Technical Summary: Spatial ML Analysis of ANC and Fertility Inequities in Ghana (1988–2022)

Problem Statement

Skilled antenatal care (ANC) is a critical preventive intervention for reducing maternal and neonatal mortality, yet sub-Saharan Africa continues to bear a disproportionate burden of maternal deaths, often concentrated in high-fertility, geographically marginalized communities. While Ghana has achieved near-universal skilled ANC coverage by 2022, significant subnational disparities persist. The core problem addressed is the complex, non-linear relationship between ANC access and Total Fertility Rate (TFR) across Ghana's 16 administrative regions over a 34-year period. Traditional regression methods often fail to capture the spatiotemporal dynamics, spatial autocorrelation, and non-linear thresholds where high fertility may suppress ANC utilization independently of supply-side improvements. Furthermore, standard coverage metrics may obscure fertility-related inequities, such as the "Workhorse" phenomenon where high coverage coexists with high fertility burdens, straining health systems.

Methodology

This study employs a longitudinal ecological design using nine waves of the Ghana Demographic and Health Surveys (DHS) from 1988 to 2022, generating 94 region-by-year observations across 16 administrative regions. The analysis integrates spatial statistics, machine learning, and novel composite metrics:

• Data Sources: DHS subnational data (skilled ANC coverage, TFR, adolescent fertility) and administrative boundary geometries (261 MMDAs).
• Inequality Analysis: The Gini coefficient was calculated to track inter-regional inequality in ANC coverage and TFR.
• Spatial Autocorrelation: Global Moran's I (with k=4 nearest neighbor weights) and Local Indicators of Spatial Association (LISA) using Rook contiguity were employed to detect clustering patterns and hotspots/coldspots.
• Machine Learning: Two supervised models were trained to predict skilled ANC coverage:
  • Random Forest (RF) Regressor: 200 estimators, max_depth=6, with 5-fold cross-validation.
  • Decision Tree (DT) Regressor: Used as a benchmark.
  • Feature Attribution: Partial dependence plots (PDP) were used to identify non-linear inflection points in the TFR-ANC relationship.
• Novel Metrics:
  • Care Efficiency Index (CEI): Defined as $CEI = \text{Skilled ANC (\%)} / \text{TFR}$, measuring service utilization efficiency relative to fertility burden.
  • Bivariate Risk Stratification: Observations were classified into four zones (Critical, Emerging, Workhorse, Resilient) based on z-scores of ANC and TFR relative to the grand mean.

Key Contributions

1. Spatiotemporal Convergence Analysis: The study quantifies the dramatic reduction in inter-regional ANC inequality (Gini coefficient decline of 87.9%) while highlighting the persistence of fertility-related spatial clustering.
2. Decoupling of Indicators: It demonstrates a divergence in spatial autocorrelation by 2022: ANC coverage clustering became non-significant (indicating uniform coverage), whereas TFR clustering intensified (Moran's I = 0.606), suggesting that demographic determinants of fertility respond over longer time horizons than health service delivery.
3. Exploratory Inflection Point: Using RF partial dependence analysis, the study identifies a potential non-linear inflection near a TFR of 5.90, above which predicted ANC coverage declines in historical data.
4. Efficiency and Risk Typologies: The introduction of the Care Efficiency Index (CEI) and bivariate risk stratification provides tools to distinguish between regions with "Resilient" (high ANC, low TFR) and "Workhorse" (high ANC, high TFR) profiles, revealing resource allocation needs that coverage percentages alone would miss.

Key Results

• Coverage Trends: National skilled ANC coverage increased from 83.1% (1988) to 97.7% (2022). The Northern Belt showed the largest absolute gains (e.g., Northern Region +43.0 percentage points), narrowing the North-South gap from 32.4 to 0.9 percentage points.
• Inequality Metrics: The inter-regional Gini for ANC dropped from 0.070 to 0.008. In contrast, the TFR Gini declined by only 27%, indicating fertility inequality is more resistant to policy intervention.
• Machine Learning Performance: The Random Forest model outperformed the Decision Tree (Test $R^2$ = 0.381 vs. -0.296). Survey year was the dominant predictor (43.7% importance), followed by TFR (38.8%).
• Inflection Point: The RF model identified an exploratory inflection at TFR ≈ 5.90. Above this threshold, predicted ANC coverage declines non-linearly, potentially reflecting demand-side barriers (e.g., partner opposition, transport costs) in super-high fertility settings.
• Regional Disparities:
  • Care Efficiency: Greater Accra led with a CEI of 31.9, while North East lagged at 14.5, representing a 2.2-fold efficiency gap despite similar coverage levels in 2022.
  • Risk Stratification: In 2022, most regions were "Resilient" or "Workhorse." However, the Northern Belt retained "Workhorse" patterns (High ANC/High TFR), implying a higher per-provider service burden compared to Southern regions.
• Spatial Clustering: By 2022, TFR exhibited strong positive spatial clustering (Moran's I = 0.606, p=0.001), particularly in the Northern Belt, while ANC clustering dissipated.

Significance and Claims

The paper claims that while Ghana has successfully achieved near-universal skilled ANC coverage and eliminated inter-regional coverage inequality, fertility-related spatial inequities persist, particularly in the Northern Belt. The study argues that conventional coverage statistics are insufficient for guiding health-system investment because they mask the "Workhorse" burden and the efficiency gap between high-fertility and low-fertility regions.

The authors posit that the Care Efficiency Index (CEI) and bivariate risk stratification offer hypothesis-generating tools for targeted investment. Specifically, they suggest that regions with high TFR and high ANC (Workhorse) require different resource allocation (e.g., staffing intensity) compared to Resilient regions. The exploratory TFR inflection near 5.90 is presented as a threshold for further validation, potentially indicating a limit where high fertility creates independent barriers to ANC utilization.

The study concludes that policy priorities must extend beyond supply-side expansion to include targeted fertility transition programs and female educational empowerment in the Northern Belt. It emphasizes that these findings are ecological and hypothesis-generating, cautioning against causal interpretations of the TFR-ANC relationship without further individual-level validation.