PaperlayerView exact revision on arXiv

Source-linked AI summary

Mixed Neural Network Approach for Temporal Sleep Stage Classification

Hao Dong, Akara Supratak, Wei Pan, Chao Wu, Paul M. Matthews, Yike Guo

arXiv:1610.06421v3q-bio.NCcs.CVcs.LGcs.NE

TL;DR

Sleep staging needs to handle person-to-person EEG variation and dependencies across successive epochs, while comfortable home recording remains difficult with conventional electrode placements. The paper combines a rectifier neural network with an LSTM and a forehead EEG/EOG configuration. Its mixed neural network outperformed SVM, RF, and MLP controls, with sequence length five producing the strongest reported performance before longer sequences degraded results.

  • Problem

    Sleep-stage classification must accommodate heterogeneous EEG patterns and temporal dependencies across epochs.

  • Method

    The method combines a rectifier neural network for feature detection with an LSTM for sequential learning, using a comfortable single electrode pair over hairless skin.

  • Results

    The MNN outperformed SVM, RF, and MLP controls overall, with performance improving through sequence length five before declining for longer sequences.

  • Takeaways & Limitations

    The proposed approach supports sleep-stage classification from a comfortable single electrode pair and a potential home-care monitoring configuration.

Abstract

from arXiv · show

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.

I. INTRODUCTION

Automated sleep staging aims to reduce the burden and cost of polysomnography, but existing electrode placements create comfort problems and single-channel methods can miss important sleep information. The paper addresses heterogeneity and temporal dependence with a mixed neural network and a forehead-based EEG/EOG configuration.

  • Sleep disorders affect quality of life, while diagnosis typically requires all-night polysomnography with EEG, EOG, and EMG followed by expert scoring.
  • Home monitoring and automatic scoring could reduce costs and increase access to diagnostic sleep studies.
  • About 10% of people lack alpha rhythm during wakefulness and another 10% show limited alpha rhythm, complicating stage classification.
  • Sleep staging depends on prior epochs because N2 and REM decisions can remain valid despite absent characteristic events in the current epoch.
  • Central, occipital, and parietal electrodes detect useful sleep features but require placement in hairy scalp regions that can restrict movement and cause discomfort.
  • The proposed MNN combines a rectifier neural network for sparse pattern detection with an LSTM for temporal patterns, using a low-frontal EEG electrode and EOG electrodes that can also serve as references.

II. METHODOLOGY

The methodology represents sleep stages using physiological EEG features extracted from standard scoring epochs. It uses time-frequency analysis, selecting Fourier-based processing to remain aligned with AASM standards while allowing neural networks to learn additional features.

  • The paper adopts the AASM standard, which defines five stages: wakefulness, N1, N2, N3, and REM.
  • Each sleep-staging decision is based on a 30- or 20-second window of physiological signals called an epoch.
  • Sleep EEG physiology is characterized using time-domain or frequency-domain features.
  • The approach uses time-frequency analysis and a conventional Fourier transform rather than complex Morlet wavelets to follow AASM standards before neural-network feature extraction.

1) Spectral power of frequency bands:

The spectral-power pipeline converts EEG epochs into time-frequency representations and sub-band power features. It uses overlapping short windows for temporal analysis, while applying whole-epoch Fourier analysis to capture slow eye movements.

  • STFT extracts frequency and phase content as signals change over time, producing a spectrogram through sliding windows, overlap, and a window function.
  • Larger window sizes improve frequency precision but reduce temporal precision, while greater overlap reduces artifacts at higher computational cost.
  • The experiment divided each 30-second EEG epoch into 5-second segments with 70% serial overlap and applied a Hamming window.
  • The selected 5-second window and 70% overlap performed best among the tested settings, while Hamming and Hanning windows showed no performance distinction.
  • PSD sub-band features were computed by summing Fourier-transformed amplitude values, with fmin and fmax defining each frequency range.
  • Because 5-second windows may miss slow eye movements, the method also applies Fourier analysis across the entire EEG epoch.
  • Averaged PSD estimates continuation of recurring rhythms, while variation statistics capture frequency fluctuations and occasional events such as K complexes and vertex waves.

