Sleep-Analysis

ML and DL-based Sleep stage classification using Actigraphy, Heart Rate Variability, and Respiratory Rate Variability

Insufficient sleep quality is directly linked to a series of physical and physiological diseases. Therefore, reliable sleep monitoring is essential for prevention, diagnosis, and treatment of such. The gold standard approach for sleep monitoring and the detection of sleep disorders is Polysomnography (PSG), which is typically performed in a sleep laboratory. However, PSG is expensive, time-consuming, and uncomfortable for the patient. To improve patient experience and prevent diseases at an early stage, wearable sensors, such as motion sensors, ECG sensors or respiration belts, offer a promising alternative. This repository presents a benchmark analysis of the performance of machine learning and deep learning algorithms for sleep stage classification using multimodal approaches including actigraphy, heart rate variability, and respiratory rate variability. Additionally, this work incorporates respiratory information through features derived via ECG-derived respiration (EDR) and compares their performance against the other approaches.

The evaluation is performed for three different experiments:

• Multistage classification (Wake / N1 / N2 / N3 / REM) & (Wake / NREM / REM)
  
• Binary classification (Wake / Sleep).

The evaluation is performed on the MESA Sleep dataset paper1, paper2, which includes sleep data of more than 1,000 participants.

ECG derived respiration (EDR)

This Repository also includes Algorithms for the extraction of EDR signals from ECG signals.

• Lindberg Algorithm paper
• Karlen Algorithm paper
• Charlton Algorithm paper
• AddisonAM Algorithm paper
• AddisonFM Algorithm paper
• Vangent Algorithm paper
• Soni Algorithm paper
• Sakar Algorithm paper

Results overview

An overview of the full results can be in the following notebook: results_notebook.
The evaluation of the sngle algorithms can be found in the following notebooks: LSTM, TCN, AdaBoost, MLP, SVM, Random Forest, XGBoost.

5 stage classification (Wake / N1 / N2 / N3 / REM, according to AASM)

3 stage classification (Wake / NREM / REM)

Binary classification (Wake / Sleep)

Using the Framework

To run the algorithms on the MESA dataset, the following steps are required:

1. Install the repository by using poetry (instructions below in the getting startet section).
2. Download the MESA dataset from here and place it in a custom folder (e.g. data/mesa).
3. Create a folder for the processed data (e.g. data/mesa_processed).
4. Adjust the MESA path and the processed path to the dataset in the study_data.json file.
5. Run the data_handling.py script to preprocess the data and extract the features.
6. Run the algorithm_scripts to run the algorithms.
   • The algorithms can be run with different modalities (actigraphy (ACT), heart rate variability (HRV), respiration rate variability (RRV), ECD-derived respiration rate variability (ED-RRV)).
   • The algorithms can also be run with custom combinations of modalities (e.g. ACT + HRV, ACT + RRV, ACT + HRV + RRV, ACT + HRV + RRV + ED-RRV, ...).
7. The jupyter notebooks in the evaluation folder can be used to evaluate the results.

