Disentangling Climate, Seasonality, and Case-History Signals for Malaria Prediction Across Nigeria

Nigeria carries the world’s highest malaria burden, accounting for approximately 27% of global malaria cases and 31% of malaria deaths [1]. Timely prediction is therefore important for surveillance, intervention planning, and allocation of limited public health resources. Previous studies have demonstrated the value of climate and historical case information for malaria prediction [2, 3, 4], but the separate contributions of climate, seasonality, and surveillance history remain unclear. This study quantifies the predictive value of each information source for malaria prediction one month in advance at the state level across Nigeria. Monthly reports of confirmed uncomplicated malaria cases from the National Malaria Elimination Programme were combined with ERA5 temperature, relative humidity, and rainfall data for all 36 states and the Federal Capital Territory. The dataset covered January 2014 to December 2024 and contained 4,884 state-month observations. After feature generation, 4,440 observations remained. Three nested predictor sets were evaluated. The first contained lagged and rolling climate variables. The second added seasonal and linear trend features. The third incorporated malaria counts from 1, 2, 3, and 12 months before the prediction month, together with a three-month rolling average. All features were computed within states using only information available before the prediction month. Six models were tuned using five-fold time series cross-validation on data from January 2015 to December 2023. These models were ridge regression, random forest, XGBoost, decision tree, k-nearest neighbours, and support vector regression (SVR). Final performance was evaluated on data from 2024, which were held out chronologically. Performance was measured on log(1 + confirmed cases) using R², RMSE, and MAE. SVR performed best with climate alone and with climate plus seasonality and trend, achieving R² values of 0.729 and 0.828, respectively. For SVR, adding seasonality and trend reduced RMSE from 0.566 to 0.451 and MAE from 0.406 to 0.333. Historical malaria information improved all six models. Ridge regression achieved the highest overall performance (R² = 0.949), followed closely by random forest (R² = 0.948). The results show that climate provides useful advance information, seasonality adds predictive value, and recent surveillance history provides the strongest signal. The strong performance of ridge regression further indicates that informative temporal features can matter more than model complexity. These findings support more timely and informed malaria surveillance planning across Nigeria.