Eur Heart J. 2026 Nov 13:ehag917. doi: 10.1093/eurheartj/ehag917. Online ahead of print.

ABSTRACT

BACKGROUND AND AIMS: A major challenge in dilated cardiomyopathy (DCM) care is the identification and management of individuals at risk before disease manifestation. An artificial intelligence (AI) model applied to electrocardiography (ECG) was developed to detect and predict the echocardiographic DCM phenotype (left ventricular dilatation with systolic dysfunction) and was evaluated for its ability to stratify the risk of incident DCM in familial cohorts.

METHODS: A deep learning model was trained to detect the echocardiographic DCM phenotype and to predict its future onset, using more than 330 000 ECG-echocardiogram pairs from the Beth Israel Deaconess Medical Center, Boston. Model performance was validated in cohorts from the UK Biobank, Zhongshan Hospital, and Antwerp University Hospital, and compared with models based on traditional ECG features and echocardiographic traits. To assess performance in familial DCM, genetically predisposed individuals identified through cascade screening at the cardiogenetics clinics of Antwerp University Hospital (n = 166, 18% DCM positive) and Maastricht University Medical Center (n = 177, 18% DCM positive) were evaluated.

RESULTS: The AI-ECG model achieved areas under the receiver operating characteristic curve of .87-.97 for diagnosing the prevalent DCM phenotype, outperforming traditional ECG parameters. AI-ECG was comparable to echocardiography for predicting future DCM (C-index .80-.87 and .80-.90, respectively). In familial cohorts, 5-year negative predictive values were 95.5% (Antwerp) and 91.5% (Maastricht), with corresponding positive predictive values of 42.4% and 39.4%, while high-risk scores were associated with a 5.1-6.4-fold increased risk of incident DCM.

CONCLUSIONS: The explainable AI-ECG model detects the echocardiographic DCM phenotype and predicts future disease, and supports the exclusion of short- to mid-term disease in genetically at-risk individuals. This low-cost tool could help reduce surveillance burden and facilitate earlier intervention.

PMID:43031488 | DOI:10.1093/eurheartj/ehag917