1 Self-supervised Learning for Label-Efﬁcient Sleep Stage Classiﬁcation A Comprehensive Evaluation

2025-04-26 0 0 1.7MB 11 页 10玖币

侵权投诉

Self-supervised Learning for Label-Efﬁcient Sleep

Stage Classiﬁcation: A Comprehensive Evaluation

Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Chee-Keong Kwoh, and Xiaoli Li

Abstract—The past few years have witnessed a remarkable

advance in deep learning for EEG-based sleep stage classiﬁcation

(SSC). However, the success of these models is attributed to

possessing a massive amount of labeled data for training, limiting

their applicability in real-world scenarios. In such scenarios, sleep

labs can generate a massive amount of data, but labeling these

data can be expensive and time-consuming. Recently, the self-

supervised learning (SSL) paradigm has shined as one of the

most successful techniques to overcome the scarcity of labeled

data. In this paper, we evaluate the efﬁcacy of SSL to boost the

performance of existing SSC models in the few-labels regime. We

conduct a thorough study on three SSC datasets, and we ﬁnd that

ﬁne-tuning the pretrained SSC models with only 5% of labeled

data can achieve competitive performance to the supervised

training with full labels. Moreover, self-supervised pretraining

helps SSC models to be more robust to data imbalance and

domain shift problems.

Index Terms—Sleep stage classiﬁcation, EEG, self-supervised

learning, label-efﬁcient learning

I. INTRODUCTION

Sleep stage classiﬁcation (SSC) plays a key role in diagnos-

ing many common diseases such as insomnia and sleep apnea

[1]. To assess the sleep quality or diagnose sleep disorders,

overnight polysomnogram (PSG) readings are split into 30-

second segments, i.e., epochs, and assigned a sleep stage. This

process is performed manually by specialists, who follow a

set of rules, e.g., the American Academy of Sleep Medicine

(AASM) [2] to identify the patterns and classify the PSG

epochs into sleep stages. This manual process is tedious,

exhaustive, and time-consuming.

To overcome this issue, numerous deep learning-based SSC

models were developed to automate the data labeling process.

These models are trained on a massive labeled dataset and

applied to the dataset of interest. For example, Jadhav et

al. [3] explored different deep learning models to exploit raw

Emadeldeen Eldele and Chee-Keong Kwoh are with the School of Com-

puter Science and Engineering, Nanyang Technological University, Singapore

(E-mail: {emad0002, asckkwoh}@ntu.edu.sg).

Mohamed Ragab and Zhenghua Chen are with the Institute for Info-

comm Research (I2R) and the Centre for Frontier AI Research (CFAR),

Agency for Science, Technology and Research (A∗STAR), Singapore (E-mail:

{mohamedr002, chen0832}@e.ntu.edu.sg).

Min Wu is with the Institute for Infocomm Research (I2R), Agency for Sci-

ence, Technology and Research (A∗STAR), Singapore (E-mail: wumin@i2r.a-

star.edu.sg).

Xiaoli Li is with Institute for Infocomm Research (I2R), Centre for Frontier

Research (CFAR), Agency of Science, Technology and Research (A∗STAR),

Singapore, and also with the School of Computer Science and Engineering at

Nanyang Technological University, Singapore (E-mail: xlli@i2r.a-star.edu.sg).

First author is supported by A∗STAR SINGA Scholarship.

Min Wu is the corresponding author.

electroencephalogram (EEG) signals, as well as their time-

frequency spectra. Also, Phyo et al. [4] attempted to improve

the performance of the deep learning model on the confusing

transitioning epochs between stages. In addition, Phan et al. [5]

proposed a transformer backbone that provides interpretable,

and uncertainty quantiﬁed predictions. However, the success

of these approaches hinges on a massive amount of labeled

data to train the deep learning models, which might not be

feasible. In practice, sleep labs can collect a vast amount of

overnight recordings, but the difﬁculties in labeling the data

limit deploying these data-hungry models. Thus, unfortunately,

the SSC works developed in the past few years have now a

bottleneck: the size, quality, and availability of labeled data.

One alternative solution to pass through this bottleneck is

the self-supervised learning (SSL) paradigm, which witnessed

increased interest recently due to its ability to learn useful

representations from unlabeled data. In SSL, the model is

pretrained on a newly deﬁned task that does not require

any labeled data, where ground-truth pseudo labels can be

generated for free. Such tasks are designed to learn the model

to recognize general characteristics about the data without

being directed with labels. Currently, SSL algorithms can

produce state-of-the-art performance on standard computer

vision benchmarks [9], [11]–[13]. Consequently, the SSL

paradigm has gained more interest to be applied for sleep stage

classiﬁcation problem [8], [14].

