🧠✨ Subject Identification Based On Cueless Imagined Speech

This repository provides tools and scripts for processing EEG data, performing signal preprocessing, and running machine and deep learning models for "Cueless EEG imagined speech for subject identification: dataset and benchmarks" paper.

🥳 Our work has been accepted by IEEE Transactions on Biometrics, Behavior, and Identity Science (T-BIOM) 🎉.

---

📜 Table of Contents

• 🛠 Usage
  • 1. Data Structuring & File Integration
    • 📥 Download the Data
    • 📁 Replace the Directory
    • 🔄 Run the Integration Script
  • 2. Signal Preprocessing
  • 3. Machine & Deep Learning Models
    • Generating Datasets
    • SVM and XGB on Statistical and Wavelet Features
    • Moment Models
    • End-to-End Architectures
• 📫 Contact
• 📄 How to cite?

---

🛠 Usage

1. Data Structuring & File Integration

📥 Download the Data

Download the "Device Output Data" directory from the Hugging Face Dataset of this project.

📁 Replace the Directory

Replace the empty Device Output Data directory in this repository with the downloaded one:

🔄 Run the Integration Script

Integrate the .tdms files with the corresponding .csv files for each session to generate the Dataset directory containing .fif files.

python data_structuring_file_integration.py --subject-num <SUBJECT_NUM> --session-num <SESSION_NUM>

Example:

python data_structuring_file_integration.py --subject-num 1 --session-num 1

⏭️ Alternatively, you can skip this step by downloading the pre-generated Dataset directory from the Hugging Face.

2. Signal Preprocessing

Preprocess the EEG signals to prepare them for analysis and modeling.

python signal_preprocessing.py --subject-num <SUBJECT_NUM> --session-num <SESSION_NUM> --gender <GENDER>

Example:

python signal_preprocessing.py --subject-num 10 --session-num 1 --gender female

• Logs: Execution logs are saved in the logs_integrate_structuring directory.
• Output: Preprocessed signals are saved in Dataset/derivatives/preprocessed_eeg.

⏭️ Alternatively, skip this step by downloading the Dataset/derivatives/preprocessed_eeg from the Hugging Face.

3. Machine & Deep Learning Models

• Generating Datasets

Navigate to the ml_dl_models directory to generate .npy datasets from the .fif files and create train/val/test splits.

cd ml_dl_models

# Generate numpy datasets from .fif files
# 🚨 Before running the following code change the data_path_template in the code to your desired path
python read_and_save.py

# Generate train and test sets with different configurations
python train_test_generate.py

This will generate datasets like train_dataset.npy (First 3 sessions of all subjects), train_dataset_ses-1,2.npy (First 2 sessions of all subjects), etc which will be used in further steps.

• SVM and XGB on Statistical and Wavelet Features

Navigate to the base_ml_features directory to replicate results using SVM and XGB with feature extraction.

cd base_ml_features

# Example: Run Optuna for XGBoost with Wavelet Features
python optuna_XGB_wavelet.py

Optuna Studies:

You can visualize hyperparameter tuning results using the Optuna dashboard.

optuna-dashboard sqlite:///study_XGB_wavelet.db

Open the provided URL in your browser to view the tuning plots.

• Moment Models

Navigate to the appropriate moment subdirectory based on the desired model size (base, large or small).

Full Fine-Tuning:

Run the following script to full fine-tune the Moment model.

python full_finetune.py

Then for training the SVM or XGB on the output embeddings of the fine-tuned or zero-shot models you could run the following script as an example.

python optuna_XGB_moment_large.py

Logs: Final results and logs are available as .md files within the respective directories.

• End-to-End Architectures

For end-to-end architectures, navigate to the directories starting with braindecode_. For example, for the EEG Conformer model:

cd braindecode_EEGConformer

# Run the training script with Optuna hyperparameter tuning
python optuna_EEGConformer.py

📫 Contact

For any questions or feedback, feel free to reach out:

• Email: ali.derakhshesh79@gmail.com
• linkedin: https://www.linkedin.com/in/ali-derakhshesh/

Citation

@ARTICLE{11251347,
  author={Derakhshesh, Ali and Dehghanian, Zahra and Ebrahimpour, Reza and Rabiee, Hamid R.},
  journal={IEEE Transactions on Biometrics, Behavior, and Identity Science}, 
  title={Cueless EEG Imagined Speech for Subject Identification: Dataset and Benchmarks}, 
  year={2025},
  volume={},
  number={},
  pages={1-1},
  keywords={biometric system;electroencephalogram;Imagined speech;machine learning},
  doi={10.1109/TBIOM.2025.3634273}
}