JEPA-Mamba: a self-supervised joint-embedding architecture with bidirectional Mamba for ECG representation learning

The electrocardiogram (ECG) is the primary non-invasive tool for early screening and clinical management of cardiovascular disease, but deep learning deployment is limited by scarce labeled data, by architectures that ignore the periodic structure of ECG signals, and by the trade-off between long-sequence efficiency and diagnostic accuracy. We propose JEPA-Mamba, a self-supervised representation learning framework that integrates the Joint-Embedding Predictive Architecture (JEPA) with a bidirectional Mamba encoder and a channel–temporal attention module (CTAM) for multi-adaptive ECG classification on downstream tasks. High-ratio contiguous-block masking trains an exponential moving average (EMA) encoder and a narrow predictor to learn diagnostic features without hand-crafted augmentation, while the bidirectional Mamba (BiM) encoder, augmented by rotary position embedding (RoPE) and a register mechanism, couples convolutional waveform extraction with linear-complexity long-sequence modeling and CTAM recalibrates the 12 leads and focuses temporal attention on clinically salient segments. Under a strict patient-level split on PTB-XL, with an encoder pretrained on MIMIC-IV-ECG for cross-dataset validation on the same tasks, JEPA-Mamba attains macro-averaged AUCs of 0.9296 ± 0.002 and 0.9306 ± 0.002 on the fully annotated task, outperforming self-supervised baselines on most downstream tasks and surpassing the most direct JEPA counterpart in the ECG domain across all tasks. Improvements over external baselines on the all, Sub-diag., and Diag. tasks are statistically significant, and internal ablation differences are confirmed by DeLong tests. Per-class analyses, attention maps, t-SNE embeddings, and robustness experiments under noise and baseline drift support these design choices. These results offer preliminary evidence for the clinical potential of JEPA-Mamba.