2) Statistics of spectral power:

The method combines statistical signal features with neural networks to address incomplete feature representations and temporal correlations. A rectifier-based MLP and LSTM are combined before softmax classification into the mixed neural network.

  • 2) Statistics of spectral power:: Low and median PSD values alongside a high maximum PSD can indicate occasional K complexes and vertex waves, while standard deviation measures frequency fluctuation.
  • 2) Statistics of spectral power:: Raw-signal maximum and minimum amplitudes provide artifact information, and Shannon entropy captures information related to amplitude distribution and signal synchronization.
  • Conventional classifiers performed poorly because they did not address temporal pattern recognition, while correlated physiological features could remain incomplete.
  • The model applies an MLP and recurrent neural network independently, then concatenates their outputs to address feature completeness and temporal correlations.
  • C. Mixed neural network: A softmax layer performs classification, and the resulting mixture of neural networks is termed a mixed neural network.
  • 1) Rectifier Neural Network:: The rectifier network is selected because EEG spectra are sparse, discontinuous, and variable across people and sleep stages.

1) Rectifier Neural Network:

The mixed neural network combines a rectifier network for hierarchical feature detection with LSTM for long-term temporal dependencies in sleep-stage classification.

  • Rectifier Neural Network:: Rectifier neurons suit sparse, discontinuous EEG spectra because they represent present frequency features as active and absent features as zero.Sleep-stage spectra contain few frequency bands, intermittent activity, and subject-dependent amplitudes.
  • Rectifier Neural Network:: The rectifier network’s linear output supports gradient backpropagation from the recurrent module and helps alleviate vanishing gradients.
  • Rectifier Neural Network:: Dropout acts as ensemble learning by training many networks that share hidden-neuron weights, addressing overfitting in deep neural networks.The selected dropout probabilities were 20%, 50%, and 50% across successive connections.
  • Rectifier Neural Network:: Without the rectifier neural network, accuracy dropped by 3%, while rectifier activation outperformed sigmoid and hyperbolic tangent functions in the experiment.
  • Rectifier Neural Network:: LSTM was selected for the recurrent module because it uses current information together with long-term information from past epochs.This addresses the long-term dependency problem relevant to sequential sleep-stage scoring.
  • Rectifier Neural Network:: The LSTM implementation uses vanilla recurrent equations with input, forget, cell-state, and output-gate operations before softmax classification over five sleep stages.The formulas define gate updates, cell-state accumulation, and block output; softmax produces five class outputs.

3) Output module:

The output module trains the network with stochastic gradient descent and evaluates predictions on recordings from healthy subjects using limited physiological data.

  • Output module:: The MNN uses stochastic gradient descent with batch size 500, learning rate 0.01, and momentum 0.9.
  • Output module:: Cross-entropy is used as the loss function based on softmax output and rectifier-network dropout, without weight decay.
  • Output module:: The evaluation dataset contains 62 healthy subjects aged 23–73, with 30-second epochs scored by one sleep expert under AASM standards.Labels include W, N1, N2, N3, REM, and unknown.
  • Output module:: The experiment uses the derivation between F4 and Left Horizontal EOG, with electrodes placed near the hairline and outer-down canthus without hair-covered skin.

B. Experimental Design

The experimental design varies MNN architectures, compares them with conventional classifiers, and evaluates performance using subject-separated cross-validation and macro F1-score and accuracy.

