Predicting individual-specific HIV survival functions - motivation, implementation, and potential applications.

Abstract

Immense progress has been made in the development and delivery of HIV prevention interventions worldwide, and that is reflected in the trends of global HIV epidemic metrics. However, such a wide perspective overlooks relapses in several regions and numerous concentrated epidemics among certain subpopulations. Risk of HIV varies substantially across geographic, temporal, and social space. That, coupled with looming budget shortfalls, has motivated calls to optimize the delivery of HIV prevention interventions. One strategy is to provide interventions to those most at risk. Typically performed through community-level geographic prioritization or social stratification within an area, advances in data generation and analysis methods have enabled individual-specific risk assessments with rich interpretations. Among these are models that predict a patient-specific survival function. Here we fit and evaluate three such individual survival models, Accelerated Failure Time Weibull Regression (AFTweib), Cox Proportional Hazards with the Kalbfleisch and Prentice estimator and Elastic Net regularization (CoxKPEN), and Random Survival Forests (RSF) on data from the dapivirine ring ASPIRE trial. Evaluated on concordance, l1-loss, integrated Brier score, and visually, the fitted models demonstrate a moderate ability to rank survival times correctly but are poor predictors of individual survival time. Highly predictive features mirror those found previously. The visual nature of the prediction may enhance strategic risk communication. Future patient-specific HIV predictive models may perform extremely well. This would present risks. Data acquisition needs may be intensive and repetitive, with implications for feasibility and opportunity cost. Privacy, equity, and righteousness should be proactively and multidimensionally safeguarded.