted average, overriding, etc. Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 3/17 Our Case Study: Heart-Pacing (1) • Permanent cardiac pacing is used as effective and reliable therapy for alterations of cardiac rhythm [Glikson, Hayes, 2001] • Cardiac pacemakers partially or totally take over the function of the Sinoatrial Node Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 4/17 Our Case Study: Heart-Pacing (1) • Permanent cardiac pacing is used as effective and reliable therapy for alterations of cardiac rhythm [Glikson, Hayes, 2001] • Cardiac pacemakers partially or totally take over the function of the Sinoatrial Node • Kinds of pacemakers [Harvey, 1995]: ◦ Asynchronous (fixed pacing frequency) ◦ Synchronous (fixed pacing frequency, activated just when needed) ◦ Adaptive (frequency is a function of physiological parameters) • The adaption aims at mimicking adequately the heart regulatory system of healthy subjects Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 4/17 Our Case Study: Heart-Pacing (2) • Currently there is no answer to the question: what is the optimal model of heart-pacing? Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 5/17 Our Case Study: Heart-Pacing (2) • Currently there is no answer to the question: what is the optimal model of heart-pacing? • The adopted technique is to combine different physiological models ◦ The model combination is based on: • Overdrive • Cross-checking • Weighted average Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 5/17 Modeling Complex Systems • Complex systems are systems that exhibit intricate interconnections, high dimensionality, multiresolution, etc. [Astrom et al., 2001] • It is commonplace in literature the idea to face complexity by using different models according to scale, operating contexts, etc. Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 6/17 Modeling Complex Systems • Complex systems are systems that exhibit intricate interconnections, high dimensionality, multiresolution, etc. [Astrom et al., 2001] • It is commonplace in literature the idea to face complexity by using different models according to scale, operating contexts, etc. • From a modeling point of view, a physiological process is similar to a complex system Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 6/17 Modeling Complex Systems • Complex systems are systems that exhibit intricate interconnections, high dimensionality, multiresolution, etc. [Astrom et al., 2001] • It is commonplace in literature the idea to face complexity by using different models according to scale, operating contexts, etc. • From a modeling point of view, a physiological process is similar to a complex system • An emerging technique is decentralized optimization where the single models are embedded in decisional makers with objective function Ji min / max Ji (u1 (t), . . . , un (t), x0 ) ∀i u s.t. f (u) ≤ 0 i Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 6/17 Anthropic Agency A multiagent system for physiological processes [Amigoni et al., 2003] able to accomplish several tasks: • Signal Processing • Decision Making • etc. Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 7/17 Anthropic Agency A multiagent system for physiological processes [Amigoni et al., 2003] able to accomplish several tasks: • Signal Processing • Decision Making • etc. Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 7/17 Modeling with Anthropic Agency • Each agent embeds a model • Each agent repeatedly performs an optimization with respect to the model it embeds (agent optimization) • The global model is the result of a distributed decision making process performed via a cooperative negotiation among the agents (agency optimization) Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 8/17 Modeling with Anthropic Agency • Each agent embeds a model • Each agent repeatedly performs an optimization with respect to the model it embeds (agent optimization) • The global model is the result of a distributed decision making process performed via a cooperative negotiation among the agents (agency optimization) • Requirements [Amigoni, Gatti, 2004]: ◦ High formal degree of the protocol • To allow an analytical study of the protocol properties: stability, optimality, etc. ◦ High parameterization degree of the protocol • To allow a fine tailoring the combination phase to the specific human being • A novel protocol is needed Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 8/17 Assumptions • n contracting agents and a mediator • Each contracting agent i embeds a partial model referring to a space of variables Ai ∈ RNi • Different Ai can overlap and the global space of variables is Sn A = i=1 Ai • Each contracting agent i is associated with a potential function Ui : Ai → [0, +∞) that represents the utility of the agent i ◦ It expresses the “distance” of a state from the optimal curve of the model embedded by the agent • pti→e is the proposal of the agent i at time t to the mediator e • pte→i is the counter-proposal of the mediator e at time t to the agent i Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 9/17 Examples of Potential Functions „ 1930.4 − QT HR = −8.2/ log 1734.7 [Sarma et al.; 1984] « HR = 26.98 · arctan (0.0995 · RR − 2.5947) + 95.02 [Voukydis et al.; 