Developing machine learning models to improve cardiovascular risk prediction for people living with HIV.

Background: As life expectancy rises for people with HIV, atherosclerotic cardiovascular disease (ASCVD) has become a major contributor to morbidity. Extant risk models understate this risk, stressing the need for better models for HIV patients. Methods: We studied new ASCVD events using Veterans Health Administration data (baseline 2010-15 and follow-up through 2020). We built four machine learning (ML) models to predict CVD: K-nearest neighbors, Random Forest, Logistic Regression and Neural Network, which were compared to two general risk models: Framingham Risk Score (FRS) and Pooled Cohort Equations (PCE). ML models were trained on all Veterans and only on HIV-positive Veterans and assessed with 5-fold validation. We measured discrimination via area under the receiver operating characteristic curve (AUC) and calibration via Hosmer-Lemeshow. Results: 20,650 Veterans with HIV and 102,654 without HIV were included. HIV patients were 97% male and 51% Black, with a mean age of 52 years. Models trained on all data had better discrimination than models trained only on HIV data. Neural Network and Logistic Regression models trained on all data, and both Random Forest models, had moderately improved discrimination compared to FRS and PCE (AUC ∼0.70 for ML models vs. ∼0.65). FRS and PCE underpredicted CVD risk with observed-to-expected ratios of 2.1 and 1.7, while ML models had ratios closer to 1. Conclusions: ML models for CVD risk can enhance predictive performance in HIV, with a notable impact on underprediction. Models developed in HIV and non-HIV mixed populations have the best performance.