EEG Seizure Detection

This project implements deep learning models for automated seizure detection using EEG (Electroencephalogram) data from the CHB-MIT Scalp EEG Database. The project explores various neural network architectures including LSTM, CNN-LSTM hybrids, and EEGNet for temporal seizure pattern recognition.

Overview

The system processes EEG recordings to automatically detect seizure events by analyzing temporal patterns in brain activity. Multiple model architectures are implemented and compared, with a focus on LSTM-based approaches that can capture the sequential nature of EEG signals.

Dataset

The project uses the CHB-MIT Scalp EEG Database which contains:

• EEG recordings from pediatric patients
• 23-channel EEG data sampled at 256 Hz
• Annotated seizure start and end times
• Data stored in EDF (European Data Format) files

Raw EEG Data (23 Channels)

Chb01 Dataset (data_org folder)

All EEG files with seizure from patient chb01 are extracted to the data_org folder for quick testing.

Data Structure

chb-mit-scalp-eeg-database-1.0.0/
├── chb01/
├── chb02/
├── ...
└── chb24/

Project Structure

├── README.md
├── seizure_times.pkl                 # Preprocessed seizure timing data
├── seizure_detection_lstm_2_seconds_model.keras  # Trained model
├── 
├── Core Scripts:
├── LSTM_all_data.py                  # Train on all patients combined
├── LSTM_one_per_patient_1_minute.py  # Per-patient model (1-minute windows)
├── LSTM_one_per_patient_2_seconds.py # Per-patient model (2-second windows)
├── LSTM_chb01.py                     # Single patient (chb01) analysis
├── 
├── Utility Scripts:
├── read_edf_dataset.py               # EDF file exploration
├── read_seisure_time.py              # Seizure timing analysis
└── predict_seizure_start_time.py     # Predicting seizure start time

Key Features

1. Data Processing

• EDF File Loading: Uses MNE library to read EEG data
• Temporal Windowing: Configurable sliding window approach
• Data Validation: Filters files with exactly 23 channels
• Seizure Labeling: Automatic labeling based on temporal overlap

2. Model Architectures

Basic LSTM Model

• Sequential LSTM layers for temporal pattern recognition
• Dropout regularization to prevent overfitting

Improved LSTM Model

• CNN Feature Extraction: TimeDistributed CNN layers for spatial feature extraction
• Bidirectional LSTM: Captures forward and backward temporal dependencies
• Batch Normalization: Stabilizes training
• Advanced Regularization: Multiple dropout layers

CNN-LSTM Hybrid

• 3D CNN for spatial-temporal pattern capture
• LSTM for sequence modeling
• Global pooling for dimensionality reduction

EEGNet

• Specialized CNN architecture for EEG analysis
• Depthwise convolutions for channel-specific feature extraction

3. Training Strategies

Cross-Patient Training

• Single model trained on all patients
• Better generalization across different seizure types

Sample one EEG file from each patient for Training (due to computational limitation)

• Individual models for different intervals (eg. 2 seconds, 1 minute)
• Accounts for patient-specific seizure patterns
• (Outliers) Excludes patients under 4 years old (chb06, chb10, chb12, chb13)

Installation

# Install required packages
pip install mne numpy tensorflow scikit-learn pickle

# For EDF file handling
pip install pyedflib

Usage

1. Data Preview

# Explore seizure timing data 
python read_seisure_time.py

# Explore EDF dataset
python read_edf_dataset.py

2. Model Training

Train on one EEG file from each patient (2-second windows):

python LSTM_one_per_patient_2_seconds.py

Train on one EEG file from each patient (1-minute windows):

python LSTM_one_per_patient_1_minute.py

Train on all data:

python LSTM_all_data.py

Single patient (chb01) analysis:

python LSTM_chb01.py

3. Model Evaluation

Models are automatically evaluated with:

• Confusion Matrix: True/False positives and negatives
• Classification Report: Precision, recall, F1-score
• Early Stopping: Prevents overfitting during training

Configuration Parameters

Data Processing

• delta: Non-seizure data length before/after seizure (default: 100 seconds)
• offset: Minimum overlap for seizure classification (default: 2-10 seconds)
• seq_len: Sliding window length (2 seconds or 60 seconds)

Model Training

• batch_size: 16
• epochs: 30-50 with early stopping
• validation_split: 0.2
• test_size: 0.2

Key Functions

Data Loading

• load_seizure_time(): Load preprocessed seizure timings
• load_edf(): Load EEG data from EDF files
• slice_data(): Create temporal windows with labels

Model Creation

• create_lstm_model(): Basic LSTM architecture
• create_improved_lstm_model(): Advanced LSTM with CNN features
• create_cnn_lstm_hybrid(): 3D CNN + LSTM combination
• create_eegnet(): EEGNet implementation

Dataset Generation

• generate_dataset(): Process multiple patients into training data

Results

• Both 2-second windows and 1-minute windows can achieve an accuracy around 85%.
• 2-seconds windows are used for predicting seizure start time and achieve an accuracy around 50%.
• Model outputs are saved as .keras files for future inference.

Future Improvements

1. Real-time Processing: Optimize for online seizure detection
2. Feature Engineering: Incorporate frequency domain features
3. Attention Mechanisms: Improve temporal focus
4. Multi-modal Data: Integrate additional physiological signals
5. Interpretability: Add model explanation capabilities

References

• CHB-MIT Scalp EEG Database: PhysioNet
• MNE Python library for EEG processing
• TensorFlow/Keras for deep learning implementation

License

This project is for research and educational purposes. Please cite the CHB-MIT database when using this code.