AI-Driven Predictive Analytics for Multi-Hazard Infectious Disease Outbreak Forecasting in Low- And Middle-Income Countries: A Systematic Review and Meta-Analysis of Multimodal Data Fusion Approaches, Implementation Barriers, and Equity Pathways

Abstract

Background: Infectious disease outbreaks in low- and middle-income countries (LMICs) emerge from intersecting biological, environmental, socioeconomic, and political hazards. AI-driven analytics and multimodal fusion offer promising early warning tools, yet their use in multi-hazard settings remains fragmented. This review assessed multimodal AI forecasting in LMICs, mapped implementation barriers, and identified equity pathways.
Methods: Following PRISMA 2020 and MOOSE guidelines (PROSPERO: CRD420251276083), 13 databases were searched (January 2005–December 2025). Eligible studies were conducted in LMICs, used AI-driven analytics across at least two data modalities, and reported quantitative performance metrics or qualitative implementation evidence. Quality was assessed using the Newcastle-Ottawa Scale, QUADAS-2, or JBI Critical Appraisal Checklist. Pooled AUC, sensitivity, and specificity were estimated using a DerSimonian-Laird random-effects model with logit-transformed outcomes. Subgroup analyses examined AI method, disease type, region, and modality; qualitative findings were synthesised thematically.
Results: Sixty-two studies (n = 518,770 records; 2005–2025) met inclusion criteria, covering COVID-19 (18, 29.0%), malaria (14, 22.6%), cholera (11, 17.7%), dengue (9, 14.5%), Ebola (5, 8.1%), and multi-disease models (5, 8.1%), predominantly from sub-Saharan Africa (38.7%), South Asia (24.2%), and Southeast Asia (19.4%). The pooled AUC was 0.846 (95% CI: 0.820–0.868; I² = 84.5%; k = 20) and the pooled sensitivity was 0.842 (95% CI: 0.817–0.864; I² = 83.4%; k = 15). Deep learning significantly outperformed classical machine learning (pooled AUC: 0.883 vs. 0.782; p < 0.001); hybrid multimodal fusion achieved the highest AUC (0.912). Primary implementation barriers were inadequate data infrastructure (74.2%), limited computational resources (67.7%), regulatory and governance gaps (48.4%), and poor health information system interoperability (43.5%). Only 29.0% addressed equity. Five equity pathways were identified: participatory co-design, federated learning, open-source architectures, capacity strengthening, and south-south knowledge transfer.
Conclusion: AI-driven multimodal outbreak forecasting in LMICs demonstrates strong discriminative performance, with effective models combining genomic, environmental, and mobility data with epidemiological surveillance. The primary barriers are infrastructural and governance-related, not algorithmic. Realizing public health benefits requires investment in digital infrastructure, cross-border data-sharing frameworks, and intentional inclusion of marginalized populations in model development, aligned with the WHO Health Emergency Preparedness and Response Agenda and the Sendai Framework for Disaster Risk Reduction.