A deep learning architecture for temporal sleep stage classification using multivariate and multimodal time series (1707.03321v2)

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.

• The paper introduces a deep learning architecture that combines spatial filtering and temporal context to enhance sleep stage prediction from multimodal signals.
• The method integrates EEG, EOG, and EMG data to improve balanced accuracy and address noise through robust channel combination techniques.
• Empirical evaluation on the MASS dataset shows superior performance over traditional methods, paving the way for real-time sleep monitoring applications.

Deep Learning for Temporal Sleep Stage Classification Using Multivariate and Multimodal Time Series

Sleep stage classification is a crucial aspect of understanding sleep physiology and diagnosing sleep disorders. Traditional methods involving polysomnography (PSG) require manual annotations by experts, an inherently time-consuming and subjective task. This paper proposes a deep learning architecture aimed at performing automatic sleep stage classification, leveraging multivariate and multimodal time series, particularly Electroencephalography (EEG), Electrooculography (EOG), and Electromyography (EMG) data.

Methodological Advancements

The proposed deep learning model is characterized by several distinct features designed to improve performance in sleep stage classification tasks. The model introduces a spatial filtering step that linearly combines input channels to enhance representation before time-domain convolutional layers extract spectral features. This spatial filtering effectively addresses the potential for electrode removal and bad channel noise, contributing to robust and reliable predictions.

A multivariate network architecture processes EEG and EOG channels together because of their comparable magnitudes and similar signal properties, while EMG data is considered separately due to its distinct characteristics. This configuration results in a versatile, integrated approach to handling multimodal data.

The architecture also allows temporal context integration, accommodating sequences of neighboring segments to account for transition rules between different sleep stages. This enables the model to use contextual cues from adjacent time segments, improving classification accuracy.

Evaluation and Performance

Empirical comparisons in this paper examine the performance of the proposed model against benchmark methods, including traditional hand-crafted features analyzed via gradient boosting and other state-of-the-art deep learning approaches. Using public datasets, particularly the Montreal Archive of Sleep Studies (MASS), the proposed architecture exhibits superior performance, especially when leveraging both EEG and additional modalities like EOG and EMG.

Specifically, the paper highlights that increasing the number of EEG channels improves performance, although the gains plateau after a certain number of channels, indicating redundancy. EMG integration further boosts accuracy, with a marked difference in balanced accuracy metrics, favoring configurations that include both EEG and EMG inputs.

The temporal context presents a nuanced impact on classification, significantly enhancing the prediction of certain stages such as N1, N2, and REM, yet an overly broad context could lead to performance degradation for stages like W and N3. This finding suggests that while leveraging temporal context is beneficial, a careful tuning of its breadth is essential for optimal performance.

Implications for Future Research

The findings from this research have several implications for future developments in automated sleep stage classification. The potential applications could extend to portable devices offering real-time analysis, opening avenues for personal sleep health monitoring and remote diagnosis of sleep disorders.

Moreover, this paper's rigorous analysis of the spatial and temporal dimensions within EEG and other signal inputs provides a foundation for further exploration into adaptive neural network architectures. Subsequent studies could explore the incorporation of recurrent neural networks to capture more complex temporal dependencies and transitions in sleep stages.

Conclusion

This paper presents a comprehensive deep learning solution for sleep stage classification using multivariate and multimodal data. The model substantially enhances classification performance by intelligently integrating spatial and temporal information across different physiological modalities. Such advances not only facilitate improved diagnostic accuracy but also pave the way for the deployment of automated systems in clinical and home settings, significantly impacting the management and treatment of sleep disorders.