AI Model EchoNext Detects Structural Heart Disease From Routine ECGs

EchoNext is an artificial intelligence model designed to identify structural heart disease from routine electrocardiograms (ECGs), potentially allowing a widely available heart test to flag conditions that traditionally require cardiac imaging for confirmation. Trained using more than 1 million ECG and imaging records, the model demonstrated strong performance across multiple health systems and identified previously undiagnosed disease in a prospective clinical evaluation. More recent research published in 2026 is also examining how well the model transfers to different clinical populations, highlighting both its promise and the need for careful validation before widespread deployment.

Turning a Routine Heart Test Into an AI Screening Tool

An electrocardiogram takes only a few minutes and is among the most commonly used tests in cardiovascular medicine. It records the heart’s electrical activity through electrodes placed on the body, helping clinicians identify abnormalities in rhythm and electrical conduction.

But an ECG does not directly produce the kind of anatomical information provided by an echocardiogram, which uses ultrasound to examine the heart’s chambers, valves and pumping function.

That distinction is where EchoNext enters the picture.

Researchers developed EchoNext as a deep-learning system that analyzes ECG information to identify patterns associated with structural heart disease — conditions affecting the heart’s valves, chambers or muscle. The model was trained using more than 1 million paired heart-rhythm and imaging records collected across a large healthcare system.

The goal is not to replace echocardiography. Instead, the technology is designed as a screening layer that could help identify people who may benefit from additional cardiac imaging.

What EchoNext Actually Detects

Structural heart disease is a broad category rather than a single diagnosis.

It can include conditions such as heart failure associated with impaired ventricular function, abnormalities of the heart valves, pulmonary hypertension and abnormal thickening of the heart muscle.

Some of these conditions can remain undiagnosed because symptoms may be absent, nonspecific or attributed to other causes. Confirming them can require imaging such as transthoracic echocardiography.

EchoNext approaches the problem differently. Its AI model takes information from an ECG, along with selected patient information, and estimates whether structural heart disease is present.

In the original research, the model was designed to detect a broad composite of clinically important structural heart diseases rather than focusing on only one condition. This broad approach is significant because a single ECG may contain subtle patterns associated with several different forms of cardiac disease.

More Than One Million Records Used for Training

One of the defining features of EchoNext is the scale of the data used to develop it.

The original model was trained using more than 1 million heart-rhythm and imaging records across a large and diverse healthcare system. Researchers then evaluated its performance in internal and external settings to examine whether its predictions remained useful outside the environment in which it was developed.

The Columbia research group reports that the broader EchoNext development program drew on 1,245,273 ECG-echocardiogram pairs from eight hospitals and 190 clinics. Its reported development and validation results included an area under the receiver operating characteristic curve of about 85% in internal testing and approximately 78–80% in external validation.

These numbers should not be interpreted as meaning that the system is correct in 85% or 80% of individual diagnoses. AUROC measures how well a model distinguishes between patients with and without a target condition across different decision thresholds.

That distinction is particularly important for medical AI, where a screening model’s usefulness depends on more than a single accuracy figure.

AI Found Signals That Can Be Difficult to See

In a controlled evaluation involving ECG interpretations, the researchers reported that EchoNext achieved higher diagnostic performance than participating cardiologists for the study’s structural-heart-disease task. The researchers also tested the system across multiple healthcare settings and reported relatively consistent performance across demographic groups.

The underlying idea is not that AI sees something mystical in an ECG.

Deep-learning models can identify statistical relationships across thousands or millions of examples that may be difficult to recognize consistently through conventional visual interpretation. In EchoNext’s case, the training process connects electrical patterns in ECGs with structural abnormalities established through echocardiographic evaluation.

Once trained, the model can examine a new ECG and produce a prediction about the likelihood of structural heart disease.

That makes the technology potentially useful as an additional screening signal, particularly in settings where access to advanced cardiac imaging is limited.

The Model Has Also Been Tested Prospectively

A particularly important step in the EchoNext research was moving beyond retrospective datasets.

The researchers conducted a prospective clinical evaluation involving patients who had not previously undergone cardiac imaging. The AI system was used to identify people whose ECGs suggested possible structural heart disease, after which appropriate cardiac evaluation could be performed.

The original study reported that the system successfully identified previously undiagnosed structural heart disease in this setting.