Most prior works aim to propose novel SSL algorithms and

show how they could improve the performance of sleep stage

classiﬁcation. Instead, in this work, our aim is to examine the

efﬁcacy of SSL paradigm to re-motivate deploying existing

SSC works in real-world scenarios, where only few labeled

samples are available. Therefore, we revisit a prominent subset

of SSC models and perform an empirical study to evaluate

their performance under the few-labeled data settings. More-

over, we explore the efﬁcacy of different SSL algorithms on

their performance and robustness. We also study the effect of

sleep data characteristics, e.g., data imbalance and temporal

relations, on the learned self-supervised representations. Fi-

nally, we assess the transferability of self-supervised against

supervised representations and their robustness to domain

shift. The overall framework is illustrated in Fig. 1. We

perform an extensive set of experiments on three sleep staging

datasets to systemically analyze the SSC models under the

few-labeled data settings. The experimental results of this

study aim to provide a solid and realistic real-world assessment

of the existing sleep stage classiﬁcation models.

arXiv:2210.06286v1 [eess.SP] 10 Oct 2022

ClsTrans

SimCLR

CPC

TS-TCC

Self-supervised Learning Algorithms

Feature Extraction

CNN branch

Dropout

BiLSTM

softmax

Temporal

Encoder

Maximize

Agreement

Feature

Extraction

Feature

Extraction

Projection

Head

Projection

Head

.SSC model .

Predictions

Feature

Extraction

Autoregressive

Model

Predictions

Feature

Extraction

Autoregressive

Model

Feature

Extraction

Autoregressive

Model

Projection

Head

Projection

Head

Maximize

Agreement

Small-kernel CNN

Wide-kernel CNN

Adaptive

Feature

Recalibration

Causal

self-attention softmax

CNN

block

Convolution Batch

Norm Max

Pooling

CNN

block CNN

block softmax

ReLU

Sleep Stage Classification Models

Evaluating

Datasets

Sleep-EDF

DeepSleepNet

AttnSleep

1D-CNN

Classifier

SHHS

ISRUC

Fig. 1. The architecture of our evaluation framework. We experiment with three sleep stage classiﬁcation models, i.e., DeepSleepNet [6], AttnSleep [7], and

1D-CNN [8]. We also include four self-supervised learning algorithms, i.e., ClsTran, SimCLR [9], CPC [10], and TS-TCC [8]. The different experiments are

performed on Sleep-EDF, SHHS, and ISRUC datasets.

II. RELATED WORK

A. Sleep Stage Classiﬁcation

A wide span of EEG-based sleep stage classiﬁcation meth-

ods have been introduced in recent years. These methods pro-

posed different architecture designs. For example, some meth-

ods adopted multiple parallel convolutional neural networks

(CNNs) branches to extract better features from EEG signals

[4], [6], [7]. Also, some methods included residual CNN layers

[15], [16], while others used graph-based CNN networks

[17]. On the other hand, Phan et al. [18] proposed Long

Short Term Memory (LSTM) networks to extract features

from EEG spectrograms. To handle the temporal dependencies

among EEG features, these methods had different approaches.

For instance, some works adopted recurrent neural networks

(RNNs), e.g., bi-directional LSTM networks as in [4], [6],

[16]. Other works adopted the multi-head self-attention as

a faster and more efﬁcient way to capture the temporal

dependencies in timesteps, as in [7], [19].

Despite the proven performance of these architectures, they

require a huge labeled training dataset to feed the deep learning

models. None of these works studied the performance of their

models in the few labeled data regime, which is our scope in

this work.

B. Self-supervised Learning Approaches

Self-supervised learning received more attention recently

because of its ability to learn useful representations from

unlabeled data. The ﬁrst SSL auxiliary tasks showed a big

improvement in the performance of the downstream task. For

example, Noroozi et al. proposed training the model to solve a

jigsaw puzzle on a patched image [20]. In addition, Gidaris et

al. proposed rotating the input images, then trained the model

to predict the rotation angle [21]. The success of these auxil-

iary tasks motivated adapting contrastive learning algorithms,

which showed to be more effective due to their ability to learn

invariant features. The key idea behind contrastive learning is

to deﬁne positive and negative pairs for each sample, then push

the sample closer to the positive pairs, and pull it away from

the negative pairs. In general, contrastive-based approaches

rely on data augmentations to generate positive and negative

pairs. For example, SimCLR considered the augmented views

of the sample as positive pairs, while all the other samples

within the same mini-batch are considered as negative pairs

[9]. Also, MoCo increased the number of negative pairs by

keeping samples from other mini-batches in a memory bank

[11]. On the other hand, some recent algorithms neglected the

negative pairs and proposed using only positive pairs such as

BYOL [13] and SimSiam [12].

C. Self-supervised learning for Sleep Staging

The success of SSL in computer vision applications mo-

tivated their adoption for sleep stage classiﬁcation. For ex-

ample, Mohsenvand et al. [22] and Jiang et al. [23] pro-

posed SimCLR-like methodologies and applied EEG-related

augmentations for sleep stage classiﬁcation. Also, Banville

et al. applied three pretext tasks, i.e., relative positioning,

temporal shufﬂing, and contrastive predictive coding (CPC)

to explore the underlying structure of the unlabeled sleep

EEG data [24]. The CPC [10] algorithm predicts the future

timesteps in the time-series signal, which motivated other

works to build on it. For example, SleepDPC solved two

problems, i.e., predicting future representations of epochs,

and distinguishing epochs from other different epochs [25].

