A predictive–prescriptive framework for operating room scheduling: Transformer-based duration estimation and system-coupled robust multi-resource optimization

Operating room scheduling (ORS) is complicated by uncertain surgical durations, downstream recovery limitations, and workforce constraints. This study develops a two-phase predictive–prescriptive decision-support framework (DSF) for tactical perioperative planning. In Phase I, heterogeneous clinical, procedural, and provider information is integrated within a Transformer-based regression model using fixed BioBERT-derived diagnostic representations. The resulting out-of-fold predictions and empirical residual distribution are used to construct residual-calibrated duration bounds, thereby creating a data-driven connection between prediction uncertainty and scheduling decisions. In Phase II, these inputs are transferred to a mixed-integer convex quadratic System-Coupled Robust Operating Room Scheduling (SC-ORS) model that coordinates OR utilization, daily aggregate PACU capacity, surgeon availability, and workload balance. The complete predictive pipeline outperforms the selected machine-learning (ML) benchmarks under leakage-controlled cross-validation, while the residual-based intervals provide internally calibrated uncertainty information for the optimization phase. Computational experiments show that SC-ORS can reduce realized overtime exposure under several moderately constrained conditions, identify emerging PACU bottlenecks, and support schedule adjustments under limited surgeon availability disruptions. The results also reveal that the value of robustness depends on the demand-to-capacity ratio and on managerial preferences regarding OR utilization, downstream capacity, and workload balance. The proposed framework contributes primarily through the explicit integration of heterogeneous surgical-duration prediction, residual-calibrated uncertainty, and multi-resource tactical optimization. It provides a structured basis for moving from OR-centric planning toward coordinated perioperative decision support, while hospital-specific calibration, intra-day sequencing, and prospective validation remain necessary before operational deployment.