TS-SEN

Two-Stream Squeeze-and-Excitation Network for Multi-modal Sleep Staging

Two-Stream Squeeze-and-Excitation Network (TS-SEN) is a deep learning model for capturing the features of electroencephalogram (EEG) and electrooculogram (EOG) for sleep staging. The TS-SEN is made up of two independent feature extraction networks for modeling the heterogeneity and a multi-modal Squeeze-and-Excitation feature fusion module for adaptively utilizing the multi-modal signals.

This repository contains the source code of the TS-SEN.

Dataset

We evaluate our model on the Montreal Archive of Sleep Studies (MASS) dataset Subset-3. The MASS dataset is an open-access and collaborative database of laboratory-based polysomnography (PSG) recordings.

Requirements

• Python 3.6.2
• CUDA 9.0
• CuDNN 7.6.4
• numpy==1.19.5
• sklearn==0.19.1
• tensorflow_gpu==1.8.0
• Keras==2.2.0

Usage

• Preprocess

  • Prepare data (raw signals) in data_dir
    • Files name: 01-03-00XX-Data.npy, where XX denotes subject ID.
    • Tensor shape: [sample, channel=26, length], where channel 0 is ECG channel, 1-20 are EEG channels, 21-23 are EMG channels, 24-25 are EOG channels.
  • Modify directory
    • Modify data_dir and label_dir in run_preprocess.py accordingly.
    • (Optional) Modify output_dir in run_preprocess.py.
  • Run preprocess program
    • cd preprocess
    • python run_preprocess.py

• Training Command Line Arguments

  • Training
    • --batch_size: Training batch size.
    • --epoch: Number of training epochs.
    • --num_fold: Number of folds.
    • --save_model: Save the best model or not.
    • --early_stop: Use the early stop trick or not.
    • --gpunums: Number of GPU(s).
    • --from_fold: Training start fold.
    • --to_fold: Training end fold.
  • Directory
    • --label_dir: The directory of labels.
    • --data_dir_eog: The directory of the EOG signals.
    • --data_dir_psd: The directory of the spectral-spatial representation of EEG signals.
    • --output: The directory for saving results.
    • --model_dir: The directory for saving the best models of each fold.

• Training

  Run run.py with the command line arguments. By default, the model can be run with the following command:

  CUDA_VISIBLE_DEVICES=0 python run.py