Assessing the Accuracy of Algorithms That Predict Ejection Fraction for Research in a Cohort of Patients With Heart Failure

Abstract

Left ventricular ejection fraction (EF) is required to differentiate heart failure (HF) phenotypes but is not collected in national datasets in Aotearoa/New Zealand. This study assessed the accuracy of models derived from the SWEDE-HF registry for predicting HF phenotypes in patients with HF. Six SWEDE-HF models were applied to the PEOPLE study: three predicted EF <50% vs ≥50% and three predicted <40% vs ≥40%. For each EF threshold, Model A used all predictors (age, sex, ischaemic heart disease, anaemia, atrial fibrillation, chronic obstructive pulmonary disease, diabetes, hypertension, valve disease, cancer, device therapy, angiotensin-converting enzyme inhibitors/angiotensin receptor blockers, beta blockers, mineralocorticoid receptor antagonists, digoxin, diuretics, New York Heart Association, mean arterial pressure, heart rate, body mass index, estimated glomerular filtration rate, and N-terminal pro B-type natriuretic peptide [NT-proBNP]); Model B excluded NT-proBNP; Model C excluded clinical measurements (last six in list). Agreement between actual and predicted EF categories was assessed using Cohen’s kappa, which assesses the degree of agreement beyond that expected by chance. Of 941 people with HF (mean age 69 years, 30% women), 610 (65%) had EF <50% and 476 (51%) had EF <40%. Model A predicting EF <40%/≥40% resulted in the greatest agreement between predicted and actual EF category beyond chance (kappa 22.9%). Removing NT-proBNP (Model B) or clinical predictors (Model C) reduced kappa to 17.2% and 19.7%, respectively. For models predicting EF <50%/≥50%, kappa for each model was 9.3%, 4.9%, and 12.5%, respectively. Models that predict EF <40% classify NZ patients into the correct EF group up to 23% better than chance. This is unlikely to be an acceptable degree of accuracy for determining EF phenotype when assessing administrative datasets involving patients with HF in NZ.