The Influence of Social Determinants of Health on Healthcare Access, Utilization, and Treatments in Atopic Dermatitis Patients Using All of Us and Machine Learning Methods

OBJECTIVES: In the United States, atopic dermatitis (AD), affects 7.3% of the adult population, or 19 million adults. Despite its prevalence, treatment is relatively challenging and time consuming, with affordability concerns raised by groups such as ICER. Recent studies have shown that social determinants of health (SDoH) affect healthcare access and utilization (HCAU). This analysis uses machine learning methods to characterize how SDoH influence HCAU in AD patients using All of Us.

METHODS: Patients with and without atopic dermatitis who filled out the All of Us HCAU and SDoH surveys were included in the analysis (n=104,667 and n=1,546, ICD-9: 691, ICD-10: L20.9). Preliminary analysis focused on the SDoH domain of neighborhood quality (n=17 questions) and the HCAU healthcare affordability (n=8 questions, including follow up care and prescription treatments), and each were aggregated into a binary composite score. Pearson, Spearman, and Kendall’s Tau correlations analyzed the relationship between neighborhood quality and healthcare affordability. Random forest classifiers were applied to predict healthcare affordability based on neighborhood quality, and confidence intervals using bootstrapping were used to test for consistency. Other decision trees and SDoH/HCAU domains will also be assessed in this analysis.

RESULTS: 67% of respondents in the cohort identified as female, and 51% of the population were over 65 years of age. Of those in the HCAU domains, approximately 92% were classified as having high healthcare affordability. Correlation analysis yielded relatively low coefficients (0.09, 0.091, 0.087). Random forest classifiers returned an approximate accuracy of 94.03%, with a precision of 70% for the poor healthcare affordability class, and 95% for the high healthcare affordability class (95% CI [92.60%, 93.48%]).

CONCLUSIONS: Preliminary results suggest that SDoH affects healthcare affordability in AD patients, and that for assessing the complex relationship between other SDoH and HCAU domains, machine learning methods may better capture other SDoH domains that impact HCAU.