A machine learning approach to predicting early and late postoperative reintubation

Abstract

Accurate estimation of surgical risks is important for informing the process of shared decision making and informed consent. Postoperative reintubation (POR) is a severe complication that is associated with postoperative morbidity. Previous studies have divided POR into early POR (within 72 h of surgery) and late POR (within 30 days of surgery). Using data provided by American College of Surgeons (ACS) National Surgical Quality Improvement Program (NSQIP), machine learning classification models (logistic regression, random forest classification, and gradient boosting classification) were utilized to develop scoring systems for the prediction of combined, early, and late POR. The risk factors included in each scoring system were narrowed down from a set of 37 pre and perioperative factors. The scoring systems developed from the logistic regression models demonstrated strong performance in terms of both accuracy and discrimination across the different POR outcomes (Average Brier score, 0.172; Average c-statistic, 0.852). These results were only marginally worse than prediction using the full set of risk variables (Average Brier score, 0.145; Average c-statistic, 0.870). While more work needs to be done to identify clinically relevant differences between the early and late POR outcomes, the scoring systems provided here can be used by surgeons and patients to improve the quality of care overall.

Keywords: Machine learning; Postoperative; Prediction; Reintubation.