Behavior-aware deep reinforcement learning for multi-objective outpatient scheduling optimization

Abstract

Abstract Outpatient departments in large hospitals face persistent scheduling inefficiencies characterized by prolonged patient waiting, underutilized resources, and high no-show rates. Existing scheduling approaches largely ignore the behavioral heterogeneity of patients, treating satisfaction as a simple proxy of waiting time rather than a psychologically grounded construct. This paper proposes MO-SAC-B, a multi-objective deep reinforcement learning framework that integrates behavioral science theory into the scheduling optimization process. We first construct a behavior-driven discrete-event simulation environment that encodes prospect-theoretic waiting disutility, nonlinear patience decay, and behaviorally calibrated no-show and abandonment dynamics. A satisfaction-aware reward shaping mechanism translates these behavioral constructs into dense learning signals, while a multi-objective Soft Actor-Critic algorithm with adaptive weight adjustment and prioritized experience replay navigates the efficiency–satisfaction Pareto frontier. Experiments calibrated with real outpatient data from a tertiary hospital demonstrate that MO-SAC-B reduces mean waiting time by 21.9%, improves composite patient satisfaction by 12.7 points, and lowers the no-show rate by 25.8% relative to the strongest baseline. Ablation studies confirm that each behavioral component contributes meaningfully, with synergistic effects amplifying performance gains under high patient flow conditions. Robustness analysis further validates the framework’s adaptability to demand surges and resource disruptions.