Abstract Body

Alliance for Public Health implements social network strategy for HIV case-finding among key populations. We developed a model to improve recruitment of undiagnosed HIV-positives in the network using machine learning (ML) algorithm.

Since 2016, 130,095 people who inject drugs and their extended risk network peers were recruited in 12 regions of Ukraine. Recruitment starts from HIV positive cases with following criteria: 14+ years, inject drugs. Participants provided with 3 coupons to invite their peers defined as an injecting or sexual partner or somebody from the social network who can be also at risk of HIV. Recruitment stops if there are two HIV-negative cases next to each other in a chain. Data on recruitment chains and participants' characteristics collected in mobile application. Additional coupons are provided to participants with certain characteristics, such as “over 10 years of injecting drugs”, “history of incarceration”, “positive HIV test result”. We implemented a ML algorithm to predict in real-time the probability of recruiting an undiagnosed HIV-positive person within onestep from the participant who receives coupons. Considering the estimated probability, recommendations on a number of additional coupons are provided.

Among participants who received coupons, 75.9%(n=35,965/47,404) recruited at least one peer and 15%(n=7,146/47,404) recruited at least one HIV-positive participant within onestep. In comparison with current recruitment algorithm ML model is 1.5-2.5 times more efficient (based on GINI index)in predicting chances of undiagnosed positive case (Fig.1). ML model with 42predictors yielded a GINI index of 34%for classification of HIV-positives and negatives. The most informative predictors of recruitment of HIV-positives included “HIV test result”, “Region”, “Years of injecting drugs”, while “Age” and “Marital Status”had the lowest contribution to prediction.

Higher level of discriminatory power (an ability to distinguish between successful and not successful recruitment)of ML model suggests that application of ML algorithm during recruitment could improve HIV-positive yield and guide HIV testing to address gap in locating undiagnosed cases. Further steps include piloting of ML algorithm with randomizing recruiters to evaluate effectiveness of ML in improving case-finding in groups connected in risk networks with high prevalence of HIV.