Artificial intelligence (AI) has recently gained significant attention due to its growing applications across various domains and the substantial improvements it has brought. Therefore, AI appears to be essential for improving health care [1,2]. A key discipline within AI is machine learning (ML). While AI is defined as the ability of a computer to perform tasks requiring human-like intelligence, ML focuses on developing statistical methods and algorithms that recognize patterns in datasets, improving prediction and classification accuracy when applied to new data. ML's increasing use in health care is justified mainly by the vast amounts of data generated and the ongoing digitalization of the industry, hospitals, and private practices, particularly through the adoption and exploration of electronic health records (EHRs) [3-6]. However, EHRs are not being fully exploited in the European Union due to technical and regulatory obstacles [7]. In response, many research projects were established to enhance the accessibility of EHR data for citizens and researchers in the European Union and facilitate better use [8,9].

Since the emergence of HIV in the 1980s, this disease has remained a major public health challenge, demanding considerable attention from both the scientific community and health care professionals. With advances in antiretroviral treatment, life expectancy for people living with HIV has increased considerably since the 1980s [10].

The management of people living with HIV remains a complex approach and still generates many inquiries from them and health care professionals. Although new therapeutic families have recently been developed, along with new forms such as intramuscular and subcutaneous injections, transforming the management of HIV infection, this only applies to certain eligible people living with HIV. Challenges remain. In addition, people living with HIV now have a near-normal life expectancy, comparable to that of the general population. The management of HIV infection, therefore, continues to pose challenges, particularly in older patients, who often experience age-related comorbidities leading to complex polypharmacy and increased risk of drug-drug interactions (DDI) and adverse drug reactions [11]. Thus, several therapeutic lines could be prescribed across the medical history of people living with HIV. Moreover, with the increasing volume and diversity of data routinely collected as part of people living with HIV care, managing the HIV landscape has become a complex task. This complexity is exacerbated by the heterogeneity of the information, which ranges from clinical data on types of treatment to biological data related to immune status and viral load to viral genetic signatures and socio-demographic factors, all of which require closer examination to improve care for this population.

Thus, comprehensive HIV care could improve both quality of life and long-term health outcomes. In recent years, ML has demonstrated its use in optimizing and revolutionizing various aspects of health care and research. For example, 2 pivotal applications of AI for clinical purposes are rule-based expert systems and clinical decision support systems. Clinical decision support systems combine clinicians’ medical expertise with recent AI advancements to enhance clinical decision-making processes. These systems leverage extensive medical knowledge derived from medical literature, complex algorithms, and patients’ EHR data to assist health care professionals in the overall quality of care while ensuring patient safety. These types of applications are time-consuming and rely on expert consensus. At first glance, it appears that there is a lack of use of ML tools in the field of HIV care. This review aims to identify the main applications of ML in HIV data management and to understand current trends. Specialized vocabulary for nonexperts in AI is summarized in Multimedia Appendix 1.

Overall, 476 studies were collected: 148 from PubMed, 200 from Embase, 16 from IEEE, and 112 from Web of Science (Figure 1).

We started our analysis with the literature trends by doing a statistical analysis of publication years, which revealed that, on average, articles were published in 2018, with a mean of 2017.7 (SD 5.63) years. The quartiles provide additional insight: the first quartile indicates that 25% (24/98) of articles were published before 2015. The median year of publication is 2020 (IQR 6.75-7), meaning that half the articles were published after this date. Finally, 25% (25/98) of articles were published after 2020, highlighting a clear trend in favor of more recent studies. These results reflect a strong temporal coverage of our literature review. Regarding the distribution of papers over time, the earliest study on HIV treatment using AI was published in 2002. Although publications remained sporadic until 2005, a noticeable uptick occurred afterwards (Figure 2).

Figure 3 illustrates the number of articles published in each category since 2002. For the comorbidities category, the first article was published in 2007. Notably, there were no publications in 2009, 2011, 2015, and 2016, with a peak of 14 articles published in 2021, indicating growing interest and progress in this area. In the HIV infection monitoring category, the first article appeared in 2005, followed by a second one after a 10-year gap. In the category of predicting drug resistance, the first article was published in 2002, marking the earliest of the selected articles. It took 4 years for the second publication, with 3 articles appearing in 2016 and 2023. The first article on predicting treatment outcomes was published in 2010. In the treatment adherence category, the first article was in 2005, with a 20-year gap before the second publication. Finally, in the treatment recommendation category, 4 articles were published between 2005 and 2008. This trend saw a significant surge post-2015, culminating in a peak in 2021. The 2021 peak may have been influenced by the COVID-19 pandemic, which intensified global research efforts and resulted in a higher volume of publications (Table S6 in Multimedia Appendix 2).

