A Task-Generic High-Performance Unsupervised Pre-training Framework for ECG

Electrocardiogram (ECG) feature extraction is becoming increasingly important in the biomedical field. In deep learning studies centred on ECG data, established methods mainly rely on labelled signals. However, obtaining high-quality labelled records from open-source datasets is challenging, whereas many unlabeled data remain underutilized. Unsupervised pre-training methods have shown significant efficacy in utilizing datasets and improving performance. Therefore, incorporating unsupervised pre-training paradigms into ECG feature extraction is a strategic choice. This paper presents an unsupervised pre-training framework based on the Masked Autoencoder for extracting ECG signal features. An encoder-decoder framework is designed to recover artificially masked ECGs. During the reconstruction process, the framework can learn knowledge from masked samples. Benefiting from the utilization of large-scale datasets and the reconstruction design, the framework is general and has a high level of performance across a wide range of tasks. In particular, the framework achieves an accuracy of 95.6% on the MITDB dataset for the ECG arrhythmia classification task and 98.8% on the ECGIDDB dataset for the human identification task. Our method introduces novel strategies for training on large-scale unlabeled data. Empirical evaluations, including multiple downstream tasks and comparisons with the state of the art, confirm the validity of our proposed approach.