AI Tool Successfully Detects Hard-to-Identify Heart Dysfunction from Standard ECGs

Researchers at Wake Forest University School of Medicine have developed ECG-based AI models that successfully detect hard-to-identify heart dysfunction, including heart failure with preserved ejection fraction. Trained on over 1 million ECGs, the models perform highly using both standard 12-lead and single-lead inputs. The single-lead model's success suggests potential adaptation for wearable devices, offering a low-cost screening tool. A clinical pilot is underway at Atrium Health Wake Forest Baptist.

ECG-based AI heart dysfunction detection

• ▪Researchers at Wake Forest University developed two ECG-based artificial intelligence models to classify subtypes of left ventricular dysfunction and detect heart failure with preserved ejection fraction.
• ▪The developed AI models analyze standard electrocardiogram data to identify reduced ejection fraction, midrange or mildly reduced ejection fraction, and heart failure with preserved ejection fraction.

Heart failure detection challenges

• ▪Heart failure with preserved ejection fraction is particularly challenging to detect in its early stages and is frequently overlooked during routine clinical evaluations.
• ▪Evaluating heart function typically requires an echocardiogram, a specialized imaging test that may not be readily available in all clinical care settings.
• ▪Heart failure affects more than 6 million Americans and stands as a leading cause of hospitalization and death.

AI model development methodology

• ▪Researchers developed the AI models using nearly 1.08 million digital ECGs from 165,243 patients at Atrium Health Wake Forest Baptist.
• ▪Combining clinical variables with ECG-AI outputs did not significantly improve performance over models built solely on ECG data.
• ▪The research team evaluated two versions of the AI tool: one utilizing 12-lead ECG inputs and another utilizing a single-lead ECG input.

Model validation across populations

• ▪Researchers validated the AI models using a separate dataset of 72,832 ECGs from 42,880 pediatric patients to assess performance in children.
• ▪The AI models were tested using a separate set of more than 72,000 ECGs from the University of Tennessee Health Science Center to evaluate performance in another patient population.

Clinical performance metrics

• ▪In pediatric patients, the 12-lead model attained areas under the curve of 0.97 for reduced ejection fraction, 0.71 for midrange ejection fraction, and 0.64 for heart failure with preserved ejection fraction.
• ▪In adults, the 12-lead AI model achieved areas under the curve of 0.90 for reduced ejection fraction, 0.81 for midrange ejection fraction, and 0.80 for heart failure with preserved ejection fraction.

Real-world implementation pilot

• ▪The study was published online on July 28, 2026, in the Journal of the American Heart Association and was partially funded by the National Heart, Lung, and Blood Institute.
• ▪The research team is piloting the AI model in a family medicine clinic at Atrium Health Wake Forest Baptist to study its performance in real-world clinical care.

2 sources