Medicine
AI can detect heart conditions from electrocardiograms with greater accuracy than cardiologists
|
Ainhoa Pérez
Alumni
|
|
MPH, MSc.
|
A recent study published in Nature evaluated the effectiveness of an artificial intelligence model trained with more than one million electrocardiograms (ECGs). The results show that AI can detect structural heart disease more accurately than cardiologists, opening up new possibilities for non-invasive screening.
PATHOPHYSIOLOGY AND MECHANISMS
Structural heart disease encompasses a heterogeneous spectrum of morphofunctional alterations affecting the valves, chambers, and walls of the heart, including ventricular dysfunction, hypertrophy, pulmonary hypertension, and valvular heart disease. These conditions affect tens of millions of people worldwide. Although diagnosis is based on imaging techniques such as echocardiography, their availability is limited by logistical and cost barriers. In this context, deep learning-based artificial intelligence models have emerged as promising tools for non-invasively inferring the presence of structural disease from electrocardiograms, which are more accessible and widely available tests. These models use convolutional neural networks that analyze electrical signals together with clinical data, which could facilitate population screening and the prioritization of additional studies.STUDY
A recent clinical study conducted in the United States (Poterucha et al., 2025) evaluated the effectiveness of a deep learning-based AI model for detecting structural heart disease from electrocardiograms. The model was trained with more than one million pairs of electrocardiograms and echocardiograms from more than 200,000 adult patients treated between 2008 and 2022 at eight hospitals in the NewYork-Presbyterian system. Structural disease was defined using established echocardiographic criteria, including left ventricular dysfunction, ventricular hypertrophy, right ventricular dysfunction, pulmonary hypertension, moderate or severe pericardial effusion, and relevant valvular heart disease. The model used raw signals from 12 electrocardiogram leads along with basic clinical variables. Its performance was validated with different clinical subgroups, including external cohorts, routine screening models without imaging, and a controlled evaluation in which a group of cardiologists interpreted 150 electrocardiograms with and without the aid of the model, comparing its diagnostic performance against the algorithm.MAIN RESULTS
The AI model achieved high accuracy in the test cohort (AUROC > 85%; AUPRC >78%), with consistent results across hospitals, clinical settings, and demographic groups, with some decrease in performance in external cohorts. In the controlled evaluation, the model outperformed cardiologists in accuracy, sensitivity, and specificity, improving their performance when used as support.CONCLUSION AND CLINICAL RELEVANCE
This study shows that a trained artificial intelligence model can accurately and generalizably detect multiple forms of structural heart disease from electrocardiograms. These findings suggest a potential role for this technology in population screening and diagnostic prioritization in cardiology. Diagnostic performance was high in internal and external validations and superior to clinical judgment under experimental conditions. Silent deployment showed a high diagnostic confirmation rate, and its prospective application allowed the identification of real cases. Among the limitations of the study are the dependence on echocardiography as the diagnostic standard and the lack of prognostic results to establish its real clinical value, among others. It should be noted that the evaluation performed by cardiologists is not carried out under complete clinical conditions that include medical history and physical examination, which limits comparison in real conditions, pointing for the moment simply to its value as a screening tool. However, high-risk ECGs in the silent cohort show high predictive value. Future studies should evaluate its applicability in other clinical settings and its real impact on healthcare practice.
#artificialintelligence #AI #cardiovasculardisease
References:
Poterucha, J. T et al, 2025. Detecting structural heart disease from electrocardiograms using AI. Nature. https://doi.org/10.1038/s41586-025-09227-0
* The news published on studies do not represent an official position of ICNS, nor a clinical recommendation.
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
