valentineghanem-bit/malaria-geospatial-ml-ghana

Geospatial clustering and ML-based malaria risk prediction across 261 Ghana districts # Geospatial Clustering and Machine Learning Prediction of Malaria Burden at 261-District Resolution in Ghana: Integrating Insecticide-Treated Net Coverage and WASH Determinants **Author:** Valentine Golden Ghanem | Ghana COCOBOD Cocoa Clinic, Accra, Ghana **ORCID:** 0009-0002-8332-0220 **Affiliation:** Ghana COCOBOD Cocoa Clinic, Accra, Ghana **Reporting standard:** STROBE **Date:** April 2026 **Status:** Manuscript in preparation --- ## 1. Abstract This study maps malaria burden across Ghana's 261 districts at subnational resolution, integrating insecticide-treated net (ITN) coverage and WASH determinants. Spatial clustering analysis identifies high-priority hotspot districts, while bivariate LISA quantifies ITN-deficit co-clustering with malaria incidence. An ensemble machine learning pipeline (XGBoost, Random Forest, CART, Logistic Regression) with Leave-One-District-Out (LODO) cross-validation produces calibrated district-level malaria risk predictions. SHAP TreeExplainer identifies ITN coverage as the dominant modifiable predictor, while water access and open defecation emerge as key structural cofactors. --- ## 2. Research Question & Aims - **Primary:** Quantify the subnational distribution of malaria burden and identify priority hotspot districts across Ghana's 261 districts. - **Secondary:** (a) Detect ITN-deficit × malaria co-clusters using bivariate LISA; (b) build a LODO-CV ensemble ML pipeline for district-level risk prediction; (c) interpret model drivers using SHAP TreeExplainer; (d) tier hotspot districts by significance for programme prioritisation. --- ## 3. Methods Summary | Method | Tool | Purpose | |--------|------|---------| | Global Moran's I (KNN k=8) | esda / libpysal | Spatial autocorrelation of malaria incidence | | Bivariate LISA (Rook contiguity) | esda | ITN deficit × malaria co-clustering | | Getis-Ord Gi* | esda | Hotspot tiering (99.9%, 99%, 95% CI) | | XGBoost (LODO-CV) | xgboost | Risk prediction with spatial cross-vali …