Deep Learning for Brain-Computer Interface (BCI)

Dr. Xiang Zhang · Prof. Lina Yao

Companion repository for the book: Deep Learning for EEG-based Brain-Computer Interface: Representations, Algorithms and Applications (World Scientific, 2021)

---

Overview

This repository contains implementable Python and Jupyter Notebook code alongside benchmark datasets for learning how to classify brain signals using deep learning. It accompanies our survey on deep learning for noninvasive brain signals and the textbook.

The materials cover:

• A taxonomy of BCI signal paradigms by acquisition method (ECoG, EEG, fNIRS, fMRI, EOG, MEG)
• Core deep learning architectures: MLP, RNN, LSTM, GRU, CNN, AE, VAE, GAN, DBN, GIN
• Guidelines for designing BCI systems across signal categories, models, and applications
• Novel BCI prototypes for authentication, visual reconstruction, language interpretation, and neurological disorder diagnosis

Distribution on signals	Distribution on DL models

---

Dataset

Brain signal collection is both financially and temporally costly. We curated 31 public datasets with download links covering most brain signal types:

Brain Signal	Dataset	Subjects	Classes	Rate (Hz)	Channels	Download
FM ECoG	BCI-C IV, Data set IV	3	5	1000	48–64	Link
MI ECoG	BCI-C III, Data set I	1	2	1000	64	Link
Sleeping EEG	Sleep-EDF Telemetry	22	6	100	2 EEG, 1 EOG, 1 EMG	Link
Sleeping EEG	Sleep-EDF Cassette	78	6	100/1	2 EEG, 1 EOG, 1 EMG	Link
Sleeping EEG	MASS-1	53	5	256	17/19 EEG, 2 EOG, 5 EMG	Link
Sleeping EEG	MASS-3	62	5	256	20 EEG, 2 EOG, 3 EMG	Link
Sleeping EEG	SHHS	5804	N/A	125/50	2 EEG, 1 EOG, 1 EMG	Link
Seizure EEG	CHB-MIT	22	2	256	18	Link
Seizure EEG	TUH	315	2	200	19	Link
MI EEG	EEGMMIDB	109	4	160	64	Link
MI EEG	BCI-C IV, Data set II a	9	4	250	22 EEG, 3 EOG	Link
Emotional EEG	AMIGOS	40	4	128	14	Link
Emotional EEG	SEED	15	3	200	62	Link
Emotional EEG	DEAP	32	4	512	32	Link
fMRI	ADNI	202	3	N/A	N/A	Link
fMRI	BRATS	65	4	N/A	N/A	Link

We also provide a well-processed, ready-to-use version of the EEGMMIDB dataset (109 subjects, 64 channels, 160 Hz). Each .npy file represents one subject with shape [N, 65]: 64 EEG channel features plus one class label column.

---

Running the Code

The tutorial notebooks cover the full BCI pipeline: data acquisition, preprocessing, feature extraction, classification, and evaluation. To run the CNN classification example:

python 4-2_CNN.py

For PyTorch beginners, we recommend Morvan Zhou's PyTorch Tutorials.

---

Chapter Resources

Companion code for the book chapters (TensorFlow implementations):

Chapter	Topic	Repository
7	Adaptive feature learning	know_your_mind
9	EEG-based user identification (MindID)	MindID
10	Visual EEG shape reconstruction	EEG_Shape_Reconstruction
11	Brain typing: EEG to text	Brain_typing
13	Neurological disorder (seizure) diagnosis	adversarial_seizure_detection

---

Requirements

Python 3.7. Install dependencies with:

pip install -r requirements.txt

Note: torch-geometric and its dependencies (cluster, scatter, sparse) require version-specific installation — follow the official guide.

---

Citation

If this repository is useful for your research, please cite the survey or book:

@article{zhang2020survey,
  title   = {A survey on deep learning-based non-invasive brain signals: recent advances and new frontiers},
  author  = {Zhang, Xiang and Yao, Lina and Wang, Xianzhi and Monaghan, Jessica JM and Mcalpine, David and Zhang, Yu},
  journal = {Journal of Neural Engineering},
  year    = {2020},
  publisher = {IOP Publishing}
}

@book{zhang2021deep,
  title     = {Deep Learning for EEG-based Brain-Computer Interface: Representations, Algorithms and Applications},
  author    = {Zhang, Xiang and Yao, Lina},
  year      = {2021},
  publisher = {World Scientific Publishing}
}

---

Contact

For questions about the code or tutorial, contact xiang.alan.zhang@gmail.com.

License

This repository is licensed under CC BY-NC-SA 4.0. Commercial use is prohibited. Derivatives must carry the same license.