Beyond Weather: Separating Climate and Surveillance Signals for Weekly Lassa Fever Prediction in Nigeria

Lassa fever remains a serious public-health challenge in Nigeria, but the separate predictive contributions of climate, seasonality, temporal trends, and previous surveillance counts remain unclear. This study evaluates these information sources for predicting weekly confirmed Lassa fever case counts across eight selected high-burden Nigerian states from 2019 to 2025. Weekly surveillance records were integrated with temperature, precipitation, relative humidity, evaporation, and soil-moisture data. The resulting state-by-week panel contained 2,616 observations, of which 53.6% reported zero cases. Three nested predictor sets represented lagged climate; climate with seasonality and trend; and a complete autoregressive set that additionally included historical case counts. Six machine-learning models were evaluated: ridge regression, random forest, XGBoost, support vector regression, k-nearest neighbours, and decision tree. Models were developed using five-fold temporally ordered cross-validation and evaluated on the chronologically held-out 2025 period against a one-week persistence baseline. Support vector regression achieved the highest climate-only performance (R² = 0.6247), showing that lagged climate provided meaningful standalone predictive information. Adding seasonality and trend improved four models, while historical case counts produced the largest gains. With the complete predictor set, support vector regression achieved the highest point estimate (R² = 0.7065), followed by ridge regression (0.7040) and random forest (0.6940). All six models exceeded the persistence baseline (R² = 0.5552) in point estimates, although only five models showed statistically significant improvement. Overall, climate provided genuine but secondary predictive information, while case history remained the strongest signal for one-week-ahead Lassa fever prediction.