  • B. Experimental Design: MNN configurations vary rectifier hidden layers from 2 to 5, rectifier units from 200 to 800, and LSTM units from 200 to 1000.
  • B. Experimental Design: SVM, Random Forest, and Multilayer Perceptron classifiers are compared with MNN using the same features.
  • B. Experimental Design: For conventional classifiers, sequence length 1 uses the current epoch, while longer sequences add preceding EEG epochs; in LSTM, sequence length counts examples per output.
  • B. Experimental Design: Performance is measured with overall accuracy and macro F1-score, the mean of per-class F1-scores across sleep stages.
  • B. Experimental Design: Training and testing data come from different subjects, using 31-fold cross-validation with two recordings held out and 60 used for training per fold.
  • B. Experimental Design: Training sequence-length-5 networks across cross-validation takes about 2 days on a single NVIDIA 630 GPU machine.

C. Result and Discussion

The Mixed Neural Network generally improved as temporal context increased to five epochs and outperformed the control classifiers across overall and sleep-stage metrics. Results also indicate that the rectifier network and MLP contribute to performance, while longer sequences or added fully connected layers provide no benefit.

  • Sequence length: MNN’s macro F1-score and overall accuracy continually improved as sequence length increased from 1 to 5, then declined for longer sequences.The authors suggest sequences longer than five epochs may introduce training noise because experts do not consider more than five epochs when labeling data.
  • Classifier comparison: MNN achieved better overall accuracy and macro F1-score than the control classifiers, with higher F1-scores for every individual sleep stage.Table VI reports the best performance of each classifier and the improvement of MNN over the control group.
  • Architecture variations: Adding fully connected layers after LSTM or increasing MLP depth did not improve performance, while dropout produced similar performance to the architecture without extra layers.The authors associate additional MLP layers with unnecessary complexity and possible overfitting.
  • Ablation findings: Removing the MLP reduced overall accuracy to 85.10% and weakened discrimination of N2 versus N3 and W versus REM.The authors suggest the MLP learned useful combinations of preprocessed features, including shared representations of sleep spindles and rapid eye movement.
  • Comparison with literature: Compared with existing studies, MNN had significantly higher macro F1-score and N1 and N2 F1-scores, although its stage-W F1-score was lower than most prior methods.The authors attribute the weaker stage-W result to poor frontal-lobe alpha-rhythm detection, while noting that MNN outperformed two studies using central- or frontal-lobe EEG.

IV. CONCLUSION AND DISCUSSION

The proposed method combines a comfortable forehead EEG configuration with a Mixed Neural Network and performs better than the cited control classifiers. The F4 placement improves information capture but is less convenient near the hairline.

  • The proposed method uses EEG from a single electrode pair positioned comfortably over hairless skin and performs well compared with SVM, RF, and MLP.
  • F4 placement near the hairline remains imperfect for convenience, while easier frontal placements provide less information about several sleep stages.
  • Fp2-EOG produces lower accuracy and macro F1-score than F4-EOG, but still outperforms control classifiers using F4-EOG.
  • The proposed home-care configuration is illustrated as a system using frontal EEG/EOG electrodes, dry electrodes, amplification, digitization, and wireless transmission.

DIFFERENT CLASSIFIERS USING SAME FEATURES

The study compares SVM, RF, and MLP using the same features and reports their best-performing sequence lengths through corresponding confusion matrices.

  • SVM, RF, and MLP were compared fairly using the same features as the proposed method and their best-performing sequence lengths.
  • SVM was evaluated at sequence length 2, RF at sequence length 3, and MLP at sequence length 4.

APPENDIX B CONFUSION MATRICES OF EXISTING STUDIES

Existing single-channel studies used different datasets, limiting direct accuracy comparisons; their confusion matrices and stage-specific metrics provide narrower reliability comparisons.

  • Accuracy cannot be compared directly across existing single-channel studies because they evaluated their methods on different datasets.
  • Recall, precision, and F1-score are presented as more useful indicators of algorithm reliability, especially for sleep stage N1.
  • The appendix reports confusion-matrix results for methods using Fpz-Cz, Cz-Pz, C3-A2 or Pz-Oz, and Fp1-Fp2 recordings.
  • One listed result reports ACC = 83.60% and MF1 = 68.11%.
Loading 1610.06421v3…