Explainable Adaptive PID Tuning with Reinforcement Learning in Anesthesia Induction

Total Intravenous Anesthesia (TIVA) automation requires rapid induction while avoiding excessive depth of anesthesia, a challenge exacerbated by strong interpatient pharmacodynamic variability. Fixed-gain PID controllers often fail to balance responsiveness and safety, whereas advanced methods such as MPC or neural networks introduce opacity that complicates regulatory certification and clinical acceptance. This work proposes a Real-Time Adaptive PID controller whose gains are scheduled through a deterministic policy derived via offline tabular Q-Learning. The learned gain-scheduling map links discretized tracking-error states to optimal PID parameters, enabling dynamic adaptation from induction to maintenance while preserving full interpretability and deterministic behavior, key requirements for European Medical Device Regulation approval. The approach was evaluated through Monte Carlo simulations on 100 virtual patients generated by randomizing Schnider and Hill model parameters. Compared to a conventional fixed-gain PID, the Adaptive RL-PID significantly reduced safety violations while maintaining deterministic, explainable behavior. The framework offers RL-driven performance improvements while remaining compatible with certification pathways for safety-critical medical controllers.