Machine Learning Targets Peer Referrals to Boost HIV Testing in South Africa

The HIV-LIFT study adds a machine-learning targeting model to an existing incentive program inside Coach Mpilo. Researchers at Harvard's Teamcore group, led by Milind Tambe, replaced an earlier complex simulation with a simpler, interpretable algorithm after field partners found the first version unusable. The model ranks existing users by predicted likelihood of successful referral and by cost-effectiveness under limited incentive budgets.

South Africa carries the world's largest HIV burden, with roughly 8 million people living with the virus. Men and adolescents remain disproportionately undiagnosed because of stigma, work schedules, and clinic access barriers. The trial therefore measures whether ML-guided nudges increase testing uptake and linkage to care compared with random or rule-based incentive allocation.

The approach builds on prior network-based HIV interventions evaluated by WHO technical staff and on earlier peer-referral trials in sub-Saharan Africa. Real-time constraints forced the team to drop theoretically optimal but non-interpretable methods, illustrating a common gap between academic optimization papers and deployable public-health tools.

Next steps include analysis of primary endpoints on referral completion rates and downstream testing yield, expected within the multiyear study window. If successful, the protocol could be adapted to other chatbot-mediated health programs facing similar resource limits.