Identification of surgical site infections using electronic health record data

Abstract

Background: The objective of this study was to develop an algorithm for identifying surgical site infections (SSIs) using independent variables from electronic health record data and outcomes from the American College of Surgeons National Surgical Quality Improvement Program to supplement manual chart review.

Methods: We fit 3 models to data from patients undergoing operations at the University of Colorado Hospital between 2013 and 2015: a similar model reported previously in the literature, a comprehensive model with 136 possible predictors, and a combination of those. All models used a generalized linear model with a lasso penalty. Several techniques for handling imbalance in the outcome were also used: Youden's J statistic to optimize the probability cutoff and sampling techniques combined with Youden's J. The models were then tested on data from patients undergoing operations during 2016.

Results: Two hundred thirty of 6,840 patients (3.4%) had an SSI. The comprehensive model fit to the full set of training data performed the best, achieving 90% specificity, 80% sensitivity, and an area under the receiver operating characteristic curve of 0.89.

Conclusions: We identified a model that accurately identified SSIs. The framework presented can be easily implemented by other American College of Surgeons National Surgical Quality Improvement Program-participating hospitals to develop models for enhancing surveillance of SSIs.

Keywords: NSQIP; lasso; postoperative; supervised learning.