EEG Preprocessing for CHB-MIT Scalp EEG Database

This project provides a pipeline for preprocessing the CHB-MIT Scalp EEG database from PhysioNet. It handles EDF file loading, extraction of ictal and pre-ictal segments, epoching, and feature extraction using covariance-based methods.

Features

• Segment Extraction:
  • Ictal: Automatically identifies and extracts seizure segments based on the patient-specific summary files.
  • Pre-ictal: Extracts segments preceding seizures with a configurable offset and duration multiplier.
• Epoching: Splits signal segments into non-overlapping fixed-length epochs (default: 5 seconds).
• Feature Extraction: Implements a CovarianceExtractor that computes channel-wise covariance matrices and vectorizes them, preserving the Frobenius norm.
• Output: Saves processed features and labels (1 for ictal, 0 for pre-ictal) as compressed .npz files for each patient in the out/data/ directory.

Prerequisites

• Python >= 3.12
• Dependencies: mne, numpy

Installation

You can install the dependencies using pip or a package manager like uv:

pip install mne numpy

Or if you are using uv:

uv sync

Dataset Structure

The project expects the CHB-MIT Scalp EEG database to be organized in its original structure as downloaded from PhysioNet. Point the PATH_ROOT_DATASET in constants.py to this root folder or pass it as a command-line argument.

chb-mit-scalp-eeg-database/
├── RECORDS-WITH-SEIZURES
├── chb01/
│   ├── chb01-summary.txt
│   ├── chb01_01.edf
│   ├── chb01_02.edf
│   └── ...
├── chb02/
│   ├── chb02-summary.txt
│   ├── chb02_01.edf
│   └── ...
└── ...

Usage

To run the preprocessing pipeline, execute the main.py script. The script uses command-line arguments to configure the pipeline:

python main.py [options]

Available Arguments

• --path, -p: Root directory of the EEG dataset (default: defined in constants.py).
• --offset_seconds, -o: Time gap (in seconds) between the pre-ictal segment end and the seizure onset (default: 300).
• --multiplier, -m: Factor used to scale the pre-ictal segment duration relative to the seizure length (default: 3).
• --epoch_duration, -e: Duration of each signal epoch in seconds for feature extraction (default: 5).

Example

python main.py -p "path/to/dataset" -o 60 -m 5 -e 2

The script will:

1. Read the RECORDS-WITH-SEIZURES file from the dataset root.
2. Process each record by identifying seizure times from summary files.
3. Extract ictal and pre-ictal segments.
4. Split segments into fixed-length epochs.
5. Extract covariance-based features.
6. Save the results as compressed .npz files in out/data/<patient_id>.npz.

Project Structure

• main.py: Main entry point for the preprocessing pipeline.
• constants.py: Configuration for dataset paths and channel selections.
• edf.py: Module for reading EDF files and parsing seizure metadata.
• signals.py: Utility functions for segment extraction and epoching.
• train_test_split.py: Utility to load and split patient data into train/test sets.
• feature_extractor/:
  • base.py: Abstract base class for feature extractors.
  • covariance.py: Implementation of covariance-based feature extraction.
• out/: Directory where logs and processed data are stored.