Project structure

sleep-analysis
│   README.md
├── sleep-analysis  # The core library folder. All project-wide helper and algorithms go here
│   ├── classification  # Classification algorithms
│       ├── algrithm_scripts # Scripts to execute the classification algorithms
│       ├── deep_learning  # Contains the architecture of the deep learning algorithms (LSTM & TCN)
│       ├── heuristic_algorithms  # Heuristic algorithms
│       ├── ml_algorithms  # Contains the architecture of the Machine learning algorithms (AdaBoost, MLP, SVM, Random Forest, XGBoost)
│       ├── utils  # Helper functions for classification algorithms e.g. scoring
│   ├── datasets # TPCP Datasets
│   ├── feature_extraction # Feature extraction algorithms
│       ├── actigraphy.py # Actigraphy feature extraction algorithms
│       ├── hrv.py # Heart rate variability feature extraction algorithms
│       ├── imu.py # IMU feature extraction algorithms
│       ├── rrv.py # Respiratory rate variability feature extraction algorithms
│       ├── utils.py # Helper functions for feature extraction algorithms
│   ├── preprocessing # Preprocessing algorithms
│       ├── mesa_dataset # Preprocessing the MESA dataset for classifying sleep stages
│       ├── utils.py # Helper functions for preprocessing algorithms
|
├── experiments  # The main folder for all experiements. Each experiment has its own subfolder
|   ├── evaluation  # Contains the evaluation of the experiments
|   |   ├── algorithm_modality_comparison  # Contains the evaluation of the algorithm and modality comparison
|   |   ├── baseline # Contains the computation of the baseline
|   |   ├── dataset_statistics # Contains the computation of the dataset statistics
|   |   ├── edr_extraction # Compares the extraction EDR to the respiration wave from the respiration belt
|   |   ├── evaluation_per_algorithm # Contains the evaluation of the algorithms in all the three experiments with a single notebook per algorithm
|   |   ├── feature_importance # Contains the computation of the feature importance for the XGBoost algorithm via SHAP 
|   |   ├── influences_on_classification performance # Illustartion of influences on the classification performance (e.g. gender, age, several diseases)
|   |   ├── latex_tables # Contains the automatically created latex tables for the paper
|   |
|   ├── data_hadling  # Contains the data handling for the preprocessing and feature extraction
|       ├── data_handling.py # Script to start the preprocessing and feature extraction
|
|   pyproject.toml  # The required python dependencies for the project
|   poetry.lock  # The frozen python dependencies to reproduce exact results
|   study_data.json  # The "hard-coded" path where to find the respective data. This needs to be manually adjusted for each user
|

Getting started

Installation

git clone https://github.com/danielkrauss2/sleep_analysis.git

For Developers

Install Python >=3.8 and poetry. Then run the commands below to get the latest source and install the dependencies:

git clone https://github.com/mad-lab-fau/sleep_analysis.git
git clone https://github.com/mad-lab-fau/mesa-data-importer.git
git clone https://github.com/empkins/empkins-io.git

Then enter the folder of the sleep-analysis project and run poetry

cd sleep-analysis
poetry install

Working with the code

Note: In order to use jupyter notebooks with the project you need to register a new IPython kernel associated with the venv of the project (poe conf_jupyter - see below). When creating a notebook, make to sure to select this kernel (top right corner of the notebook).

To run any of the tools required for the development workflow, use the poe commands of the poethepoet task runner:

$ poe
docs                 Build the html docs using Sphinx.
format               Reformat all files using black and isort.
lint                 Lint all files with Prospector.
test                 Run Pytest with coverage.
check                Run all checks (format, lint) - does not change the respective files.
update_version       Bump the version in pyproject.toml and empkins_io.__init__ .
conf_jupyter         Register a new IPython kernel named `sleep-analysis` linked to the virtual environment.
remove_jupyter       Remove the associated IPython kernel.

Note: The poe commands are only available if you are in the virtual environment associated with this project. You can either activate the virtual environment manually (e.g., source .venv/bin/activate) or use the poetry shell command to spawn a new shell with the virtual environment activated.

To add new dependencies you need for this repository:

poetry add <package_name>

To update dependencies after the pyproject.toml file was changed (It is a good idea to run this after a git pull):

poetry update

For more commands see the official documentation.

Jupyter Notebooks

To use jupyter notebooks with the project you need to add a jupyter kernel pointing to the venv of the project. This can be done by running:

poe conf_jupyter

Afterwards a new kernel called sleep-analysis should be available in the jupyter lab / jupyter notebook interface. Use that kernel for all notebooks related to this project.

You should also enable nbstripout, so that only clean versions of your notebooks get committed to git

poe conf_nbstripout

All jupyter notebooks should go into the notebooks subfolder of the respective experiment. To make best use of the folder structure, the parent folder of each notebook should be added to the import path. This can be done by adding the following lines to your first notebook cell:

# Optional: Auto reloads the helper and the main sleep-analysis module
%load_ext autoreload
%autoreload 2

from sleep-analysis import conf_rel_path
conf_rel_path()

This allows to then import the helper and the script module belonging to a specific experiment as follows:

import helper
# or
from helper import ...

Format and Linting

To ensure consistent code structure this project uses prospector, black, and isort to automatically check the code format.

poe format  # runs black and isort
poe lint # runs prospector

If you want to check if all code follows the code guidelines, run poe check. This can be useful in the CI context