South African trial tests AI selection for HIV peer referrals

A trial in KwaZulu-Natal, South Africa, is testing whether AI can improve HIV outreach by choosing which existing service users to ask for referrals. HIV-LIFT compares a fixed model-generated priority list with random selection. Its intended beneficiaries include men and adolescents who are harder to reach through conventional outreach.[1]

The work uses Coach Mpilo, an established WhatsApp-based AI coach offering free, confidential support around HIV testing, treatment and other health questions. Harvard researchers described the collaboration with the University of the Witwatersrand and health partners on 20 October. Their new task is to improve referral selection within that service, rather than to evaluate a new diagnostic model.[1]

According to Harvard’s account, Coach Mpilo has enrolled more than 150,000 users, around 15% of them men. That imbalance helps define the outreach challenge. Existing users may be able to introduce the service through relationships it cannot easily reach with a general campaign.[1]

Allocating a limited outreach budget

The programme has limited resources for encouraging users to refer friends and relatives. The model estimates which recipients of a nudge are likely to produce successful referrals, then uses that information to establish priorities. This makes the decision unit specific: whom to contact with an invitation, given a finite outreach budget.[1]

Researchers analysed roughly 16,000 historical nudges sent over two months. Their retrospective comparison suggested that model-based selection could improve referral yield. The current randomised trial tests whether that advantage appears when the selection policy actually changes who receives an invitation. The published account does not yet report a trial result for completed HIV tests or treatment uptake.[1]

Historical data can help identify a promising policy, but it cannot reproduce every consequence of putting that policy into use. People selected repeatedly may respond differently, and users who were easy to reach in an earlier campaign may differ from those needed for the next. Random assignment supplies a direct comparison under the new campaign’s conditions.

Partners changed the design

The team initially explored a more complex simulation approach, but community partners needed rules they could understand and explain. The resulting design uses a fixed priority policy and treats profiles judged equivalent in the same way. Those requirements shaped the technical solution rather than being appended after the model had been selected.[1]

That makes the fairness question concrete. Maximising the number of referrals could favour people already well connected to the service. A programme intended to reach underserved groups also needs to examine who arrives through those referrals. Total enrolments and access for the intended groups are related outcomes, but they are not interchangeable.

The evaluation should also preserve the distinction between contact and care. An invitation, an enrolment and a completed clinical step each describe a different part of the pathway. Understanding where referrals help—and where people still drop out—would give programme staff more useful information than a single aggregate success rate.

If the trial supports the policy, its contribution would be a practical way to spend outreach resources more effectively while retaining rules that local partners can explain. The immediate milestone is that a selection method developed from past records is being tested prospectively in the service where it would be used.[1]

Report an error or suggest a correction