A deep learning architecture for temporal sleep stage classification using multivariate and multimodal time series.

Stanislas Chambon, Mathieu Galtier, Pierrick Arnal, Gilles Wainrib, Alexandre GramfortarXiv 2017 · OpenReview

Abstract

Sleep stage classification constitutes an important preliminary exam in the diagnosis of sleep disorders. It is traditionally performed by a sleep expert who assigns to each 30s of signal a sleep stage, based on the visual inspection of signals such as electroencephalograms (EEG), electrooculograms (EOG), electrocardiograms (ECG) and electromyograms (EMG). We introduce here the first deep learning approach for sleep stage classification that learns end-to-end without computing spectrograms or extracting hand-crafted features, that exploits all multivariate and multimodal Polysomnography (PSG) signals (EEG, EMG and EOG), and that can exploit the temporal context of each 30s window of data. For each modality the first layer learns linear spatial filters that exploit the array of sensors to increase the signal-to-noise ratio, and the last layer feeds the learnt representation to a softmax classifier. Our model is compared to alternative automatic approaches based on convolutional networks or decisions trees. Results obtained on 61 publicly available PSG records with up to 20 EEG channels demonstrate that our network architecture yields state-of-the-art performance. Our study reveals a number of insights on the spatio-temporal distribution of the signal of interest: a good trade-off for optimal classification performance measured with balanced accuracy is to use 6 EEG with 2 EOG (left and right) and 3 EMG chin channels. Also exploiting one minute of data before and after each data segment offers the strongest improvement when a limited number of channels is available. As sleep experts, our system exploits the multivariate and multimodal nature of PSG signals in order to deliver state-of-the-art classification performance with a small computational cost.

Introduction

Sleep stage classification—the process of creating a "hypnogram" to map a patient's sleep cycle—is the cornerstone of diagnosing disorders like insomnia and sleep apnea. Traditionally, this is a manual task where a sleep expert visually inspects hours of polysomnography (PSG) signals (EEG, EOG, EMG) and assigns a stage to every 30-second window. This process is slow, expensive, and subjective. Existing automated methods often rely on hand-crafted features or focus on single EEG channels, potentially missing subtle patterns and spatial information.

Methods

The proposed Time Distributed Multivariate Network learns features directly from raw data. The architecture features several key innovations:

  • Learned Spatial Filtering: The first layer acts as a "virtual sensor" builder, learning to linearly combine input channels to maximize the signal-to-noise ratio (SNR).
  • Multimodal Pipelines: EEG and EOG signals are processed in one pipeline, while EMG signals, which have different statistical properties, are processed in a parallel pipeline before merging.
  • Temporal Context: The model exploits the sequential nature of sleep by using a sequence of segments (the current 30s window plus preceding and following windows) to inform classification.
  • End-to-End Learning: Temporal convolutions extract spectral features directly from the time series, eliminating the need for pre-computed spectrograms.
Proposed deep learning architecture for sleep stage classification.

The study used the MASS dataset (session 3), comprising 61 night records. Signals were low-pass filtered at 30Hz and downsampled to 128Hz to retain relevant sleep frequencies such as Delta, Theta, Alpha, Sigma, and Beta.

Key Results

The model achieved state-of-the-art performance, matching or outperforming existing deep learning benchmarks like DeepSleepNet. A key finding was the trade-off between spatial and temporal context: rich spatial context can compensate for limited training data. The authors identified an optimal sensor configuration of 6 EEG, 2 EOG, and 3 EMG channels.

Comparison of classification performance and computational efficiency across different models.

Notably, the architecture is highly efficient, utilizing only $\sim10^5$ parameters—orders of magnitude fewer than competing models. This efficiency, combined with high sensitivity to stages like "Wake" (0.85), makes it a strong candidate for portable medical device deployment.

Discussion and Future Work

While the model successfully leverages temporal context, future improvements could involve integrating Recurrent Neural Networks (RNNs) like LSTMs to explicitly model complex transition sequences. Additionally, testing against a consensus of multiple experts would better account for human subjectivity in ground truth labels. While currently optimized for offline analysis due to its use of "future" context windows, the model could be adapted for real-time monitoring by restricting context to past segments.

Summary updated

Abstract

Sleep stage classification constitutes an important preliminary exam in the diagnosis of sleep disorders. It is traditionally performed by a sleep expert who assigns to each 30s of signal a sleep stage, based on the visual inspection of signals such as electroencephalograms (EEG), electrooculograms (EOG), electrocardiograms (ECG) and electromyograms (EMG). We introduce here the first deep learning approach for sleep stage classification that learns end-to-end without computing spectrograms or extracting hand-crafted features, that exploits all multivariate and multimodal Polysomnography (PSG) signals (EEG, EMG and EOG), and that can exploit the temporal context of each 30s window of data. For each modality the first layer learns linear spatial filters that exploit the array of sensors to increase the signal-to-noise ratio, and the last layer feeds the learnt representation to a softmax classifier. Our model is compared to alternative automatic approaches based on convolutional networks or decisions trees. Results obtained on 61 publicly available PSG records with up to 20 EEG channels demonstrate that our network architecture yields state-of-the-art performance. Our study reveals a number of insights on the spatio-temporal distribution of the signal of interest: a good trade-off for optimal classification performance measured with balanced accuracy is to use 6 EEG with 2 EOG (left and right) and 3 EMG chin channels. Also exploiting one minute of data before and after each data segment offers the strongest improvement when a limited number of channels is available. As sleep experts, our system exploits the multivariate and multimodal nature of PSG signals in order to deliver state-of-the-art classification performance with a small computational cost.
BibTeX
@misc{chambon2017deep,
  author = {Stanislas Chambon and Mathieu Galtier and Pierrick Arnal and Gilles Wainrib and Alexandre Gramfort},
  title = {{A deep learning architecture for temporal sleep stage classification using multivariate and multimodal time series.}},
  year = {2017},
  eprint = {1707.03321},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1707.03321},
}
Figure