A Systematic Review of Artificial Intelligence and Machine Learning Methods and Deployment Challenges for Public Health Predictions Using Electronic Health Records in Low- and Middle-Income Countries

The growing availability of electronic health records (EHRs) has accelerated the use of artificial intelligence (AI) and machine learning (ML) in public health. Yet, how well these methods work in resource-limited settings, particularly low- and middle-income countries (LMICs), remains poorly understood. This systematic review synthesizes evidence from 64 peer-reviewed studies (2018–2025) on ML-based predictive analytics using EHRs, with LMICs as the primary focus and high-income country studies as a methodological reference. Following PRISMA guidelines, searches across five major databases identified 64 eligible studies published between 2018 and 2025. Of these, 12 (18.8%) were conducted exclusively in LMIC settings, 44 (68.8%) in high-income countries, and 8 (12.5%) drew on mixed or multi-setting data. Retrospective designs predominated (81.3%). Disease progression (40.6%), mortality (34.4%), and treatment response (25.0%) were the most common prediction targets. Deep learning architectures were the most frequently applied category overall (39.1%, n = 25), driven by high-income country studies with access to large curated datasets; among LMIC-focused studies, traditional ML and ensemble methods were each applied in 33.3% of studies. Evaluation practices were dominated by discrimination metrics, particularly AUROC; external validation was reported in only 5 studies (7.8%) and calibration in only 4 (6.2%). Explainability assessment was reported in 1 of 12 LMIC studies (8.3%) compared with 16 of 44 high-income studies (36.4%), with governance and ethical considerations inconsistently documented in LMIC settings. This review highlights key methodological and contextual gaps and offers actionable guidance for developing interpretable, reliable, and context-appropriate AI tools for public health decision-making in resource-constrained settings.