AI Maps Atrial Cardiomyopathy Using Heart Signals

ARTIFICIAL intelligence could enable non-invasive characterisation of atrial cardiomyopathy by analysing electrical signals recorded from the body surface, according to a computational study.

Researchers developed a graph neural network capable of identifying the location and severity of atrial cardiomyopathy using simulated body surface potential maps, achieving accuracies of 89% and 84%, respectively.

Investigating Non-Invasive Atrial Assessment

Atrial cardiomyopathy involves structural and electrical changes within the atria and plays an important role in the development and progression of atrial fibrillation.

However, detailed assessment of the atrial substrate frequently relies on invasive electroanatomical mapping or advanced cardiac imaging, which may not be appropriate or accessible for every patient.

Researchers therefore investigated whether artificial intelligence could extract clinically relevant information from body surface potential maps.

These recordings capture electrical activity across multiple electrodes positioned around the torso, providing more extensive spatial information than a conventional 12-lead electrocardiogram.

The researchers developed a graph neural network to analyse the relationships between electrical signals and identify patterns associated with atrial tissue abnormalities.

Model Achieved 89% Accuracy

The study used a computational database containing 14,400 simulated body surface potential maps generated using multiple atrial and torso models.

The simulations represented different locations and densities of atrial cardiomyopathy.

The model was trained to perform two classification tasks: identifying the affected atrial region and estimating disease severity.

When evaluated using previously unseen atrial and torso anatomies, the model achieved 89% accuracy for localising atrial cardiomyopathy across six anatomical regions.

Accuracy for classifying disease density across three severity categories reached 84%.

These findings suggest that electrical patterns recorded non-invasively could potentially provide information about both the distribution and extent of atrial abnormalities.

Potential Implications for Atrial Fibrillation

The researchers also investigated which electrode positions contributed most strongly to the model’s predictions.

Electrodes positioned on the back provided particularly important information, suggesting that conventional anterior chest recordings may not capture all relevant spatial features.

Performance improved as more electrodes were incorporated, highlighting the potential advantages of comprehensive body surface mapping.

However, the research remains a proof-of-concept study.

All training and testing were conducted using simulated data, and the model has not yet been validated using clinical recordings from patients with confirmed atrial cardiomyopathy.

The findings therefore do not establish that the technology can currently diagnose atrial disease or guide treatment decisions.

Prospective clinical validation will be necessary to determine whether this approach could eventually support atrial fibrillation assessment and personalised treatment planning.

Reference

Macarulla-Rodríguez M et al. A graph neural network framework for characterizing atrial cardiomyopathy from body surface potential maps. Discov Computing. 2026;29:569.