Also, TS-TCC proposed temporal and contextual contrasting

approaches to learn instance-wise representations about the

sleep EEG data [8]. In addition, SSLAPP developed a con-

trastive learning approach with attention-based augmentations

in the embedding space to add more positive pairs [26].

Last, CoSleep [14] and SleepECL [27] are yet another two

contrastive methods that exploit information, e.g., inter-epoch

dependency and frequency domain views, from EEG data to

obtain more positive pairs for contrastive learning.

III. EVALUATION FRAMEWORK

A. Preliminaries

In this section, we describe the SSL-related terminologies,

i.e., pretext tasks, contrastive learning, and downstream tasks.

1) Problem Formulation: We assume that the input is

single-channel EEG data in Rd, and each sample has one label

from one of Cclasses. The supervised downstream task has

an access to the inputs and the corresponding labels, while

the self-supervised learning algorithms have access only to

the inputs.

The SSC networks consist of three main parts. The ﬁrst

is the feature extractor, which maps the input data into the

embedded space fφ:Rd→Rm1parameterized by neural

network parameters φ. The second is the temporal encoder

(TE), which is another intermediate network to improve the

temporal representations. The TE may change the dimension

of the embedded features fθ:Rm1→Rm. Finally, the

classiﬁer fγ:Rm→RC, which produces the predictions.

The SSL algorithms learn φfrom unlabeled data, while ﬁne-

tuning learns θand γwith also updating φ.

2) Pretext tasks: Pretext tasks refer to the pre-designed

tasks to learn the model generalized representations from the

unlabeled data. Here, we describe two main types of pretext

tasks, i.e., auxiliary, and contrastive tasks.

a) Auxiliary tasks: This category includes deﬁning a new

task along with free-to-generate pseudo labels. These tasks

can be deﬁned as classiﬁcation, regression, or any others. In

the context of time-series applications, a new classiﬁcation

auxiliary task was deﬁned in [28], [29] by generating several

views to the signals using augmentations such as: adding noise,

rotation, and scaling . Each view was assigned a label, and the

model was pretrained to classify these transformations. This

approach showed success in learning underlying representa-

tions from unlabeled data. However, it is usually designed

with heuristics that might limit the generality of the learned

representations [8].

b) Contrastive learning: In contrastive learning, repre-

sentations are learned by comparing the similarity between

samples. In speciﬁc, we deﬁne positive and negative pairs

for each sample. Next, the feature extractor is trained to

achieve the contrastive objective, i.e., push the features of the

sample towards the positive pairs, and pull them away from

the negative pairs. These pairs are usually generated via data

augmentations. Notably, some studies [9], [30] relied on strong

successive augmentations and found them to be a key factor

in the success of their contrastive techniques.

Formally, given a dataset with Nunlabeled samples, we

generate two views for each sample x, i.e., {ˆ

xi,ˆ

xj}using data

augmentations. Therefore, in a multiviewed batch with Nsam-

ples for each view, we have a total of 2Nsamples. Next, the

feature extractor transforms them into the embedding space,

and a projection head h(·)is used to obtain low-dimensional

embeddings, i.e., zi=h(fφ(ˆ

xi)) and zj=h(fφ(ˆ

xj)). As-

suming that for an anchor sample indexed i∈I≡ {1. . . 2N},

and A(k)≡I\{k}. The objective of contrastive learning is to

encourage the similarity between positive pairs and separate

the negative pairs apart using the NT-Xent loss, deﬁned as

follows:

LNT-Xent =−1

2NX

i∈I

log exp (zi·zj/τ)

Pa∈A(i)exp (zi·za/τ),(1)

where ·symbol denotes the inner dot product, and τis a

temperature parameter.

3) Downstream tasks: Downstream tasks are the main tasks

of interest that lacked a sufﬁcient amount of labeled data

for training the deep learning models. In this paper, the

downstream task is sleep stage classiﬁcation, i.e., classifying

the PSG epochs into one of ﬁve classes, i.e., W, N1, N2,

N3, and REM. However, in general, the downstream task

can be different and deﬁned by various applications. Notably,

different pretext tasks can have a different impact on the

same downstream task. Therefore, it is important to design a

relevant pretext task to the problem of interest, to learn better

representations. Despite the numerous proposed methods in

self-supervised learning, identifying the proper pretext task is

still an open research question [31].

B. Sleep Stage Classiﬁcation Models

We perform our experiments on three sleep stage classi-

ﬁcation models, i.e., DeepSleepNet [6], AttnSleep [7], and

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

1Self-supervisedLearningforLabel-EfcientSleepStageClassication:AComprehensiveEvaluationEmadeldeenEldele,MohamedRagab,ZhenghuaChen,MinWu,Chee-KeongKwoh,andXiaoliLiAbstractThepastfewyearshavewitnessedaremarkableadvanceindeeplearningforEEG-basedsleepstageclassication(SSC).However,thesuccessofthesem...

展开>> 收起<<

1 Self-supervised Learning for Label-Efﬁcient Sleep Stage Classiﬁcation A Comprehensive Evaluation.pdf

共11页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

1 Self-supervised Learning for Label-Efﬁcient Sleep Stage Classiﬁcation A Comprehensive Evaluation

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: