A Comparative Study of Few-Shot Learning Methods for 1-D ECG Time-series Classification

Employing deep learning models to classify time-series data from Internet of Things (IoT) devices requires a significant quantity of labeled data. Nevertheless, the limited resources available in IoT devices often pose challenges in accommodating training with large datasets. This work studies and demonstrates the implementation of well-established few-shot learning(FSL)networks for 1-dimensional ECG arrhythmia classification, employing a matching network, prototype network, and relation network. Few-shot learning addresses the challenge of data insufficiency by classifying with only a small number of labeled examples. It effectively mitigates the data imbalance problem inherent in classical machine learning models. It has been shown by our experiments that a matching network can do a 2-class classification with an accuracy of 85 % and a 5-class classification with an accuracy of 72.9 % on the MIT- BIH dataset. Additionally, accuracies of 81 % for 2-class classification and 74.3 % for 5-class classification are obtained on the IN CART dataset using matching networks. We additionally assess the performance of the matching network in comparison to other established few-shot learning networks, including prototype network and relation network, using the same amount of data. We conclude that the matching network outperforms, particularly when dealing with a limited dataset size.