Abstract:

The operating room constitutes a critical unit within hospital operations， with its efficiency exerting a direct impact on surgical patient throughput and overall healthcare service quality. Consequently， advance development of a well-designed surgery schedule is essential for optimizing resource utilization and enhancing patient satisfaction. A particular challenge arises from the fact that patients typically require postoperative care in intensive care units or general wards， necessitating the explicit incorporation of downstream resource capacities into surgical planning. Failure to do so may result in surgery delays， cancellations， or suboptimal allocation of hospital resources. Motivated by these considerations， the surgical scheduling problem under uncertainty with downstream resource constraints is investigated， where surgery durations， postoperative recovery durations， and emergency arrivals are uncertain. Traditional sample average approximation （SAA） approaches often become computationally intractable as the number of scenarios increases， limiting their applicability to large-scale problems. To address this challenge， a hybrid algorithm is developed that integrates machine learning with stochastic programming. Specifically， a stochastic programming model is first formulated and solved on training instances to obtain benchmark solutions. A heuristic method is then applied to identify a representative scenario such that the objective value of the solution obtained by solving the model in this scenario closely aligns with the benchmark. Using instance features and the representative scenario as input， a training dataset is constructed to develop a machine learning model. Once trained， the model can directly predict representative scenarios for new problem instances， effectively converting the stochastic optimization problem into a deterministic one. This transformation enables the rapid generation of high-quality surgical schedules. Extensive computational experiments are conducted using realistic hospital data sourced from the literature. The results demonstrate that the proposed hybrid algorithm achieves a remarkable reduction in computation time while preserving solution quality. Compared with the traditional SAA method， the proposed approach reduces average solution time from 1635.27 seconds to 1.83 seconds—an improvement of nearly three orders of magnitude—while keeping the optimality gap within 1%. These findings not only demonstrate a novel methodological framework for surgery scheduling under uncertainty but also represent a significant advancement in computational efficiency. The study underscores the potential of hybrid data-driven optimization techniques in addressing complex healthcare scheduling challenges.

Key words: surgery scheduling, downstream resource constraint, stochastic programming, machine learning, heuristic algorithm