Machine Learning Applications for Prediction of Postoperative Complications

Abstract

Machine learning has moved from a methodological curiosity to an active area of perioperative research, offering a way to exploit the dense, high-dimensional data generated in modern surgical care rather than relying on the handful of variables used by conventional risk scores. This review surveys the current state of machine learning applications for predicting postoperative complications, covering general morbidity, mortality, acute kidney injury, and delirium as the four categories with the most mature published evidence. Across these categories, models trained on automatically curated electronic health record data, continuous intraoperative vital sign streams, and even unstructured clinical text have generally outperformed conventional logistic regression-based scores on measures of discrimination. The strongest work exploits data that traditional scores cannot use at all, such as intraoperative hemodynamic waveforms for kidney injury prediction and free-text preoperative documentation for mortality risk, rather than simply re-fitting familiar preoperative variables with more complex algorithms. We conclude that external validation across institutions, consistent outcome definitions, and prospective evidence that model-guided decisions actually reduce complications remain the principal barriers separating this literature from routine clinical adoption.