Hierarchical Reinforcement Learning for Fair Operating Room Scheduling in Plastic Surgery

Equitable operating room (OR) scheduling is critical for surgical training and resource usage. Traditional, hand-crafted schedules frequently perpetuate hidden biases. We present a hierarchical reinforcement learning (HRL) framework, coupled with statistical fairness analysis, to optimize OR assignment while maintaining institutional constraints. A retrospective dataset of 1.5 years (July 2023 to December 2024) of weekly OR assignments from a plastic surgery service (n = 24 surgeons) was analyzed. A random forest model predicted future OR demand; feature importance was quantified with Shapley Additive Explanations. Fairness metrics (Gini, Theil, coefficient of variation, disparate-impact ratios) were computed at baseline. An HRL architecture-Proximal Policy Optimization at a group level and Deep Q-Learning at an individual level-reallocated OR days subject to slot, training, and part-time constraints. Performance was assessed by pre- and postoptimization fairness metrics, mean absolute deviation, and constraint-compliance rate. Baseline scheduling was highly unequal (Gini = 10.39; variance = 1615.8). Past workload dominated prediction (94.6% importance), whereas hierarchy (3.1%) and employment status (2.3%) were minor predictors. HRL reduced the Gini to 0.22, variance to 865.4, and mean absolute deviation from 33.6 to 25.9, whereas qualitative fairness (case complexity, teaching value) remained to be evaluated prospectively and achieved 100% rule compliance. Fellows gained +50% OR exposure (88 versus 58 d), and junior residents, -14% (postgraduate year 2), while preserving the total annual OR capacity. The reproducible HRL framework produces data-driven, fair, and constraint-respecting OR schedules, balancing efficiency with educational equity. Multicenter prospective validation is warranted.