Arrhythmia classification method based on knowledge distillation and 2D ECG images

Arrhythmia classification is crucial for identifying and preventing the potential risks associated with cardiovascular diseases. Currently., using deep learning for arrhythmia classification faces challenges of model complexity and deployment difficulty. To address this., We present a novel approach for classifying arrhythmias using knowledge distillation and 2D ECG images. In this study., the classification of the MIT-BIH dataset is performed according to AAMI criteria. Preprocessing., such as denoising., interception., and 2D grayscale image conversion., is first performed. Then., high-performance 2D-CNN models are pre-trained using 2D ECG images as the teacher model., and a lightweight 1D-CNN is constructed as the student model. Finally., we align raw and 2D ECG data., utilizing knowledge distillation to allow the student model to learn from the teacher model during the training process., thus enabling model compression. The final results show that the method significantly improves the performance of the student model in arrhythmia classification. For overall performance., when using Vgg16 as the teacher model., Acc and P+ improve by 1.62% and 6.64%., and Se is comparable. In addition., the F1-score values have been improved in all categories. The proposed method provides a feasible approach for low computational power device deployment for arrhythmia classification.