Using Machine Learning to Predict Operative Time and Enhance Operating Room Scheduling for Robotic Hysterectomies

Abstract

<h3>Study Objective</h3> To create a machine learning based predictive analytic model in order to improve estimation of operative time for robotic hysterectomies. <h3>Design</h3> Four machine learning models were tested including linear regression, XGboost, random forest, and catboost. The models incorporated patient characteristics (age, body mass index, surgical history, uterine size, diagnosis, and history of pelvic adhesive disease) and surgical characteristics (presence of trainees, number of procedures performed concomitantly, and surgeon median times for prior 5 and 50 surgeries). Shapely values were used to interpret feature importance. Hyperparameter tuning determined that the XGBoost model performed best. The XGBoost model was compared to current institutional practice of scheduling based on the median time of last 50 surgeries. <h3>Setting</h3> N/A. <h3>Patients or Participants</h3> A dataset of N = 1090 patients undergoing robotic hysterectomy from January to December 2019 across three hospital campuses was utilized. Seventy percent of the data were used for training, 15% for validation and 15% for testing. <h3>Interventions</h3> N/A. <h3>Measurements and Main Results</h3> The XGBoost model outperformed current practice as noted in Table 1. The features that most contributed to this difference were uterine size, body mass index, and number of concomitant procedures. <h3>Conclusion</h3> Machine learning is a valuable tool that may be useful in predicting operative times with the potential to reduce operating room inefficiencies for robotic hysterectomies. Limitations of this model such as the small dataset will be addressed in future work to strength the model's predictive capability.