- Final project of "CS598. Deep Learning for Healthcare" course at University of Illinois at Urbana-Champaign
- Team:
[email protected](1 member)
In this experiment, I construct two deep neural networks to score the sleep stage automatically. Both models are working on single-channel signals from Sleep Cassette study dataset in publicly available Sleep-EDF 2018 dataset. The first model comprises multiple convolutional neural networks (CNN) and bidirectional long short-term memory (Bi-LSTM) networks achieving overall accuracy of 75.4%, an average macro F1-score of 67.8% across all folds in 20-fold cross-validation and inter-rater reliability coefficient Cohen's kappa κ = 0.66. The second model is based on convolutional neural networks predicting with an average accuracy of 79% and average macro F1-score of 75% and κ =0.70 in the same cross-validation procedure. The result shows that my model is comparable to state-of-the-art methods with hand-engineered features.
Read the full report here
I evaluate the model with data from Sleep Cassette study of Sleep-EDF 2018 Dataset. Run the command below to download all data:
cd data
chmod +x download_sleep_edfx.sh
./download_sleep_edfx.sh
Then run the following script to extract specified EEG channels and their corresponding sleep stages. (this script was forked from DeepSleepNet)
python prepare_physionet.py --data_dir data/sleep-cassette --output_dir data/sc_eeg_fpz_cz --select_ch 'EEG Fpz-Cz'
python prepare_physionet.py --data_dir data/sleep-cassette --output_dir data/sc_eeg_pz_oz --select_ch 'EEG Pz-Oz'
usage: main.py [-h] [--run RUN] [--data_dir DATA_DIR]
[--output_dir OUTPUT_DIR] [--fold_ids FOLD_IDS]
[--model MODEL] [--total_fold TOTAL_FOLD]
optional arguments:
-h, --help show this help message and exit
--run RUN running mode: train / summarize.
+ train: Training. Need to specfiy data_dir, output_dir, fold_ids, model and total_fold
+ summarize: View result of the last run in which the result is stored in `output_dir`
--data_dir DATA_DIR directory contain extracted channels
--output_dir OUTPUT_DIR
Directory to store model, progress, result
--fold_ids FOLD_IDS fold/folds to train, each valid fold from [0 to total_fold-1].
+ Can be a single fold, for example: 0 - the first fold, or 19 - the last fold in 20-fold cross validation
+ Can contain multiple folds, separate by comma `,`, for example: 0,1,2.
+ -1 to train on all folds
--model MODEL model to train and evaluate performance
+ mod_sleep_eeg: Modified SleepEEG - using CNNS
+ mod_deep_sleep: Modified DeepSleep - using CNNs + Bi-LSTMs
--total_fold TOTAL_FOLD
Number of fold
- Train & evaluate result on the first fold of Modified Sleep EEG Net
python main.py --model mod_sleep_eeg --fold_ids 0 --data_dir data/sc_eeg_fpz_cz --output_dir output/sc_eeg_fpz_cz
- Train & evaluate result on the first two folds of Modified SleepEEG Net
python main.py --model mod_sleep_eeg --fold_ids 0,1 --data_dir data/sc_eeg_fpz_cz --output_dir output/sc_eeg_fpz_cz
- Train & evaluate result on the last fold of Modified DeepSleep Net
python main.py --model mod_deep_sleep --fold_ids 19 --data_dir data/sc_eeg_fpz_cz --output_dir output/sc_eeg_fpz_cz
- Train and evaluate result on all folds of Modified SleepEEG Net
python main.py --model mod_sleep_eeg --fold_ids -1 --data_dir data/sc_eeg_fpz_cz --output_dir output/sc_eeg_fpz_cz
- View summary of performance of the last run of modified_sleep_eeg model (Can only view summary if you finished training)
python main.py --run summarize --output_dir output/sc_eeg_fpz_cz/modified_sleep_eeg
-
My hardware:
- AMD (R) Ryzen 7 3700x 8 cores 16 threads
- 32 GB RAM
- GPU GeForce GTX 1660 6GB
- 1TB SSD.
-
The software environment is as follows:
- python
3.7.10 - tensorflow/tensorflow-gpu
2.4.1 - numpy
1.19.2 - pandas
1.2.4 - scikit-learn
0.24.1 - mne
0.23.0
- python
Notice: I have faced some runtime issues with tensorflow 2.1.x, and numpy 1.20, so it's advised to create a conda enviroment from environment.yml to make sure you have the same libraries as mine.
You can use the script below:
conda env create -f environment.yml