Machine Learning in HIV Care and Antiretroviral Therapy: Systematic Review

Background: Artificial intelligence (AI) is expanding across various medical fields, with machine learning (ML) being increasingly used to enhance patient management in diagnosis, prevention, and therapeutic care. Objective: This study aims to provide an overview of ML applications in HIV care, focusing on real clinical data to improve health care for people living with HIV and on antiretroviral therapy, while highlighting unexplored areas. Methods: Following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 reporting guidelines, we analyzed four databases: PubMed, Embase, IEEE, and Web of Science until August 31, 2024. The keywords used were: “Machine Learning,” “HIV,” and “Antiretroviral Therapy.” We excluded from this review studies (1) that were not directly focused on HIV or those that did not apply ML to real clinical data, (2) that focused on pre-exposure prophylaxis, (3) studies involving in silico antiretroviral drug development, and (4) studies on the biological mechanisms underlying HIV diagnosis. Three experts (TB, MBVR, and JLR) screened each article independently. Results: Overall, 476 studies were identified, and after eligibility assessment, 98 were finally analyzed in detail. Three experts (TB, MBVR, and JLR) identified 6 major categories of ML applications used in the clinical field of HIV: consideration of comorbidities for people living with HIV, predicting drug resistance of the virus, monitoring HIV infection itself, predicting treatment outcomes for people living with HIV, treatment adherence for people living with HIV, and treatment recommendation for clinicians. Random forests emerged as the most used algorithm with 17.49% (43/247), proving effective in identifying biomarkers of metabolic syndrome, genetic features of the HIV envelope, and predicting neurocognitive impairment. Random forests model has several advantages: (1) handle linear, nonlinear data, and missing data, (2) reduce overfitting compared to single trees, (3) robust to noise and outliers, (4) provide feature importance measures, and (5) good generalization ability. Support vector machines demonstrated strong abilities in analyzing the associations between HIV-1 genotypes and resistance phenotypes, predicting virological response to therapy based on HIV genotype, detecting mutations associated with HIV drug resistance , and enhancing computational predictions of resistance from genotype data. Logistic regression appears to be most powerful in predicting various treatment outcomes, including virological failure, adverse events, immune changes in people living with HIV receiving antiretrovirals, and biomarkers of mitochondrial toxicity. Conclusions: Depending on the field of application, some ML methods are more suitable and adapt better to certain HIV concerns. However, some areas, such as treatment recommendations, treatment adherence, and treatment optimization, still lack AI algorithms and need further exploration, such as therapeutical optimization. The development of new clinical decision-support systems for people living with HIV is the new challenge for the years ahead, and AI represents one of the most promising tools to address it.