A graph neural network framework for characterizing atrial cardiomyopathy from body surface potential maps

Abstract Purpose Atrial cardiomyopathy (ACM) plays a key role in the development and progression of atrial fibrillation, but its assessment currently relies on invasive procedures or imaging techniques that are not suitable for all patients. While body surface potential maps (BSPMs) are non-invasive and provide rich spatial–temporal information, their effective exploitation for atrial tissue characterization remains an open computational challenge. Methods We propose a graph-based learning framework that models BSPMs as spatial–temporal graphs and employs graph neural networks (GNNs) to non-invasively characterize ACM patterns. As no consensus clinical reference exists for the atrial substrate at the sub-regional granularity targeted here, the framework is developed and evaluated on a large-scale in silico database comprising 14,400 simulated BSPMs generated using multiple biatrial and torso models with varying ACM locations and densities. The framework addresses two supervised classification tasks: ACM localization and ACM density estimation. Results The proposed method achieved an accuracy of 89% for ACM localization and 84% for density classification on BSPMs derived from previously unseen atrial and torso anatomies. In addition, attention mechanisms within the GNN enabled the identification of electrode regions contributing most to tissue characterization, providing insight into the spatial relevance of BSPM measurements. Conclusion As a proof of concept trained and evaluated entirely on simulated data, these results demonstrate the potential of graph-based spatial–temporal analysis of BSPMs for non-invasive atrial tissue characterization. Prospective validation on clinical recordings is required before the approach can be translated to treatment-planning applications.