Development and Validation of Models for Preoperative Risk of Postoperative Infections Using Electronic Health Records Data

Abstract

Introduction: The American College of Surgeons (ACS) NSQIP and the National Health and Safety Network (NHSN) provide risk-adjusted postoperative infection rates for hundreds of hospitals across the US. However, manual review of electronic health records (EHR) is burdensome, expensive, and inefficient. The purpose of this study was to develop and validate preoperative prediction models for infectious adverse events using machine learning techniques which does not require manual chart review. Methods: We developed models for postoperative infection surveillance using EHR data linked to ACS NSQIP outcomes data. Combined with the preoperative risk models presented here, we will estimate risk-adjusted postoperative infection rates. EHR predictors included laboratory data, procedures, diagnosis phenotypes, wound class, surgeon specialties, comorbidities, and other operative variables. We used the knockoff filter to select predictors for the following outcomes: surgical site infection, urinary tract infection, sepsis/septic shock, and pneumonia. Predictors were first selected by the knockoff filter in a training dataset, then logistic regression models were fit using the selected predictors. Performance was evaluated in the test dataset. Results: Each model contained between 6-12 predictors, and achieved >66% sensitivity, >75% specificity, and 0.75-0.9 area under the curve in the test dataset. This is similar performance compared with the ACS NSQIP risk calculator and other previously published preoperative risk calculators that require user manual input. Conclusion: Preoperative risk was estimated with good performance in a large healthcare system using EHR data. These models can be combined with postoperative surveillance models to estimate risk-adjusted rates of postoperative infections at the surgeon and subspecialty levels for all operations.