amibeuret/spindle

This is a simple framework to preprocess and classify sleep stages based on recorded EEG/EMG data

DeepSleep

This is a simple framework to preprocess and classify sleep stages based on recorded EEG/EMG data. It replicates and extends on top of the Spindle method.

Updates

• 10.08.2023 This code base was last modified on May 7 2021 and was an attempt to replicate the paper's results as close as possible in Pytorch. A very similar implementation is used in the website.

Installation

First make a new python virtual environment and then install the package using pip:

pip install deepsleep

Spindle sleep scoring

You could use both preprocessing routines and CNN network for the classification demonstrated in the paper

Spindle preprocessing routine

Below is an example of how to use the preprocessing routine:

import numpy as np
import pyedflib
from deepsleep.data import SpindlePreproc
from deepsleep.utils import setup_logger_to_std


# Set the logger to print to stdout. You could skip this and ignore logs.
setup_logger_to_std()

# Adapt this for your own .edf data
eeg_file = '/path/to/eeg.edf'

# preprocessing parameters. Below are defaults used as in the paper
SPINDLE_PREPROCESSING_PARAMS = {
    'name': 'SpindlePreproc',
    'target_srate': 128,
    'spectrogram-stride': 16,
    'time_interval': 4,
    'num_neighbors': 4,
    'EEG-filtering': {'lfreq': 0.5, 'hfreq': 12},
    'EMG-filtering': {'lfreq': 2, 'hfreq': 30}
}

print("Loading EEG/EMG...")
all_signals, signal_header, header = pyedflib.highlevel.read_edf(eeg_file)

print("Preprocessing data...")
preprocessing = SpindlePreproc(SPINDLE_PREPROCESSING_PARAMS)
data = preprocessing(all_signals, signal_header, np.array([0, 1]), np.array([]), np.array([2]))

Spindle CNN network

You could import the pytorch CNN network for training:

from deepsleep.models import SpindleGraph

myGraph = SpindleGraph(input_dim=(3, 24, 160), nb_class=3, dropout_rate=0.5)

Making prediction with Spindle

There are two ways of making predictions with our pretrained model, either by importing the model, preprocessing routine etc. from the library, or simply using the terminal and commandline:

Predictions using command line

In the terminal you could simply provide the path to your .edf file, a path where you would like to save predictions .csvs, and the path to the directory where the pre-trained weights are stored as demonstrated below:

predict.py /path/to/my/AWESOME_EEG.edf --predictions /path/to/csv_folder/ --weight_dir /path/to/weights/

Two .csv files /path/to/csv_folder/AWESOME_EEG_ad_predictions_prob.csv and /path/to/csv_folder/AWESOME_EEG_predictions_without_artifacts.csv will be stored as a result.

The directory /path/to/weights/ should contain two provided weights named checkpoint_replicatePaperAD_1563189723.278988epoch3.pth and checkpoint_replicatePaperSS_1563188754.9532504epoch10.pth

Predictions by importing the library

Very similar to predict.py, you could follow this example: (Make sure you set all the paths)

from pathlib import Path

from deepsleep.data import SpindlePreproc
from deepsleep.data import ProdData
from deepsleep.models import SpindlePredictModel
from deepsleep.utils import setup_logger_to_std
from deepsleep.utils import set_up_paths


# Set parameters for preprocessing
SPINDLE_PREPROCESSING_PARAMS = {
    'name': 'SpindlePreproc',
    'target_srate': 128,
    'spectrogram-stride': 16,
    'time_interval': 4,
    'num_neighbors': 4,
    'EEG-filtering': {'lfreq': 0.5, 'hfreq': 12},
    'EMG-filtering': {'lfreq': 2, 'hfreq': 30}
}

# Set parameters for pytorch data loader
DATALOADER_PARAMS = {
    'num_workers': 4,
    'batch_size': 100,
    'do_shuffle': False,
    'batch_prefetch': 10,
    'hdf5': '',
    'folds': ['fold1']
}

# Set parameters to set up the Spindle model
MODEL_PARAMS = {
    'name': 'SpindlePredictModel',
    'artefact_threshold': 0.5,  # The probability threshold more than which the sample is considered as a noise sample
    'weights': ['checkpoint_replicatePaperAD_1563189723.278988epoch3.pth',
                'checkpoint_replicatePaperSS_1563188754.9532504epoch10.pth']
}

# Set the logger to print to stdout
setup_logger_to_std()

# Set the required paths
root_path = Path('/path/to/my/results_folder')
data_path = Path('/path/to/eeg.pdf')  # Set this to your input .edf file
weights_path = Path('/path/to/weights/')  # Set this to the folder in which weights are located
set_up_paths(root_path=root_path, data_path=data_path, weights_path=weights_path)

# Load data and preprocess
preprocessing = SpindlePreproc(SPINDLE_PREPROCESSING_PARAMS)
pred_handlers = ProdData(DATALOADER_PARAMS, preprocessing)

# Make model
params = dict({'model': MODEL_PARAMS})
params['model']['predictions_path'] = Path('/path/to/csv_folder/preds_')  #  Set this to the folder in which .csv prediction files will be saved
model = SpindlePredictModel(params)

# Set inputs and predict
model.set_inputs(prediction=[pred_handlers])
model.prediction()

spindle's Issues

Incomplete installation

Unable to install deepsleep package properly, ends up providing an error about pytorch version, as well as numpy version.
does not seem to be able to read deepsleep package.