Omnibus artificial intelligence-based quantitative electrocardiographic biomarkers better predict heart failure: an external validation study

Abstract Background Heart failure (HF) is a heterogeneous disease entity that has various etiology. However, current biomarkers for HF mostly stand for systolic dysfunction in advanced stage, which limits the applicability in general populations. Purpose This study aimed to assess the additional value of omnibus artificial intelligence (AI)-based quantitative electrocardiographic (omni-QCG) biomarkers covering various aspect of cardiac structure and function for predicting heart failure in general population. Methods Participants who underwent electrocardiography (ECG) in UK Biobank were enrolled. ECGs were analyzed using a deep-learning system, which was previously trained to derive QCG scores for sixteen cardiac conditions and imaging phenotypes: left and right ventricular dysfunction, acute coronary syndrome, ST-segment elevation myocardial infarction, myocardial injury, obstructive coronary artery disease, pulmonary hypertension, pericardial effusion, atrial fibrillation, valvular heart disease, left ventricular hypertrophy, left atrial enlargement, left ventricular ejection fraction, and global longitudinal strain of left ventricle, right ventricle, and left atrium. The incremental value of omni-QCG scores for predicting incident HF at 1- and 5-year was assessed by area under the receiver operating characteristic curve (AUROC), using traditional ECG parameters (P wave duration, PR interval, QRS duration, corrected QT interval, and RR interval) and cardiac magnetic resonance (CMR)-derived cardiac structural and functional parameters (end-diastolic and end-systolic volumes of both ventricles, maximal and minimal volumes of both atria, ejection fractions of 4 chambers, left ventricular cardiac output, mass, and global wall thickness). Results Among 50,133 participants (age 64.6 ± 7.8 years; male 48.2%), HF occurred in 70 (0.1%) and 293 (0.6%) individuals at 1- and 5-year follow-up, respectively. Comparing to the model with traditional ECG parameters (AUROC 0.660 for 1-year, 0.653 for 5-year), the model with omni-QCG scores presented better prediction performance for HF (AUROC 0.836 for 1-year [P<0.001], 0.802 for 5-year [P<0.001]). When added to the model with CMR parameters for structure and function of 4 cardiac chambers (AUROC 0.862 for 1-year, 0.793 for 5-year), the omni-QCG scores significantly improved the predictive ability for 5-year HF (AUROC 0.870 for 1-year [P=0.521], 0.827 for 5-year [P<0.001]). Conclusions Omni-QCG scores improves HF prediction in general population, which highlights the utility of AI-derived ECG analysis reflecting comprehensive cardiac conditions and imaging phenotypes.Abstract Figure