Development of a Human Immunodeficiency Virus Risk Prediction Model Using Electronic Health Record Data From an Academic Health System in the Southern United States.

This article focuses on the development and evaluation of electronic health record (EHR)-based machine learning models to predict incident human immunodeficiency virus (HIV) diagnoses as a proxy for identifying candidates for pre-exposure prophylaxis (PrEP) in a large southern U.S. healthcare system. Using data from nearly one million patients at Duke University Health System, the study compared Extreme Gradient Boosting (XGBoost) and least absolute shrinkage and selection operator (LASSO) logistic regression models, finding that XGBoost performed better overall (AUROC 0.89) while LASSO performed better for the female-only cohort (AUROC 0.86). Key predictive variables included race, sex, male sexual partner status, and, among women, history of pelvic inflammatory disease, drug use, and tobacco use. The findings demonstrate the feasibility of using EHR-based predictive models to enhance identification of individuals, including women, who may benefit from PrEP in the southern United States, a region with high HIV incidence but low PrEP uptake.