Forecasting malaria dynamics based on causal relations between control interventions, climatic factors, and disease incidence in western Kenya.

Background: Malaria remains one of the deadliest diseases worldwide, especially among young children in sub-Saharan Africa. Predictive models are necessary for effective planning and resource allocation; however, statistical models suffer from association pitfalls. In this study, we used empirical dynamic modelling (EDM) to investigate causal links between climatic factors and intervention coverage with malaria for short-term forecasting.

Methods: Based on data spanning the period from 2008 to 2022, we used convergent cross-mapping (CCM) to identify suitable lags for climatic drivers and investigate their effects, interaction strength, and suitability ranges on malaria incidence. Monthly malaria cases were collected at St. Elizabeth Lwak Mission Hospital. Intervention coverage and population movement data were obtained from household surveys in Asembo, western Kenya. Daytime land surface temperature (LSTD), rainfall, relative humidity (RH), wind speed, solar radiation, crop cover, and surface water coverage were extracted from remote sensing sources. Short-term forecasting of malaria incidence was performed using state-space reconstruction.

Results: We observed causal links between climatic drivers, bed net use, and malaria incidence. LSTD lagged over the previous month; rainfall and RH lagged over the previous two months; and wind speed in the current month had the highest predictive skills. Increases in LSTD, wind speed, and bed net use negatively affected incidence, while increases in rainfall and humidity had positive effects. Interaction strengths were more pronounced at temperature, rainfall, RH, wind speed, and bed net coverage ranges of 30-35°C, 30-120 mm, 67-80%, 0.5-0.7 m/s, and above 90%, respectively. Temperature and rainfall exceeding 35°C and 180 mm, respectively, had a greater negative effect. We also observed good short-term predictive performance using the multivariable forecasting model (Pearson correlation coefficient = 0.85, root mean square error = 0.15).

Conclusions: Our findings demonstrate the utility of CCM in establishing causal linkages between malaria incidence and both climatic and non-climatic drivers. By identifying these causal links and suitability ranges, we provide valuable information for modelling the impact of future climate scenarios.