Mixed Neural Network Approach for Temporal Sleep Stage Classification

Hao Dong, Akara Supratak, Wei Pan, Chao Wu, Paul M. Matthews, Yike GuoIEEE Transactions on Neural Systems and Rehabilitation Engineering 2016 · DOI · OpenReview

Abstract

This paper proposes a practical approach to addressing limitations posed by using of single-channel electroencephalography (EEG) for sleep stage classification. EEG-based characterizations of sleep stage progression contribute the diagnosis and monitoring of the many pathologies of sleep. Several prior reports explored ways of automating the analysis of sleep EEG and of reducing the complexity of the data needed for reliable discrimination of sleep stages at lower cost in the home. However, these reports have involved recordings from electrodes placed on the cranial vertex or occiput, which are both uncomfortable and difficult to position. Previous studies of sleep stage scoring that used only frontal electrodes with a hierarchical decision tree motivated this paper, in which we have taken advantage of rectifier neural network for detecting hierarchical features and long short-term memory network for sequential data learning to optimize classification performance with single-channel recordings. Evaluation of data from 62 people (with 494 hours sleep) demonstrated better performance of our analytical algorithm than is available from existing approaches with vertex or occipital electrode placements. Use of this recording configuration with neural network deconvolution promises to make clinically indicated home sleep studies practical.

Introduction

Sleep disorders such as insomnia and apnea affect up to a third of the population, yet clinical diagnosis typically requires Polysomnography (PSG)—an expensive and often uncomfortable overnight stay in a laboratory. While automated scoring systems have been developed to alleviate this burden, they face significant hurdles including individual heterogeneity in EEG patterns, the inherent temporal dependency of sleep stages, and the discomfort associated with traditional electrode placements on the vertex or occiput.

This research introduces a Mixed Neural Network (MNN) architecture designed to bridge the gap between the comfort of a forehead-mounted wearable and the accuracy of full clinical PSG. By utilizing a single-channel EEG recorded from a practical forehead position, the proposed model achieves state-of-the-art performance, making home sleep monitoring more accessible.

Methodology

Mixed Neural Network Architecture

The MNN processes EEG data in two distinct stages to handle both spectral complexity and temporal context:

  1. Feature Extraction (MLP): A "Rectifier Neural Network" (a Multi-Layer Perceptron using ReLU activation functions) detects hierarchical features from the frequency spectrum. ReLUs are chosen because EEG spectral data is sparse; unlike sigmoid functions, they do not saturate and are better at representing the activation of specific frequency bands.
  2. Sequential Learning (LSTM): To address the temporal nature of sleep, features from the MLP are fed into a Long Short-Term Memory (LSTM) network. This allows the model to "remember" previous epochs, mimicking the contextual decision-making process of human experts.
The Mixed Neural Network (MNN) architecture, combining an MLP module for feature extraction and an LSTM module for sequential data learning.

Signal Configuration

To maximize user comfort, the study utilizes a single-channel EEG (F4) combined with an electro-oculogram (EOG) reference. This configuration is placed near the hairline, capturing sufficient signal to identify eye movements and brain waves without the need for messy adhesive pastes in hair-covered areas.

Proposed electrode placement on the forehead (F4, Fp2) and near the eye (EOG Left) to optimize for both signal quality and user comfort.

Experimental Setup

The study utilized an open-access dataset of 62 healthy subjects (29 men, 33 women) aged 23 to 73, comprising approximately 494 hours of sleep data. Each 30-second epoch was manually scored by a sleep expert according to AASM standards. The model was validated using 31-fold cross-validation to ensure generalizability across different individuals.

Key Results and Discussion

The MNN achieved an overall accuracy of 85.92% and a Macro F1-score of 80.50%, significantly outperforming baseline models such as SVM (79.70%) and standard MLPs (81.43%).

The Importance of Temporal Context

A critical insight from the study is the power of context. Traditional classifiers like Random Forests often see decreased performance when given longer sequences. In contrast, the MNN’s performance improved steadily as the sequence length increased to 5 epochs, proving that LSTMs are uniquely suited for the sequential logic of sleep stage progression.

Performance Across Stages

The model showed particular strength in identifying Stage N1 and N2, which are notoriously difficult to distinguish in single-channel setups. While frontal electrodes are less sensitive to alpha rhythms than occipital ones—leading to slightly lower F1-scores for the "Wake" stage—the MNN is powerful enough to extract proxy information that allows it to outperform traditional methods using more "ideal" but less comfortable electrode locations (e.g., Cz-Pz).

Conclusion

The Mixed Neural Network approach demonstrates that high-accuracy sleep stage classification is possible using a single-channel forehead EEG. By automatically learning complex, non-linear rules from raw spectral data and incorporating temporal context via LSTMs, this method provides a practical solution for clinical-grade home sleep studies. Future work will focus on validating the model for pathological sleep patterns and integrating it with dry-electrode hardware.

Summary updated

Abstract

This paper proposes a practical approach to addressing limitations posed by use of single active electrodes in applications for sleep stage classification. Electroencephalography (EEG)-based characterizations of sleep stage progression contribute the diagnosis and monitoring of the many pathologies of sleep. Several prior reports have explored ways of automating the analysis of sleep EEG and of reducing the complexity of the data needed for reliable discrimination of sleep stages in order to make it possible to perform sleep studies at lower cost in the home (rather than only in specialized clinical facilities). However, these reports have involved recordings from electrodes placed on the cranial vertex or occiput, which can be uncomfortable or difficult for subjects to position. Those that have utilized single EEG channels which contain less sleep information, have showed poor classification performance. We have taken advantage of Rectifier Neural Network for feature detection and Long Short-Term Memory (LSTM) network for sequential data learning to optimize classification performance with single electrode recordings. After exploring alternative electrode placements, we found a comfortable configuration of a single-channel EEG on the forehead and have shown that it can be integrated with additional electrodes for simultaneous recording of the electroocuolgram (EOG). Evaluation of data from 62 people (with 494 hours sleep) demonstrated better performance of our analytical algorithm for automated sleep classification than existing approaches using vertex or occipital electrode placements. Use of this recording configuration with neural network deconvolution promises to make clinically indicated home sleep studies practical.
BibTeX
@article{dong2016mixed,
  author = {Hao Dong and Akara Supratak and Wei Pan and Chao Wu and Paul M. Matthews and Yike Guo},
  title = {{Mixed Neural Network Approach for Temporal Sleep Stage Classification}},
  journal = {IEEE Transactions on Neural Systems and Rehabilitation Engineering},
  year = {2016},
  doi = {10.1109/tnsre.2017.2733220},
}
Figure