AI-powered ECG model outperforms doctors in detecting hidden heart disease

By Vijay Kumar MalesuReviewed by Susha Cheriyedath, M.Sc.Jul 21 2025

A breakthrough AI model can spot silent structural heart disease from a simple ECG, promising to catch dangerous conditions earlier, streamline patient care, and close the diagnostic gap missed by traditional screening.

Study: Detecting structural heart disease from electrocardiograms using AI. Image Credit: DC Studio / Shutterstock

In a recent study published in the journal Nature, a group of researchers investigated whether an artificial intelligence (AI) electrocardiogram (ECG) model can reliably detect diverse structural heart diseases (SHDs) across various hospitals and care settings, outperforming standard physician review. The model, called EchoNext, was developed as a multitask classifier to address collinearity among different SHD component labels.

Background

Every minute, another United States (US) patient enters the hospital with symptoms that may mask underlying SHD. Treating SHD already drains the nation of more than 100 billion dollars each year. Yet, an estimated 6.4% of older adults carry clinically significant valvular heart disease (VHD) that has never been diagnosed, in addition to 4.9% already diagnosed, making the total prevalence over 11%.

Early echocardiography saves lives, but ultrasound labs, trained readers, and patient travel costs remain barriers, leaving busy clinicians guessing whom to scan.

Large-scale digital ECG archives and modern AI offer a low-cost alternative: if one ten-second ECG could reliably uncover silent disease, scarce imaging resources could be directed to those who need them most.

Further research is needed to determine whether algorithm-guided screening improves survival and equity. Additionally, the paper discusses potential deployment strategies for such models, including both “gatekeeper” and “safety net” applications, each with unique benefits and trade-offs for clinical practice.

About the study

Investigators assembled 1,245,273 paired ECG-echocardiogram records from 230,318 adults treated between 2008 and 2022 at eight NewYork-Presbyterian (NYP) hospitals, reserving patient-level splits for training, validation, and testing.

SHD was labeled when any guideline defined abnormality was present with left ventricular ejection fraction (LVEF) ≤ 45%, left ventricular wall thickness ≥ 1.3 cm, moderate or worse right ventricular dysfunction, pulmonary artery systolic pressure (PASP) ≥ 45 mm Hg, or tricuspid regurgitation jet velocity ≥ 3.2 m/s as an alternative pulmonary hypertension definition, moderate or worse regurgitation/stenosis of any valve, or a moderate/large pericardial effusion.

The authors note these thresholds are somewhat arbitrary, as different studies and guidelines may use varying cutoffs.

A convolutional neural network named EchoNext ingested the raw 12-lead waveform, along with seven routine ECG parameters and age/sex data. Performance was first measured on a held-out NYP test set, and then on external cohorts from Cedars-Sinai, the Montreal Heart Institute, and the University of California, San Francisco.

Generalization across age, sex, race, ethnicity, and clinical context was assessed. Silent “shadow” deployment ran EchoNext on 84,875 consecutive ECGs from patients without previous echocardiography, storing scores but not influencing care.

Finally, a single-site pilot, Detecting Structural Heart Disease Using Deep Learning on an Electrocardiographic Waveform Array (DISCOVERY), prospectively invited adults with no recent imaging to undergo echocardiography stratified by a predecessor model’s risk score; EchoNext was analyzed post hoc.

Study results

EchoNext, an AI-powered ECG model, excelled in retrospective analysis. Within the eight-hospital NYP test set, it detected composite SHD with an area under the receiver operating characteristic (AUROC) of 85.2% and an area under the precision–recall curve (AUPRC) of 78.5%. Accuracy remained consistent across academic and community campuses and did not falter when training and test sites were exchanged, demonstrating generalization.

External validation at Cedars-Sinai Medical Center, the Montreal Heart Institute (MHI), and the University of California, San Francisco, yielded AUROC values of 78 to 80%, despite higher disease prevalence.

Disease-specific performance: LVEF ≤ 45% achieved AUROC 90.4%, while PASP ≥ 45 millimeters of mercury reached 82.7%. The authors emphasize that AUPRC values for component diseases are highly dependent on the underlying disease prevalence and should not be directly compared across conditions or use cases.

A 150-trace reader study compared EchoNext with thirteen cardiologists. Reviewing wide age, sex, waveform, and ECG intervals, physicians identified SHD correctly in 64% of cases. The AI alone achieved 77% accuracy, and when clinicians were shown the algorithmic risk score, their accuracy increased modestly to 69%, underscoring that the model captured prognostic patterns that were hidden from expert eyes. It is important to note that cardiologists in this assessment had access only to de-identified ECGs and routine parameters, without any clinical context, which is not typical of standard clinical care.

To estimate clinical opportunity at scale, the team silently ran EchoNext on 124,027 ECGs recorded in 2023 from 84,875 adults who had never undergone echocardiography. The model flagged nine percent of traces as high risk. Usual care, nevertheless, left 45% of these individuals without follow-up imaging, suggesting that an estimated 1,998 cases of silent SHD might have been intercepted had the alert been live, based on modelled prevalence and sensitivity scenarios provided in the paper.

Among the 15,094 patients who eventually received echocardiography, EchoNext preserved accuracy (AUROC 83%; AUPRC 81%) and delivered a positive predictive value of 74%, reinforcing its reliability in a contemporary workflow. The paper also provides modelled performance estimates at different prevalence scenarios and sensitivity thresholds, underscoring the practical implications for population-wide screening.

Prospective evidence came from the DISCOVERY pilot, which recruited 100 imaging-naive adults. Post hoc EchoNext scoring revealed clear tiers, with previously unrecognized SHD present in 73% of high-risk participants, 28% of moderate-risk participants, and 6% of low-risk participants; moderate to severe left-sided VHD followed a similar gradient.

These results illustrate the model’s capacity to triage scarce echocardiography resources toward those most likely to benefit, while sparing low-risk individuals unnecessary testing. The original trial used a predecessor model (ValveNet) to stratify risk and recruit participants, and the EchoNext model was applied retrospectively to these participants for further analysis.

Conclusions

To summarize, EchoNext demonstrates that an AI-enhanced ECG can detect SHD associated with LVEF reduction, elevated PASP, and significant VHD, with AUROC and AUPRC metrics superior to those of cardiologists. By flagging high-risk patients for timely echocardiography, the algorithm promises to shrink diagnostic delay and the billion-dollar burden of SHD while maintaining equity across sites and demographics. However, the authors caution that AI-based screening may also carry potential risks, including patient anxiety from false positives or bias in clinical adoption, and highlight the need for further study of these aspects.

The public release of code and data encourages independent validation; however, large pragmatic trials must verify that AI-guided ECG screening truly improves survival, quality of life, and healthcare value. Notably, the authors have released a large de-identified dataset and a benchmark AI model (the Columbia mini-model) to support further research and enable transparent comparison of future algorithms.