Georgios-Ak/EEGNet_Identification_Model

EEGNet Identification Model

Model Description

A deep learning model for EEG-based subject identification using the EEG Motor Movement/Imagery Dataset. Based on the EEGNet architecture implemented in TensorFlow/Keras.

• Input: 3-second EEG segments (64 channels, 160 Hz sampling rate)
• Output: Predicted subject ID (109 possible subjects)
• Architecture: Compact CNN with:
  • Temporal/spatial convolutions
  • Batch normalization
  • Dropout regularization
  • Softmax classification layer

Dataset

EEG Motor Movement/Imagery Dataset (EEGMMIDB)
109 subjects, 64 channels (10-20 system)

Preprocessing Pipeline:

1. Bandpass filtering (1–40 Hz)
2. Resampling to 160 Hz
3. Segmentation into 3-second non-overlapping windows
4. Channel name normalization
5. Standard 10-20 montage alignment

Training

Parameter	Value
Loss Function	Categorical Crossentropy
Optimizer	Adam (lr=0.001)
Batch Size	64
Max Epochs	100 (early stopping)
Validation Split	15%
Test Split	15%

Callbacks:

• EarlyStopping (patience=8)
• ModelCheckpoint (best weights)
• ReduceLROnPlateau (factor=0.2, patience=5)
• TensorBoard logging

Evaluation

Primary Metric: Classification Accuracy
Test Performance: 93.97%

Analysis Tools:

• Overall/per-subject accuracy
• Confusion matrix (top 15 subjects)
• Classification report (precision/recall/F1 per subject)
• Training/validation curves
• Filter visualizations (temporal + spatial)
• Model attention maps for sample segments

Usage

Inference Requirements

Input Shape: (64, 480) (channels × time points)
Critical: Input must match preprocessing:

• Bandpass filtered (1-40 Hz)
• Resampled to 160 Hz
• Standard 10-20 montage channels
• Identical normalization as training data

Visualization

Code templates available for:

1. Training history plots
2. Confusion matrix generation
3. Spatial filter visualization
4. Attention heatmaps for EEG segments

Limitations

1. Trained/tested exclusively on EEGMMIDB
2. Performance not guaranteed for:
   • Different EEG systems
   • Alternative montages
   • Varying preprocessing pipelines
3. Requires strict adherence to original preprocessing for reliable inference

Citation

@article{lawhern2018eegnet,
  title={EEGNet: A Compact Convolutional Neural Network for EEG-based Brain–Computer Interfaces},
  author={Lawhern, Vernon J. and Solon, Amelia J. and Waytowich, Nicholas R. and Gordon, Stephen M. and Hung, Chou P. and Lance, Brent J.},
  journal={Journal of Neural Engineering},
  volume={15},
  number={5},
  pages={056013},
  year={2018},
  publisher={IOP Publishing}
}

@article{goldberger2000physiobank,
  title={PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals},
  author={Goldberger, Ary L. and Amaral, Luis A. N. and Glass, Leon and Hausdorff, Jeffrey M. and Ivanov, Plamen Ch. and Mark, Roger G. and Mietus, Joseph E. and Moody, George B. and Peng, Chung-Kang and Stanley, H. Eugene},
  journal={Circulation},
  volume={101},
  number={23},
  pages={e215--e220},
  year={2000}
}