Assessing the impact of social determinants on pharmacy visits and access: A machine learning approach

Community pharmacies provide preventive, urgent, and traditional pharmacy services and thus serve as key access points for care in the U.S. health system. This article examines how social determinants of health and built environment factors influence pharmacy visits. We integrate two data sources: U.S. Census socioeconomic data and pharmacy foot traffic data from Placer.ai. Hence, our dataset consists of origin–destination pairs between residential block groups and pharmacies, with crime data added at both origin and destination levels. Several machine learning models, including Random Forest, Support Vector Regression, and Neural Networks, are evaluated for predicting pharmacy visits. Extensive computational tests, using Jefferson County, KY, as a case study, demonstrate that the Neural Network model yields the best performance. To address partial repetition of features inherent in origin–destination paired data, we propose a Diversity Subsampling with Simulated Annealing (DSSA) method. Results show that the algorithm improves the representativeness of subsamples and the computational efficiency. Sensitivity analyses suggest that neighborhoods with a higher percentage of adults with a college education or higher and a higher percentage of the uninsured population are associated with fewer visits in some clusters. A higher crime rate at a pharmacy location reduces visits, while transit access shows mixed effects. Overall, this paper contributes a data-driven framework for predicting pharmacy visits, evaluating the role of social determinants in these visits, and informing strategies to improve equitable pharmacy access.