valentineghanem-bit/ghana-child-mortality-261-districts

Subnational spatial distribution and ensemble ML predictors of neonatal and under-five mortality across Ghana 261 health districts # Subnational Spatial Distribution and Ensemble ML Predictors of Neonatal and Under-Five Mortality Across Ghana's 261 Health Districts **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:** May 2026 **Status:** Manuscript in preparation --- ## 1. Abstract This study characterises the subnational spatial distribution of neonatal mortality (NMR) and under-five mortality (U5MR) across Ghana's 261 health districts, identifies high-risk spatial clusters using global and local spatial autocorrelation, and develops ensemble machine learning models to predict district-level mortality risk. The analysis integrates 2022 Ghana DHS regional data, 2021 Census district estimates, and Ghana Statistical Service geospatial boundaries to generate actionable district-level risk maps. A stacked ensemble (Random Forest + Gradient Boosting with logistic meta-learner) under 10-fold stratified cross-validation is used for risk classification. --- ## 2. Research Question & Aims - **Primary:** Quantify the subnational distribution of NMR and U5MR across Ghana's 261 health districts. - **Secondary:** (a) Detect spatial clusters using Global Moran's I, LISA, and bivariate Moran's I (NMR × U5MR); (b) rank predictors using Random Forest Gini importance; (c) build a stacked ensemble classifier with logistic meta-learner; (d) produce district-level risk maps for programmatic targeting. --- ## 3. Methods Summary | Method | Tool | Purpose | |--------|------|---------| | Global Moran's I (KNN-4, 999 permutations) | esda / libpysal | Spatial autocorrelation of U5MR and NMR | | LISA | esda | Local spatial cluster detection | | Bivariate Moran's I | esda | NMR × U5MR co-clustering | | Random Forest | scikit-learn | Predictor importance (Gini impurity) | | Gradient Boosting | scikit-learn | Mortality risk prediction | | Stacked ensemble (log …