Algorithm Predicts Ejection Fraction Subphenotypes in Heart Failure

human heart and stethoscope
Educational model of human heart
An algorithm that uses routine clinical values was able to accurately predict heart failure subphenotypes.

An algorithm that uses routine clinical values was able to accurately predict heart failure (HF) subphenotypes, according to study findings published in ESC Heart Failure.

In this multivariable, multinomial analysis and external validation study, data from 42,061 patients with HF from the Swedish Heart Failure Registry collected between 2000 and 2012 were examined. Regression models including 22 variables were used to predict whether a patient had high vs low EF and to determine the HF subtypes (ie, HF with reduced ejection fraction [HFrEF], HF with mid-range EF [HFmrEF], or HF with preserved HF [HFpEF]). Data from an independent cohort (n=10,627) from the Chronic Heart Failure European Society of Cardiology-guideline based Cardiology Practice Quality project in the Netherlands, which were collected between 2013 and 2016, were used to validate the prediction algorithm.

In the Swedish cohort, 55.6% of patient had HFrEF, 21.4% had HFmrEF, and 22.9% had HFpEF. The following baseline characteristics varied for the 3 groups: age (P <.001), sex (P <.001), all HF measurements (P <.001), all clinical variables (P ≤.001), all prescribed therapies (P <.001), and all comorbidities (P ≤.035) except for the rate of peripheral artery disease (P =.338).

The factors that most strongly predicted an EF ≥50%, with an odd’s ratio (OR) of >1.5, were age, sex, hypertension, anemia, and atrial fibrillation. Factors that predicted an EF <50% (with an OR <0.5) were therapeutic device, renin angiotensin system inhibitor use, and high N-terminal pro-B-type natriuretic peptide levels. This model had a C-statistic of 0.775 (95% CI, 0.77-0.78).

Patients with HFrEF or HFpEF vs HFmrEF were older, more often women, had a higher body mass index, and atrial fibrillation. Results for HFrEF vs HFmrEF were C-statistic = 0.758 (95% CI, 0.754-0.763) and for HFpEF vs HFmrEF C-statistic = 0.775 (95% CI, 0.770-0.780).

The validation cohort had similar results as the Swedish cohort. Specifically, the C-statistic for discerning between EF ≥50% and <50% was 0.76 (95% CI, 0.75–0.76).

Study limitations included the fact that the Swedish cohort collected EF as a categorical variable, preventing the calculation of linear associations between predictors. In addition, the best derived model from this study had difficulty in classifying HFmrEF, so the investigators recommended pooling HFmrEF individuals with either HFpEF or HFrEF to avoid misclassifications.

“We created an algorithm based on patient demographics, clinical characteristics and use of treatments to identify EF subphenotypes in HF patients without an available EF assessment,” noted the study authors. “Accuracy was good for the prediction of HFpEF and HFrEF but lower for HFmrEF, perhaps due to the heterogeneity that characterizes this subphenotype. Our model could significantly support more effective research in the ‘big data’ setting.”

Disclosure: Multiple authors declared affiliations with industry. Please refer to the original article for a full list of disclosures.


Uijl A, Lund L H, Vaartjes I, et al. A registry-based algorithm to predict ejection fraction in patients with heart failure. [Published online June 17, 2020] ESC Heart Fail. doi:10.1002/ehf2.12779