Data-Based Prediction of the Duration of the Postoperative Stay of Patients

Bed management plays a crucial role in hospitals. Up to now, available bed management software mainly allows for recording occupied beds, but there is no forecast about when beds will be unoccupied. This study aims to train models using machine learning (ML) algorithms to predict the length of patients’ postoperative stay in Upper Austrian hospitals following 27 different procedures. This includes data from over 85,000 surgical procedures and tested algorithms such as genetic programming (GP), artificial neural networks (ANNs), random forest (RF), support vector machines (SVM), gradient-boosting (GB), and light gradient-boosting machines (LGBM). Additionally, the inclusion of postoperative data up to seven days after surgery was examined. Both open and closed-box, also known as white and black-box ML techniques were employed. The accuracy of these models was tested using unseen data, representing 30% of the entire dataset. When predicting the length of stay at the day of different surgeries, especially surgeries with low rates of acute cases such as nasal septum correction (accuracy is 87.50% and 96.88% with ± one-day tolerance) show better results than procedures with higher rates of acute cases such as appendectomy (accuracy is 41.81% and 76.70% with ± one day) or cesarean section (accuracy is 39.29% and 85.33% with ± one day). Further, the overall accuracy of most of the used algorithms correlates with the decreasing amount of data and increasing variance in the number of performed services in the days after surgery. However, the determined results suggest promising potential for an AI-based system for postoperative bed occupancy management in hospitals.