yasa_classifier

Repository to train a sleep staging classifier. We will use:

• An open source sleep staging dataset with ground truth labels (see dataset)
• yasa to extract features
• lightgbm to train a classifier

Dataset

The dataset was collected from OSF. It contains mouse EEG/EMG recordings (sampling rate: 512 Hz) and sleep stage labels (epoch length: 2.5 sec).

Training was performed using extracted features from 24h recordings.

Procedure to reproduce

Dataset can be downloaded using this link

https://files.osf.io/v1/resources/py5eb/providers/osfstorage/?zip=

or downloaded manually from OSF. See Dataset Structure for details of expected file structure.

Dataset structure

Datasets have the following structure

.
├── Mouse01
│   ├── Day1_dark_cycle
│   │   ├── EEG.mat
│   │   ├── EMG.mat
│   │   └── labels.mat
│   ├── Day1_light_cycle
│   │   ├── EEG.mat
│   │   ├── EMG.mat
│   │   └── labels.mat
│   ├── Day2_dark_cycle
│   │   ├── EEG.mat
│   │   ├── EMG.mat
│   │   └── labels.mat
│   └── Day2_light_cycle
│       ├── EEG.mat
│       ├── EMG.mat
│       └── labels.mat

Extract Features

01-extract_features.qmd was run to extract features. An important note is that it uses a local version of SleepStaging() (from staging import SleepStaging) that differs from the implementation in yasa. This was included for reproducibility, though we have plans to include this version in yasa itself and will be no longer needed.

Train the model

02-train.qmd was run to train on the 24 hour recordings. The outputs of this notebook are saved into /output.

Output structure

Model output will be stored as

output/
├── classifiers/
│   ├── eeg_full/
│   │   ├── model.joblib
│   │   └── feature_importances.csv
│   ├── eeg_no_kurt/
│   │   ├── model.joblib
│   │   └── feature_importances.csv
│   ├── eeg_no_std/
│   |   ├── model.joblib
│   |   └── feature_importances.csv
|   |
|   | ...

Evaluate the model

03-evaluate.qmd was run to evaluate and produce accuracy and cohen's kappa metrics. The outputs of this notebook are saved into /output.

Contribute

This is a preliminary release, file issues to enhance functionality.