Men's Health

AI Shows Predictive Utility in HIV Care but Lacks Real-World Clinical

AI Shows Predictive Utility in HIV Care but Lacks Real-World Clinical

A comprehensive systematic review synthesizing research on artificial intelligence applications in human immunodeficiency virus care for men who have sex with men has determined that while computational models demonstrate promising predictive performance, empirical evidence confirming real-world improvements in clinical or population-level outcomes remains limited. The synthesis, published in Discover Artificial Intelligence, evaluated ten peer-reviewed studies published between 2015 and 2025 across diverse global settings. The international research team examined the application of machine learning, deep learning, network-based architectures, and conversational agents across selected stages of the HIV care continuum. Across the analyzed literature, researchers identified substantial diagnostic and prognostic capabilities for tasks such as HIV status classification, pre-exposure prophylaxis adherence tracking, and clinic attendance forecasting. However, the authors emphasized that high technical accuracy within specific historical datasets does not establish generalizable clinical utility or documented improvements in patient health outcomes.

The clinical motivation for evaluating computational tools arises from persistent global disparities in HIV transmission, prevention, and treatment engagement among men who have sex with men. In the epidemiological context established by the review, men who have sex with men continue to bear an elevated burden of infection globally, with estimated relative acquisition risks twenty-three to twenty-six times higher than in the general population. Worldwide median HIV prevalence among gay men and other men who have sex with men reached 7.6 percent across eighty reporting countries, with transmission sustained by high-probability biological factors alongside dense sexual networks. Despite international prevention initiatives, service coverage remains suboptimal; global median data indicate that only approximately 40 percent of men who have sex with men access at least two combination prevention options, while antiretroviral therapy coverage among those living with HIV exhibits a median of 65 percent. Disparities in pre-exposure prophylaxis uptake, regular testing frequency, linkage to medical care, and continuous treatment adherence are compounded across diverse settings by pervasive stigma, privacy concerns, criminalization, and structural barriers within traditional healthcare delivery systems.

To investigate how computational methodologies are deployed against these healthcare challenges, the authors registered their protocol in the International Prospective Register of Systematic Reviews under registration number CRD420261292658 and conducted the evaluation in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 guidelines. Systematic literature searches across the PubMed and Scopus databases were completed on December 15, 2025, using structured search strings combining terms for artificial intelligence, machine learning, neural networks, chatbots, and predictive models with terms specific to HIV and men who have sex with men. Eligible publications were restricted to peer-reviewed English-language empirical studies that evaluated data-driven computational algorithms rather than purely rule-based systems and reported measurable outcomes across the HIV care continuum. Two reviewers independently screened titles, abstracts, and full texts, resolving discrepancies through consensus and formal third-party adjudication, and utilized backward citation tracking to supplement database queries.

The review included ten studies conducted across six

The review included ten studies conducted across six countries: China, the United States, Australia, Malaysia, Zimbabwe, and Switzerland. Participant sample sizes within the analyzed cohort ranged from 14 individuals in small digital feasibility assessments to 6,354 individuals in large epidemiological datasets. Methodological quality and risk of bias were independently evaluated using the Joanna Briggs Institute Critical Appraisal Tool and the Mixed Methods Appraisal Tool, applying review-defined operational thresholds based on the percentage of satisfied criteria. Five of the ten included studies achieved scores of 80 percent or higher and were classified as having a low risk of bias, four studies met 60 to 79 percent of applicable criteria and were rated as moderate risk, and one study satisfied less than 60 percent of criteria, designating it as high risk of bias. Because of substantial methodological, algorithmic, and outcome heterogeneity, statistical meta-analysis was not feasible, prompting the investigators to conduct a narrative synthesis.

Computational techniques evaluated across the studies were categorized into four primary technical frameworks: conventional machine learning algorithms, deep learning neural networks, network-based models, and interactive digital interventions. Conventional machine learning approaches encompassed supervised models such as logistic regression, decision trees, support vector machines, random forests, gradient-boosting machines, and extreme gradient boosting. These algorithms were trained on behavioral surveys, clinical records, and administrative data to execute targeted tasks, including identifying candidates for pre-exposure prophylaxis, stratifying sexual risk, identifying social network influencers for prevention programs, and forecasting whether patients would attend scheduled clinic visits following automated reminder messages.

Deep learning and advanced graph-based learning frameworks showed notable classification performance within dataset-constrained evaluations. In Bulawayo and Harare, Zimbabwe, investigators evaluated a recurrent neural network designed for HIV status prediction that achieved an accuracy metric of 0.98. The systematic review authors noted that this high metric reflects isolated classification performance within a specific training dataset rather than evidence of 98 percent real-world clinical efficacy or transportable diagnostic precision. In another technical implementation, graph convolutional networks were employed to integrate complex social network structures into risk prediction pipelines, successfully identifying previously unrecognized HIV infections among young men who have sex with men by mapping network connectivity patterns alongside individual risk profiles.

