HIV Spatial Epidemiology Ghana — Conference Poster

Background

HIV spatial heterogeneity in sub-Saharan Africa reflects complex interactions between behavioural factors, socioeconomic determinants, and health system access. Ghana's national prevalence of 1.7% masks substantial subnational variation that standard aggregate summaries fail to capture.

Objective: Characterise the spatial distribution and socioeconomic determinants of HIV prevalence across Ghana's 260 health districts using spatial statistics and machine learning.

Study Design: Ecological cross-sectional analysis integrating DHS 2014/2022, Ghana Census 2021, and Ghana Statistical Service data across 260 districts in 16 administrative regions.

Methods

Spatial Analysis:

Global Moran's I (KNN k=4)
LISA with Rook contiguity (999 permutations, p<0.05)
Getis-Ord Gi* hotspot detection
Spatial lag regression (ML estimation, Queen contiguity)

Machine Learning:

LASSO with 10-fold CV (feature selection)
Random Forest with leave-one-region-out spatial CV (16 folds)
SHAP TreeExplainer (feature importance)

Software: Python (PySAL, scikit-learn, SHAP, GeoPandas). CRS: EPSG:32630.

Fig 1. Study Area & HIV Prevalence

Bivariate choropleth: HIV prevalence (left) and poverty incidence (right) across 260 districts. Northern districts show high poverty but low HIV; southern/central districts show inverse pattern.

Fig 2. LISA Cluster Map

Local Indicators of Spatial Association. 113 significant clusters identified: 59 Low-High (blue), 51 High-Low (red), 3 non-significant. Boundary-transition effects dominate. p<0.05, 999 permutations.

Fig 3. Pearson Correlation Matrix

Full 19-variable symmetric matrix. Strongest correlates of HIV prevalence: VCT uptake (r=0.84), Female secondary education (r=0.81), Wife beating acceptance (r=−0.82), Illiteracy (r=−0.68).

Fig 4. SHAP Feature Importance

Top 10 Random Forest predictors via SHAP TreeExplainer. VCT Women (0.639) and Wife Beating Acceptance (0.241) dominate. Remaining 8 features contribute <0.03 each.

Key Findings

Global Moran's I: 0.907 (z=22.3, p<0.001)

Strong positive spatial autocorrelation confirmed.

Spatial Lag Regression:

ρ = 0.596 (p<0.001)
Pseudo-R² = 0.927
VCT uptake: β=0.039 (p<0.001)
Modern contraception: β=0.037 (p=0.002)

ML Model Comparison:

LASSO: R² = 0.933 (13/19 features)
RF Spatial CV: R² = 0.611

LASSO captures global linear structure; RF captures local spatial heterogeneity. Performance gap reflects 16 discrete regional strata in the data.

Conclusions

Spatial clustering is substantial — district-level HIV prevalence exhibits strong neighbourhood dependence (I=0.907)
Boundary-transition effects dominate LISA patterns (HL/LH), suggesting regional-scale epidemic geography rather than focal hotspots
VCT uptake and gender norms are the two most powerful predictors across all models
Ecological interpretation required — district-level associations do not imply individual-level causation

Policy Implications

Targeted VCT scale-up in low-testing districts (Northern, Savannah, North East regions)
Gender-based violence programmes in districts with high wife-beating acceptance (>50%)
Education investment — female secondary education shows strong protective association
Cross-boundary surveillance — spillover effects (HL/LH clusters) demand coordinated inter-district response

STROBE compliant — Ecological cross-sectional study. Reporting follows STROBE guidelines for observational studies.

Key References: 1. Ghana AIDS Commission. Ghana's HIV and AIDS Status Report 2022. Accra: GAC; 2023. 2. Moran PAP. Biometrika. 1950;37(1–2):17–23. doi:10.1093/biomet/37.1-2.17 3. Anselin L. Geogr Anal. 1995;27(2):93–115. doi:10.1111/j.1538-4632.1995.tb00338.x 4. Lundberg SM, Lee SI. NeurIPS. 2017;30:4765–4774. arXiv:1705.07874 5. Jewkes RK et al. Lancet. 2010;376(9734):41–48. doi:10.1016/S0140-6736(10)60548-X 6. Cuadros DF et al. Sci Rep. 2017;7:9093. doi:10.1038/s41598-017-09464-y 7. Sia D et al. BMC Public Health. 2016;16:1136. doi:10.1186/s12889-016-3783-5