Optimizing surgical efficiency: predicting case duration of common general surgery procedures using machine learning
Michelle Kwong1,2, Mohammad Noorchenarboo3, Katarina Grolinger3
1Department of Anesthesiology and Pain Medicine, University of Alberta, Edmonton, Canada.
Background:
Accurate prediction of surgical duration is critical to optimizing use of operating room resources. Currently, cases are scheduled using subjective estimates of length by surgeons, relying heavily on prior experience. This study aims to develop and compare various predictive models-from conventional statistics to machine learning-based algorithms-to accurately and objectively predict case duration for common elective general surgical procedures.
Methods:
Electronic health record data across three academic tertiary centers were used to train models to predict "case time duration," defined as the time between patient entry to and departure from the operating room. Model performance was evaluated based on predictive accuracy as well as residual analysis, and ultimately benchmarked against "scheduled duration," defined as case time estimated preoperatively by primary surgeons.
Results:
Predictive models, including simple linear regression, Ridge regression, Lasso regression, Support Vector Regression, Random Forest, Gradient Boosting Machine, XGBoost, and Artificial Neural Network (ANN), were trained on a cohort of 16,159 patients [mean age, 56.85 ± 15.95; 47.48% male] having undergone 17,246 elective general surgery procedures. The ANN model demonstrated superior predictive accuracy (Root Mean Squared Error, 49.7 min [95% CI 47.5 to 52.0]; Mean Absolute Error, 31.8 min [95% CI 30.6 to 33.0]). Residual analysis showed that the ANN resulted in an average residual of -0.37 min [95% CI - 40.42 to 39.68, p = 0.34], while the scheduled duration produced an average residual of - 18.52 min [95% CI - 55.24 to 18.2, p < 0.01], demonstrating that the ANN provided a more accurate case time estimation by more than 18 min.
Conclusion:
The ANN model estimates of case time were meaningfully more accurate than provider knowledge-based estimates. By eliminating the subjective bias and dogma inherent in the traditional scheduling methods, future applications of machine learning to predict case duration may improve healthcare resource utilization.
