Multi-Source Feature Fusion with Self-Supervised Contrastive Learning for AF Detection under Label Scarcity

Accurate atrial fibrillation (AF) screening from short, single-lead wearable ECG remains challenging due to noise contamination, limited computation budgets, and scarce annotations. We present a multi-source feature fusion framework that combines rhythm, morphology, and implicit representations to improve AF detection under label scarcity. The pipeline performs denoising, robust R-wave localization, and heartbeat segmentation to construct rhythm sequences and beat-level waveform inputs for downstream modeling. Complementary representations are learned from rhythm variability statistics, morphology-aware encoders, and self-supervised contrastive pretraining on external data with subsequent adaptation to the target domain. The resulting embeddings are integrated using attention-guided fusion for classification. Experiments on multiple public datasets demonstrate consistent performance in low-label settings, indicating the potential of the proposed approach for wearable AF screening. Further evaluation of computational efficiency and on-device performance is still required.