SleePyCo

By Seongju Lee, Yeonguk Yu, Seunghyeok Back, Hogeon Seo, and Kyoobin Lee

This repo is the official implementation of "SleePyCo: Automatic Sleep Scoring with Feature Pyramid and Contrastive Learning", accepted to Expert Systems With Applications (I.F. 8.5).

[Paper] [CodeOcean]

Model Architecture

Training Framework

Updates & TODO Lists

• (2023.03.03) Official repository of SleePyCo is released
• Script for preprocessing Sleep-EDF
• Config files for training from scratch
• (2023.11.09) Config files for ablation studies
• Add TinySleepNet baseline and Fix minor errors
• You can download checkpoints more conveniently!
• (2023.11.18) 🎉Online publication is available!🎉
• Scripts for preprocessing MASS, Physio2018, SHHS

Getting Started

Environment Setup

Trained and evaluated on NVIDIA GeForce RTX 3090 with python 3.8.5.

1. Set up a python environment

conda create -n sleepyco python=3.8.5
conda activate sleepyco

2. Install PyTorch with compatible version to your develop env from PyTorch official website.

3. Install remaining libraries using the following command.

pip install -r requirements.txt

Dataset Preparation

Sleep-EDF dataset

1. Download Sleep-EDF-201X dataset via following command. (X will be 3 or 8)

cd ./dset/Sleep-EDF-201X
python download_sleep-edf-201X.py

2. Check the directory structure as follows

./dset/
└── Sleep-EDF-201X/
    └── edf/
        ├── SC4001E0-PSG.edf
        ├── SC4001EC-Hypnogram.edf
        ├── SC4002E0-PSG.edf
        ├── SC4002EC-Hypnogram.edf
        ├── ...

3. Preprocess .edf files into .npz.

python prepare_sleep-edf-201X.py

4. Check the directory structure as follows

./dset/
└── Sleep-EDF-201X/
    ├── edf/
    │   ├── SC4001E0-PSG.edf
    │   ├── SC4001EC-Hypnogram.edf
    │   ├── SC4002E0-PSG.edf
    │   ├── SC4002EC-Hypnogram.edf
    │   ├── ...
    │
    └── npz/
        ├── SC4001E0-PSG.npz
        ├── SC4002E0-PSG.npz
        ├── ...

Train & Evaluation (SleePyCo Training Framework)

Contrastive Representation Learning

python train_crl.py --config configs/SleePyCo-Transformer_SL-01_numScales-1_{DATASET_NAME}_pretrain.json --gpu $GPU_IDs

When one GeForce RTX 3090 GPU is used, it may requires 22.3 GB of GPU memory.

Multiscale Temporal Context Learning

python train_mtcl.py --config configs/SleePyCo-Transformer_SL-10_numScales-3_{DATASET_NAME}_freezefinetune.json --gpu $GPU_IDs

When two GeForce RTX 3090 GPU is used, it may requires 16.7 GB of GPU memory each.

If you use PyTorch $\geq$ 2.0.0, it may requires only 5.4 GB of GPU memory.

Train & Evaluation (From Scratch)

python train_mtcl.py --config configs/SleePyCo-Transformer_SL-10_numScales-3_{DATASET_NAME}_scratch.json --gpu $GPU_IDs

Main Results

Dataset	Subset	Channel	ACC	MF1	Kappa	W	N1	N2	N3	REM	Checkpoints
Sleep-EDF-2013	SC	Fpz-Cz	86.8	81.2	0.820	91.5	50.0	89.4	89.0	86.3	Link
Sleep-EDF-2018	SC	Fpz-Cz	84.6	79.0	0.787	93.5	50.4	86.5	80.5	84.2	Link
MASS	SS1-SS5	C4-A1	86.8	82.5	0.811	89.2	60.1	90.4	83.8	89.1	Link
Physio2018	-	C3-A2	80.9	78.9	0.737	84.2	59.3	85.3	79.4	86.3	Link
SHHS	shhs-1	C4-A1	87.9	80.7	0.830	92.6	49.2	88.5	84.5	88.6	Link

How to reproduce results

1. Download and extract checkpoint using python download_checkpoints.py.

• You can download all checkpoints using following command:

  cd checkpoints
  python download_checkpoints.py
  

• You can also select checkpoints as follows:

  cd checkpoints
  python download_checkpoints.py --datasets 'Sleep-EDF-2013' 'Sleep-EDF-2018'
  

2. Evaluate the dataset using the following command.

python test.py --config configs/SleePyCo-Transformer_SL-10_numScales-3_{DATASET_NAME}_freezefinetune.json --gpu $GPU_IDs

Authors

• SeongjuLee [GoogleScholar] [GitHub]
• Yeonguk Yu [GoogleScholar] [GitHub]
• Seunghyeok Back [GoogleScholar] [GitHub]
• Hogeon Seo [GoogleScholar] [GitHub]
• Kyoobin Lee (Corresponding Author) [GoogleScholar]

License

The source code of this repository is released only for academic use. See the license file for details.

Citation

@article{lee2024sleepyco,
title = {SleePyCo: Automatic sleep scoring with feature pyramid and contrastive learning},
journal = {Expert Systems with Applications},
volume = {240},
pages = {122551},
year = {2024},
issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2023.122551},
url = {https://www.sciencedirect.com/science/article/pii/S0957417423030531},
author = {Seongju Lee and Yeonguk Yu and Seunghyeok Back and Hogeon Seo and Kyoobin Lee}
}

Acknowledgments

This research was supported by a grant from the Institute of Information and Communications Technology Planning and Evaluation (IITP) funded by the Korean government (MSIT) (No. 2020-0-00857, Development of cloud robot intelligence augmentation, sharing and framework technology to integrate and enhance the intelligence of multiple robots). Furthermore, this research was partially supported by the Korea Institute of Energy Technology Evaluation and Planning (KETEP) grant funded by the Korean government (MOTIE) (No. 20202910100030).