ECG-based artificial intelligence for classifying left ventricular dysfunction and heart failure with preserved ejection fraction

Authors

• Ibrahim Karabayir, Department of Cardiovascular Medicine Wake Forest School of Medicine Winston-Salem NC USA.
• Olivia Gilbert, Department of Cardiovascular Medicine Wake Forest School of Medicine Winston-Salem NC USA.
• Ali Valika, Advocate Health - MidwestFollow
• Turgay Celik, Department of Cardiovascular Medicine Wake Forest School of Medicine Winston-Salem NC USA.
• Arunkumar Krishnan, Department of Medicine Wake Forest University School of Medicine Winston-Salem NC USA.
• Lokesh Chinthala, Center for Biomedical Informatics University of Tennessee Health Science Center Memphis TN USA.
• Ambarish Pandey, Department of Internal Medicine, Division of Cardiology and Geriatrics UT Southwestern Medical Center Dallas TX USA.
• Elsayed Z. Soliman, Department of Cardiovascular Medicine Wake Forest School of Medicine Winston-Salem NC USA.
• John L. Jefferies, School of Public Health University of Memphis Memphis TN USA.
• Dalane Kitzman, Department of Cardiovascular Medicine Wake Forest School of Medicine Winston-Salem NC USA.
• David Herrington, Department of Cardiovascular Medicine Wake Forest School of Medicine Winston-Salem NC USA.
• Robert L. Davis, Center for Biomedical Informatics University of Tennessee Health Science Center Memphis TN USA.
• Oguz Akbilgic, Department of Cardiovascular Medicine Wake Forest School of Medicine Winston-Salem NC USA.

Recommended Citation

Karabayir I, Gilbert O, Valika A, et al. ECG-Based Artificial Intelligence for Classifying Left Ventricular Dysfunction and Heart Failure With Preserved Ejection Fraction. J Am Heart Assoc. 2026;15(15):e041948. doi:10.1161/JAHA.124.041948

Affiliations

Advocate Good Samaritan Hospital, Wake Forest School of Medicine, Atrium Health Levine Cancer Center

Abstract

Background: Left ventricular (LV) dysfunction and heart failure with preserved ejection fraction (HFpEF) often present with early signs that are frequently overlooked or attributed to other conditions. This study proposes a novel, externally validated artificial intelligence (AI) tool using ECG data (ECG-AI) for the simultaneous detection of subtypes of LV dysfunction and HFpEF.

Methods: Two ECG-AI models, using 12-lead or single-lead ECG, were developed using data from Atrium Health Wake Forest Baptist and University of Tennessee Health Science Center (UTHSC) to classify ECGs into 4 categories: reduced LV ejection fraction (rEF; EF<40), midrange EF (mEF; 40≤mEF<50), HFpEF, and controls (no LV dysfunction or HFpEF). Next, a boosting algorithm was used to incorporate clinical risk factors. Finally, the models were independently validated on pediatric data from UTHSC.

Results: The Atrium Health Wake Forest Baptist cohort included 1 078 198 digital ECGs from 165 243 patients (5% rEF, 8% mEF, 2% HFpEF), and the UTHSC external validation cohort comprised 72 832 ECGs from 42 880 patients (1% rEF, 1% mEF, 1% HFpEF). For the 12-lead (Lead-I) ECG-AI model, areas under the curve for rEF/mEF/HFpEF were 0.90/0.81/0.80 (0.89/0.78/0.75) in the Atrium Health Wake Forest Baptist holdout data and 0.92/0.76/0.73 (0.90/0.75/0.74) in the UTHSC data. Clinical data-only machine-learning models lacked generalizability; clinical + ECG-AI models did not show any significant improvement compared with ECG-alone models. With 8418 ECGs (142 cases, 8276 controls) from UTHSC pediatric data, the 12-lead model achieved areas under the curve of 0.97/0.71/0.64 for rEF/mEF/HFpEF, and the Lead I model had areas under the curve of 0.94/0.77/0.67.

Conclusions: ECG-AI can accurately detect LV dysfunction and HFpEF even from single-lead ECG, enabling potential low-cost screening.

Document Type

Article

PubMed ID

42517500