[BIBM 2025] Bidirectional Time-Frequency Pyramid Network for Enhanced Robust EEG Classification

BITE is a novel end-to-end deep learning architecture designed to address cross-dataset generalization challenges in EEG classification. The framework combines time-frequency analysis with bidirectional processing, featuring:

1. Dual-stream processing for synchronized time-frequency analysis
2. Pyramid attention mechanism for multi-scale feature extraction
3. Bidirectional temporal learning for comprehensive signal understanding

• Python == 3.10.16
• PyTorch == 2.7.1
• CUDA == 12.4
• Additional dependencies in requirements.txt

Install dependencies:

pip install -r requirements.txt

1. BCI Competition IV Dataset 2a

   • Data: BCICIV_2a_gdf.zip
   • Labels: true_labels.zip

2. BCI Competition IV Dataset 2b

   • Data: BCICIV_2b_gdf.zip
   • Labels: true_labels.zip

3. SDSSVEP Dataset

   • Available at: SSVEP-SanDiego on Kaggle

After downloading, organize the datasets in the data/ directory with the following structure:

data/
├── 2a/              # BCICIV-2a dataset
│   ├── 2a_pre/      # Preprocessed data
│   ├── A01E.gdf     # Subject 1 evaluation data
│   ├── A01T.gdf     # Subject 1 training data
│   ├── A01T.mat     # Subject 1 training labels
│   └── ...          # Similar files for other subjects
│
├── 2b/              # BCICIV-2b dataset
│   ├── 2b_pre/      # Preprocessed data
│   ├── B0101T.gdf   # Subject 1 session 1 training data
│   ├── B0101T.mat   # Subject 1 session 1 labels
│   ├── B0102T.gdf   # Subject 1 session 2 training data
│   └── ...          # Similar files for other subjects
│
└── sdssvep/         # SDSSVEP dataset
    ├── sdssvep_pre/ # Preprocessed data
    ├── S01testEEG.mat    # Subject 1 test data
    ├── S01trainEEG.mat   # Subject 1 training data
    ├── S02testEEG.mat    # Subject 2 test data
    └── ...               # Similar files for other subjects

Each dataset will be automatically preprocessed according to the specifications in get_data.py when running the experiments.

Modify config.yaml to customize your experiments. The configuration supports:

• Multiple datasets (2a, 2b, sdssvep, HGD)
• Different paradigms (Within-Subject, Cross-Subject)
• Model variants and hyperparameters

Default configuration includes:

• Datasets: BCICIV-2a, BCICIV-2b, sdssvep, HGD
• Paradigms: Within-Subject and Cross-Subject
• Model: BiTE

python main.py

This will automatically:

• Load and preprocess datasets
• Train and evaluate models
• Save results in the results/ directory

Results are saved in results/timestamp/ with:

• Detailed logs
• Model weights
• Performance metrics
• Visualizations:
  • Confusion matrices
  • Learning curves
  • PR curves
  • ROC curves
  • t-SNE visualizations

1. How to add a new model?

   • Create your model class in model/ directory
   • Add model import in get_model.py
   • Update model initialization in get_model() function

2. How to use a different dataset?

   • Add dataset configuration in config.yaml
   • Implement data loading function in get_data.py
   • Add preprocessing steps if needed

3. How to run specific experiments?

   • Modify the experiments section in config.yaml
   • Comment out unwanted dataset configurations
   • Adjust model variants as needed