This retrospective, multicenter study asked whether a two-stage machine learning–based pipeline, which first predicts the duration of surgery (DOS) and then optimizes operating room schedules, could reduce overutilization and underutilization in total hip and knee arthroplasty scheduling compared with traditional mean-based scheduling. Using a multilayer perceptron (MLP) model trained on ACS NSQIP data, researchers predicted DOS for 302,490 TKA and 196,942 THA cases, then fed those predictions into three alternative integer-linear-programming scheduling formulations to construct weekly elective surgery schedules under realistic clinical constraints. The models were trained on 2014-2017 data, tuned on 2018, and evaluated on 2019 procedures. Preprocessing included extracting 33 preoperative patient features and rounding predicted completion times to 10- or 15-minute blocks. Schedules generated with ML predictions, historical mean DOS values, and true DOS (hindsight upper bound) were compared across >100 simulated weekly cycles. The best-performing formulation (MSSP: one surgeon per room per day) achieved consistent reductions in overtime of approximately 300–500 minutes per simulated week (P<.001) relative to mean-based schedules, though with modest increases in underutilization. Across schedule parameters, the two-stage method outperformed mean scheduling in >80% of simulated weeks and >90% of weeks when using 15-minute granularity. The analysis showed that the “Any” formulation yielded the poorest performance across parameter sets, while the MSSP and Split formulations produced significantly less overtime and avoided cascading errors arising from DOS underestimation. Secondary analyses demonstrated that schedule granularity and surgeon waitlist size significantly influenced underutilization, with less unused time observed when using 15-minute blocks and waitlists ≥500 cases. Although hindsight schedules, using perfect DOS information, generated near-ideal performance, neither the ML-based nor the mean-based methods matched this upper bound. Limitations included reliance solely on preoperative features, potential selection bias, lack of surgeon-specific operative patterns, absence of bed and staffing constraints, and incomplete fairness or subgroup analyses. External validation was not performed, and findings primarily reflect simulation-based performance rather than real-world operational outcomes.