Background
Effective antiretroviral therapy to maintain durable viral suppression is key to ending the HIV epidemic in the United States. We evaluated the ability of machine learning (ML) algorithms to predict people with HIV (PWH) who are at risk of falling out of care.
Methods
This was a retrospective study among all PWH in San Diego County. The study used mandatorily reported de-identified public health HIV data between 2017 and 2022. The data was partitioned temporally into two training (training and refinement datasets) and one validation dataset to accurately predict performance for predicting future observations. The outcome of interest was developing unsuppressed viral load (>200 copies/ml) during a year-long prediction window. Variables included socio-demographic data, baseline HIV infection information including stage and transmission risk, and longitudinal HIV viral load and CD4 data. HIV viral load testing patterns and trends were incorporated into the analysis. A random forest model was used to generate outcome predictions for the overall population and was stratified by race. Mediation analyses were undertaken to assess underlying causality.
Results
A total of 18,919 PWH were included in the analysis from which 13.1%, 13.0% and 12.1% had unsuppressed viral load during the observation window in the training, refinement and validation datasets respectively. The model had an area under the receiver operating characteristic (AUROC) curve of 82.2 (95%CI 79.3-85.0), a sensitivity of 33.8% (95%CI 28.6-39.0), and specificity of 96.9% (95%CI 95.7-97.2) corresponding to a positive predictive value of 55.7% (95%CI 48.7-62.8) and negative predictive value of 91.7% (95%CI 90.6-92.8). The AUROC was similar across race/ethnicities. Prior viral load testing pattern/trends such as linear trend, proportion of unsuppressed viral loads, and slope of viral load were identified as the most important variables by the model; however, they partially acted as mediators of underlying demographic (e.g., race) and HIV infection risk (e.g., injection drug use). Relative importance of variables and outcome prediction by race are shown in Figure 1.
Conclusions
ML algorithms using mandatory reported public health HIV data can predict which PWH will have future unsuppressed viral load and would be easily generalizable to other U.S. jurisdictions. Future work will assess its applicability in other settings and its clinical utility compared to existing data-to-care initiatives.