Automated Machine Learning Approaches for Surgery Duration Prediction in Orthopaedics

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Operations Research

Background:

• Accurate surgical case duration prediction is crucial for operating room efficiency.
• Traditional estimation methods have limited accuracy, with mean absolute errors (MAE) of 30-70 minutes.
• Optimizing theatre utilization requires precise surgical time forecasting.

Purpose of the Study:

• To evaluate the performance of AutoGluon, an automated machine learning (AutoML) framework, for predicting elective orthopaedic surgery duration.
• To compare AutoML performance against traditional statistical and machine learning models.
• To identify key predictors of surgical duration and overrun risk.

Main Methods:

• Retrospective analysis of 94,502 elective orthopaedic procedures.
• Utilized AutoGluon framework for surgery duration prediction.
• Compared AutoGluon against linear regression, XGBoost, and a feed-forward neural network using standardized preprocessing and optimization.

Main Results:

• AutoGluon achieved a mean absolute error (MAE) of 15.70 minutes, outperforming XGBoost by 26%.
• Extended training with AutoGluon reduced MAE to 11.84 minutes, a 46% improvement over surgeon estimates.
• SHAP analysis identified procedure type, inpatient status, and anesthetic type as key predictors.

Conclusions:

• Automated machine learning frameworks like AutoGluon offer state-of-the-art predictive performance for surgical duration.
• AutoML requires minimal technical expertise, facilitating its adoption in healthcare operations.
• Improved prediction accuracy enhances operating room utilization and reduces costly overruns.