Deep learning for localisation of arrhythmia origins using intracardiac echocardiography

Background/Introduction Accurate localization of arrhythmia origin is essential for guiding effective intervention, yet current practice, creation of a voltage map of the heart, is invasive, labor-intensive, and prolongs procedure time. Intracardiac echocardiography (ICE) provides high-resolution, real-time imaging during electrophysiology procedures but remains underused for automated localization of arrhythmogenic sites. Leveraging ICE for this purpose could reduce mapping time and improve procedural efficiency. Purpose To develop and evaluate a deep-learning framework that localizes arrhythmia origins from ICE videos, specifically distinguishing left-sided from right-sided arrhythmias to narrow the search region for ablation. Methods Patients undergoing elective electrophysiology procedures at a single center were prospectively enrolled. ICE recordings were acquired in four standard views (tricuspid valve, mitral valve, left pulmonary veins, crista terminalis) during normal sinus rhythm and during controlled pacing at two coronary sinus sites, yielding three classes: sinus rhythm, distal pacing, and proximal pacing. After removal of non-informative regions, heartbeat-level cine loops were segmented using electrocardiogram-guided annotations, normalized, temporally standardized to 32 frames, and augmented in the training set. A three-dimensional convolutional neural network was trained independently for each view using patient-level splits. Ten-fold cross-validation was performed across 39 unique patients with non-overlapping training, validation, and test cohorts. Clip-level predictions were obtained by majority voting across heartbeats, and patient-level predictions were derived using cross-view majority voting. Results Across ten folds, mean patient-level accuracies for the three-class task were 76.27% (validation) and 66.2% (test) when fusing all available views. View-specific mean test accuracies ranged from 54.7% to 65.1%, with the MV view achieving the highest performance, followed by LPV, TV, and CT. Gradient-based class activation mapping highlighted regions most influential to the model’s predictions and will be further examined to assess correspondence with clinically meaningful structures. Conclusion(s) This study demonstrates the feasibility of using deep learning applied to ICE videos for automated localization of arrhythmia origins. The findings represent a promising step toward reducing procedure time and improving clinical efficiency, while underscoring the need for expanded patient cohorts and further optimization of model robustness and generalizability. AI-assisted ICE analysis may help streamline mapping, standardize interpretation, and support more targeted ablation strategies; validation in larger and more diverse populations is warranted.