gracekitonyi-bit/cholera-early-warning-kenya

Explainable machine learning framework for county-level cholera early warning in Kenya. # Cholera Early Warning in Kenya This repository contains the reproducible analysis for the MSc thesis: **Explainable Machine Learning Models for Early Warning of Cholera Outbreaks in Kenya Using Climate and Epidemiological Indicators** The study develops and evaluates a county-level early-warning framework for predicting cholera outbreaks in Kenya two to four weeks in advance using epidemiological, climatic, environmental, demographic and water, sanitation and hygiene (WASH) indicators. ## Study Overview The analysis covers all 47 counties in Kenya using weekly observations from 2015 to 2025. The study aimed to: - describe temporal patterns in cholera cases, deaths, incidence and outbreak occurrence; - compare ARIMA, Random Forest and XGBoost for three-week-ahead outbreak prediction; - evaluate outbreak detection and false-alert burden; - identify important predictive contributors using SHAP; - assess robustness across two-, three- and four-week prediction horizons. The primary analytical outbreak definition was a county-week with at least five reported cholera cases. A broader sensitivity definition additionally classified county-weeks with incidence of at least one case per 100,000 population as outbreaks. ## Data Structure The complete surveillance panel contained: - 47 counties; - 574 weekly observations per county; - 26,978 county-week observations; - 23 predictors used in the machine-learning analysis. After removing observations with structurally unavailable lagged predictors or future outcomes, the primary three-week modelling sample contained: **26,508 county-weeks** The predictor domains included: - epidemiological surveillance; - temporal indicators; - rainfall and temperature; - lagged climatic variables; - environmental indicators; - population; - WASH vulnerability. ## Prediction Framework For county \(i\) at epidemiological week \(t\), the study estimated the probability of a future outbreak at prediction horizon \(h\): \[ P(Y^{(h) …