1967] Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 10/17 Examples of Potential Functions „ 1930.4 − QT HR = −8.2/ log 1734.7 [Sarma et al.; 1984] « HR = 26.98 · arctan (0.0995 · RR − 2.5947) + 95.02 [Voukydis et al.; 1967] 8 d(p)2 > > < if d(p)2 < 20 10 „„ « « Definition of potential U : 2 d(p) > > − 20 + 1 if d(p)2 > 20 :20 + 10 ∗ log 10 Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 10/17 Examples of Potential Functions „ 1930.4 − QT HR = −8.2/ log 1734.7 [Sarma et al.; 1984] « HR = 26.98 · arctan (0.0995 · RR − 2.5947) + 95.02 [Voukydis et al.; 1967] 8 d(p)2 > > < if d(p)2 < 20 10 „„ « « Definition of potential U : 2 d(p) > > − 20 + 1 if d(p)2 > 20 :20 + 10 ∗ log 10 Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 10/17 The Negotiation Protocol (1) 1. WHILE true 2. R EAD CurrentState pi 3. Each agent operates a local search: p0i→e ← LocalSearch (pi , Ui (·)) 4. WHILE Agreement has not been reached 5. Each agent sends to the mediator: hpti→e , Ui (pti→e )i 6. The mediator computes the agreement mt 7. The mediator sends the agreement mt to each agent 8. The agents computes their new proposal pt+1 i→e 9. t = t + 1 10. END WHILE 11. END WHILE Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 11/17 The Negotiation Protocol (2) • Succession of proposals: p0i→e p0e→i p1i→e · · · pτe→i Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 12/17 The Negotiation Protocol (2) • Succession of proposals: p0i→e p0e→i p1i→e · · · pτe→i Pn t t p · N · [1 + U (p i i i→e i→e )] i=1 t t t • Agreement: m = Pn , p = m e→i t i=1 Ni · [1 + Ui (pi→e )] Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 12/17 The Negotiation Protocol (2) • Succession of proposals: p0i→e p0e→i p1i→e · · · pτe→i Pn t t p · N · [1 + U (p i i i→e i→e )] i=1 t t t • Agreement: m = Pn , p = m e→i t i=1 Ni · [1 + Ui (pi→e )] • New proposal: t+1 t+1 t t t = p + kp − p k(α (p )u + β (p )w ) pt+1 i i i i i→e e→i i→e i→e i i ◦ ut+1 i ◦ vit+1 pte→i − pti→e = kpte→i − pti→e k (accommodate the tendency of society with weight αi ) ∇Ui (pti→e ) = k∇Ui (pti→e )k ◦ wit+1 = ut+1 i kut+1 i − − vit+1 vit+1 · · t+1 ut+1 v i i t+1 ut+1 k v i i (keep close to optimal curve with weight βi ) Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 12/17 A Negotiation Example Two agents negotiate on the HR value 410 405 QT Negotiation LocalSearch 400 395 390 60 70 80 90 100 110 120 130 140 150 120 130 140 150 HR 30 25 RR Negotiation LocalSearch 20 15 10 60 70 80 90 100 110 HR Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 13/17 Experimental Setup in Heart-Pacing • 5 sequences 24-hours long ECGs taken from PhysioNet (http://www.physionet.org/ ) • Signal processing of ECG to extract temporal series: QT, RR, HR Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 14/17 Neg. Prot. Evaluation in Heart-Pacing Indexes • e: absolute value of the difference between the esteemed value of HR and the real value of HR • E[e]: the average of e over time (24 hours) • var[e]: the variance of e Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 15/17 Neg. Prot. Evaluation in Heart-Pacing Indexes • e: absolute value of the difference between the esteemed value of HR and the real value of HR • E[e]: the average of e over time (24 hours) • var[e]: the variance of e Results • Averaging HR values of two partial models: E[e] = 4.6 ± 1.3, var[e] = 8.6 ± 3.0 • Negotiating HR values of two partial models: E[e] = 3.9 ± 1.3, var[e] = 6.8 ± 1.2 Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 15/17 Negotiation Protocol Analysis • Our negotiation protocol generalizes other techniques for model combination in heart-pacing (by using particular values for parameters): ◦ Weighted average ◦ Overdrive ◦ Cross-checking Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 16/17 Negotiation Protocol Analysis • Our negotiation protocol generalizes other techniques for model combination in heart-pacing (by using particular values for parameters): ◦ Weighted average ◦ Overdrive ◦ Cross-checking • Advantages: ◦ Improvement of E[e] and var[e] over other techniques ◦ Adoption of βi improves results ◦ Less sensitivity to weights with respect to weighted average Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 16/17 Negotiation Protocol Analysis • Our negotiation protocol generalizes other techniques for model combination in heart-pacing (by using particular values for parameters): ◦ Weighted average ◦ Overdrive ◦ Cross-checking • Advantages: ◦ Improvement of E[e] and var[e] over other techniques ◦ Adoption of βi improves results ◦ Less sensitivity to weights with respect to weighted average • Drawbacks: ◦ Increased complexity of tailoring the model on a particular patient Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 16/17 Conclusions and Future Works • Cooperative negotiation is an interesting paradigm to combine physiological partial models • The proposed protocol improve effectively the current heart-pacing model combination techniques • Future works: ◦ Enrich heart-rate modeling ◦ Apply our approach to model other physiological processes ◦ Study analytical properties of the protocol ◦ Develop learning techniques for automatically tailoring parameters Beda, Gatti and Amigoni, Heart-Pacing Simulation and Control via Multiagent Systems – p. 17/17