Abstract
OBJECTIVE: HIV molecular transmission network typologies have previously demonstrated associations to transmission risk; however, few studies have evaluated their predictive potential in anticipating future transmission events. To assess this, we tested multiple models on statewide surveillance data from the Florida Department of Health. DESIGN: This was a retrospective, observational cohort study examining the incidence of new HIV molecular linkages within the existing molecular network of persons with HIV (PWH) in Florida. METHODS: HIV-1 molecular transmission clusters were reconstructed for PWH diagnosed in Florida from 2006 to 2017 using the HIV-TRAnsmission Cluster Engine (HIV-TRACE). A suite of machine-learning models designed to predict linkage to a new diagnosis were internally and temporally externally validated using a variety of demographic, clinical, and network-derived parameters. RESULTS: Of the 9897 individuals who received a genotype within 12 months of diagnosis during 2012-2017, 2611 (26.4%) were molecularly linked to another case within 1 year at 1.5% genetic distance. The best performing model, trained on two years of data, was high performing (area under the receiving operating curve = 0.96, sensitivity = 0.91, and specificity = 0.90) and included the following variables: age group, exposure group, node degree, betweenness, transitivity, and neighborhood. CONCLUSIONS: In the molecular network of HIV transmission in Florida, individuals' network position and connectivity predicted future molecular linkages. Machine-learned models using network typologies performed superior to models using individual data alone. These models can be used to more precisely identify subpopulations for intervention.