Neuro-causal multimodal transformer for interpretable arrhythmia subtyping in congenital and pediatric ECG with intracardiac fusion

Abstract

Automated arrhythmia subtyping in pediatric and congenital heart disease (CHD) patients is challenging. We propose a Neuro-Causal Multimodal Transformer (NCMT) that, for the first time, fuses 12-lead surface ECG with 7-lead intracardiac electrograms (IEGM) for high-fidelity classification. Using a causal attention mechanism, our model ensures predictions are based only on past and present data. Evaluated on the Leipzig Heart Center dataset of 39 pediatric/CHD patients, our model achieved 93.4% accuracy and an F1-score of 0.912 in subtyping challenging arrhythmias like AVNRT, AVRT, and AFL, outperforming existing baselines. Integrated interpretability via SHAP and attention maps provides transparent, clinically relevant insights into the model's decisions, establishing a framework for trustworthy AI in pediatric electrophysiology.