A multi-state climate epidemiology analysis of Lassa fever using interpretable machine learning and an early warning framework

Lassa fever (LF) remains hyperendemic in Nigeria, yet the environmental factors associated with short-term outbreak escalation are not well characterised. We examined climate–disease associations across six endemic Nigerian states using multi-year surveillance and satellite-derived environmental data. Weekly LF incidence data (2018–2025; n = 6,801 cases) from six Nigerian states were analysed alongside five climate variables derived from Google Earth Engine: soil moisture, relative humidity, rainfall, temperature, and NDVI. Seasonal-Trend decomposition (STL; 52-week periodicity), lag-response analysis (1–8 weeks), state-stratified Negative Binomial generalized linear models, and Random Forest models with SHAP interpretation were applied. Predictive performance was compared with a climate-naïve SARIMA benchmark using a common temporally held-out 52-week validation period. In the Negative Binomial analyses, higher soil moisture was significantly associated with lower weekly Lassa fever incidence in Ebonyi (IRR = 0.836, 95% CI 0.797–0.876; p < 0.001), while lag-response analyses identified consistent inverse associations between soil moisture and incidence across multiple states (Spearman ρ = −0.46 to − 0.66; FDR-adjusted q < 0.05). State-specific SHAP analyses identified soil moisture as the dominant climatic predictor in five of six states, whereas rainfall predominated in Bauchi. Predictive performance varied geographically: Random Forest outperformed the climate-naïve SARIMA model in Edo, Ondo, and Plateau, whereas SARIMA performed better in Bauchi, Ebonyi, and Taraba. Soil moisture and rainfall were consistently associated with temporal variation in weekly Lassa fever incidence across endemic Nigerian states, although their predictive value differed geographically. Climatic influences on weekly Lassa fever incidence differed across ecological settings, and the predictive value of climate-informed models varied by state. These findings support the development of locally tailored climate-informed surveillance strategies rather than a single national forecasting approach.