One ECG to Identify Who Should Get an Echocardiogram: EchoNext Receives FDA Clearance

Electrocardiograms are inexpensive, fast, and widely available, but they are not the primary tool for examining the structure of the ventricles, heart muscle, and valves. Now, the artificial intelligence software EchoNext seeks to detect clues in the heart’s electrical signals that are difficult for the human eye to discern, first identifying people who may need an echocardiogram. The U.S. Food and Drug Administration (FDA) has cleared the software through the traditional 510(k) pathway as a clinical notification and triage tool for structural heart disease.

EchoNext was developed by a research team at NewYork-Presbyterian Hospital and Columbia University and later commercialized by the spinout company Pathwai. It analyzes digital waveforms from resting 12-lead ECGs in adults and, in addition to an overall signal for structural heart disease, can flag six categories of abnormalities: left and right ventricular systolic dysfunction, severe left ventricular hypertrophy, moderate-or-greater aortic stenosis, valvular regurgitation, and pulmonary hypertension. Patients with pacemakers are outside the cleared intended-use population.

The development team said the model was trained on more than 700,000 paired ECG and echocardiogram records from over 230,000 patients spanning 14 years. FDA documents also state that formal performance validation included 102,219 patients across four healthcare systems in the United States and Canada. Each patient underwent an ECG and transthoracic echocardiogram within 30 days, with the latter serving as the reference standard for determining structural abnormalities.

In this validation population, the prevalence of structural heart disease was 34.8%. EchoNext had 73.7% sensitivity, 75.2% specificity, a positive predictive value of 61.6%, and a negative predictive value of 84.6% for overall disease. Performance varied across the six subcategories; for example, sensitivity reached 80.1% for left ventricular systolic dysfunction, while it was 54.9% for aortic stenosis and 48.2% for left ventricular hypertrophy. This means the model is better suited to providing clues for the next diagnostic step, and a single negative result cannot be used as a basis for ruling out disease.

The research team also prospectively deployed the model in nearly 85,000 patients who had never previously undergone an echocardiogram, with about 9% flagged as high risk. About half of those flagged subsequently received their first echocardiogram, and nearly three-quarters of them were diagnosed with structural heart disease. However, outcomes remain unknown for those who did not undergo follow-up imaging, and those willing or scheduled to be examined may already have had higher clinical risk. This high diagnostic yield therefore cannot be interpreted directly as the effectiveness of general screening.

Another comparison involving 3,200 ECG interpretations found that the model identified structural heart disease with 77% accuracy, compared with 64% for cardiologists reviewing the ECG alone; physicians using AI achieved 69%. The result suggests that the algorithm can capture signals not necessarily targeted by conventional interpretation, but it also shows that placing the model within a clinical workflow does not necessarily produce the same performance as testing the model alone. How alerts are presented and how physicians act on them are both part of whether deployment succeeds or fails.

The core of the FDA’s decision was its determination that EchoNext is “substantially equivalent” to an existing legally marketed device, not proof that it can replace echocardiography or improve survival. In addition to plans to sell the product to hospitals, the company is partnering with the clinical information platform OpenEvidence to allow physicians to submit ECGs and receive risk predictions. However, whether the platform’s image-upload workflow fully corresponds to the compatible digital waveform input reviewed by the FDA still needs to be confirmed against the product labeling and actual deployment. The more critical questions for the next stage are whether performance can be maintained across institutions, how many unnecessary examinations it will add, and whether earlier disease detection will truly change treatment and patient outcomes.