Hybrid CNN-Transformer for OSA diagnosis on PSG

This repository contains the source code of the following paper:

A novel deep learning model for obstructive sleep apnea diagnosis: hybrid CNN-Transformer approach for radar-based detection of apnea-hypopnea events
Sleep 2024 Dec 11;47(12):zsae184. doi: 10.1093/sleep/zsae184.

Installation

conda create -n osaai python=3.9
conda activate osaai

git clone https://github.com/jwc-rad/osa-ai.git
cd osa-ai
pip install -r requirements.txt

This repo is based on Python 3.9 and PyTorch 2.2.2 + CUDA 11.8. You may have to change to the PyTorch version that suits your workstation.

Usage

Data format

Currently, all the data including PSG signals and labels are expected to be in 8Hz.

The the event labels are 0 (normal), 1 (hypopnea), and 2 (apnea), and the sleep stages are 0 (wake) and 1+ (other sleep stages).

The sampling frequency and labels can be changed according to custom datasets, then the config (config/data/dataset/2311_clsHA_2ch_v0.yaml) should be changed also.

Dataset format

The base dataset directory DATASET_DIR contains input (e.g. Radar), label (e.g. Event), and sleep stage (e.g. Stage) directories with {PATIENT_ID}.npy files, and a dataset split file.

DATASET_DIR/
├── Event
│   ├── PSG_0001.npy
│   ├── PSG_0002.npy
│   ├── ...
├── Radar1
│   ├── PSG_0001.npy
│   ├── PSG_0002.npy
│   ├── ...
├── Radar2
│   ├── PSG_0001.npy
│   ├── PSG_0002.npy
│   ├── ...
├── Stage
│   ├── PSG_0001.npy
│   ├── PSG_0002.npy
│   ├── ...
├── split.json

A sample code for generating split file for 5-fold CV based on sklearn.model_selection.StratifiedKFold on OSA classes:

# suppose you already have list of cases and OSA classes for dev and test sets

kf = StratifiedKFold(n_splits=5, random_state=12345, shuffle=True)

splits = {}

for i, (train_idx, valid_idx) in enumerate(kf.split(cases_trainvalid, osa_trainvalid)):
    splits[i] = {
        'train': [x for j, x in enumerate(cases_trainvalid) if j in train_idx],
        'valid': [x for j, x in enumerate(cases_trainvalid) if j in valid_idx],
        'test': cases_test,
    }

ds_json = {
    'split': splits,
}

with open(os.path.join(DATASET_DIR, 'split.json'), 'w') as f:
    json.dump(ds_json, f)

Run training

python train.py --multirun experiment=psgradar_paper data.dataset.cv_fold=0,1,2,3,4

For different options, look at config/experiment/psgradar_paper.yaml and config/data/dataset/2311_clsHA_2ch_v0.yaml for details.

Inference

Refer to inference.ipynb.

Citation

If you find this repo useful for your work, please consider citing:

@Article{10.1093/sleep/zsae184,
    author = {Choi, Jae Won and Koo, Dae Lim and Kim, Dong Hyun and Nam, Hyunwoo and Lee, Ji Hyun and Hong, Seung-No and Kim, Baekhyun},
    title = {A novel deep learning model for obstructive sleep apnea diagnosis: hybrid CNN-Transformer approach for radar-based detection of apnea-hypopnea events},
    journal = {Sleep},
    volume = {47},
    number = {12},
    pages = {zsae184},
    year = {2024},
    month = {08},
    issn = {0161-8105},
    doi = {10.1093/sleep/zsae184},
    url = {https://doi.org/10.1093/sleep/zsae184},
    eprint = {https://academic.oup.com/sleep/article-pdf/47/12/zsae184/60774106/zsae184.pdf},
}

If you find our preliminary study useful, please also consider citing:

@Article{s22197177,
    AUTHOR = {Choi, Jae Won and Kim, Dong Hyun and Koo, Dae Lim and Park, Yangmi and Nam, Hyunwoo and Lee, Ji Hyun and Kim, Hyo Jin and Hong, Seung-No and Jang, Gwangsoo and Lim, Sungmook and Kim, Baekhyun},
    TITLE = {Automated Detection of Sleep Apnea-Hypopnea Events Based on 60 GHz Frequency-Modulated Continuous-Wave Radar Using Convolutional Recurrent Neural Networks: A Preliminary Report of a Prospective Cohort Study},
    JOURNAL = {Sensors},
    VOLUME = {22},
    YEAR = {2022},
    NUMBER = {19},
    ARTICLE-NUMBER = {7177},
    URL = {https://www.mdpi.com/1424-8220/22/19/7177},
    PubMedID = {36236274},
    ISSN = {1424-8220},
    DOI = {10.3390/s22197177}
}