AI Model Detects Hidden Heart Disease from ECGs, Outperforming Cardiologists
Researchers at Columbia University and New York-Presbyterian have developed a new AI model, EchoNext, that can detect structural heart disease (SHD) from standard 12-lead electrocardiograms (ECGs).
Trained on more than 1.2 million ECG-echocardiogram pairs from over 230,000 patients, the deep learning system achieved an area under the ROC curve (AUROC) of 85.2%, outperforming cardiologists in controlled evaluations.
SHD includes serious conditions such as heart failure, valvular heart disease, pulmonary hypertension, and left ventricular hypertrophy. These are often underdiagnosed due to limited access to echocardiography, which is typically required for definitive diagnosis.
Model Training and Clinical Validation
The EchoNext model was trained using a diverse multicenter dataset from eight hospitals affiliated with NewYork-Presbyterian. To define SHD, the study used echocardiographic findings such as reduced ejection fraction, valve abnormalities, and right or left ventricular dysfunction.
The model integrated both ECG waveform data and standard tabular ECG features (e.g., age, sex, heart rate). Validation on independent external datasets from Cedars-Sinai, UCSF, and Montreal Heart Institute confirmed strong generalization (AUROC 78–80%) despite differing patient demographics and SHD prevalence.
Head-to-Head: AI vs. Cardiologists
In a blinded survey of 150 ECGs, EchoNext achieved a diagnostic accuracy of 77.3%, surpassing the average cardiologist’s 64% accuracy. Even when cardiologists were assisted by the model’s risk score, they still trailed the AI’s standalone performance. This highlights the model’s potential to support — or even surpass — expert interpretation in diagnostic settings.
In a silent deployment across a new patient cohort with no prior cardiac imaging, EchoNext identified 3,444 high-risk patients who had not yet received echocardiograms. The model achieved a 74% positive predictive value in the follow-up subset, signaling its effectiveness in flagging undiagnosed SHD.
A separate prospective trial, DISCOVERY, tested an earlier version of the model (ValveNet) and confirmed its ability to detect SHD in 53% of high-risk patients — results that improved further when reanalyzed using EchoNext.
Public Release of Dataset and Model Weights
To accelerate research, the team has publicly released a large annotated dataset of 100,000 ECGs with corresponding echocardiographic labels, along with model weights and preprocessing code. This open-access resource enables benchmarking and further development of ECG-based AI diagnostics.
Implications for Clinical Practice
EchoNext could serve as a scalable front-line screening tool, particularly in underserved settings where echocardiography access is limited. Whether used for opportunistic screening or to inform decisions on further imaging, this AI model may help close the gap in early SHD detection.
As healthcare systems face growing cardiovascular burdens, EchoNext represents a step forward in integrating AI into routine diagnostics, offering speed, accuracy, and scalability without additional imaging costs.
