Open Collections

Abstract

The automation of intravenous anesthesia is a complex control problem hindered by significant inter-individual variability, model uncertainty, and surgical disturbances. While classical systems improve precision, their reliance on generalized population-based models limits adaptation to individual patient responses. This thesis systematically overcomes these challenges, progressing from an analysis of population-based models to the development and validation of a novel, adaptive framework for multi-variable anesthesia. We critically evaluate whether more complex Pharmacokinetic (PK) models are inherently superior for control system design due to their ability to capture demographic-based variability in response. A rigorous comparison reveals that the choice of the PK model does not significantly reduce control-relevant variability. This highlights the fundamental need for robust, adaptive feedback systems rather than sole reliance on achieving model perfection. To facilitate the development of such a system, we designed and validated the Anesthesia Response Simulator (AReS). AReS provides a realistic in-silico environment by incorporating state-of-the-art models for nonlinear drug interactions, simulating surgical disturbances, and including a diverse set of patient-specific models to simulate variability beyond demographics. The primary contribution of this thesis is a novel, hierarchical control framework that merges model-based control with data-driven intelligence. This architecture features a safety-constrained Generalized Predictive Controller (GPC) to manage the multi-variable administration of drugs for hypnosis, nociception, and hemodynamics. A high-level Reinforcement Learning agent acts as an intelligent supervisor, learning an adaptive policy to dynamically tune the GPC’s parameters. Evaluated using AReS across a diverse cohort of virtual patients and surgical scenarios, this adaptive framework demonstrated superior performance over a fixed-parameter baseline. It achieved faster induction and maintained physiological stability more consistently despite surgical stimuli. The system successfully adapts to different phases of anesthesia despite inter-individual response variability, demonstrating a capacity for personalization that presents a viable pathway toward safer, more reliable automated anesthesia. Furthermore, a theoretical investigation within a multi-agent framework demonstrated that formulating the problem to guarantee convergence of data-driven algorithms would require an overly simplistic formulation of this complex problem. This finding justifies our focus on rigorous experimental validation for the proposed hierarchical framework.