A random forest model to predict malaria outbreaks A case study of Kisii county

Abstract

In Sub-Saharan Africa, malaria has remained as a major source of morbidity and mortality, and a constant menace to the population health and socio-economic balance. Seasonal epidemics in highlands areas like Kisii County of Kenya create a huge burden to the healthcare systems, interfere with livelihood, and hamper development in the area. Despite the major advances achieved in reducing malaria rates in the country, the fact that it still is spread in highlands indicates the weakness of the existing control and monitoring measures. Conventional forecasting methods such as logic regression and ARIMA models have been actively used in predicting trends in malaria but their assumptions of linearity, normality, and stationarity makes them less useful to the study of the non-linear and dynamic relationships that exist between environmental, epidemiological, and socio-economic variables that define malaria transmission in highland micro-ecologies. Machine learning, especially ensemble-based ones, is a promising new direction in the development of malaria prediction. Ensemble models can learn more of complex interactions using heterogeneous data, hence able to adjust to changing epidemiological situations and produce more precise and useful information. This paper thus builds and compares an ensemble learning model to forecast malaria outbreaks in Kisii County that incorporates Random Forest (RF), XGBoost and RF-XGBoost hybrid model. The framework uses ecological, climatic, socio-economic, and epidemiological predictors to show that highland malaria occurs in a multifactorial manner. To resolve the important data issues, methodological advances like the Synthetic Minority Over-Sampling Technique (SMOTE), adaptive class weight, and cost-sensitive learning are used to curb the issue of class imbalance, and temporal feature weighting and sliding-window retraining are used in order to overcome concept drift due to climatic and intervention variability. Besides the accuracy of predictions, the interpretability and operational relevance are a priority of the study because of the usage of explainable AI tools like analysis of feature importance, SHAP values, and LIME visualization. To assess model performance, we measure such metrics as accuracy, F1-score, ROC-AUC, and precision-recall and complemented by ablation studies to evaluate the strength of and contribution of a specific methodological improvement. The hybrid ensemble approach will be based on the ability of the generalization of the Random Forest with the bias-correcting power of the XGBoost to create a transparent/scalable/adaptive ecologically specific system in the highland environments. Besides technical improvement, the research advances the principles of responsible and ethical AI by focusing on interpretability, fairness, and data governance. The expected deliverable is the locally calibrated predictive system, which can be used to support timely, evidence-based public health decision-making in the Kisii County and can be used to offer a transferable methodological framework of malaria surveillance in ecologically similar highland areas across Sub-Saharan Africa.