Among the 6 broad categories described above, the number of publications devoted to comorbidities amounts to 38% (n=37), closely followed by studies on the prediction of drug resistance at 23% (n=23), then monitoring HIV infection itself at 13% (n=13), predicting treatment outcomes at 11% (n=11), treatment adherence at 8% (n=8), and treatment recommendation at 6% (n=6).

More than half of the data used in the included studies came from America (more than 50% [53/98]), followed by Europe (20/98, 20%), Africa (18/98, 18%), and Asia (9/98, 9%). This imbalance revealed a strong dominance of data from Western regions, while data from Africa and Asia remained underrepresented. Overall, 52 studies were conducted in America [6,13-63], 20 in Europe [64-83], 11 in Africa [84-94], and 7 in Asia [95-101]. In addition, 8 studies are classified as collaborative [102-109], as the authors’ affiliations span multiple continents. The number of first authors is 56 in America [6,13-63,102,103,107,108], 24 in Europe [64-83,104-106,109], 11 in Africa [84-94], and 7 in Asia [95-101] (Figure 4). Interestingly, the density of articles by region showed a slightly different distribution (Figure 5).

Fairness-aware causal paths decomposition around 2020, Africa and Asia presented the highest concentration of research activity on this topic, with Asia reaching the most significant peak, followed by Africa. Europe and America also make notable contributions, but with less intensity. The datasets used in these studies come from a variety of countries and are highly heterogeneous. To deepen the analysis, the various data types used in the reviewed articles are summarized in Figure 6.

The three main data types are first numerical (n=75), then categorical (n=69), and finally textual (n=3). The 5 most used databases are cohort data (n=70), the Stanford database (n=15), the EuResist database (n=4), the Los Alamos National Laboratory database (n=3), and the Akwa Ibom database (n=2) (from a southern Nigerian state). Of these, HIV cohort data are the most widely used, particularly in studies of comorbidities, predicting drug resistance, predicting treatment outcomes, monitoring HIV infection, treatment adherence, and treatment recommendations. The Stanford database, which specializes in HIV drug resistance, comes second and is mainly used for research into resistance mutations. Finally, the EuResist database, the fruit of collaboration between several European sources, is used in 4 separate studies. A wide range of ML statistical methods were used in the studies we reviewed. The most used methods are random forest (RF), support vector machine (SVM), and logistic regression, each playing a key role in different aspects of HIV care research (Figure 7). Regarding method and algorithm trends, RF emerges as the most widely used algorithm, with 17.49% (43/247), reflecting its growing popularity in recent years. RF appears around 2010, with increasing use and peaking significantly in 2021. Artificial neural network (ANN), SVM, and logistic regression appeared after 2005, with increased use since 2017.

An in-depth analysis of the most used ML methods in the 6 categories mentioned above (Figure 8), as well as the tasks for which they are most effective, is detailed in this section. Logistic regression appears in all categories, whereas the ANN model was not used in monitoring HIV infection, and the other 3 methods were not used in the treatment recommendation category. We can observe that RF is the most used algorithm in each category where it appears. SVM is the second most widely used algorithm within the comorbidities category, ANN is used in predicting drug resistance, and logistic regression is used in predicting treatment outcomes and monitoring HIV infection.

