Artificial Intelligence for the Diagnosis of Heart Failure

초록

Heart failure (HF) remains a major global health burden with high morbidity and mortality, yet its diagnosis is often challenging due to heterogeneous presentations and overlap with pulmonary or systemic diseases. Despite advances in biomarkers and imaging, current strategies frequently face limitations in accuracy and accessibility. Artificial intelligence (AI) has recently emerged as a transformative tool to address these challenges by analyzing multimodal data, including electrocardiography (ECG), echocardiography, imaging, and laboratory parameters. Several studies have demonstrated the clinical utility of AI in HF diagnosis. In Korea, an AI-based clinical decision support system (AI-CDSS) combining expert-driven knowledge with machine learning achieved diagnostic concordance exceeding 98% with HF specialists, significantly outperforming non-specialists in both retrospective and prospective cohorts. AI-ECG algorithms based on deep learning and transformer architectures have also shown excellent diagnostic accuracy. In a Korean cohort of over 3,000 patients with acute dyspnoea, a transformer-based AI-ECG distinguished cardiac from pulmonary causes with an AUC of 0.938 and 88.1% accuracy, outperforming NT-proBNP. Similarly, another clinical study of emergency department patients identified left ventricular systolic dysfunction with an AUC of 0.89 and accuracy of 85.9%, again exceeding NT-proBNP performance. Importantly, attention map analyses provide interpretability by highlighting ECG segments contributing to classification, addressing the “black box” concern. Despite these advances, challenges remain, including generalizability across populations, need for robust external validation, clinical workflow integration, and ethical considerations. Nonetheless, AI holds great promise for HF diagnosis.