Characterising cardiac tissue properties with graph neural networks

Details

Identifying ablation targets for arrhythmia treatment requires characterizing cardiac electrophysiological properties from sparse intracardiac measurements — a task that is both time-sensitive and clinically high-stakes. The paper introduces a GNN framework trained on synthetic 2D flat-surface electrogram data to detect fibrosis, rapid depolarization, and high excitability, achieving 0.96, 0.97, and 0.95 average precision respectively. The critical practical result: the model trained on flat synthetic surfaces can transfer to curved 3D surfaces with few-shot fine-tuning, avoiding the need to retrain on expensive clinical data. The framework targets premature ventricular complex (PVC) ablation specifically. Presented at the STACOM workshop 2026 (a satellite event of MICCAI), this is early-stage applied research. The authors explicitly note that future work is needed before clinical use. The value is demonstrating sim-to-real and flat-to-curved generalization for cardiac mapping.