In the comorbidity category [6,13-35,61,64-67,84,95-99,102,103] (Figure S1 in Multimedia Appendix 3), which constitutes the largest part of our study data (37/98, 38%; Figure 4), ML methods were used for various tasks, including identifying risk factors for cardiac complications, predicting vaccine responses in the context of HIV co-infection, testing for retinal damage in people living with HIV that may lead to subtle visual field defects, and predicting increasing comorbidity risks using EHR data [6]. Cardinal [13] conducted a study that identified significant associations between carotid artery plaques and selected features in a population of HIV-positive and HIV-negative individuals with low to intermediate cardiovascular risk. The RF algorithm effectively distinguished between individuals with and without carotid artery plaques by combining traditional cardiovascular risk factors with strain elastography features, using 5 most discriminant combinations of features achieved area under the receiver operating characteristic (AUROC) between 0.76 and 0.80 as classification performance. For predicting vaccine response in the context of coinfection, nonlinear models estimated by regression tree and RF were more accurate than generalized linear models in predicting humoral responses to SARS-CoV-2 mRNA vaccination in people living with HIV (classification and regression trees: R²=0.795; root mean square error=0.451 and RF: R²=0.845; root mean square error=0.412) [64]. This is linked to the fact that the different feature selection strategies used in linear models often tend to exclude important variables which, taken in isolation, might have a more significant predictive role [110]. Kozak et al [14] and Goldbaum et al [15,61] investigated whether people living with HIV had retinal lesions resulting in subtle visual field defects, and whether ML classification methods could distinguish these defects from those of HIV-negative individuals. The results confirm that HIV has an effect on the retina, and the eyes of people living with HIV with low cluster of differentiation 4 (CD4) counts show visual field defects and retinal lesions, while those of people living with HIV with high CD4 counts can appear normal. The SVM method was able to distinguish the visual fields of people living with HIV from those of HIV-negative people, even for people living with HIV with high CD4 counts, a task that is more challenging for a human expert. The model achieved an AUROC of 0.843 for the subset of people living with HIV with low CD4 counts, and an AUROC of 0.695 for those with high CD4 counts. Another study was also conducted on predicting the risk of multidrug-resistant enterobacterial (MDR-E) infections among people living with HIV [16]. Among 4734 study participants, MDR-E was isolated from 1.6% (95% CI 1.2%‐2.1%). In unadjusted analyses, MDR-E was strongly associated with nadir CD4 cell count ≤200 cells/mm³ (prevalence ratios [PR], 4.0; 95% CI, 2.3‐7.4), history of an AIDS-defining clinical condition (PR, 3.7; 95% CI 2.3‐6.2), and hospital admission in the prior 12 months (PR, 5; 95% CI 3.2‐7.9). Searches were also made to identify biomarkers and assess the patterns of T lymphocyte cell activation associated with the development of tuberculosis following the initiation of ART in people living with HIV displaying high CD4 counts [95,102]. Furthermore, ML methods were able to identify biomarkers associated with metabolic syndrome and HIV-associated neurocognitive disorders. They also pinpointed genetic features on the HIV envelope, accurately predicted neurocognitive impairment, and recognized molecular signatures linked to HIV-associated neurocognitive disorders diagnosis [17-23,25,29-31,33-35,65,84,96,97,103].

Another example of application: a multivariable logistic regression model was more suitable to predict the risk of serious falls among an older population of people living with HIV. However, including ART classes taken by these individuals did not improve the algorithm’s prediction, with the C-statistic increasing only slightly from 0.725 to 0.732 after the inclusion of ART classes [32]. The most frequently used method in this category is RF, as seen in (Figure S1 in Multimedia Appendix 3), valued for its robustness and use of the bootstrap principle.

For predicting or classifying drug resistance to antiretrovirals and identifying genetic signatures, 23% (23/98) of the articles focused on this area (Figure 4; Table S2 in Multimedia Appendix 2; Figure S2 in Multimedia Appendix 3) [36-43,62,63,68-77,85,86,104]. The classification and regression trees algorithm was able to predict drug resistance phenotypes from HIV-1 genotypes with good accuracy; some key sequence positions that are associated with drug resistance were identified using a mutual information analysis [68]. RF algorithm shows strongest correlation between predicted and actual virological responses, outperforming SVM and ANN. The mean absolute error ranged in [0.494-0.644] for RF, in [0.50-0.790] for SVM, and in [0.677-0.903] for ANN. Combining these methods further enhanced prediction accuracy, indicating that ensemble approaches are particularly effective in forecasting virological response [62]. ML methods were also used to analyze correlations between HIV-1 genotypes and resistance phenotypes, to predict virological responses to therapy based on HIV genotype, to detect mutations associated with HIV drug resistance without requiring expert knowledge, and to improve computational prediction of resistance from genotype data [36-43,63,69-77,85,86,104]. Both RF and SVM classified reverse transcriptase mutants with known resistance or sensitivity to nevirapine, achieving high accuracy in predicting susceptibility to the drug.

