Research Fellow Mayo Clinic Rochester, Minnesota, United States
Introduction: Estimating operative time for planned spine surgeries typically relies on the experience of the surgical team or retrospective case studies. Efficient management of operative time and operating room utilization is crucial for hospitals to deliver timely and cost-effective care. This study aims to develop and compare predictive machine learning models using institutional and nationwide datasets to improve the accuracy of operative time estimation for elective spine surgeries.
Methods: Two datasets were used: an institutional cohort and the ACS-NSQIP database, both containing patient and spine procedure details. Due to skewed distributions of operative time, the top 5% of outliers were excluded from both datasets. Several machine learning models were tested, including Linear Regression (LR), Support Vector Regression (SVR), Deep Neural Networks (DNN), and Extreme Gradient Boosting Regression (XGBR), with extensive hyperparameter tuning. Performance was evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R² metrics.
Results: Models built on the institutional dataset yielded more accurate predictions (DNN RMSE: 53.44, MAE: 38.33, R²: 0.74; XGBR RMSE: 52.24, MAE: 37.25, R²: 0.76) than those trained on the ACS-NSQIP dataset (XGBR RMSE: 61.98, MAE: 47.90, R²: 0.38). The performance gap underscores the impact of dataset characteristics on predictive accuracy. The XGBR model consistently outperformed others across datasets, demonstrating robustness in managing complex data.
Conclusion : The study highlights that while machine learning models can enhance operative time prediction for spine surgeries, dataset selection plays a pivotal role in their accuracy. Institutional datasets, with greater clinical granularity, provide more reliable predictions, supporting the notion that institution-specific models may be more effective for optimizing operating room scheduling in spine surgery.