Spatial and Machine Learning Analysis of Breast and Cervical Cancer Screening Uptake in Ghana: Evidence from the 2022 Ghana Demographic and Health Survey
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Breast and cervical cancer screening in Ghana remains low, and several analyses of the Ghana Demographic and Health Survey (GDHS) have already shown that wealth, education, and place of residence pattern who gets screened [1–3]. Whether this patterning clusters geographically below the level of administrative region has not been tested for this population, and whether cluster-aware machine learning adds anything to the standard regression approach used so far remains open.
Methods
We analyzed the 2022 Ghana Demographic and Health Survey women’s file (N = 15,014; primary sample of women aged 25-49 years, n = 9,510) linked to cluster geographic coordinates for 618 enumeration areas. Clinical breast examination and cervical cancer testing were the two outcomes. We estimated survey-weighted prevalence across demographic and socioeconomic strata, tested global spatial autocorrelation with Moran’s I, mapped local clustering with Getis-Ord Gi* statistics, and separately fitted gradient-boosted classifiers on individual-level socioeconomic covariates, validated under cluster-held-out five-fold cross-validation to prevent within-cluster information leakage. Feature contributions to the breast-screening model were interpreted with an additive, feature-level explanation technique, and socioeconomic inequality was quantified with both the ordinary and Erreygers-corrected concentration index. The predictive models did not include geographic coordinates or survey weights; both are noted as limitations.
Results
Weighted prevalence among women aged 25-49 years was 22.5% (standard error 0.77) for breast examination and 6.9% (standard error 0.46) for cervical testing. Both rose with education and wealth and were roughly double in urban areas relative to rural ones. Moran’s I was positive and significant for both outcomes (breast: 0.217, z = 11.69, p < 0.001; cervical: 0.125, z = 6.77, p < 0.001), and local cluster statistics located discrete hotspots around Greater Accra and parts of Ashanti and Bono, with coldspots concentrated across the north, though these local tests were not adjusted for multiple comparisons. Cross-validated discrimination reached an area under the curve of 0.716 for breast examination and 0.730 for cervical testing, without confidence intervals or calibration assessment; education, wealth, and age were the dominant predictors for both. The Erreygers index put breast examination as the more wealth-concentrated outcome (0.231 versus 0.095 for cervical testing), reversing the ranking implied by the uncorrected index.
Conclusions
Screening uptake in Ghana is spatially clustered at a resolution that regional reporting cannot show, and this clustering is compositionally associated with, though not formally shown to be mediated by, the socioeconomic makeup of individual clusters. Cluster-level spatial analysis and a model-based risk ranking may offer a useful complement to regional targeting, but calibration, external geographic validation, and comparison against a regional-allocation baseline are needed before any operational use.
Trial registration
Not applicable. This is a secondary, hypothesis-generating cross-sectional analysis of existing, publicly available survey data and was not prospectively registered.
