Could referral networks help target HIV resources without mapping every contact?

Two networks, different meanings

Network-based HIV interventions face a practical obstacle: the connections relevant to transmission are difficult and expensive to document. Peer referral programmes offer a more accessible record. Participants invite people they know, who then invite others, potentially through digital channels.

But a referral network is not a transmission-risk network. An invitation may follow a social relationship without sexual contact or shared injection equipment. And because each participant accepts only one recruitment referral, connections between people already enrolled disappear from the referral record.

In an arXiv preprint, Akseli Kangaslahti, Milind Tambe, Alastair van Heerden and Cheryl Johnson introduce R2C-Agent, a system designed to plan interventions despite those gaps. The work reports retrospective simulations, not a clinical trial or demonstrated reductions in real-world HIV infections.

Reconstruct possibilities, not certainty

The system’s central component, Referral-to-Contact Diffusion, learns to reverse a process that transforms a contact network into a referral network.

During training, the researchers progressively replace contact connections with referral connections, following the recorded order in which participants were prompted to recruit peers. A small graph neural network learns to reverse these changes step by step. Its inputs include referral connections, HIV status and selected participant characteristics.

At inference, the model starts with a referral network and samples a possible underlying contact network. Repeating this produces multiple plausible networks rather than a single definitive reconstruction.

R2C-Agent then uses existing intervention-planning algorithms to choose resource recipients across these possibilities. The distinction matters: a decision that works well on the most likely network might perform poorly on another plausible one. Planning across samples aims to account for that uncertainty.

These inferred connections are model outputs, not evidence that particular individuals share transmission-risk relationships.

What the researchers tested

The evaluation used an anonymized network extracted from an existing HIV dataset: 1,386 people, including 248 living with HIV, with 3,793 recorded connections. Of those connections, 1,381 were classified as contact ties involving sex or needle sharing; the remainder were social ties.

Crucially, the referral networks were simulated, not collected alongside the contact network in a real referral programme. Participants in the simulation recruited along all their recorded social and contact connections, until everyone was enrolled.

The researchers divided the data into two halves while keeping connected components intact. Each half served once as training data and once as test data, leaving roughly 693 people for training in each split. Multiple simulated referral networks captured variation in recruitment.

They tested two resource-allocation tasks: intensive treatment support for people living with HIV, and pre-exposure prophylaxis, or PrEP, for people without HIV. R2C-Agent supplied 50 sampled contact networks to the planners. Decisions were evaluated through separate simulations of spread on the recorded contact network.

Better decisions, not better maps

The main metric was total regret: the gap in simulated new infections relative to a planner given the true contact network, integrated across intervention budgets. Lower values were better. This is not a direct percentage reduction in infections.

For intensive treatment, R2C-Agent scored 53.8, compared with 119.2 for the strongest non-omniscient baseline, which prioritized people with the most referral connections. For PrEP, it scored 35.8, compared with 50.3 for the best baseline, which treated referral connections as contact connections. Those represent regret reductions of 54.9% and 28.8%, respectively.

Yet the model’s sampled networks were less accurate by average edge-edit distance than simply using the referral network unchanged. The result supports the paper’s central distinction: the most accurate-looking reconstruction need not support the best decisions.

In another experiment, reducing the number of sampled networks from 50 to one—while keeping the total number of planning spread simulations fixed—raised intensive-treatment regret from 53.8 to 240.4. Exploring structural uncertainty mattered substantially in this setup.

The assumptions still matter

The spread model assigned every contact the same fixed transmission probability of 0.75. Treatment recipients were assumed to achieve complete suppression of onward transmission, while PrEP recipients were assumed fully protected. All initially HIV-positive participants were assumed to be experiencing first-line treatment failure. These are experimental assumptions, not measured clinical outcomes.

The study also relied on one dataset and recruitment assumptions that may not hold in practice. Adding synthetic missing connections tested some departures from those assumptions, but did not establish performance in real referral programmes.

The authors propose local calibration before deployment and expert-led decision support rather than autonomous allocation. They also identify an equity risk: minimizing average infections could divert resources away from smaller, isolated groups. Better simulated targeting is promising; deciding whether it is accurate, fair and safe in practice remains a separate challenge.