At the Conference on Retroviruses and Opportunistic Infections (CROI 2026) held in Denver this week, scientists discussed how machine learning and generative AI can be used to improve various outcomes for HIV.
Exactly how these technologies are leveraged to do this remains questionable and often perplexing. Dr. Ravi Goyal, who moderated one of the AI sessions at the University of California, San Diego, clarified some of this doubt. “We’ve been told that AI will revolutionize public health, that it will revolutionize our healthcare system. But if you’re like me, you don’t know. Maybe you don’t quite believe the hype, maybe you haven’t looked closely yet. And don’t get me wrong: Machine learning and generative AI is very impressive in demos, but it doesn’t always improve patient outcomes.”
The arrival of the AI HIV era
Machine learning refers to the use of algorithms to detect patterns in large datasets and make predictions and classifications based on them. Rather than following rules programmed by humans, these systems learn from data and identify relationships and regularities that may not be obvious to human analysts. For example, spam filters learn to distinguish between junk and legitimate messages based on characteristics of thousands of previous emails.
Dr. Joseph Hogan of Brown University in Rhode Island shared the results of a simple application of machine learning to guide outreach calls to clients at high risk of missing HIV care visits. Because treatment retention is critical to HIV outcomes and machine learning can accurately predict the risk of missed visits based on past behavior, this intervention tested how effective this type of predictive model is at improving treatment retention.
The pilot study was conducted in western Kenya, where care is provided to approximately 130,000 people living with HIV. Risk prediction models were integrated with medical record systems. Therefore, healthcare professionals can instantly see each customer’s predicted risk score and indicate the likelihood of missing their next visit. The goal here was to contact the client before their next scheduled visit and encourage attendance.
In this pilot, 27% (12,696 people) of customers were flagged as being at high risk of missing their next appointment. Of those reported, 54% received a call from the outreach team, while the remainder did not. Hogan explained that this was due to constraints such as staff capacity and broadcast time. Of those who called, 64% were reached by phone, while the remainder did not answer the phone or respond to messages.
The researchers then looked at return-to-care rates for each group. This was a very strict standard. Clients had to return to care on time based on their appointments. As expected, those who were not flagged as high risk had the highest response rate at 62%. Of the customers flagged by the machine learning algorithm, those who called and contacted had the highest return rate at 43%, those who never called had a 32% return, and those who called but never heard back had a 29% return. Somewhat interestingly, however, while successful contacts were 77% more likely to return (odds ratio 1.77, 95% confidence interval 1.55 to 2.03), calls that were unsuccessful were still 22% more likely to return than no calls at all (OR 1.22, 95% CI 1.10 to 1.35).
The machine learning model showed relatively good performance. One measure of the ability to discriminate between high- and low-risk clients had a value of 0.72, indicating an acceptable discrimination rate.
However, Hogan reminded the audience that “human behavior is very difficult to predict.” One of the aspects he raised was the need for client involvement both in model building and in subsequent interventions. For example, does the client prefer to be notified by phone? Do they have other preferences?
In another machine learning study, Dr. Peter Cairo of the International Center for Health Education and Biosecurity (CIHEB) presented findings on the ability of machines to accurately identify people at high risk of HIV infection in eastern Kenya.
She noted that current HIV test eligibility screening tools lack precision in risk stratification, which limits their ability to accurately identify those at highest risk of HIV infection.
The retrospective study was conducted from January 2024 to March 2025, analyzing electronic medical records at 58 healthcare facilities. In this example, the machine learning risk model took into account age, gender, STI history, past HIV testing, number of sexual partners, condom use, and partners’ HIV status. Using the cumulative risk score, customers were classified into different risk categories, from very high to low.
In total, approximately 40,000 customers with a median age of 27 years were offered HIV tests during this period, and 620 or 1.6% tested positive. Regarding machine predictions:
- 84 (0.5%) of 17,519 customers classified as low risk tested positive
- 164 (1.2%) of 13,155 customers classified as moderate risk tested positive
- Of 6,362 customers classified as high risk, 156 (2.5%) tested positive
- Of the 2,919 customers classified as very high risk, 216 (7.4%) tested positive.
The increase in percentage indicates that the model performed well in predicting people infected with HIV. For example, the odds of being HIV positive in the very high-risk group were 22 times higher than that in the low-risk group (OR 22.2, 95% CI 17.03-29.01).
Protective factors that reduce the likelihood of being classified as high risk include attending a maternal and child testing service, having repeated tests, and, interestingly, being a member of a key population such as sex workers. Increasing awareness of the need for HIV testing among these groups may be contributing to increased testing frequency. Kyalo said ideally this type of model could be used to inform more targeted testing efforts in the future.
The CROI audience expressed ethical concerns about the use of some of these technologies, particularly regarding confidentiality. Some researchers say they have struggled to fully and accurately anonymize data before it is processed by AI systems, but this still has to be done manually. If a client consents to their data being used by a machine learning system or generative AI, it is doubtful that the client (and perhaps the researcher) was fully aware of how this data could be stored, used, and repurposed for future uses, as AI continues to evolve rapidly and there are few robust ethical safeguards or solutions in place.
References
Hogan, J.W. Effectiveness of machine learning-based pre-visit outreach for patients at high risk of missing an HIV visit. Conference on Retroviruses and Opportunistic Infections, Denver, Poster 1076, 2026.
Cairo, P. Optimizing HIV risk prediction and case identification in eastern Kenya using machine learning. Conference on Retroviruses and Opportunistic Infections, Denver, Abstract 161, 2026.
