Abstract 26: Validity of Cardiovascular Data From Electronic Data Research Networks: The Multi-Ethnic Study of Atherosclerosis (MESA) and HealthLNK

Abstract

Background: Understanding the validity of data from electronic clinical data research networks compared to population-based CVD cohorts, the gold standard for epidemiological research, is essential for conducting epidemiological research across large, diverse populations more efficiently. Methods: We linked individual-level data from MESA with HealthLNK, a 2006-2012 database of electronic health records (EHRs) from 6 major Chicago hospitals. To evaluate for correlation and bias for blood pressure (BP) and BMI between HealthLNK and in-person MESA examinations, we used Pearson Correlation Coefficients and Bland-Altman plots. Median BMI and BP values between 2006-2008 in HealthLNK were compared to MESA exam 4 and between 2009-2012 were compared to MESA exam 5. Using diagnoses in MESA as the gold standard, we also calculated the performance of HealthLNK queries for hypertension (HTN) and obesity using ICD9 codes alone and with the addition of clinical data. Results: Of the 1164 MESA participants at the Chicago field center, 802 participants had data in HealthLNK. There was a high correlation between BMI in MESA and HealthLNK (0.94). HealthLNK only slightly underestimated BMI by 0.5 kg/m 2 (Figure). The correlation was lower for systolic BP and diastolic BP (0.3 and 0.4, respectively). Compared to MESA, HealthLNK significantly overestimated SBP and DBP by 6.5 and 3.6 mmHg, respectively. Using ICD9 codes alone, the sensitivity and specificity for HTN were 62.2% and 72.7% and for obesity were 28.1% and 97.4%. The addition of BP and medications to ICD9 codes increased sensitivity and decreased specificity for HTN to 73.2% and 53.2%. The addition of BMI increased sensitivity and decreased the specificity for obesity to 64.3% and 95.4%. Conclusion: Significant disagreement between risk factor status and BP exists between HealthLNK and MESA, while BMI is highly correlated. Identifying areas of concordance and discordance informs our understanding of the strengths and limitations of using EHR data for epidemiological research.