Binary versus multiclass deep learning models for left ventricular ejection Fraction classification from the electrocardiogram: The intermediate range problem
LVEF is classified into reduced (≤40%), mildly reduced (41–49%), and preserved (≥50%) categories. ECG-based deep learning models may be formulated as binary or multiclass classifiers, but their comparative performance across all three strata in a shared pipeline has not been established. In the present study, we trained an autoencoder (AE) on 12-lead median-beat ECGs from 10,430 ECG-echocardiography pairs from 8,226 patients at the University Medical Centre Groningen to learn latent representations, which were subsequently classified by a multilayer perceptron (MLP) classifier. We systematically compared binary and multiclass LVEF classification across three clinically relevant LVEF thresholds: reduced (≤40%), mildly reduced (41–49%), and preserved (≥50%). Binary models achieved AUROCs of 0.897, 0.664, and 0.849 for ≤40%, 41–49%, and ≥50%, respectively, with comparable multiclass one-vs-rest AUROCs of 0.896, 0.660, and 0.845 (all P > 0.05). Both formulations performed poorest in the intermediate stratum. Our findings demonstrate that binary and multiclass AI–ECG LVEF models performed similarly across strata within the same AE–MLP framework, but both were weakest in the intermediate (41–49%) range.
Authors
- Michiel Rienstra (ORCID: https://orcid.org/0000-0002-2581-070X)
- Jan Walter Benjamins (ORCID: https://orcid.org/0000-0002-5701-0539)
- Rajiv S. Rama (ORCID: https://orcid.org/0000-0001-7562-8650)
Institutions
- University Medical Center Groningen (NL)
- University of Groningen (NL)
Publication Details
- Journal
- European Heart Journal - Digital Health
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1093/ehjdh/ztag183
- Primary Topic
- ECG Monitoring and Analysis
- Type
- article
- Field-Weighted Citation Impact
- 0.00