Malaria incidence, severity and mortality in children under five in Ghana: evidence from generalised additive models

Abstract

Abstract Background Under-five malaria remains a public-health priority in Tarkwa-Nsuaem, Ghana. This study analysed 132 monthly observations (2013–2023) to characterise trends, weather sensitivity, and near-term risk. Methods Surveillance counts (incidence, severe cases, and deaths) were linked with monthly rainfall and temperature. Negative-binomial generalized additive models (NB-GAMs) with smooth long-term trends and cyclic monthly effects were fitted. Weather influences were evaluated as contemporaneous, lagged (1–3-month), and accumulated (3–6-month) rainfall indices. Model selection employed AIC, adjusted R2, deviance explained, dispersion, and residual autocorrelation. A 12-month forecast was generated using the selected model. Results Monthly incidence was high (median = 847 cases), while severe cases (median = 2) and deaths were rare. Adding rainfall “memory” (1–3-month lags; 3–6-month accumulations) improved the incidence NB-GAM (AIC 1,269 → 1,234; ΔAIC = − 35; deviance explained ≈ 80%; adjusted R2 ≈ 0.88). In the pruned incidence model, trend (χ2≈334, p < .001), Rain_lag1 (χ2≈15.4, p = .002) and Rain_roll6 (χ2≈9.7, p = .002) remained significant. Out-of-sample errors were MAE/RMSE = 96/116 (train) and 190/232 (test). Severe malaria showed a secular decline with weak weather effects (deviance explained ≈ 0.85). Deaths were best modeled with zero-inflated NB (AIC = 104.1 vs 113.3 for NB), with a trend-only signal. Pearson dispersions indicated acceptable fit (incidence 0.95; severe 1.32; deaths 0.76). Twelve-month projections centered at ~ 600–670 incidence cases/month (95% PI ≈ 250–1,350), with deaths ~ 0–1/month. Conclusions Transmission reflects recent rainfall history rather than concurrent totals. Declines in severity and mortality likely mirror improved care. NB-GAMs with rainfall-memory terms yield interpretable, operational forecasts for early warning and resource planning.