At the Conference on Retroviruses and Opportunistic Infections (CROI 2026) held in Denver, US this week, scientists discussed how machine learning and generative AI can be used to improve various HIV outcomes.
There are lingering, and often perplexing, questions as to exactly how these technologies will be mobilised to do this. Dr Ravi Goyal, a moderator for one of the AI sessions from the University of California, San Diego, aired some of this doubt: “We’ve been told that it’s going to revolutionise public health, it’s going to revolutionise our healthcare system. But if you’re like me, I don’t know, maybe you don’t quite believe the hype, maybe you haven’t quite seen it yet. And don’t get me wrong, machine learning and generative AI are very impressive in demos and in labs, but that doesn’t mean that it always translates into better patient outcomes.”
Entering the AI HIV era
Machine learning refers to the use of algorithms to detect patterns in large datasets and make predictions or classifications based on them. Rather than following rules programmed by humans, these systems learn from data – identifying relationships and regularities that may not be obvious to human analysts. For example, a spam filter learns to distinguish junk email from legitimate messages based on the characteristics of thousands of previous emails.
Dr Joseph Hogan from Brown University in Rhode Island shared the results of a simple application of machine learning: guiding outreach calls to clients who are at a high risk of missing HIV care visits. As retention in care is essential to HIV outcomes, and machine learning can accurately predict the risk of missed visits based on past behaviour, this intervention tested how effective this type of prediction model could be at increasing retention in care.
This pilot study took place in western Kenya, where care is provided for approximately 130,000 people living with HIV. The risk prediction model was integrated with the medical record system. Thus, health workers could instantly see a predicted risk score for each client, showing the likelihood that they would miss their next visit. The aim here was to contact the client prior to their next scheduled visit to encourage attendance.
In this pilot, 27% of clients (12,696 people) were flagged as high-risk for defaulting on their next appointment. Of those flagged, 54% were called by the outreach team, while the remainder were not called. Hogan explained that this was due to constraints such as staff capacity and airtime. Of those called, 64% were reached by phone while the rest did not answer the call or respond to messages.
Researchers then considered each group’s return to care rate. This was quite a strict criterion: clients had to return to care on time, based on their appointment. As expected, those who were not flagged as high-risk had the highest return rate at 62%. Of those who were flagged by the machine learning algorithm, clients called and reached had the highest return rate at 43%, those not called at all had a return rate of 32%, while those called and not reached had a 29% return rate. However, and somewhat interestingly, even while a successful contact had a 77% higher return rate (Odds Ratio 1.77, 95% Confidence Interval 1.55-2.03), even those who were called and not reached had a 22% higher return rate than those not called at all (OR 1.22, 95% CI 1.10-1.35).
The machine learning model performed relatively well. One measure indicating its ability to discriminate between high- and low-risk clients indicated a value of 0.72, indicating an acceptable discrimination rate.
However, Hogan reminded the audience that, “predicting human behaviour is very difficult.” One aspect he raised was the need for client engagement during both model building and the subsequent interventions. For instance, do clients want to be reminded via phone call? Do they have other preferences?
In another machine learning study, Dr Peter Kyalo from the Centre for International Health, Education and Biosecurity (CIHEB) presented findings on a machine’s ability to correctly identity people at a high risk of acquiring HIV in eastern Kenya.
He noted that current HIV testing eligibility screening tools lack precision in risk categorisation – this limits their ability to correctly identify those at highest risk of getting HIV.
A retrospective study was done across 58 health facilities for the period from January 2024 to March 2025, analysing electronic medical records. In this instance, the machine learning risk model considered age, sex, STI history, prior HIV testing, number of sexual partners, condom use and partners’ HIV statuses. Using a cumulative risk score, clients were categorised in varying risk categories, ranging from very high to low.
In total, HIV testing was offered to nearly 40,000 clients with a median age of 27 over this period, with 620 or 1.6% testing positive. In terms of the machine predictions:
- 84 of the 17,519 clients classified as low risk tested positive, 0.5%
- 164 of the 13,155 clients classified as moderate risk tested positive, 1.2%
- 156 of the 6362 clients classified as high risk tested positive, 2.5%
- 216 of the 2919 clients classified as very high risk tested positive, 7.4%.
The increasing percentages indicate that the model worked well to predict those who went on to acquire HIV. For instance, the odds of HIV positivity for the very high-risk group were 22 times higher than that of the low-risk group (OR 22.2, 95% CI 17.03-29.01).
Protective factors that would lower the chances of being classified as high risk included being connected to mother-to-child testing services, being a repeat tester, and, interestingly, being a member of a key population, such as sex workers. Increased awareness of the need to test for HIV among these groups is likely a factor that contributes to more frequent testing. Kyalo noted that ideally this type of model can be used to help inform more targeted testing drives in future.
Audience members at CROI expressed ethical concerns over some uses of these technologies, especially in relation to confidentiality. While some researchers said that pains were taken to de-identify data fully and correctly before they were processed by AI systems, this still needed to be done manually. In cases where clients consented to their data being used by machine learning systems or generative AI, it’s doubtful that they – and possibly the researchers – knew the full extent of how this data can be stored, used and repurposed for future uses, as AI continues to develop swiftly, with few solid ethical safeguards and solutions in place.
References
Hogan, J.W. Effect of Machine-Learning-Guided Previsit Outreach to Patients at High Risk to Miss HIV visits. Conference on Retroviruses and Opportunistic Infections, Denver, poster 1076, 2026.
Kaylo, P. Using Machine Learning to Optimize HIV Risk Prediction and Case Identification in Eastern Kenya. Conference on Retroviruses and Opportunistic Infections, Denver, abstract 161, 2026.
