Self-Supervised Learning for Sleep Stage Classification with Predictive and Discriminative Contrastive Coding

Abstract

The purpose of this paper is to learn efficient representations from raw electroencephalogram (EEG) signals for sleep stage classification via self-supervised learning (SSL). Although supervised methods have gained favorable performance, they heavily rely on manually labeled datasets. Recently, SSL arrives comparable performance with fully supervised methods despite limited labeled data by extracting high-level semantic representations. To alleviate the severe reliance of labels, we propose SleepDPC, a novel sleep stage classification algorithm based on SSL. By incorporating two dedicated predictive and discriminative learning principles, SleepDPC discovers underlying semantics from raw EEG signals in a more efficient manner. We thoroughly evaluate the performance of our proposed method on two publicly available datasets. The experimental results show that our method not only learns meaningful representations but also produces superior performance versus various competing methods despite limited access of labeled data.

SleepPriorCL: Contrastive Representation Learning with Prior Knowledge-based Positive Mining and Adaptive Temperature for Sleep Staging

“…• T-Loss (Franceschi, Dieuleveut, and Jaggi 2019): Unsupervised scalable representation learning for multivariate time series. SleepDPC (Xiao al. 2021): Self-supervised learning for sleep staging by predicting future representations and distinguishing epochs from different epoch sequences.…”

“…However, both of the two paradigms require fully labelled datasets, which are laborious to acquire in health care area. Recent progress of self-supervised learning (SSL) (Jing and Tian 2019;He et al 2020;Chen et al 2020) has gained promising performance for physiological time series, with competitive performance compared with supervised methods (Franceschi, Dieuleveut, and Jaggi 2019;Xiao et al 2021). SSL methods can extract representations with semantic information from raw physiological signals, which are promising to alleviate the burden of manual labeling works.…”

“…Specifically for physiological time series, Banville et al (2020) construct a binary classification task, identifying whether the selected samples are adjacent to each other. SleepDPC (Xiao et al 2021) learn representations for sleep staging by predicting near future positives and discriminating temporal neighbors. All of the above time series SSL methods can only mine temporally neighboring positives, ignoring all nontemporally neighboring semantic-similar positives.…”

“…However, supervised methods rely heavily on labels, which are laborious to obtain in medical field. Recent SSL methods demonstrate promising results when few labels are accessible (Xiao et al 2021;Banville et al 2020;Eldele et al 2021).…”

neuro2vec: Masked Fourier Spectrum Prediction for Neurophysiological Representation Learning

“…Thus, learning generic representations for neurophysiological signals via self-supervised training remains challenging. Recently, inspired by the success of contrastive learning, some researchers [15,45] proposed contrastive based pretraining methods on neurophysiological signals such as electroencephalogram (EEG). The success of contrastive learning relies on two essential assumptions: (i) augmented views of the same data sample should be semantically consistent in the latent representation space; (ii) strong augmentations are required to learn useful representation.…”