From twelve leads to single: Cross-lead knowledge distillation for domain-generalized atrial fibrillation detection

Atrial fibrillation (AF) detection from long-term ambulatory electrocardiography (ECG) is critical because AF is a leading cause of stroke yet is frequently missed by conventional diagnostics, particularly in wearable settings that demand strong cross-domain generalization. Deep learning has improved AF identification by exploiting rhythm patterns, but methods such as BMIRC and SJTU-ECGNet still struggle to generalize across heterogeneous single-lead wearable datasets. To address this issue, we propose X-LeadGen, a two-stage cross-lead distillation framework for single-lead ECG AF detection that explicitly targets domain generalization. First, a 12-lead teacher network distills rich multi-lead information into a single-lead student through cross-lead knowledge transfer and a simulated self-distillation branch within the teacher, improving cross-lead representation alignment and AF discrimination. Second, we introduce a domain-generalization strategy that applies domain-adversarial training in both teacher and student stages, complemented by GANs-based augmentation in the student phase to learn domain-invariant features and enhance robustness to distribution shifts. Across seven public ECG datasets under real and simulated Lead-I settings, X-LeadGen consistently outperforms existing approaches in accuracy and F1-score, demonstrating strong generalization for real-world wearable AF monitoring. Our implementation is publicly available at https://github.com/Aveiro-Lin/X-LeadGen , enabling further research and validation in the ECG community.