Hierarchical Reinforcement Learning for Fair Operating Room Scheduling in Plastic Surgery

Area of Science:

• Surgical Operations Research
• Machine Learning in Healthcare
• Health Equity

Background:

• Equitable operating room (OR) scheduling is vital for surgical training and resource optimization.
• Traditional scheduling methods often contain hidden biases, leading to unequal resource distribution.
• A hierarchical reinforcement learning (HRL) framework is proposed to address these biases.

Purpose of the Study:

• To develop and evaluate an HRL framework for optimizing OR assignment.
• To analyze and improve fairness in surgical scheduling.
• To maintain institutional constraints while enhancing efficiency and equity.

Main Methods:

• Retrospective analysis of 1.5 years of OR assignments for a plastic surgery service (24 surgeons).
• Utilized a random forest model with Shapley Additive Explanations to predict OR demand and quantify feature importance.
• Implemented an HRL architecture (Proximal Policy Optimization and Deep Q-Learning) to reallocate OR days, considering various constraints.

Main Results:

• Baseline scheduling exhibited significant inequality (Gini = 10.39).
• HRL reduced inequality (Gini = 0.22) and mean absolute deviation (33.6 to 25.9).
• OR exposure increased for fellows (+50%) and junior residents (-14%), preserving total OR capacity with 100% rule compliance.

Conclusions:

• The HRL framework provides data-driven, fair, and constraint-adherent OR schedules.
• This approach balances operational efficiency with educational equity in surgical training.
• Multicenter prospective validation is recommended for broader applicability.