EIA-CKD: An End-to-End Internal Augmentation and Curriculum-based Knowledge Distillation framework for lightweight ECG classification

Automated ECG classification is increasingly important for wearable and Internet of Medical Things (IoMT) applications, where deep learning models must provide accurate predictions under strict computational and memory constraints. Knowledge distillation can improve model efficiency, but conventional approaches typically rely on separately pre-trained teacher networks. To address these limitations, this study proposes End-to-End Internal Augmentation and Curriculum-based Knowledge Distillation (EIA-CKD), a lightweight end-to-end framework that generates and distills supervisory knowledge within a single network. During training, multiple views are generated internally from each ECG image and processed through a shared backbone along a frozen-consensus path, where gradient tracking is disabled. The resulting logits are aggregated by an attention-based consensus module to form a stable consensus representation that serves as an internal source of supervisory knowledge. A curriculum-based training strategy progressively increases the contribution of the internal consensus, allowing the model to gradually transition from ground-truth supervision to internally generated knowledge during training. The proposed framework relies on a lightweight CNN backbone that combines depthwise separable convolutions(DSC) with convolutional block attention modules (CBAM) to reduce computational overhead while preserving representational power. Experimental evaluation on the PTB-XL and Chapman ECG datasets confirms the effectiveness of the proposed EIA-CKD approach. With only 102 K trainable parameters, the proposed framework achieves AUROC scores of 95.60% and 99.16% and classification accuracies of 84.67% and 94.23% on PTB-XL and Chapman, respectively. These results show that EIA-CKD provides competitive classification performance with a substantially lower computational footprint, supporting its potential for real-time ECG analysis in wearable and resource-constrained IoMT environments.