Validation of VA Administrative Data Algorithms for Identifying Cardiovascular Disease Hospitalization

To the Editor: We assembled a cohort of US veterans initiating treatment for type 2 diabetes to evaluate the comparative effectiveness of diabetes treatments on cardiovascular outcomes. To inform our strategies for outcome, exposure, and covariates, we determined the validity of algorithms to identify stroke or acute myocardial infarction (MI) hospitalization, and pharmacy and administrative claims for antidiabetic drugs and covariates. The institutional review boards of the Veterans Health Administration and Vanderbilt University approved the study. The population included veterans >18 years old receiving Veterans Health Administration healthcare between 1 January 2000 and 31 December 2007. Diagnoses were coded according to the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) system.1 The cohort was limited to patients with known sex, birth date, and a filled prescription for an antidiabetic drug. We randomly selected a sample of 400 ICD-9-CM coded hospitalizations (352 patients) indicating stroke or acute MI for medical record review. Definitions of stroke, acute MI, exposures, and covariates are available in e Table 1 (https://links.lww.com/EDE/A652).2–4TABLE: PPV and CIs for Stroke and Acute MI ICD-9-CM CodesMedical records were abstracted using structured forms, specific for stroke and acute MI. Outcomes were adjudicated using the following definitions. Stroke was defined by the rapid onset of a persistent neurologic deficit attributed to an obstruction or rupture of the arterial system lasting >24 hours unless death supervened or by a lesion compatible with an acute stroke on computed tomography or magnetic resonance imaging scan.5 An acute MI was classified as “definite or probable” using an adaption of the Women’s Health Initiative definition based on chest pain, electrocardiogram pattern, and abnormal cardiac enzymes.3 We considered the medical record as the referent, and we calculated the positive and negative predictive values (PPV/NPV) with 95% confidence intervals (CIs) for binomial proportions using Wilson’s formula.6 The Table presents the results for each ICD-9-CM code used to identify stroke or acute MI. Of 226 stroke hospitalizations in 194 patients (20 patients with two to five hospitalizations), 183 (PPV = 81% [CI = 75–86]) met criteria for an acute stroke. Of the 174 acute MI hospitalizations among 158 patients (12 patients with two to four hospitalizations), there were 148 definite and eight probable AMIs (PPV = 90% [84–93]). Filled prescriptions recorded in Veterans Health Administration pharmacy databases were compared with medications recorded in patient charts among the 400 hospitalizations. Of the 94 patients with a metformin prescription fill within 90 days before hospitalization, 83 had metformin listed as a medication at hospitalization (PPV = 88% [80–93]). The results were similar for sulfonylurea and insulin (eTable 2; https://links.lww.com/EDE/A652). Algorithms for selected covariates showed similar performance. The PPVs for aspirin (80% [72–87]) and indicators of smoking (72% [63–80]) were also high; however, sensitivity for aspirin and smoking were 79% and 56%, respectively (eTable 3; https://links.lww.com/EDE/A652). Although our algorithms can identify covariates, it is essential to assess whether validity is similar across antidiabetic exposures. eTable 4 (https://links.lww.com/EDE/A652) shows that the age-adjusted probability of smoking, as documented in the medical record, was lower among insulin users compared with those with no antidiabetic drug fill (22 vs. 45%, P = 0.007). The age-adjusted probability of aspirin by a prescription fill was slightly higher among insulin users than those with no antidiabetic drug fill (68% vs. 45%, P = 0.002). In summary, using a sample of 400 Veterans Health Administration cardiovascular hospitalizations from a large retrospective cohort of diabetic veterans, our algorithms identified stroke with a PPV of 81% and acute MI with a PPV of 89.7%. Our algorithms have reasonable validity in identifying stroke and acute MI hospitalizations in Veterans Health Administration and provide an efficient strategy to identify cardiovascular outcomes for cohort studies. Measurements of exposures and covariates through pharmacy and administrative data had good validity as well. Kurt Niesner VA Tennessee Valley Geriatric Research Education Clinical Center (GRECC) HSR&D Center Nashville, TN Harvey J. Murff VA Tennessee Valley Geriatric Research Education Clinical Center (GRECC) HSR&D Center Department of Medicine Vanderbilt University Nashville, TN Marie R. Griffin VA Tennessee Valley Geriatric Research Education Clinical Center (GRECC) HSR&D Center Departments of Medicine and Preventive Medicine Vanderbilt University Nashville, TN Brian Wasserman Department of Medicine Vanderbilt University Nashville, TN Robert Greevy VA Tennessee Valley Geriatric Research Education Clinical Center (GRECC) HSR&D Center Department of Biostatistics Vanderbilt University Nashville, TN Carlos G. Grijalva VA Tennessee Valley Geriatric Research Education Clinical Center (GRECC) HSR&D Center Department of Preventive Medicine Vanderbilt University Nashville, TN Christianne L. Roumie VA Tennessee Valley Geriatric Research Education Clinical Center (GRECC) HSR&D Center Department of Medicine Vanderbilt University Nashville, TN [email protected]