Sci Rep. 2026 Feb 6. doi: 10.1038/s41598-026-36932-1. Online ahead of print.
ABSTRACT
Electrocardiography is a cornerstone in the diagnosis of cardiovascular diseases; however, accurate interpretation demands expert knowledge and is often impeded by data scarcity and annotation costs. To address these challenges, we propose SimCardioNet, a hybrid self-supervised and supervised deep learning framework for multi-class electrocardiography image classification. SimCardioNet leverages a custom multi-scale convolutional neural network backbone enhanced with residual connections and multi-head self-attention, pretrained via a modified SimCLR contrastive learning strategy that integrates a hybrid loss combining InfoNCE and cosine similarity. Following self-supervised pretraining, the model undergoes supervised fine-tuning with progressive layer unfreezing to mitigate overfitting and preserve meaningful representations. We evaluate SimCardioNet across three distinct ECG image datasets: (1) a 4-class Pakistani clinical ECG dataset (Dataset I), (2) an external Kaggle electrocardiography dataset for out-of-distribution validation (Dataset II), and (3) the large-scale PTB-XL benchmark (Dataset III) covering five diagnostic superclasses. On Dataset I, SimCardioNet achieves 0.975 accuracy, 0.973 precision, 0.973 recall, and 0.972 F1-score under 3-fold cross-validation. On Dataset II, the model demonstrates perfect classification performance (1.00 accuracy, precision, recall, and F1-score), highlighting strong generalization. On the PTB-XL dataset (Dataset III), SimCardioNet attains 0.921 accuracy and 0.921 F1-score, outperforming current state-of-the-art models including dual-branch CNNs, entropy-enhanced CNNs, and Bi-GRU architectures. Ablation studies confirm the critical contributions of self-supervised pretraining, attention mechanisms, and domain-specific augmentations. Grad-CAM visualizations further validate the model's focus on clinically relevant Electrocardiography regions. Our results underscore SimCardioNet's potential to reduce reliance on labeled data while delivering robust, interpretable, and clinically viable Electrocardiography classification especially valuable in resource-constrained settings.
PMID:41644577 | DOI:10.1038/s41598-026-36932-1