Developing HIV risk prediction tools in four African settings.

OBJECTIVE:HIV risk prediction tools are a critical component of efforts to end the HIV pandemic. We aimed to create and validate tools for identifying individuals at highest risk of prevalent and incident HIV in an African setting. METHODS:We used Logistic regression and Poisson regression to determine risk factors for HIV prevalence and incidence in a multi-country HIV vaccine trial preparedness cohort study among individuals at high risk of HIV, and used the identified factors to create and validate tools that predict HIV risk. We also assessed the performance of the VOICE risk score in predicting HIV incidence among women in the cohort. RESULTS:The prevalent HIV prediction tool created had good predictive ability [area under the curve (AUC) = 0.70, 95% CI 0.66-0.74]. It included the following participant variables: age, sex, recreational drug use, unprotected male-to-male anal sex, a sexual partner who had other partners, transactional sex and having a partner who was a long-distance truck driver/miner. It was not possible to create a valid HIV incidence prediction tool. Participants with high VOICE risk scores (≥7) had slightly higher HIV incidence but this tool performed poorly within our study (AUC = 0.58, 95% CI 0.51-0.64: Harrell's concordance index = 0.59). CONCLUSION:We created a prevalent HIV prediction tool that could be used to increase efficiency in diagnosis of HIV and linkage to care in sub-Saharan Africa. Existing incident HIV prediction tools may need modification to include context-specific predictors such as calendar period, participant occupation, study site, before adoption in settings different from those in which they were developed.