An Interpretable Deep Actor–Critic Framework for Automated Propofol Dosing During General Anesthesia

Administering general anesthesia demands anesthesiologists to concurrently regulate various physiological functions, often under time-critical conditions. The automation of hypnotic drug delivery offers the potential to enhance precision in maintaining optimal sedation levels while allowing clinicians to focus on more complex intraoperative decisions. In this study, we propose a reinforcement learning (RL) framework to guide the real-time administration of propofol, a widely used anesthetic agent. Building upon earlier discrete-action approaches, our model adopts a continuous-action actor-critic architecture. It incorporates a policy network that generates a continuous probability distribution over infusion rates based on dynamic anesthetic state observations, along with a value network that evaluates the desirability of those states. The RL agent is trained on synthetic patient simulations derived from pharmacokinetic and pharmacodynamic (PK/PD) models with randomized parameter sets to ensure robustness across diverse patient profiles. We examine three distinct reward formulations tailored to different clinical objectives and evaluate agent performance in both simulated scenarios and retrospective clinical data from nine surgical cases. To enhance transparency and support clinical adoption, we apply Shapley Additive Explanations (SHAP) to interpret the agent’s dosing rationale. The analysis revealed that dosing decisions were primarily influenced by the level of unconsciousness (LoU) error and predicted effect-site concentration, with LoU target dominating steady-state control. These clinically consistent patterns demonstrate that the agent’s recommendations align with established pharmacological principles, thereby improving clinician trust and interpretability. Experimental results demonstrate that the proposed RL agent surpasses traditional PID controllers in maintaining target sedation levels and aligns closely with anesthesiologist-administered dosages in real cases. This work represents a novel advancement in automated anesthesia control, showcasing the adaptability of reward functions and the integration of interpretability tools to support the development of clinically relevant AI-based drug delivery system.