Modeling the Impact of Social Determinants of Health on COVID Behaviors in Older Adults using the All of Us Dataset

Non-pharmaceutical interventions such as hand-washing hygiene, avoiding large gatherings, and avoiding visiting nursing homes remain important in mitigating risks of COVID infection among at-risk populations such as older adults. The NIH's All of Us Research Program offers a unique dataset which contains detailed survey data that medical records often lack. Leveraging this dataset and impact scores, we were able to compare deep neural network (DNN) models to more conventional logistic regression and XGBoost models in the task of examining the relationships between social determinants of health and COVID-related behaviors in older adults. LR and DNN models found that African American participants were more likely than White participants to report adherence to guidelines regarding attending large social gatherings, abiding by stay-at-home recommendations and practicing pandemic-related hygiene. Both models also showed that respondents who were employed were less likely than their unemployed/retired counterparts to avoid large social gatherings or participate in activities outside their homes but were more likely to report practice pandemic-related hygiene. DNN models combined with impact scores to explain their output present an alternate approach to modeling outcomes in large, multi-variate cohorts which can outperform conventional statistical modeling.