Harvard Tests AI to Boost HIV Testing Referrals in South Africa

A field trial in KwaZulu-Natal

According to Harvard SEAS, a research team led by Milind Tambe, the Gordon McKay Professor of Computer Science, is testing whether AI can enhance a digital health intervention aimed at generating peer referrals for HIV services. The focus is on populations that have traditionally been harder to reach. The field trial is now underway in the KwaZulu-Natal province of South Africa, which the source says has about 8 million people living with HIV, the highest of any country in the world.

The trial is part of HIV-LIFT, short for Leveraging Innovation and Friendships to Improve HIV Testing. The Gates Foundation funds the initiative, which includes Wits University in South Africa, Harvard's Teamcore research group led by Tambe, partners at the World Health Organization, and others. Alastair van Heerden of Wits University is the principal investigator of the multi-year study. It targets groups such as men and adolescents, who are less likely to seek testing and more likely to remain undiagnosed or to attend clinics.

How the AI prioritization works

The trial centers on Coach Mpilo, a free, confidential AI health coach and chatbot available through WhatsApp and designed specifically for South Africans. It offers information and support on HIV testing and treatment, as well as other health issues. It has more than 150,000 enrolled users, but only about 15% are men. Men are historically less likely to seek HIV testing and treatment, according to the source, because of factors like social stigma or long working hours.

The app already includes a financial incentive for existing users who refer others who might need HIV services. The trial asks a practical question: when a program can afford to encourage only a limited number of people to recruit contacts, whom should it prioritize?

The Harvard team built a machine-learning and optimization approach to target existing users who are then nudged with an incentive. Their model estimates which users are most likely to refer people in underserved groups, including men and adolescents. The planned study compares two groups. In one, eligible users are selected for outreach at random. In the other, an AI filter further narrows the pool of incentive-eligible users. The AI system supplies a fixed priority list developed in advance from data collected in earlier incentive campaigns.

In retrospective evaluations of approximately 16,000 historical nudges collected over two months, the approach was estimated to generate more successful referrals than any other strategy, according to the researchers. The work is intended to supplement, not replace, clinics, health workers or the social relationships behind peer referrals.

Transparency over complexity

The team says real-world constraints shaped the design. Graduate student Akseli Kangaslahti said the group first used a more complex simulation algorithm, but its results were not easily interpretable for regular users. "I think we learned a lot from that work, even if we weren't able to directly employ the algorithms that we proposed in them," Kangaslahti said.

Postdoctoral fellow Davin Choo, who has led the algorithm development, said: "There's no point in having fancy things look good on paper but cannot be used in practice." Tambe said partners proposed policy restrictions to ensure fairness and equal treatment for users with identical profiles, while keeping the framework simple, transparent and practical.

What it could mean next

With the algorithms generating the target groups, the researchers will determine whether the priority-group approach could be implemented permanently within Coach Mpilo or potentially in other applications. The source gives no results from the live trial yet.

The project builds on more than a decade of Teamcore research into AI and social networks for HIV prevention. Tambe began focusing on HIV prevention in 2013, initially working with youth experiencing homelessness in Los Angeles. "We hope that our work will ultimately contribute to some improvement in how HIV-affected people may be cared for, or how people can get access to HIV information," Tambe said. The source notes that his separate Google Research AI for Social Good role is not part of HIV-LIFT.