Thesis title: Data-Driven Control of Safety-Critical Biomedical Systems
This thesis investigates data--driven control methodologies for safety--critical biomedical systems, addressing the fundamental challenge of ensuring safety, performance, and adaptability in the presence of modeling uncertainty, nonlinear dynamics, and inter--subject variability. Four complementary methodological directions are explored, each contributing to the unification of control theory and modern machine learning under a safety--aware and data--efficient paradigm.
First, a data--driven Model Predictive Control framework is developed, where Neural Network models are trained to reproduce the nonlinear dynamics of complex biological processes. The approach is validated on both linearized and nonlinear tumor–immune dynamics, demonstrating that neural MPC controllers can recover the stabilizing and constraint--satisfying properties of model--based formulations while relying solely on data. The results highlight the potential of learning--based predictors to retain control--theoretic guarantees even in regimes where analytical models are unavailable or unreliable.
Second, the thesis constructs safe Deep Reinforcement Learning for biomedical control, with applications to automated insulin delivery for glycemic regulation in type--1 diabetes. Several policy optimization algorithms—ranging from deterministic and stochastic actor–critic methods (DDPG, TD3, SAC, PPO) to constrained formulations—are examined and validated within a physiologically realistic in--silico environment. Building upon these foundations, two safe controllers are designed: a Lagrangian Constrained Markov Decision Process controller, and a Barrier--Lyapunov Actor–Critic agent. The former enforces probabilistic safety via dual updates on constraint functions, whereas the latter integrates Lyapunov--stability priors and Control Barrier Functions into the actor–critic loop. Simulation results across multiple virtual patients confirm that embedding control--theoretic structure within Deep Reinforcement Learning markedly improves robustness and inter--patient generalization, thereby bridging the gap between model--based safety and model--free adaptability.
Third, the thesis extends data--driven safe closed--loop control to systems with high relative degree by introducing discrete--time high--order Control Barrier Functions. A probabilistic formulation is developed to bound safety--violation risk under model uncertainty that is modeled through Gaussian Processes, enabling forward invariance guarantees for complex biological dynamics whose input appears after multiple discrete steps. The theoretical results unify Lyapunov-- and barrier--based reasoning in discrete time and offer a principled path toward learning--based safety certification.
Fourth, a Fractional--Order Neural Network identification and control framework is proposed to model systems with long--memory effects, again with application to type--1 diabetes. By embedding Fractional Calculus within recurrent neural architectures, the approach captures complex dependencies across extended horizons, enabling more accurate multi--step prediction and safer closed--loop regulation when integrated in a data--driven Model Predictive Control formulation. This line of research lays the foundation for hybrid fractional--learning architectures in biomedical control.
Beyond the mentioned core methodological directions, the thesis includes complementary studies that highlight the versatility of the proposed approaches across different domains. These contributions comprise a Machine Learning framework for passivity-preserving Control Barrier Functions, unifying energy-based control and neural approximation; Dynamic Mode Decomposition, providing interpretable and computationally efficient patient-specific models; a Deep Reinforcement Learning framework for medical-image augmentation, improving classifier robustness under data scarcity; a neural regression model for non-invasive diagnosis of portal hypertension, reducing reliance on invasive procedures; Behavioural Cloning for pediatric neurorehabilitation, enabling adaptive therapy personalization; and a two-step explainable-AI system for surgical timing in pediatric oncology. Together with the author’s participation in national and European R$\&$D projects, these works demonstrate the broad applicability and translational potential of the proposed methodologies toward clinical decision support.
Overall, this thesis contributes to the emerging field of safe data--driven control of biomedical systems by harmonizing the theoretical rigor of control theory with the adaptability of modern learning--based paradigms. The presented methodologies advance the practical feasibility of intelligent closed--loop therapies and lay the groundwork for clinically deployable implementations.