Risk Factors for HIV sero-conversion in a high incidence cohort of men who have sex with men and transgender women in Bangkok, Thailand

Abstract

Background: We measured Human Immunodeficiency (HIV) incidence, retention, and assessed risk factors for seroconversion among two previously unreported cohorts of men who have sex with men (MSM) and Transgender Women (TGW) in Bangkok, Thailand between 2017 and 2019.

Methods: We conducted an 18-month prospective cohort study of HIV-uninfected Thai cisgender men and TGW aged between 18 and 35 years who reported sex with men in the past six months and at least one additional risk factor for HIV infection. HIV and syphilis testing and computer-based behavioral questionnaires were administered at each visit. We utilized Poisson regression to calculate HIV incidence rates. A survival random forest model identified the most predictive risk factors for HIV sero-conversion and then used in a survival regression tree model to elucidate hazard ratios for individuals with groups of selected risk factors. Cox proportional hazards (pH) regression evaluated the strength of association between individual covariates and risk of sero-conversion.

Findings: From April 2017-October 2019, 1,184 participants were screened, 167 were found ineligible, and 1,017 enrolled. Over the 18-month study, visit retention was 93·4% (95% CI 91·6%-94·8%) and HIV incidence was 3·73 per 100 person-years (95% CI 2·79-5·87). Utilizing survival regression tree modeling, those who were 18-20 years of age, reported sexual attraction to mostly or only men, and had five or more lifetime sexual partners were 4·9 times more likely to seroconvert compared to other cohort participants. Factors associated with HIV incidence utilizing Cox pH regression included sexual attraction to mostly or only men (adjusted hazard ratio (aHR) 14·9 (95% CI 20·1-107·9), younger age (18-19 years, aHR 10·88 (95% CI 4·12-28·7), five or greater lifetime sexual partners (aHR 2·0, 95%CI 1·1-3·6), inconsistent condom use with casual partners (aHR 2·43, 95% CI 1·3-4·5), and prior HIV testing (adjusted HR 2·0, 95% CI 1·1-3·5).

Interpretation: Interpretation HIV incidence remains high among Bangkok-based MSM and TGW. These key populations expressed high interest in participating in efficacy evaluation of future prevention strategies and had high retention in this 18 month study.

Funding: Funding US National Institute of Allergy and Infectious Diseases (NIAID), Division of AIDS Interagency Agreements (DAIDS) and U.S. Department of the Army.

Fig. 1

Participant Flow Diagram. This participant flow diagram displays the progression of participants through screening, enrollment, follow-up and analysis. A total of 1184 potential participants were screened and 1017 were eligible and enrolled. Twenty-one participants were lost to follow-up and 55 participants withdrew from the study. In total, 1017 participants were included in the study analysis.

Fig. 2

Retention rate by visit and site. Retention was defined as attending and completing the study visit. The x-axis shows the visit number and corresponding month of the visit. The Y axis shows the percentage of visits completed for all eligible cohort participants at that visit. Overall retention was 93·4% (95% CI 91·6%−94·8%) and did not differ significantly between Mahidol University (VTC) and Royal Thai Army Clinical Research center (RTA).

Fig. 3

Determination of the optimal number of variables to include in survival tree model. In the plot of the out of bag (OOB) error rate versus model complexity (K), the dots reflect the OOB error for tree models built with K candidate variables and the blue line (along with grey confidence interval) is a LOWESS smoother which attempts to smooth out the randomness and estimate the mean OOB error rate curve.

Fig. 4

Risk Factors for HIV Sero-conversion. Figure 4a ranks the top 20 random forest variables in terms of Variable Importance (VIMP). VIMP measures the change in the predictiveness of the random forest model when the variable is randomly permuted; large scores indicate importance. Figure 4b ranks the top 20 variables in terms of minimal depth. The minimal depth indicates the average depth of the variable among all survival trees in the forest; smaller values of depth indicate greater importance.

Fig. 4

Risk Factors for HIV Sero-conversion. Figure 4a ranks the top 20 random forest variables in terms of Variable Importance (VIMP). VIMP measures the change in the predictiveness of the random forest model when the variable is randomly permuted; large scores indicate importance. Figure 4b ranks the top 20 variables in terms of minimal depth. The minimal depth indicates the average depth of the variable among all survival trees in the forest; smaller values of depth indicate greater importance.

Fig. 5

Survival tree model with hazard ratios for risk of HIV sero-conversion. Survival tree model with hazard ratios for risk of HIV sero-conversion utilizing top seven variables from the union of depth and variable importance lists. The “splitting” decisions made in survival regression trees are based on the log-rank statistic and each split in the tree represents a log rank test with a statistically significant difference. Thus, the first (dichotomous) variable that a tree model splits by is arguably the most important as it makes the largest difference in the hazard ratio. Below the root node at the top of the survival tree model, if the criterion below the node equals yes, the tree splits to the left; if it equals no, it splits to the right. Within each node, the first number is the hazard ratio relative to the overall study population. The second row in each node describes the number of people who seroconverted by the end of the study out of the number of people at risk in that node subgroup. The last row depicts the percentage of the total population represented by the node. The color of the node depicts the strength of the hazard ratio, with higher hazard ratios having darker shades.

Fig. 6

Kaplan-Meier Curves for Risk of HIV Sero-conversion by Covariate. This figure shows the Kaplan-Meier Curves for Risk of HIV Sero-conversion by Covariate. Panel A, Age. Panel B, Sexual Attraction Men. Panel C, Number of Lifetime Sexual Partners. Panel D, Frequency of Condom Use during Receptive Anal Sex with Casual Male Partners. Panel E. History of Prior HIV Testing; Panel F. History of Receptive Anal Intercourse in the Past Twelve Months. Note: to facilitate interpretation of distinctions between subgroups, we chose an upper Y-axis limit of 0.02, with the exception of Age, which has an upper limit of 0.05.

Fig. 6

Kaplan-Meier Curves for Risk of HIV Sero-conversion by Covariate. This figure shows the Kaplan-Meier Curves for Risk of HIV Sero-conversion by Covariate. Panel A, Age. Panel B, Sexual Attraction Men. Panel C, Number of Lifetime Sexual Partners. Panel D, Frequency of Condom Use during Receptive Anal Sex with Casual Male Partners. Panel E. History of Prior HIV Testing; Panel F. History of Receptive Anal Intercourse in the Past Twelve Months. Note: to facilitate interpretation of distinctions between subgroups, we chose an upper Y-axis limit of 0.02, with the exception of Age, which has an upper limit of 0.05.