Time-Frequency Transform Based Cross-Subject EEG Data Augmentation

📰 News: We've released a new version of DWTaug, namely DWTaug-reverse, which is more effective for within-subject data augmentation. By reverse the inputs, DWTaug-reverse augments EEG signals by three times. More information can see in code.

This repository contains the original Python code for our paper Time-Frequency Transform Based Cross-Subject EEG Data Augmentation (KBS, 2025), featuring three key implementations:

• DWTaug: Discrete Wavelet Transform-based EEG data augmentation.
• HHTaug: Hilbert-Huang Transform-based EEG data augmentation.
• DWTaug-ML: A multi-level version of DWTaug.
• DWTaug-reverse: A more effective way of implementation, for the within-subject scenario.

<img width="908" alt="image" src="https://github.com/user-attachments/assets/92343b72-8a22-4075-91fe-9765f0f8955f" />

Overview

This work aims to tackle key challenges in BCI applications, including data scarcity, EEG signal non-stationarity, as well as individual differences. The proposed DWTAug and HHTAug follow three steps: time-frequency domain signal decomposition, cross-subject sub-signal reassembling, and time domain reconstruction. Augmenting data expands the pool of labeled training samples, alleviating the data scarcity problem; time-frequency decomposition captures the non-stationary properties of EEG signals more effectively; finally, cross-subject reassembling of sub-signals handles individual differences.

Applications

The proposed approaches are effective for motor imagery (MI), P300, and SSVEP, especially when significant individual difference exists. Notably, DWTaug and HHTaug demonstrate high effectiveness in the SSVEP paradigm, particularly on Benchmark dataset, achieving an accuracy improvement exceeding 10%.

Results

The proposed methods have been tested on 17 EEG datasets across multiple BCI paradigms, consistently outperforming existing data augmentation approaches.

<img width="970" alt="image" src="https://github.com/user-attachments/assets/cfdb26e9-fb84-405f-a3dc-e8214109b308" />

Visualizations

(1) Visualizations of EEG trials before (blue lines) and after (orange lines) ten different data augmentation approaches:

<img width="701" alt="image" src="https://github.com/user-attachments/assets/a55d7dbf-3c07-4bab-b309-bb69cfdcda2f" />

(2) $t$-SNE feature visualizations of the original, DWTaug, and HHTaug data from the source and target subjects on four MI datasets.

<img width="653" alt="image" src="https://github.com/user-attachments/assets/33fc7789-aedc-4194-b6c6-87b69c00c88e" />

Citation

If you find this repo helpful, please cite our work:

@Article{Wang2025CSDA,
  title={Time-frequency transform based EEG data augmentation for brain-computer interfaces},
  author={Wang, Ziwei and Li, Siyang and Chen, Xiaoqing and Wu, Dongrui},
  journal={Knowledge-Based Systems},
  pages={113074},
  year={2025},
  volume={311}
}

Contact

For any questions or collaborations, please feel free to reach out via vivi@hust.edu.cn or open an issue in this repository.