Development and Preliminary Validation of a Medicare Claims-Based Model to Predict Left Ventricular Ejection Fraction Class in Patients With Heart Failure.

Abstract

Ejection fraction (EF) class is an important predictor of treatment response in heart failure (HF); however, administrative claims databases lack information on EF, limiting their usefulness in clinical and health services research of HF. We linked Medicare claims data to electronic medical records containing EF measurements for a cohort of 11 073 patients with HF from 2 academic medical centers. A a claims-based model predicting EF class was constructed using data from center 1 ("training sample") and validated using data from center 2 ("testing sample). Linear and logistic regression models with least absolute square shrinkage operator and Bayesian information criteria were developed to select the relevant predictor variables out of the total 57 candidate variables in the training sample. Higher accuracy was noted in the testing sample with models classifying patients into 2 EF classes (reduced EF <0.45) versus preserved EF (≥0.45) when compared with classifying patients into 3 EF classes (reduced, <0.40, moderately reduced, 0.40-0.49, or preserved, ≥0.50). In the testing sample, the most efficient model had 35 predictors and resulted in 83% of patients being correctly classified (95% CI, 82%-84%). The model had positive predictive value of 0.73 (95% CI, 0.68-0.78) and 0.84 (95% CI, 0.83-0.86) and sensitivity of 0.29 (95% CI, 0.25-0.32) and 0.97 (95% CI, 0.97-0.98) for reduced and preserved EF, respectively. In addition to HF-specific diagnosis codes, other factors including age, sex, medication use, and comorbidities, such as myocardial infarction and valve disorders, were important discriminators between EF classes. The claims-based model developed in this study may be used to identify patient subgroups with specific EF class in studies evaluating the health outcomes, utilization patterns, and cost, of HF patients in routine care when EF measurements are not available.