(149) Two Component Claims-Based Algorithm to Identify Heart Failure Ejection Fraction Phenotypes

Publication Number: 1194

Background: SGLT2is have been proven to have beneficial effects for HF patients in recent trials but EF cutoffs in the inclusion criteria have limited use across the HF spectrum (HFrEF, HFmrEF and HFpEF).Previous claims algorithms have been distinguished two HF phenotypes.

Objectives: An effective algorithm for distinguishing the three HF phenotypes in claims data could be beneficial in estimating population sizes, disease burden, and identifying differences and commonalities across the phenotypes.

Methods: Heart failure patients were identified as HFrEF (EF<=40%), HfmrEF (40% < EF < 50%), and HFpEF (EF>= 50%) according to the first EF measurement in a US claims-EMR data set from January 2013 to September 2019. In the 20,608 patients with EF measures (4,172 HFrEF, 1,805 HFmrEF, 14,631 HFpEF), separate logistic regression algorithms were defined for HFrEF vs. not HFrEF and HFpEF vs not HFpEF in 400 bootstrap samples of 900 patients of each phenotype. Covariates included comorbidities and treatments prior to index (EF measurement date). Optimal logistic models (minimum AIC) were used to calculate ROC curves to define the probability (pscore) cut point for optimum sensitivity and specificity for HFrEF vs. not HFrEF (HFrEF optimum) and HFpEF vs not HFpEF (HFpEF optimum) separately. HFrEF and HFpEF pscores were predicted in a population without EF measurements. Patients were categorized as predicted HFrEF (pHFrEF) (HFrEF pscore ≥ HFrEF optimum and ≤ HFpEF optimum; pHFpEF (HFpEF pscore ≥ HFpEF optimum and ≤ HFrEF optimum); pHFmrEF-lo (pscores < HFrEF and HFpEF optima) or pHFmrEF-hi (pscores > HFrEF and HFpEF optima).

Results: The HFrEF and HFpEF bootstrap models had maximum AUC of 0.70 and 0.72 and median AUC of 0.68 and 0.70, respectively. The distributions of sensitivities (IQR [0.58, 0.67] and [0.62,0.70]), specificities (IQR [0.59, 0.67] and [0.60,0.67]), and positive predictive values (PPV) (IQR [0.45, 0.48] and [0.47, 0.49]) were narrow. In the 94,458 patients without EF measurements 34.2% were pHFrEF, 13.4% were pHFmrEF, and 52.4% were pHFpEF patients. Predicted classification agreed with the original diagnosis in 51.4% of HFrEF and 63.1% of HFpEF patients. Patients with an unspecified EF diagnosis code (HFuEF) were classified as 30.8% pHFrEF, 12.2% pHfmrEF, 55.5% pHFpEF. pHFrEF and pHFpEF have distinct covariate patterns; pHFmrEF-lo is intermediate to pHFrEF and pHFpEF; pHFmrEF-hi have high prevalence of many covariates.

Conclusions: A two component model to distinguish EF phenotypes in claims data produces acceptable models ( AUC ~0.70),identifying three distinct HF phenotypes. Algorithm predicts a higher proportion of HFpEF patients than indicated by diagnosis codes, but less than indicated by EF measurements.