Electrocardiogram (ECG) analysis is crucial for the diagnosis of cardiovascular disease. However, real-world deployment faces three key challenges: (1) the need for large amounts of manually labeled ECG data, which is time-consuming to collect. (2) the necessity of flexible addressing issues such as variations in clinical data distribution, noise, and low data quality. (3) the requirement for models to generalize across ECG data from different application scenarios. To address these issues, we propose ECGMind, a self-supervised foundational model for various ECG classification tasks. ECGMind is trained on unlabeled data and utilizes a masked autoencoder (MAE) backbone to selectively mask critical regions (e.g., R waves), thereby promoting the reconstruction of diagnostically important features. A dynamic energy score adjusts the importance of each ECG patch based on reconstruction difficulty, prioritizing hard-to-reconstruct areas to enhance feature learning. Our experiments are conducted on both static and dynamic ECG datasets, involving different classification tasks. The results demonstrate ECGMind 's superior performance and scalability, offering a solution for both clinical and ambulatory ECG analysis. ECGMind has deployed as an intelligent diagnostic service at http://heartmind.ecgdb.com. The core code can be evaluated at https://github.com/tianti1/ECGMind.

Electrocardiogram (ECG) analysis is a critical tool for early diagnosis and cardiac monitoring. However, the high cost of expert annotations and the complexity of physiological variability pose significant challenges to supervised learning approaches, motivating a growing interest in self-supervised methods. Existing models often struggle to handle both static and dynamic ECG signals, limiting their applicability. We propose 3S-ECG, a novel self-supervised framework for ECG analysis that focuses on the fundamental characteristics of ECG waveforms and integrates reconstruction and contrastive learning within a three-stage pretraining paradigm. We further introduce an R-peak-weighted random masking strategy to enhance learning of periodic cardiac patterns and employ tai-lored data augmentation to improve discriminative performance. Extensive experiments demonstrate that 3S-ECG outperforms state-of-the-art methods by 5% to 10 % on multiple public ECG datasets, while significantly reducing the reliance on labeled data.

Electrocardiogram (ECG) classification plays a pivotal role in computer-aided cardiovascular disease diagnosis. However, current models often underperform in real-world applications due to the scarcity of annotated data and their limited capacity for capturing high-level semantic information. To address these challenges, we propose MTECG, a novel multimodal self-supervised learning framework that bridges raw ECG signals with corresponding clinical text reports. At its core is MACFormer, an ECG encoder based on masked autoencoding and contrastive learning, designed to capture physiologically meaningful features from ECG signals without relying on labels. To enable cross-modal alignment, MTECG integrates MACFormer with pretrained language models using three self-supervised objectives: ECG-to-text reconstruction, text-to-ECG generation, and cross-modal contrastive learning. These objectives encourage the model to learn semantically grounded and generalizable representations without relying on manual labels. Extensive experiments on actual open ECG datasets have shown that MTECG significantly improves classification accuracy. Compared with the existing ECG classification baselines, MTECG achieves state-of-the-art in the macro-averaged F1-score metric on open datasets, and also has good accuracy in the zeroshot scenario, which cannot be achieved by the existing models.