Few-shot learning for cardiac arrhythmia detection based on electrocardiogram data from wearable devices

Abstract

Wearable devices have dramatically developed over the past decade as their functions extended from the simple posture analysis to non-invasive condition monitoring for early warning and proactive healthcare, which are especially significant for the dangerous disease such as cardiac arrhythmia. However, it is difficult for the wearable devices to collect plentiful and high-quality training samples so as to meet the fundamental requirements for the learning-based methods. To address this challenge, we propose a meta-transfer based few-shot learning method to handle arrhythmia classification with the ECG signal from the wearable devices. First, the original ECG signals are converted into spectrograms applicable to the 2D-CNN models. Second, we propose the special large-training scheme to pre-train the feature extractor to emphasize the meaningful information for classification, and the feature output dimension is reshaped to reduce the influence of irrelevant and redundant information. Then, the meta-transfer scheme is developed to avoid the training from scratch, which is prone to overfitting without the adequate samples. Finally, we conduct the extensive experiments to assess the performance of our method. The experimental results illustrate that the proposed method outperforms in accuracy than other comparative methods when handling the various few-shot tasks under the same training samples.