TAKD: A Temporal Attention-Based Knowledge Distillation Framework for Efficient Multi-Lead ECG Diagnosis

Arrhythmia is a leading cause of cardiovascular-related mortality, and early detection through ECG analysis is crucial. However, current deep learning models for ECG classification are computationally expensive, hindering their deployment on resource-constrained devices. This paper introduces a Temporal Attention-based Knowledge Distillation (TAKD) framework, which leverages temporal attention mechanisms to enhance feature transfer from a teacher model to a lightweight student model. TAKD improves the student's focus on critical temporal features and multi-lead interactions, significantly reducing model complexity without sacrificing performance. Experiments on the ICBEB2018 dataset show that TAKD achieves high accuracy with a much smaller model, making it suitable for large-scale screening and remote monitoring in resource-constrained environments.