UC San Diego

BACKGROUND: Days before surgery, add-ons may be scheduled to fill unused surgical block time at an outpatient surgery center. At times, outpatient surgery centers have time limitations for end of block time and discharge from the postanesthesia care unit (PACU). The objective of our study was to develop machine learning models that predicted the following composite outcome: (1) surgery finished by end of operating room block time and (2) patient was discharged by end of recovery room nursing shift. We compared various machine learning models to logistic regression. By evaluating various performance metrics, including F1 scores, we hypothesized that models using ensemble learning will be superior to logistic regression. METHODS: Data were collected from patients at an ambulatory surgery center. The primary outcome measurement was determined to have a value of 1 (versus 0) if they met both criteria: (1) surgery ends by 5 pm and (2) patient is discharged from the recovery room by 7 pm. We developed models to determine if a procedure would meet both criteria if it were scheduled at 1 pm, 2 pm, 3 pm, or 4 pm. We implemented regression, random forest, balanced random forest, balanced bagging, neural network, and support vector classifier, and included the following features: surgery, surgeon, service line, American Society of Anesthesiologists score, age, sex, weight, and scheduled case duration. We evaluated model performance with Synthetic Minority Oversampling Technique (SMOTE). We compared the following performance metrics: F1 score, area under the receiver operating characteristic curve (AUC), specificity, sensitivity, precision, recall, and Matthews correlation coefficient. RESULTS: Among 13,447 surgical procedures, the median total perioperative time (actual case duration and PACU length stay) was 165 minutes. When SMOTE was not used, when predicting whether surgery will end by 5 pm and patient will be discharged by 7 pm, the average F1 scores were best with random forest, balanced bagging, and balanced random forest classifiers. When SMOTE was used, these models had improved F1 scores compared to no SMOTE. The balanced bagging classifier performed best with F1 score of 0.78, 0.80, 0.82, and 0.82 when predicting our outcome if cases were to start at 1 pm, 2 pm, 3 pm, or 4 pm, respectively. CONCLUSIONS: We demonstrated improvement in predicting the outcome at a range of start times when using ensemble learning versus regression techniques. Machine learning may be adapted by operating room management to allow for a better determination whether an add-on case at an outpatient surgery center could be appropriately booked.