Development and Validation of Models for Preoperative Prediction of Risk and Postoperative Detection of Non-Infectious Complications Using Interpretable Machine Learning and Electronic Health Record Data

Objective: To apply interpretable machine learning methodology to electronic health record (EHR) data to develop models for preoperative risk estimation and postoperative detection of non-infectious postoperative complications. Summary Background Data: We previously developed preoperative risk and postoperative detection models for surveillance of postoperative infections. The purpose of the present study was to develop and validate similar models for the non-infectious complications of the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP). Methods: Preoperative and postoperative EHR data from five hospitals across one healthcare system (University of Colorado Health), 2013-2019, including diagnoses, procedures, operative variables, patient characteristics, and medications were obtained. Lasso and the knockoff filter were used to perform controlled variable selection to develop preoperative risk models and postoperative detection models of 30-day non-infectious outcomes of mortality, overall morbidity, bleeding, cardiac, pulmonary, renal, and venous thromboembolism morbidity, non-home discharge, and unplanned readmission. Results: Among 30,639 patients included, postoperative complication rates for each outcome ranged from 0.1% (stroke) to 10.4% (overall morbidity). Area under the receiver operating characteristic curve for preoperative risk models ranged from 0.68-0.91 and from 0.92-0.97 for postoperative detection models. Between 6-22 predictor variables were included in each model. Conclusions: We developed parsimonious models for estimating risk of and detection of postoperative non-infectious complications. Our models showed good to excellent performance suggesting that these models could be used to augment manual surveillance.