EEG_classification

Description of the approach : https://towardsdatascience.com/sleep-stage-classification-from-single-channel-eeg-using-convolutional-neural-networks-5c710d92d38e

Sleep Stage Classification from Single Channel EEG using Convolutional Neural Networks

Photo by Paul M on Unsplash

Quality Sleep is an important part of a healthy lifestyle as lack of it can cause a list of issues like a higher risk of cancer and chronic fatigue. This means that having the tools to automatically and easily monitor sleep can be powerful to help people sleep better.
Doctors use a recording of a signal called EEG which measures the electrical activity of the brain using an electrode to understand sleep stages of a patient and make a diagnosis about the quality if their sleep.

In this post we will train a neural network to do the sleep stage classification automatically from EEGs.

Data

In our input we have a sequence of 30s epochs of EEG where each epoch has a label {“W”, “N1”, “N2”, “N3”, “REM”}.

Fig 1 : EEG Epoch

Fig 2 : Sleep stages through the night

This post is based on a publicly available EEG Sleep data ( Sleep-EDF ) that was done on 20 subject, 19 of which have 2 full nights of sleep. We use the pre-processing scripts available in this repo and split the train/test so that no study subject is in both at the same time.

The general objective is to go from a 1D sequence like in fig 1 and predict the output hypnogram like in fig 2.

Model Description

Recent approaches [1] use a sub-model that encodes each epoch into a 1D vector of fixed size and then a second sequential sub-model that maps each epoch’s vector into a class from {“W”, “N1”, “N2”, “N3”, “REM”}.

Here we use a 1D CNN to encode each Epoch and then another 1D CNN or LSTM that labels the sequence of epochs to create the final hypnogram. This allows the prediction for an epoch to take into account the context.

Sub-model 1 : Epoch encoder

Sub-model 2 : Sequential model for epoch classification

The full model takes as input the sequence of EEG epochs ( 30 seconds each) where the sub-model 1 is applied to each epoch using the TimeDistributed Layer of Keras which produces a sequence of vectors. The sequence of vectors is then fed into a another sub-model like an LSTM or a CNN that produces the sequence of output labels.
We also use a linear Chain CRF for one of the models and show that it can improve the performance.

Training Procedure

The full model is trained end-to-end from scratch using Adam optimizer with an initial learning rate of 1e⁻³ that is reduced each time the validation accuracy plateaus using the ReduceLROnPlateau Keras Callbacks.

Accuracy Training curves

Results

We compare 3 different models :

• CNN-CNN : This ones used a 1D CNN for the epoch encoding and then another 1D CNN for the sequence labeling.
• CNN-CNN-CRF : This model used a 1D CNN for the epoch encoding and then a 1D CNN-CRF for the sequence labeling.
• CNN-LSTM : This ones used a 1D CNN for the epoch encoding and then an LSTM for the sequence labeling.

We evaluate each model on an independent test set and get the following results :

• CNN-CNN : F1 = 0.81, ACCURACY = 0.87
• CNN-CNN-CRF : F1 = 0.82, ACCURACY =0.89
• CNN-LSTM : F1 = 0.71, ACCURACY = 0.76

The CNN-CNN-CRF outperforms the two other models because the CRF helps learn the transition probabilities between classes. The LSTM based model does not work as well because it is most sensitive to hyper-parameters like the optimizer and the batch size and requires extensive tuning to perform well.

Ground Truth Hypnogram

Predicted Hypnogram using CNN-CNN-CRF

Source code available here : https://github.com/CVxTz/EEG_classification

I look forward to your suggestions and feedback.

[1] DeepSleepNet: a Model for Automatic Sleep Stage Scoring based on Raw Single-Channel EEG

How to cite:

@software{mansar_youness_2020_4060151,
  author       = {Mansar Youness},
  title        = {CVxTz/EEG\_classification: v1.0},
  month        = sep,
  year         = 2020,
  publisher    = {Zenodo},
  version      = {v1.0},
  doi          = {10.5281/zenodo.4060151},
  url          = {https://doi.org/10.5281/zenodo.4060151}
}