Using Machine Learning to Predict-Then-Optimize Elective Orthopedic Surgery Scheduling to Improve Operating Room Utilization: Retrospective Study.

Background: Total knee and hip arthroplasty (TKA and THA) are among the most performed elective procedures. Rising demand and the resource-intensive nature of these procedures have contributed to longer wait times despite significant health care investment. Current scheduling methods often rely on average surgical durations, overlooking patient-specific variability. Objective: To determine the potential for improving elective surgery scheduling for TKA and THA, respectively, by using a 2-stage approach that incorporates machine learning (ML) prediction of the duration of surgery (DOS) with scheduling optimization. Methods: In total, 2 ML models (one each for TKA and THA) were trained to predict DOS using patient factors based on 302,490 and 196,942 patients, respectively, from a large international database. In total, 3 optimization formulations based on varying surgeon flexibility were compared: Any (surgeons could operate in any operating room at any time), Split (limitation of 2 surgeons per operating room per day), and multiple subset sum problem (MSSP; limit of 1 surgeon per operating room per day). Two years of daily scheduling simulations were performed for each optimization problem using ML prediction or mean DOS over a range of schedule parameters. Constraints and resources were based on a high-volume arthroplasty hospital in Canada. Results: The TKA and THA prediction models achieved test accuracy (with a 30 min buffer) of 78.1% (mean squared error 0.898) and 75.4% (mean squared error 0.916), respectively. Any scheduling formulation performed significantly worse than the Split and MSSP formulations with respect to overtime and underutilization (P<.001). The latter 2 problems performed similarly (P>.05) over most schedule parameters. The ML prediction schedules outperformed those generated using a mean DOS for most scheduling parameters, with overtime reduced on average by 300-500 minutes per week (12-20 min per operating room per day; P<.001). However, there was more operating room underutilization with the ML prediction schedules, with it ranging from 70-192 minutes more underutilization (P<.001). Using a 15-minute schedule granularity with a waitlist pool of a minimum of 1 month generated the ML schedule that outperformed the mean schedule 97.1% of times. Conclusions: Assuming a full waiting list, optimizing an individual surgeon's elective operating room time using an ML-assisted predict-then-optimize scheduling system improves overall operating room efficiency, significantly decreasing overtime. This has significant potential implications for health care systems struggling with pressures of rising costs and growing operative waitlists.