SleepStages runs on Windows and Mac machines with Python 3 installed. Required packages can be installed using pip (Python package installer).
pip install torch pytorch-ignite torchsummary pyserial PyQt5 tqdm scipy matplotlib sklearn
If the user chooses to obtain EEG signals using National Instruments DAQ, NI-DAQmx-Base software library should be downloaded and installed. It is freely available at the National Instruments website. NI-DAQmx-Base is not necessary when predicting stages for already-recorded EEG signals offline only and not using the application online (real-time).
http://www.ni.com/download/ni-daqmx-base-15.0/5648/en/
SleepStages was tested using National Instruments USB-6210 (Multifunction I/O Device), but other NI devices will probably work too. A Python package for running DAQmx is also necessary if NI-DAQ is used for data acquisition.
pip install PyDAQmx
The source code and data files can be downloaded from GitHub by:
git clone [email protected]/tarotez/sleepstages
Move to the code directory before running the program.
cd sleepstages/code
The program can be run in a demo mode (mockup mode) by
python app.py m
"m" is for mock-up. In this mode, the application reads a short pickled wave file stored as "data/pickled/eegAndStage.sample.pkl". The demo file is very short and all predictions will be "wake".
The application can be started in online (real-time) mode by
python app.py
Since the default prediction model requires past EEG epochs, the classifier starts to predict after a 10-epochs "burn-in" period. Before that, the classifier outputs '?' because it has not accumulated enough information to make predictions.
The default setup only accepts input signals sampled at 128 Hz. In order to predict stages using signals sampled at a different frequency, the user should modify the neural network model defined in "code/deepClassifier.py" and train a classifier using that model and signals sampled at that sampling frequency. It requires some knowledge on PyTorch to design and implement an appropriate new neural network model, but basically it amounts to changing a few parameters designating the size of input.
The predicted sleep stage can be sent to an Arduino-based external device connected by USB.
The system can predict EEG stages for already-recorded EEG signals stored in a text file. It runs offline without acquiring signals from DAQ by
python app.py o
or
python app_offline.py
The latter is an alias to the former.
In the offline mode, the program obtains EEG wave data from the "data/aipost" directory. The wave data should be provided as a text file. There is a short sample file in the "data/aipost" directory.
The predicted stages are written out as a text files in the "data/prediction" directory.
There is a sample wave file in the "data/aipost" directory. In this text file, each row (line) corresponds to a time point sampled at 128 Hz. The columns are (from left to right) the time stamp, EEG amplitude, Channel 2 amplitude, in this order, separated by commas.
If EMG nor mouse movement is not recorded, the column for Channel 2 can be left blank.
Channel 2 is used to overwrite the predicted sleep stage to "Wake" when an excessive mouse movement is detected. The algorithm for judging a mouse movement is as follows. Each 10-seconds epoch is divided into 80 segments (that is, each segment is 0.125 second-long). Within each segment, the amplitude of Channel 2 is averaged. When the maximum of these 80 averaged amplitudes exceeds the threshold set by the scrollbar on the GUI, the predicted sleep stage is changed to "Wake". In "data/prediction", files ending with "_pred_before_overwrite.txt" list predicted stages before overwriting, and files ending with "_pred.txt" list predicted stages after overwriting.
Sleep stages for wave files in the "data/aipost" directory can be predicted without activating the GUI. Instead of running app.py, run
python offline.py
The predicted result is written out as files in the "data/prediction" directory.
To evaluate the result of prediction, use
python eval_offline.py PREDICTION_FILE JUDGE_PATH
where PREDICTION_FILE is the name of the prediction file in "data/prediction", and JUDGE_PATH is the path to the Judge file that contain ground-truth labels for each epoch.
