Automated Machine Learning Approaches for Surgery Duration Prediction in Orthopaedics

Accurate prediction of surgical case duration is essential for reducing operating room overruns and maximising theatre utilisation. Traditional estimation methods provide limited accuracy, with reported mean absolute errors (MAE) of 30-70 minutes. This study evaluated AutoGluon, an automated machine learning (AutoML) framework, for surgery duration prediction in a retrospective dataset of 94,502 elective orthopaedic procedures. AutoGluon's performance was compared against linear regression, XGBoost, and a feed-forward neural network under standardised preprocessing and identical optimisation protocols. AutoGluon outperformed all comparators, achieving an MAE of 15.70 minutes (26% improvement over XGBoost) and 11.84 minutes with extended training (46% improvement over surgeon estimates). SHAP-based feature analysis identified procedure type, inpatient status, and anaesthetic type as the strongest predictors of duration and overrun risk. These findings demonstrate that AutoML frameworks can deliver state-of-the-art predictive performance with minimal technical input, supporting their adoption in healthcare operations.