A Reinforcement Learning Framework for Personalized Anticoagulation Dosing in Critical Care: Integrating Batch-Constrained Policy Optimization and Off-Policy Evaluation

Precise medication dosing in the intensive care unit (ICU) is vital for patient survival. Heparin, a widely used anticoagulant, requires careful administration due to patient-specific variability, and inappropriate dosing can cause severe complications such as stroke or hemorrhage. This study introduces a reinforcement learning (RL)-based decision-support framework for heparin dosing, integrating offline RL algorithms with rigorous evaluation. We employ Batch-Constrained deep Q-Learning (BCQ) to learn an optimal dosing policy from retrospective data, addressing distributional shift inherent in offline settings. The dosing policies are trained on the MIMIC-III database and evaluated on the MIMIC-IV database, and vice versa. Policy effectiveness is assessed through multiple off-policy evaluation (OPE) methods, demonstrating higher expected returns than clinician-derived strategies. Interpretability is enhanced through t-SNE visualization, showing that Q-values are well aligned with therapeutic aPTT targets. To our knowledge, this is the first study to combine BCQ, multi-metric OPE, and interpretability analysis for anticoagulation management across two large-scale ICU cohorts. By advancing both methodological rigor and clinical relevance, this work provides a foundation for reliable RL-based decision-support systems in critical care.