Multi-label Few-Shot Classification of Abnormal ECG Signals Using Metric Learning

Accurate and rapid classification of electrocardiogram (ECG) signals is crucial for correct diagnosis and treatment by doctors. However, existing deep learning techniques often require large-scale sample data for training, making research on few-shot classification tasks has attracted considerable attention. Despite some progress in few-shot ECG classification, most methods are still limited to single-label classification. The paper primarily proposes a multi-label few-shot ECG signal classification method based on metric learning. The model learns the differences between samples by computing the similarity of features for pairs of samples and infers the category of the test sample based on a small number of known category samples. Additionally, the paper introduces a specific category-guided multi-task training strategy. This strategy effectively alleviates the limitations of the model in learning sample pair similarities by introducing embedded features of specific categories. The proposed method is verified on the publicly available PTB-XL database. The results indicate significant improvements over existing methods, with F1 scores of 0.794 and 0.466 achieved for the 5-class and 20-class classification tasks. The proposed method has achieved a relatively significant improvement in classification performance, providing new insights and approaches for future multi-label few-shot ECG signal classification tasks. This advancement offers more accurate and rapid support for ECG signal classification for medical professionals.