Phase-specific climatic sensitivities of under-five malaria in Ghana using epidemic modelling, NB-GAMs and SHAP explainability

Abstract

Malaria transmission in Ghana remains perennial and climate-sensitive, yet the long-term transmission phases underlying disease progression and their differential climatic sensitivities are poorly understood. This study delineated latent, accelerated, and delayed phases of under-five malaria transmission and evaluated the phase-specific influence of rainfall and temperature on malaria incidence, severity, and mortality. Monthly under-five malaria incidence, severe admissions, and malaria-attributed deaths recorded in Tarkwa-Nsuaem Municipality, Ghana, from January 2013 to December 2023 (132 months) were linked with municipality-wide rainfall and temperature data. Epidemic phases were identified using logistic growth modelling and changepoint detection. Incidence, severity, and mortality trajectories were characterized using logistic, exponential–quadratic, and saturating epidemic functions, respectively. Climatic associations were evaluated using phase-specific negative binomial generalized additive models (NB-GAMs) with cyclic seasonal smooths. Model robustness was assessed through bootstrap resampling, cross-validation, basis-dimension diagnostics, and extreme-rainfall sensitivity analyses. SHapley Additive exPlanations (SHAP) were used to quantify predictor importance. The cumulative incidence trajectory exhibited an asymptotic burden of 115,052 cases (95% CI 107,390–125,844), with an inflection point at approximately 75 months and a growth scale of 23.9 months. Severe malaria declined rapidly, reaching functional decline around month 44, whereas mortality exhibited a slower decline, with functional reduction occurring near month 90. Climatic effects varied across transmission phases. Incidence was most climate-sensitive during the accelerated phase, where both temperature and rainfall were strongly associated with malaria transmission during this phase. Sensitivity analyses demonstrated that the accelerated-phase temperature association remained statistically significant after exclusion of extreme-rainfall months, indicating a robust climatic signal, whereas rainfall associations remained statistically detectable in the expanded-k incidence model but demonstrated greater sensitivity to extreme-rainfall observations than the corresponding temperature associations. Severe malaria showed no consistent climatic associations across alternative model specifications. Mortality exhibited temperature sensitivity primarily during the accelerated phase, with limited evidence of rainfall effects. SHAP analyses identified transmission phase, rainfall, and temperature as the most influential predictors of model-predicted incidence. Under-five malaria transmission in Tarkwa-Nsuaem follows distinct latent, accelerated, and delayed phases characterized by differing climatic sensitivities. Temperature demonstrated a stable association with transmission during the accelerated phase, whereas rainfall effects were strongly dependent on extreme hydrological conditions. The integration of epidemic phase delineation, phase-specific NB-GAMs, and explainable machine learning provides a robust framework for identifying climate-sensitive transmission periods and may inform phase-targeted malaria surveillance, preparedness, and intervention planning.