valentineghanem-bit/hiv-tb-ml-ghana-261districts

Ensemble geospatial intelligence for HIV-TB co-infection mapping across 261 Ghana districts # Spatial Distribution, Determinants, and Machine Learning-Based Risk Prediction of HIV-TB Co-infection Across Ghana's 261 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 **Release DOI:** 10.5281/zenodo.21351614 **Concept DOI:** 10.5281/zenodo.21351613 --- ## 1. Abstract A nationwide district-level analysis of HIV-TB co-infection in Ghana combining spatial statistics, geographically weighted regression, and ensemble machine learning across all 261 health districts (post-2018 Local Governance Act). Univariate and bivariate Moran's I characterise spatial autocorrelation; LISA and Getis-Ord Gi* delineate hotspots; GWR estimates spatially varying determinants; and a stacked ensemble (Random Forest + XGBoost + LightGBM) with SHAP interpretation yields district-level risk predictions. Analysis reveals moderate spatial clustering of co-infection (global Moran's I = 0.472, p DHS data accessed under signed Data Use Agreement (ICF International). No individual participant data redistributed. --- ## 5. Key Findings | Metric | Value | |--------|-------| | Global Moran's I (HIV-TB co-infection) | 0.472 (p **Note on the 10-fold vs. spatial-CV gap:** the ~20-point AUC drop and large SD (+/-0.25-0.27) between standard 10-fold CV and leave-one-region-out spatial CV reflects genuine, strong spatial autocorrelation in Ghanaian district-level health data (neighbouring districts have highly correlated HIV/TB rates), not a modelling error. The 10-fold number should not be quoted as the model's ability to generalise to an unseen region; the spatial-CV number is the more defensible generalisation estimate, and its wide SD signals real fold-to-fold instability that any downstream use of this model should account for. --- ## 6. Repository Structure ``` hiv-tb-ml …