(ASPOT36) A Data-driven Approach to Unravelling the Contextual Social Determinants of Racial/Ethnic Health Disparities in Nonadherence to Antihypertensive Medications

Background: The lack of comprehensive spatio-temporal datasets with measures of the natural, built, social, economic and political environments of the places where people live has stymied efforts towards unraveling the contextual social determinants that drive disparities in medication nonadherence.

Objectives: To apply a combinatorial approach to discover contextual social determinants of nonadherence to antihypertensive medications from complex and highly heterogenous spatio-temporal data sources.

Methods: Data source and sample: Our team created the Public Health Exposome (PHE) dataset which contains over 50,000 descriptors of the natural, built, social and political environments that span over 30 years (1980–2020). We restricted our analysis to counties (and parishes) because measures of AHM nonadherence in the PHE were available only at these geographic units and those counties with non-zero nonadherence events.

Analysis: We applied our combinatorial toolchain to create graph-theoretical algorithms for discovering contextual social determinants of medication nonadherence from the PHE dataset. To implement this approach, we started by building three subgraphs, overall, Black/African American (BAA) and non-Hispanic White (nHW), each containing county-level AHM nonadherence as anchors and its neighbors. Next, we performed factor analysis and Bayesian inference on the extracted subgraphs to discover any potential latent networks. Bayesian network analysis was used to determine directionality within bicliques of interest. To visualize a biclique, first a biclique and a pseudorandom subset of vertices from each partite set were selected. This pseudorandom subset was selected to serve as background from which the biclique was visualized. Then both subsets were visualized using the Networkx python package. Finally, we applied domain expertise to elucidate both known and previously unrecognized variable relationships between social determinants constructs and AHM nonadherence.

Results: We generated three county lists, one with 613 counties for BAAs, another with 813 counties for nHWs and a third with 814 counties for all. Through the factor analysis and Bayesian inference we observed evidence of latent networks specific to extreme heat and lack of health insurance for BAAs, indicators of poverty for nHWs, and a mix of extreme heat and poverty for all populations. The Bayesian Information Criterion (BIC) scores were -91473.29 for BAAs, -86686.03 for nHWs, and 51986.14 for overall; where the lower the score, the more likely it is believed that a model is true.

Conclusions: The successful application of our combinatorial approach facilitated the discovery of contextual social determinants of AHM nonadherence.