A Deep Learning Model for Automated Sleep Stages Classification Using PSG Signals (2019)

irehste·2023년 8월 7일
Medical AI
0

Paper Review

목록 보기
1/1
  1. Introduction

  • Sleep disorder (수면 장애)

    • Insomnia, hypersomnias, parasomnias, sleep-related breathing, narcolepsy, circadian rhythm disorders, and sleep-related movement disorders, etc.
    • 건강한 수면을 취하지 못하거나, 충분한 수면을 취하고 있음에도 낮 동안에 각성을 유지하지 못하는 상태, 또는 수면 리듬이 흐트러져 있어 잠자거나 깨어 있을 때 어려움을 겪는 상태를 포함하는 매우 폭넓은 개념

  • Polysomnogram(PSG) recordings (수면다원검사)

    • recordings of physiological signals that are collected during an entire night of sleep
    • multivariate system; EEG + ECG + EOG + EMG
    • sleep stage scoring is performed on PSG records
    • sleep experts visually evaluate the PSG signals for a specific time frame -> determine scores according to various criteria

  • Sleep stages (AASM guidelines)

    • W(wake), NREM(non-rapid eye movement), REM(rapid eye movement)
    • W: awakening before sleep
    • NREM S1: brain activity slows down, muscles are relaxed
    • NREM S2: actual sleep begins, eye movements stop
    • NREM S3: deep sleep; brain function significantly reduced
    • NREM S4: deep sleep continued
    • REM: eyes closed but move rapidly

  • Why automated detection is needed?
    Visual inspection of PSG signals and manual determintation of sleep stages is a complex, costly and problematic process + visually hard to detect EEG signal variations

  • Why EEG?
    Most commonly used PSG signal for sleep stage classification; can be easily obtained with wearable technologies and consist useful information
    - EEG signal processing: Feature extraction, feature selection, and classification steps are commonly used

  • Proposed Model & Contribution

    • 1D-CNN for automated sleep stage classification
    • End-to-end structure; no handcrafted feature is used for sleep stage recognition with raw PSG signals
    • Can be used without changing any of its layer parameters for two to six sleep classes and other types of PSG signals
  1. Materials (Datasets)
  • Sleep-edf dataset
    - records of 8 males and females' PSG records(two EEG channels) + hypnogram files(that contain annotations of sleep patterns)
  • Sleep-edfx dataset
    - PSG records of 61 subjects + scored recordings for six sleep stages
  • W >> S1 > ...
  1. Method
  • 1D CNN
    - Input: 1D sequence
    - Convolutional layers: feature map as output
    (S∗W)n=∑i=1∣W∣W(i)S(i+n−1)(S*W)_n = \sum_{i=1}^{|W|} W(i)S(i+n-1) (convolution operation)
    - Activation function: applied element-wise to the convolved features
    - Pooling layers: downsample spatial dimensions of the feature map; prevents overfitting
    - Fully connected layers
    - Output layer: produces classification/regression output
  • Input: preprocessed PSG signals with 3000 samples
  • First convolution layer (64x5 filters, 3 stride ratios) -> 64x999 feature maps
  • Second convolution layer(128x5 filters) -> 128x997 feature maps
  • MaxPool layer: maximum values in two unit regions of the two output vectors are reduced to a single value -> 129x498
  • Consecutive layers in similar fashion but with different sizes of filters
  • Dropout layers: to prevent overfitting
  • Flattened layer: converted to the appropriate dimenstions for the dense layers
  • Softmax layer: input signals are mapped to the output signals

  1. Experiment Results
    Highest recognition rate: 98.33% (for C=2 using single-channel EEG signals with the sleep-edf dataset)

  2. Conclusion

  • Advantages of proposed study
    : Complete end-to-end recognition structure is developed without the need of any manual feature extraction stages, detection performance is significantly improved for five classes (2 to 6) using different PSG signals without changing any parameters of layer values
  • Disadvantage of proposed study
    : Elimination of noise from the PSG, improvement of performance of S1 is needed
  • Future Work
    : Different deep learning approaches such as CNN + LSTM to increase recognition performance, use of patient level criteria instead of signal level criteria to evaluate the performances
profile
irehste

1개의 댓글

comment-user-thumbnail
happy
2023년 8월 7일

정보에 감사드립니다.

답글 달기