This type of testing matters because an AI system can perform well on historical data without necessarily delivering the same value when introduced into real clinical workflows.

Prospective evaluation provides a more realistic test of whether predictions can help identify disease that was not already known.

EchoNext Is Not a Replacement for an Echocardiogram

Despite the potential of ECG-based AI, the technology has an important limitation: an AI prediction is not the same thing as a definitive anatomical examination.

An echocardiogram can directly visualize heart structures and measure characteristics such as ventricular function, chamber dimensions and valve behavior.

EchoNext instead infers the possibility of structural disease from electrical information.

That makes it better understood as a screening or decision-support technology rather than an autonomous diagnostic replacement.

A positive AI prediction may indicate that additional testing should be considered, while a negative prediction does not necessarily eliminate the need for clinical evaluation when symptoms or other evidence point toward heart disease.

New Research Is Testing Whether the AI Travels Well

Another important question for medical AI is whether a model developed in one healthcare environment performs reliably somewhere else.

A September 2026 study published in the European Heart Journal – Digital Health evaluated the transportability and recalibration of EchoNext-Mini using ECG data from the MIMIC-IV database. The researchers examined whether the released model could detect echocardiographically defined structural heart disease in a different patient population and clinical environment.

This type of external validation is essential because differences in patient populations, hospitals, ECG equipment, clinical workflows and disease prevalence can influence AI performance.

The EchoNext research ecosystem has also made a substantial portion of its work available for further investigation. PhysioNet now hosts a 100,000-ECG dataset containing ECG recordings paired with structural-heart-disease labels derived from echocardiography, supporting reproducibility and additional benchmarking.

From Screening to Earlier Detection

The broader significance of EchoNext lies in how it could change the role of a test that already exists in many clinical settings.

If an ECG can be used not only to examine electrical rhythm but also to flag hidden structural abnormalities, the same test could potentially provide an additional layer of cardiovascular screening without requiring every patient to undergo imaging immediately.

That could be particularly relevant in primary care, emergency medicine and other settings where ECGs are already routinely collected.

However, whether such a system improves patient outcomes at scale is a separate question from whether it can detect disease in a research setting. Implementation would require prospective studies, appropriate clinical workflows, evaluation of false positives and false negatives, regulatory considerations and evidence that AI-assisted screening leads to meaningful improvements in care.

The Next Phase of AI in Cardiology

EchoNext illustrates a broader direction in medical artificial intelligence: extracting additional clinical information from tests that physicians already use.

Rather than introducing an entirely new diagnostic device, researchers are asking whether machine learning can uncover signals hidden within familiar measurements.

For cardiovascular medicine, the attraction is clear. ECGs are comparatively inexpensive, fast and widely available, while echocardiography requires specialized equipment and trained personnel.

EchoNext does not eliminate that difference, nor does it turn an ECG into an echocardiogram. What it demonstrates is that electrical signals from the heart may contain more information about structural disease than conventional interpretation alone can reveal.

The challenge now moves from proving that AI can detect these patterns to establishing where such systems can be safely and effectively integrated into real-world healthcare.

If those questions are answered successfully, AI-powered ECG screening could become another layer in the effort to find structural heart disease earlier — before an undiagnosed condition becomes a more serious clinical problem.

FAQs

1. What is EchoNext AI?

EchoNext is a deep-learning model designed to identify structural heart disease from electrocardiograms by learning patterns associated with abnormalities confirmed through cardiac imaging.

2. How many records were used to develop EchoNext?

The original EchoNext model was trained using more than 1 million heart-rhythm and imaging records. The broader development program included more than 1.24 million ECG-echocardiogram pairs across eight hospitals and 190 clinics.

3. Can EchoNext replace an echocardiogram?

No. EchoNext is designed as an AI-based screening and decision-support tool. Echocardiography remains an important method for directly evaluating the structure and function of the heart.

4. What heart conditions can EchoNext detect?

The model was developed to identify a broad group of structural heart diseases, including abnormalities involving heart valves, ventricular function, heart-muscle thickness and pulmonary hypertension.

5. Why is AI-powered ECG screening important?

Because ECGs are already widely used and comparatively accessible, AI analysis could potentially help identify people who may need further cardiac evaluation without requiring every person to undergo advanced imaging as the first screening step.