The "sleep stages" repository contains the source code and text files that set up default hyperparameters. The directory structure is as follows:
sleepstages/
code/
data/
params
pickled
waves
WAVEDIR/
Raw/
Judge/
"WAVEDIR" is a directory with an arbitrary name in which text files containing EEG raw data signals and ground truth stage labels are stored. These directories are not needed if the user chooses to use an already trained classifier (provided as a part of this software) and not training a new classifier. For example, a user might want to train a new classifier to make the system predict sleep stages for epochs with different lengths.
The training program reads this directory to train classifiers. The user can place more than one WAVEDIR directory under the "data" directory. In other words, the user can prepare a directory for each dataset.
We included a sample "WAVEDIR" named "data/sampledata". Its "Raw" and "Judge" directories contain sample wave and judge files.
The code in the "code" directory uses the file named "path.json" to find the "data" directory. It can be changed to another directory name by rewriting "path.json".
The "waves" directory is where EEG recordings are written out in case of online classification. By default, data/params/params.json is set to "writeWholeWaves" : 0, suppressing this output. By setting to "writeWholeWaves" : 1, whole EEG recordings are written out to the "waves" directory.
The "data/params" directory should contain "params.json" for setting up parameters for feature extraction, training, and prediction.
Be editing "params.json", the behavior of the GUI and also of training can be altered.
A new classifier can be trained and validated by the following procedure:
python readOfflineEEGandStageLabels2pickle.py WAVEDIR
python extractFeatures.py
python trainClassifier.py
python computeTrainingError.py
python computeTestError.py
readOfflineEEGandStageLabels2pickle.py reads text files containing EEG raw data signals and ground truth stage labels from the WAVEDIR directory. It writes files starting with "eegAndStage" into the "data/pickled" directory. These files are in Python's pickle format to enable faster access.
extractFeatures.py reads "eegAndStage" files and write files starting with "features". These files contain feature vectors used for training classifiers. They will be generated in "data/pickled".
trainClassifier.py reads "features" and writes files starting with "weights", "params", and "files_used_for_training". These files contain randomly generated six-character IDs (i.e., classifier IDs) in their file names.
The "weights" file contains weight parameters obtained from training. The "params" file is a copy of "params.json" in "data/pickled" that is intended to save the parameters used for training the classifier. "files_used_for_training" indicates which recordings were used for training that classifier. These files are excluded when testing the classifier.
To have a newly trained model in applicational programs (offline.py, online.py and app.py), files params.json and weights.pkl must be copied from data/params to data/finalclassifier.
data/params : used for training and testing
data/finalclassifier : used for making predictions by offline.py, online.py, and app.py
Edit or revise data/finalclassifier/classifierTypes.csv so that the newly trained model can be used by applicational programs (offline.py, online.py and app.py). For each line, the format is
classifierID, network type, sampling frequency, epoch time length.
In summary, the pipeline is as follows:
- Train using data/params and data/pickled
- Choose a good trained model and copy params.CLASSIFIER_ID.json and weights.CLASSIFIER_ID.pkl to to data/finalclassifier
- Run either offline.py, online.py or app.py
Some results of experiments using this software are provided in our paper:
Taro Tezuka, Deependra Kumar, Sima Singh, Iyo Koyanagi, Toshie Naoi, Masanori Sakaguchi, Real-time, automatic, open-source sleep stage classification system using single EEG for mice, Scientific Reports, 11:11151, May 2021. [DOI:10.1038/s41598-021-90332-1]
https://www.nature.com/articles/s41598-021-90332-1
Iyo Koyanagi, Taro Tezuka, Jiahui Yu, Sakthivel Srinivasan, Toshie Naoi, Shinnosuke Yasugaki, Ayaka Nakai, Shimpei Taniguchi, Yu Hayashi, Yasushi Nakano and Masanori Sakaguchi, Fully automatic REM sleep stage-specific intervention systems using single EEG in mice, Neuroscience Research, 186:51-58, January 2023. [DOI:10.1016/j.neures.2022.10.001]
https://www.sciencedirect.com/science/article/pii/S0168010222002620
A usage video is available from the following link.