KwabenaAsabere/ghana-meningitis-spatiotemporal-modeling

Spatiotemporal machine learning for predicting district-level meningitis cases in Ghana using epidemiological, environmental, and spatial data. # Spatiotemporal Machine Learning of Meningitis Cases in Ghana, 2016–2024 ## Project Overview This project applies **spatiotemporal feature engineering and machine learning** to model district-level meningitis cases across Ghana from **2016 to 2024**. The analysis integrates meningitis surveillance data with environmental, demographic, geographic, and vaccination-related predictors to examine how disease occurrence varies across both **space and time**. A **LightGBM Poisson regression model** is used to predict district-level meningitis case counts, with particular attention to temporal validation and spatially informed predictors. --- ## Objectives The main objectives are to: * Model district-level meningitis cases across Ghana. * Integrate epidemiological, environmental, demographic, and geographic information. * Engineer temporal and spatial-neighborhood features that capture disease dynamics. * Develop a machine-learning model appropriate for count outcomes. * Evaluate the model on a completely held-out future year. * Identify geographic and environmental patterns that may help support meningitis surveillance and public health decision-making. --- ## Data The final modeling dataset contains: * **2,349 district-year observations** * **261 districts** * **9 years (2016–2024)** The outcome is: * `total_cases` — annual number of reported meningitis cases in each district. ### Predictor Variables The analysis includes several groups of predictors. **Environmental** * Rainfall * Temperature * Relative humidity * NDVI * Dust emission * Built surface **Population and Prevention** * Population * Vaccination coverage **Geographic** * District * Region * Ecological zone * Longitude * Latitude --- ## Spatiotemporal Feature Engineering Several additional features were created to capture temporal and geographic patterns that may not be represented by the original covariates. ### Temporal Features These include: * Previous-year meningitis cases * Pr …