Application of Artificial Intelligence for Predicting HIV Prevention: A Systematic Review and Meta-Analysis

There has been substantial progress in reducing HIV-related mortality globally, yet persistently high HIV incidence rates remain a challenge. Emerging technologies, particularly artificial intelligence (AI) including machine learning (ML) and deep learning, offer promising avenues to predict clinical outcomes and enhance intervention effectiveness. Hence, it is crucial to critically evaluate current efforts and explore future directions in applying AI for HIV prevention interventions. This systematic review and meta-analysis aim to assess the performance of AI prediction tools in the prevention of HIV. Following PRISMA guidelines, we systematically searched PubMed, Scopus, Embase, and ScienceDirect for primary articles on AI prediction tools, including ML and deep learning, for HIV prevention. Eligible studies were full-text, published in English, and dated between December 2013 and December 2023. Two independent reviewers followed the study protocol, extracted data, and assessed quality using the CASP Tool. Model performance was evaluated using sensitivity, specificity, f1-score, and AUC to determine the effectiveness of AI-based HIV prevention tools. Out of 377 original studies, 23 were included, with most conducted in the United States (n=9) and Africa (n=5). A total of 13 AI-based prediction tools targeted key populations. Studies were categorized into four prevention methods: monitoring, surveillance, and evaluation of interventions (n=7); identifying suitable candidates for Pre-Exposure Prophylaxis (PrEP) (n=4); health promotion and awareness (n=4); and HIV screening and testing (n=10). A meta-analysis of 11 studies showed an average model sensitivity of 79.7% and specificity of 81.6%, with a pooled AUC of 72.0% across 15 studies. The top-performing algorithms were Random Forest and LASSO. AI prediction tools hold promise for enhancing HIV prevention, particularly among key populations. However, standardizing and evaluating data quality, outcome definitions, predictive performance, and tool usability are essential for maximizing their effectiveness and reliability in real-world applications.