We combined the entire cohort of self-identified European American individuals identified across the five eMERGE sites (n = 13,835 individuals) into one analysis. To define diseases, we queried all ICD9 codes from the respective EMRs from the five eMERGE sites. The PheWAS software then used these ICD9 codes to classify each person as having one of the 1,358 possible clinical phenotypes belonging to >25 patients in the populations (as noted above). For each disease, the PheWAS code defined relevant control groups for each disease or finding, such that patients with related diseases do not serve as controls for that disease (e.g., a patient with Graves disease cannot serve as a control for an analysis of thyroiditis).
We have previously found that the positive predictive value for some algorithms to establish a diagnosis from EMR data is improved by requiring the presence of multiple instances of disease-associated ICD9 codes44 (link). For example, to be considered a case for tuberculosis, a patient is required to have at least two ICD9 codes in the ranges of 10–18 (tuberculosis infections of different sites), 137 (late effects of tuberculosis) or V12.01 (personal history of tuberculosis). Accordingly, for the present study, we used a threshold of relevant ICD9 codes on two distinct days to establish that person as a “case” for a given phenotype. Controls are patients without any ICD9 codes in the corresponding control range; thus, patients with a single ICD9 case code are excluded for the analysis as neither a case nor a control. Each SNP-phenotype association test was run independently with PLINK43 (link), using logistic regression adjusted for age, gender, site (e.g., Vanderbilt, Marshfield Clinic), and the first three principal components as calculated by EIGENSTRAT, using ancestry informative markers as above41 (link). Analysis was performed assuming an additive genetic model. These data were aggregated and analyzed using Perl scripts and the R statistical package.
We have previously found that the positive predictive value for some algorithms to establish a diagnosis from EMR data is improved by requiring the presence of multiple instances of disease-associated ICD9 codes44 (link). For example, to be considered a case for tuberculosis, a patient is required to have at least two ICD9 codes in the ranges of 10–18 (tuberculosis infections of different sites), 137 (late effects of tuberculosis) or V12.01 (personal history of tuberculosis). Accordingly, for the present study, we used a threshold of relevant ICD9 codes on two distinct days to establish that person as a “case” for a given phenotype. Controls are patients without any ICD9 codes in the corresponding control range; thus, patients with a single ICD9 case code are excluded for the analysis as neither a case nor a control. Each SNP-phenotype association test was run independently with PLINK43 (link), using logistic regression adjusted for age, gender, site (e.g., Vanderbilt, Marshfield Clinic), and the first three principal components as calculated by EIGENSTRAT, using ancestry informative markers as above41 (link). Analysis was performed assuming an additive genetic model. These data were aggregated and analyzed using Perl scripts and the R statistical package.
