Abstract MP20: The Discriminatory Characteristics of Neighborhood Socioeconomic Status in Predicting Cardiovascular Disease in Electronic Health Record Based Studies Differs by Age

Abstract

Introduction: Recent studies report an association between neighborhood residence and health outcomes. There is less information on the relative utility of neighborhood socioeconomic status (nSES) in models that predict future health outcomes and the impact that age may have on this. Objective: To quantify if nSES data alone or in concert with electronic health record (EHR) data can improve risk prediction for myocardial infarction (MI) and stroke beyond current models. Methods: Neighborhood SES was derived using the AHRQ SES index. Clinical and demographic data was obtained from the EHR of patients seen at the Duke University Health System from 2009-2015; it was split into a training set (2009-2012) and testing set (2012-2015). Age (in yrs) was categorized as young (18-44), middle age (45-64), and old (≥65). Logistic regression models were fit for each outcome over 6 time horizons (30, 90, & 180 days; 1, 2, & 3 years) using machine learning methods (least absolute shrinkage and selection operator [LASSO]) for model selection to determine if nSES improved discrimination, as measured by the c-statistic. Results: Of 106703 patients, 63% were female, 41% were Black, 2.6% had CVD, 12% had diabetes, and 29% had hypertension at baseline with mean age of 47 years. The majority of the correlation between EHR variables and nSES (r 2 =0.31) was explained by demographic information within the EHR (r 2 =0.29; p<0.01). In LASSO models incorporating EHR and SES data, EHR variables (e.g., comorbidities) were frequently selected while nSES variables (e.g., AHRQ SES index) were rarely chosen. The c-statistic for predicting MI and stroke when using nSES data with and without EHR data was higher in middle aged and older patients as compared to younger patients (Figure). Conclusions: The added value of nSES was less than expected as much variability in nSES may be phenotyped through demographic information in the EHR. In discrete instances, nSES can improve risk prediction but varies by age, clinical outcome, and time horizon.