Physics-informed Koopman-constrained implicitQ-learning for safe offline reinforcement learning in mechanical...

Mechanical ventilation management in intensive care units requires continuous optimization of ventilator parameters while adhering to clinical safety constraints. This paper presents physics-informed Koopman-constrained implicitQ-learning (PIK-IQL), a novel offline reinforcement learning framework that integrates respiratory mechanics with data-driven policy optimization. The proposed approach addresses three fundamental challenges: learning from retrospective clinical data without online interaction, incorporating physical laws governing respiratory dynamics, and ensuring policy robustness against unmeasured confounding. We develop a PIK world model that lifts nonlinear respiratory dynamics into a linear representation while preserving mechanical constraints. The Koopman model guides policy learning by providing physics-consistent state representations and enabling model-based reward computation, while the IQL policy learns conservative ventilation strategies through expectile regression with hierarchical action regularization. Experiments on the MIMIC-IV database (15,446 patients, 1.58 million hourly observations) demonstrate that PIK-IQL achieves a policy value improvement of 0.088 (95% CI: [0.078, 0.096]) over behavior policy, with zero aggressive actions exceeding safety thresholds. A simplified Rosenbaum sensitivity analysis suggests robustness to unmeasured confounding with criticalΓ∗=3.0, providing an approximate indicator of causal validity for the observed improvements.