Do not 😴 on traditional ML

Simple and Interpretable Techniques Are Competitive to Deep Learning for Sleep Scoring

Code from the paper Do Not Sleep on Traditional Machine Learning: Simple and Interpretable Techniques Are Competitive to Deep Learning for Sleep Scoring.

Preprint: https://arxiv.org/abs/2207.07753
Published article: https://doi.org/10.1016/j.bspc.2022.104429

Citation:

@article{vanderdonckt2023donotsleep,
  title={Do not sleep on traditional machine learning: Simple and interpretable techniques are competitive to deep learning for sleep scoring},
  author={Van Der Donckt, Jeroen and Van Der Donckt, Jonas and Deprost, Emiel and Vandenbussche, Nicolas and Rademaker, Michael and Vandewiele, Gilles and Van Hoecke, Sofie},
  journal={Biomedical Signal Processing and Control},
  volume={81},
  pages={104429},
  year={2023},
  publisher={Elsevier}
}

How is the code structured?

For each dataset you can find a separate notebook in the notebooks folder.

The notebooks allow to reproduce the results as they contain;

data loading (see code in src folder)
pre-processing & feature extraction
(seeded) machine learning experiments

notebook	dataset
SleepEDF-SC +- 30min.ipynb	`SC-EDF-20` & `SC-EDF-78`
SleepEDF-ST	`SC-EDF-ST`
MASS-SS3	`MASS SS3`

Additional experiments

The notebooks/other folder contains some additional experiments;

notebook	experiment description
inputs_SleepEDF-SC +- 30min.ipynb	evaluate impact of signal combination on performance for `SC-EDF-20` & `SC-EDF-78`
inputs_SleepEDF-ST.ipynb	evaluate impact of signal combination on performance for `SC-EDF-ST`
inputs_SleepEDF-MASS.ipynb	evaluate impact of signal combination on performance for `MASS SS3`
feature_selection.ipynb	show the (little to no) impact of feature selection on performance
feature_space_visualization.ipynb	`PCA` and `t-SNE` visualization of the feature vector for `SleepEDF-SC +/- 30min`

A table showing the impact of signal combination on performance can be found in notebooks/other/signal_combination_impact.md.

How to install the requirements?

This repository uses poetry as dependency manager.
A specification of the dependencies is provided in the pyproject.toml and poetry.lock files.

You can install the dependencies in your Python environment by executing the following steps;

Install poetry: https://python-poetry.org/docs/#installation
Install the dependencies by calling poetry install

How to download the datasets?

This work uses 4 (sub)sets of data;

SC-EDF-20: first 20 patients (40 recordings) of Sleep-EDFx - Sleep Cassette
SC-EDF-78: : all 78 patients (153 recordings) of Sleep-EDFx - Sleep Cassette
ST-EDF: all 22 patients (44 recordings) of Sleep-EDFx - Sleep Telemetry
MASS SS3: all 62 patients (62 recordings) of the MASS - SS3 subset

Sleep-EDFx

Contains the the SC-EDF-20, SC-EDF-78, and ST-EDF subset.

You can download & extract the data via the following commands;

mkdir data
# Download the data
wget https://physionet.org/static/published-projects/sleep-edfx/sleep-edf-database-expanded-1.0.0.zip -P data
# Extract all data
unzip data/sleep-edf-database-expanded-1.0.0.zip -d data

MASS

Contains the MASS SS3 subset.

In order to access the data you should submit a request as is described here; http://ceams-carsm.ca/mass/

predict-idlab / sleep-linear