We used data from the completed Botswana Combination Prevention Project (BCPP), also known as Ya Tsie [16, 17] (NCT01965470). BCPP was a pair-matched, cluster-randomized trial in communities across Botswana that tested whether a combination of measures in the intervention communities reduced population-level cumulative incidence of HIV infection [17]. The intervention included intensive HIV testing campaigns, rapid universal ART initiation with active linkage to care, and enhanced voluntary male circumcision services. The trial was conducted from 2013–2018 in 30 rural and peri-urban communities (15 intervention, 15 standard of care) across Botswana. BCPP enrolled a prospective cohort of all consenting persons aged 16–64 years residing in a random 20% sample of households in each of the 30 communities. A baseline participant survey collected extensive sociodemographic information, self-reported HIV risk behavior, HIV testing and treatment history, and health history. All participants without documentation of a prior positive HIV test underwent HIV testing. We followed each cohort participant prospectively at approximately 12 and 29 months after enrollment, and re-tested HIV-negative persons for HIV at each visit.

Participants 16–18 years of age provided written assent (with parent/guardian written permission). Adult participants provided written informed consent. The study was approved by the Botswana Health Research and Development Committee and US Centers for Disease Control and Prevention IRB.

We used information on individual HIV infection status over time in the HIV incidence cohort to build an interpretable prediction model for predictors of HIV seroconversion based on 110 potential predictors collected in the baseline household survey in the study (Supplemental Table S1). We excluded persons documented as living with HIV at baseline from this analysis. Missing data were processed by multiple imputation using missForest [18] with maximum number of iterations of 10 and 50 trees to grow in each forest. We stacked 5-fold imputed datasets together to build the random survival forest model, resulting in 25,995 females and 16,760 males included in the machine learning analysis.

We then considered decision trees and random forest to build risk scores, using randomForestSRC [19] and following rpart function in the rpart [20] R package to grow our decision tree model. A natural risk score was defined as the estimated probability of HIV seroconversion during the follow-up period. We used a two-step tree-based algorithm to build our prediction model. The first step was to run a random survival forest to obtain the variable importance of each covariate. We first screened the variables by selecting the top-K variables (10% of the variables with highest variable importance for predicting the risk of HIV acquisition). The variable importance is defined as the difference between prediction error when a variable is “permuted” versus the prediction error otherwise [21]. The second step was to build a classification tree based on the top-K important variables to predict the risk score. The obtained tree is interpretable and provides a flowchart-like structure in which each internal node represents a homogeneous group.

We stratified analyses by sex, as the applicability and importance of some potential HIV incidence predictor variables differs by sex (e.g., male circumcision is relevant only to men, and hormonal contraception only to women). We defined the risk of an individual instance as the percentage of participants within the same terminal node who acquired HIV.

We then performed univariate logistic regression to further validate our tree model, using glm function [22], including variables that were identified by our tree models (Table 1, Supplemental Tables S2 and S3). The regression analysis was done using the full dataset.

We enrolled 12,610 individuals in the prospective cohort (in randomly selected households from the 30 communities): 3,596 (29%) people with HIV and 8,974 (71%) people without HIV at enrollment (the HIV incidence cohort). Most HIV incidence cohort participants were female (60%) with median age 29.6 years (IQR 22.5, 43.2). Retention of the HIV incidence cohort was high: 8551 (95%) of the 8974 HIV-negative participants were-tested for HIV at 12 and/or 29 months after enrollment [16].

During the median 29 months of follow-up of the HIV incidence cohort, 147 HIV seroconversions were observed (i.e. 1.72% of the HIV incidence cohort members seroconverted during follow-up). Annual HIV incidence was lower in the intervention (0.59%) than in the standard of care arm (0.92%), and higher in females (0.71%) than in males (0.40%) [16].

The top 11 (10%) variables, ranked by variable importance, are shown in Figures 1A and 1B for females and males, respectively.

For females, the top 11 variables that most strongly predicted HIV acquisition risk were: education level, age category, alcohol use during last sexual encounter, community HIV prevalence, use of injectable hormonal contraception, type of relationship with 2^nd most recent partner in past 12 months, energy source, frequency of alcohol use in past 12 months, marital status, perception about whether people work together to solve neighborhood problems, and frequency of sex in last 3 months with most recent sexual partner (Figure 1A).

For males, the top 11 variables that most strongly predicted HIV acquisition risk were: type of relationship with most recent sexual partner, community HIV prevalence, education level, type of relationship with 3rd most recent sexual partner in past 12 months, region of country, frequency of sex in past 3 months with second most recent partner, age of most recent partner, reason for most recent HIV test, whether second most recent partner lived within community, age, and timing of most recent HIV test (Figure 1B).

Figures 2A and 2B present the tree models for females and males, respectively. The decimal numbers refer to the estimated probability of acquiring HIV over the follow-up period in the node, and percentages refer to the percentage of the population that is represented in that node. For example, in Figure 2A, the root node includes 100% of female participants, and the estimated probability of overall HIV seroconversion in females is 2.3%.

For females, a combination of using injectable hormonal contraception, frequency of sex with the most recent partner in the past 3 months, and community HIV prevalence of 29% or higher were the most important variables for predicting risk of HIV acquisition.

In the tree model for females, the estimated probability of acquiring HIV was 5.1% for females using injectable hormonal contraception (who comprised 10% of female participants) (Figure 2A). Among the 90% of women who did not use injectable hormonal contraceptives, who had a lower risk of HIV, the most important variables identified were energy source (not gas or electricity) and being less than 35 years of age (Supplemental Figure S1). Among females using injectable hormonal contraception who reported having sex less than 1.5 times in the past 3 months with their most recent sexual partner, their estimated probability of acquiring HIV was 13%. Among females using injectable hormonal contraception who reported having sex less than 1.5 times in the past 3 months with their most recent sexual partner and who lived in a community with HIV prevalence of at least 29%, their estimated probability of acquiring HIV was 34% over the 29-month study period (which translates to approximately 14.1% annual HIV incidence). It is important to note that this latter group that had all 3 of these characteristics comprised only 0.28% of female participants.

For males, reporting that they were in a non-long-term relationship with their most recent partner and living in a community with HIV prevalence >34% were the most important variables for predicting HIV risk.

In the male group tree model, the estimated probability of acquiring HIV was 1.7% for males reporting that they were in a non-long-term relationship with their most recent partner (Figure 2B). Among males reporting that they were in a non-long-term relationship with their most recent partner, their estimated probability of acquiring HIV during the 29-month follow-up period was 5.1% if their community HIV prevalence was at least 34% (approximate annual incidence 2.1%).