AI Model Predicts Patients’ Risk of Developing and Worsening Disease from ECGs

An electrocardiogram (ECG) is a test that records the electrical activity of the heart and is among the most frequently performed medical assessments globally. ECGs illustrate the flow of electrical signals within and between the heart's various chambers, including the atria and ventricles. Additionally, ECGs gather extensive information from the body, as conditions like diabetes, which affect organs such as the kidneys or liver, can also impact heart function. Cardiologists rely on their expertise and established guidelines to interpret ECGs, categorizing them into ‘normal’ and ‘abnormal’ patterns to aid in diagnosing various diseases. While artificial intelligence (AI)-enhanced ECGs are recognized for their accuracy in diagnosing heart conditions, they have not previously been employed to inform clinicians about an individual patient's risk of developing a range of specific, treatable diseases in the future. A new AI model now enables the prediction of patients’ risks for developing and worsening diseases, as well as their risk of early mortality, utilizing an ECG. This model empowers doctors to detect diseases earlier and prioritize urgent cases for intervention.

Researchers at Imperial College London (London, UK) utilized extensive datasets from international sources, encompassing millions of ECGs collected as part of routine care, to train their AI model to analyze ECGs and accurately predict which patients would go on to develop new diseases, experience disease progression, or ultimately die. The AI model was trained to interpret the flow of electrical signals within and between the atria and ventricles, identifying patterns in the electrical signals. According to the researchers, the model can discern ECG patterns with greater complexity and nuance than a cardiologist. Their findings, published in Lancet Digital Health, indicate that the AI model—referred to as AI-ECG risk estimation, or AIRE—successfully identified the risk of death within ten years following the ECG with an accuracy of 78%. In the cases where the model was incorrect, researchers suggest that unknown factors, such as subsequent treatment or unforeseen causes of death, may have played a role.

The system is capable of predicting future health risks, including heart rhythm issues, heart attacks, and heart failure, as well as the likelihood of dying from non-heart-related causes. The researchers reported a high level of accuracy in these predictions. They also analyzed imaging and genetic data, which supported their findings by confirming that the AI predictions were associated with actual biological factors in the heart's structure and function. This aspect is critical for establishing the model’s credibility with clinicians, as it demonstrates the model’s ability to detect subtle changes in the heart’s structure over time, which are early indicators of disease or mortality risk.

“Our work has shown that this AI model is a credible and reliable tool that could, in future, be programmed for use in different areas of the NHS to provide doctors with relevant risk information,” said Dr. Fu Siong Ng, senior author of the study and Reader in Cardiac Electrophysiology at the National Heart & Lung Institute atImperial College London. “This could have a positive impact on how patients are treated, and ultimately improve patient longevity and quality of life. It could also reduce waiting lists and allow more efficient allocation of resources. We believe this could have major benefits for the NHS, and globally.”