Abstract Body:
Introduction: Social risk factors have been shown to be related to overweight/obesity and developing related poor health outcomes. Prediction models for obesity-related outcomes in people with obesity or overweight developed within clinical care settings can be used in decision making to identify individual patient risk for poor outcomes. There have been calls to link person-level or census tract level social risk factors to Electronic Health Record (EHR) data for use in clinical decision-making. However, such methods have not been fully developed and more information is needed regarding which social risk factors are relevant for predicting risk.
Objective: To examine whether adding primary payer type and publicly available zip code level composite (Area Deprivation Index; National Walk Index; Rural-urban Commuting Area) or non-composite (poverty rate, low-income tract, median family income, population density) risk factors improve risk prediction model fit and accuracy.
Methods: EHR data for 837,373 eligible patients (18-80 years, BMI 25-80 kg/m², with 5-digit zip code to link to census data) were utilized to determine the value of adding social risk factors to validated prediction models developed for use in clinical care settings. Outcomes were times to event for type 2 diabetes (T2DM), atherosclerotic cardiovascular disease, heart failure, sleep apnea, and obesity-related cancer. Models were fit using a multi-level Cox regression with a competing risk of death. Adaptive LASSO was applied to add social risk factors to existing models. Full models with all significant social risk variables and reduced models in which highly correlated social risk variables were removed, were examined. Assessments of model fit were done utilizing the AIC/BIC and Harrel’s C-index was calculated to assess predictive ability.
Results: Several models retained some social risk factors. The model for T2DM was the only one in which social risk factors did not improve the EHR-only models, though increases in Harrel’s C-index values were small (<0.01) for the other models with social risk factors added.
Conclusions: Individual and composite zip code level social risk factors were statistically significantly associated with all outcomes and improved the models, except for T2DM, but did not meaningfully improve predictive ability. Additional studies are warranted examining effects of social risk closer to, or at, the individual level, such as census block or neighborhood.