In contrast to predictive modeling, evidence evaluating real-world

In contrast to predictive modeling, evidence evaluating real-world interactive digital interventions remained considerably more circumscribed. Chatbot-based conversational systems, such as an artificial intelligence interface assessed in a mixed-methods investigation in Malaysia, were primarily evaluated through usability metrics, demonstrating preliminary technical feasibility and user acceptability for promoting HIV testing and pre-exposure prophylaxis consultation. The study did not demonstrate measurable changes in long-term clinical endpoints or confirmed diagnostic increases. Similarly, mobile smartphone-based machine learning tools evaluated in the United States successfully identified momentary situational and contextual factors associated with condomless anal intercourse using real-time predictive modeling, yet the research design did not measure sustained subsequent reductions in sexual risk behavior or incident HIV infections.

Evaluations of algorithmic secondary distribution strategies for HIV self-testing kits revealed mixed empirical results when compared directly against conventional public health outreach. In China, machine learning models were deployed across social networks to identify key peer influencers capable of distributing HIV self-testing kits within their communities. While one randomized trial reported positive behavioral and testing-related engagement outcomes, a companion quasi-experimental investigation found no statistically significant improvement in identifying first-time testers when comparing algorithmic selection against standard empirical outreach approaches. These comparative discrepancies demonstrated that predictive algorithmic optimization does not uniformly outperform human-delivered or standard empirical public health strategies.

The authors highlighted significant methodological constraints across the evidence base that preclude pooled conclusions regarding algorithmic superiority over conventional biostatistical methods. While selected machine learning and deep learning models achieved higher discrimination metrics than standard multivariable logistic regression in specific datasets, the studies varied widely in validation protocols, feature sets, and performance metrics, including sensitivity, specificity, area under the receiver operating characteristic curve, and F1 scores. The authors emphasized that high accuracy metrics are frequently confounded by underlying class imbalance and outcome prevalence, and cannot serve as proxies for model calibration, generalizability, or downstream clinical utility. Furthermore, reliance on self-reported behavioral inputs in several studies introduced potential recall and reporting biases, while small sample sizes in digital interventions constrained statistical precision.

Ethical, legal, and operational considerations form an essential

Ethical, legal, and operational considerations form an essential dimension of deploying artificial intelligence among key populations experiencing social marginalization. The review underscored that men who have sex with men face ongoing discrimination, social stigma, and legal criminalization in numerous international jurisdictions, making the handling of digital footprint, behavioral, and network data exceptionally sensitive. Unregulated data aggregation introduces severe risks of privacy breaches, algorithmic surveillance, and unintended disclosure. Additionally, algorithmic training on unrepresentative cohorts risks propagating systemic biases that could misallocate healthcare resources or restrict access to vital preventive care. The review concluded that future clinical integration requires transparent governance, participatory design involving local communities, rigorous external validation across diverse healthcare environments, and prospective randomized trials measuring hard endpoints such as viral suppression, treatment retention, and verified reductions in HIV transmission.

Subject of Research: Systematic review of artificial intelligence applications across the HIV care continuum for men who have sex with men

Article Title: Artificial intelligence applications in HIV Care domains among men who have sex with men (MSM): a systematic review

Article References: Okesanya, O

Article References: Okesanya, O. J., Adigun, O. A., Hassan, M. M., Adebayo, U. O., Oso, T. A., Markus, M., Ephraim, O. E., Lamem, M. F. H., Ayelaagbe, O. B., Obadeyi, K. B., Agboola, A. A., & Prisno III, D. E. L. (2026). Artificial intelligence applications in HIV Care domains among men who have sex with men (MSM): a systematic review. Discover Artificial Intelligence, 6(1), Article 1910. https://doi.org/10.1007/s44163-026-02801-z

DOI:10.1007/s44163-026-02801-z

Keywords: artificial intelligence, machine learning, deep learning, HIV care continuum, men who have sex with men, predictive modeling, pre-exposure prophylaxis, HIV risk prediction, secondary distribution, HIV self-testing, recurrent neural networks, graph convolutional networks, chatbots, systematic review, PRISMA 2020, clinical effectiveness, health informatics

News. AI Shows Predictive Utility in HIV Care but Lacks Real-World Clinical Evidence. Scienmag

Leave a Reply

Your email address will not be published. Required fields are marked *