seizyml 0.1.0

A CLI tool for seizure detection using interpretable ML

• SeizyML uses interpretable machine learning models to detect 🕵️‍♂️ seizures from EEG recordings coupled with manual verification in user-friendly GUI.
• 📖 To reference SeizyML, or view the manuscript, please refer to the following publication (To be updated soon).
• You can access the data and code used to reproduce the experiments and figures from the accompanying paper on Zenodo.

---

📚 Contents

• ⚙️ Hardware Requirements
• 💾 Installation
• 🛠️ App Configuration
• 📈 Model Training
• 🚀 How To Use

📄 Additional Resources

• Configuration settings
• SeizyML Processing Pipeline
• Custom Model Integration

---

Hardware requirements

• SeizyML is a lightweight application that utilizes Gaussian Naive Bayes (GNB) models to predict seizure events from EEG data.
• Any modern CPU with sufficient RAM to load your EEG recordings should work effectively.
• For example, a quad-core CPU with 16 GB RAM can efficiently handle 24-hour long EEG recordings with 2 channels sampled at 4000 Hz.
• No GPU is required for SeizyML's operation.

---

Installation

Conda (Recommended)

1. Download and install miniconda.

2. Clone or Download SeizyML on your machine.

3. Start Anaconda's prompt, navigate to the downloaded /seizy_ml to create the conda environment:

    conda env create -f environment.yml
   

4. Activate environment

    conda activate seizyml
   

5. Launch App

    seizyml
   

Pip

1. Download and install Python 3.9.

2. In the terminal

    pip install seizyml
   

3. Launch App

    seizyml
   

If this works you should see the SeizyMl CLI interface.

---

App Configuration

All settings are stored in the config.yaml file.

• This file will be created in the SeizyML folder from a template file (temp_config.yaml) after you use the setpath command for the first time. ⚠️ The temp_config.yaml file should not be edited by the user.

To edit the config.yaml use any text editor such as notepad:

• The only setting that requires editing before training a model and using the app is the channels field.

•     **channels** : List containing the names of LFP/EEG channels, e.g. ["hippocampus", "frontal cortex"]
  

• All other settings can be left at default, given that the data were prepared in the recommended format (.h5 files with shape [Nsegments, 500 (1 segment), channels]).
• For data conversion check the accompanying app seizy_convert or the h5_conversion script for more help.
• An explanation of all other settings can be found here.

---

Model Training

• Before using SeizyML for seizure detection a model should be first trained on ground truth (hand-scored) data.

1. Launch App.

For conda:

# In anaconda prompt
cd ./seizy_ml
conda activate seizyml
seizyml

For pip:

# In terminal
seizyml

2. Set path for data processing.

seizyml setpath 'path'

• This is the folder path where the training data in .h5 format along with the corresponding training labels in .csv format are stored.

• The training data consist of each recording in .h5 format [Nsegments, 1 segment, Nchannels]. Where a segment is 500 (win*fs).

• The training labels consist of a corresponding .csv file containing the ground truth labels (1 for seizure, 0 for non seizure) with length [Nsegments].

• Training data and labels for each recording need to have a matching name.

  

• The win, fs, channels fields should be set in config.yaml to match the shape of the data. Defaults are win=5, fs=100.

• The config.yaml is created when the path is first set in SeizyML set from temp_config.yaml.

• This folder should be kept in one location as the trained models will be stored here.

• If the folder is moved, then the training_path field in config.yaml should be updated to reflect the new location.

3. Model Training

seizyml train

• This is a multi-step process:
  • a) Data preprocessing (high pass filter and exterme outlier removal).
  • b) Feature extraction.
  • c) Find six best feature sets.
  • d) Train a GNB model on these feature sets and select the one with highest F1 score.
  • The model_id will be stored in the config.yaml file and will be used to load that model.

4. Feature Contributions Features contribution to the GNB model can be visualized using the following command.

seizyml featurecontribution

---

How To Use

⚠️ Note: A model must be trained ☝️ before using the app for seizure detection.

1. Launch App.

For conda:

# In anaconda prompt
cd ./seizy_ml
conda activate seizyml
seizyml

For pip:

# In terminal
seizyml

2. Set path for data processing.

seizyml setpath 'path'

• This is the parent path where the child folder with h5 data resides and where all subsequent folders will be created. Check configuration settings for more information.
• The h5 data should be added in a child folder called data.

3. Run file check.

seizyml filecheck

• ⚠️ This step checks that the h5 files have the correct dimensions. For help on how to convert files to h5 have a look at the h5_conversion script.

4. Preprocess data.

• This is the step where the h5 data files will be filtered and large outliers will be removed.

seizyml preprocess

5. Generate model prections.

seizyml predict

• Here selected features will be extracted and model predictions will be generated using the selected model (model id can be found in the configuration settings file).

6. Verify seizures and adjust seizure boundaries.

• This will launch a prompt to allow for file selection for verification.
• After file selection, a GUI will be launched for seizure verification and boundary adjustment.

seizyml verify

7. Get seizure properties. -This step will generate a csv file with seizure properties for each h5 file.

seizyml extractproperties

---

Contributions

We welcome all project contributions including raising issues and pull requests!

---

-> Back to Top.