Machine learning-enhanced prediction of operating room occupation time and length of stay: a retrospective cohort study on emergency surgery care pathways
Résumé
Emergency surgeries are resource-intensive procedures with high variability in operating room occupation time (OT) and hospital length of stay (LOS), complicating scheduling and capacity planning. Manual estimates by surgeons are frequently inaccurate, especially in emergency settings. Machine learning models (MLMs) have shown good predictive performance in elective surgery, but their applicability to emergency contexts remains underexplored. We conducted a retrospective, single-center study on 3,117 emergency procedures performed at the Pitié-Salpêtrière hospital, a major trauma center, between 2015 and 2018. Preoperative data available at the time of surgical scheduling were used to train four regression models for OT and LOS prediction: Ridge Regression, Random Forest, XGBoost, and a Multi-Layer Perceptron. Model performance was evaluated using Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, and operational metrics: proportion of OT predictions within 20% of actual value (Within20) and LOS within fixedday thresholds. RF and XGB outperformed manual estimates for OT, with RF achieving a MAE of 32 min and Within20 of 60%, improving surgeon estimates by 13%. For LOS, XGB was the best performing model with a MAE of 5 days and RMSE of 12 days. As measured through MAPE, prediction performance varied across specialties, with better accuracy in digestive and maxillofacial procedures. As for elective cases, MLMs can improve OT and LOS predictions in emergency surgery, though predictive performance remains moderate. Future work should refine models through enriched data, clinically relevant thresholds, and integration into decision-support tools to enhance emergency surgical care coordination.