Development of a Process-Based and Machine Learning Early Warning System for Malaria Transmission Dynamics under Multi-Scenario Climate Change: A Subnational Assessment of the North Bank Region of The Gambia

Abstract Background Malaria is a major health challenge in the West African Sahel, driven by variable hydroclimatic conditions. Traditional surveillance relies on retrospective clinical data, limiting proactive resource allocation. This study aims to: (1) quantify nonlinear biological and statistical lags through which meteorological factors influence malaria in the North Bank Region (NBR); (2) compare machine learning models with traditional biological models to identify the best early warning system; and (3) project malaria risk and economic impacts under climate change using the MPI-ESM1-2-HR and NCAR-CESM2 climate models across SSP2-4.5 and SSP5-8.5 scenarios to 2055. Methodology : Using an 11-year longitudinal database (2011–2021) comprising 132 consecutive months of clinical surveillance records from the National Malaria Control Programme (NMCP), this study integrated satellite-derived meteorological data (CHIRPS precipitation, MODIS land surface temperature, and relative humidity). I ran a process-based biological simulation with the Liverpool Malaria Model (LMM), mapped continuous relative risks with a Distributed Lag Non-linear Model (DLNM), and conducted an out-of-sample machine learning tournament (SVR, XGBoost, and Random Forest) within a strict 24-month pandemic-era validation window (2020–2021). Optimized models were projected across a cross-model matrix pairing MPI-ESM1-2-HR with NCAR-CESM2 climate surfaces under SSP2-4.5 and SSP5-8.5 scenarios through 2055, and results were linked to a health-economic valuation matrix. Results The baseline Random Forest model demonstrated exceptional explanatory power, accounting for 95.56% of continuous transmission variability. Lagrangian feature mapping identified delayed, multi-month rainfall patterns as key drivers, with three-month-lagged rainfall (RFt3) showing the highest predictive importance at 24.66%. In validation, Random Forest was the top performer, limiting the average absolute error to 85.13 cases. Projections indicate that risks will expand nonlinearly across the inland savannah, with risk levels rising to 90.15% under the MPI-ESM1-2-HR model and 87.65% under the NCAR-CESM2 model under the extreme SSP5-8.5 scenario, effectively closing the historic protective gap near the estuarine coast. A cross-model comparison reveals that MPI-ESM1-2-HR offers a more geographically precise and physically constrained view of regional climate feedbacks than NCAR-CESM2, which tends to overestimate aridity during transitional decades. Economic analysis shows that interventions guided by this early warning system could prevent up to 42,100 cases and save approximately $189,450 USD in healthcare costs within the interior savannah by mid-century. Conclusion Shifting from reactive disease management to proactive, climate-informed targeting through robust tree-based machine-learning early-warning frameworks is a vital humanitarian and fiscal imperative for protecting vulnerable Sahelian communities from accelerating climate change, across multiple models.