In the category of monitoring HIV infection, 13% (13/98) of (Table S3 in Multimedia Appendix 2; Figure S3 in Multimedia Appendix 3) studies involved applications of ML for tracking viral load and infection status [44-50,78,79,87,88,100,105]. For example, some studies identify biomarkers of viral rebound in people living with HIV from independent cohorts' pretreatment interruption [47] or tend to determine the viral compartmentalization dynamics to document a specific genetic signature of virus [44]. Mahto and Sood [49] evaluate various ML methods for predicting HIV progression and patient outcomes, identifying RF and extreme gradient boosting as the most effective, with accuracy rates of 0.88 and 0.89, respectively. They highlight the potential for further optimization through hyperparameter fine-tuning, feature engineering, and incorporating additional data sources to enhance algorithm robustness and generalizability. Kagendi and Mwau [88] developed an RF algorithm to predict HIV viral load hotspots in Kenya as an early warning system, achieving an accuracy of 0.78 and correctly identifying 434 viral load hotspots in December 2019, demonstrating its potential to optimize ART programs by supporting proactive resource allocation. A study found that the risk of HIV infection was higher for African Americans compared to non-African Americans. By using the fairness-aware causal paths decomposition method, researchers identified several social determinants of health that contribute to this racial disparity, including education, income, violent crime, drinking, smoking, and rurality. Performance evaluation of predictive models demonstrated comparable results across methods: boosted logistic regression achieved an AUROC of 0.79, decision tree-based methods yielded an AUROC of 0.77, and RF outperformed with an AUROC of 0.80 [50]. The SVM was the top-performing algorithm for predicting CD4/CD8 ratios in patients with baseline CD4 counts below 200 cells/ml. The RF followed with R² values of 0.438 and 0.519 for SVM [100]. ML methods also monitored ART adherence and retention in care for people living with HIV.

For predicting treatment outcomes, 11% (11/98) of the articles explored this topic (Table S4 in Multimedia Appendix 2; Figure S4 in Multimedia Appendix 3) [51-53,80,81,89-92,101,106]. Among 7 classifiers, the RF algorithm achieved high accuracy in predicting CD4 count changes, with precision, sensitivity, and recall values close to 0.99 (sensitivity=1.00, precision=0.987, F₁-score=0.993, AUC=0.998), and it outperformed other algorithms in predicting virological failure and identifying key predictors [89]. The logistic regression excelled in predicting virological suppression [90]. ML methods were also able to predict adverse events [53], immune changes in ART recipients [52,54], biomarkers linked to mitochondrial toxicity [51,52], viral load [81,92], and highly active antiretroviral therapy (HAART) regimens recommendations [80].

For treatment adherence and predicting retention in care, 8% (8/98) of the articles examined ML applications (Table S5 in Multimedia Appendix 2; Figure S5 in Multimedia Appendix 3) [54-57,93,94,107,108]. A study conducted in Nigeria by Ogbechie et al [94] was able to predict interruptions in ART treatment at 30 days among people living with HIV, using routine program data. The method used integrated boosting tree and extreme gradient boosting techniques and demonstrated a strong performance, with a sensitivity of 0.81, specificity of 0.88, and a positive predictive value of 0.83. After its integration into the national electronic medical records system, the proportion of people living with HIV interrupted ART treatment cases decreased from 58.6% to 14.2%. Health workers reported that the model enabled proactive interventions, underscoring the potential of ML to enhance HIV treatment adherence and improve patient outcomes. Stockman et al [108] used ML methods to predict loss to follow-up among people living with HIV on ART treatment in Nigeria and Mozambique, using data from health facilities, geospatial sources, and satellite imagery. For the Mozambique dataset, the RF was the best method with an area under the precision-recall curve of 0.65, while in Nigeria, the boosted tree method achieved an area under the precision-recall curve of 0.52. Lu et al [57] demonstrated that SVM effectively models factors influencing adherence to HAART in people living with HIV, outperforming neural networks.

In total, 6 articles (6%) [58-60,82,83,109] discussed the use of ML for treatment recommendations in people living with HIV (Table S6 in Multimedia Appendix 2; Figure S6 in Multimedia Appendix 3. Fuzzy discrete event system-based HIV regimen selection system demonstrated high self-learning accuracy in predicting treatments, with more than 80% agreement with actual prescribed regimens for 35 patients under practical conditions. The system outperformed neural networks in transparency and interpretability and showed an accuracy of 0.84 to 1 in predicting new treatment regimens. Additionally, it achieved 82.9% agreement with non-expert physicians and 100% agreement with expert physicians. By considering clinical, demographic, and genotypic data, the combined system provided the best prediction performance for HAART therapy recommendations based on genotype (Figure S6 in Multimedia Appendix 3) [58-60,82,109]. Another approach was to assess the relevance of clinical drug interaction (DDI) with ART. Two DDI prediction algorithms: DeepARV-ChemBERT and DeepARV-sim were developed by Pham et al [83] for predicting DDIs between ARVs and comedications. Those algorithms achieved a weighted accuracy of 0.729 and 0.776, respectively.