A Novel Data Augmentation Approach Using Mask Encoding for Deep Learning-Based Asynchronous SSVEP-BCI.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Deep learning (DL) methods are effective for steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCI).
• Limited electroencephalography (EEG) data often causes overfitting in DL models for SSVEP-BCI.
• Overfitting hinders the generalization and robustness of BCI systems.

Purpose of the Study:

• To introduce EEG mask encoding (EEG-ME) as a data augmentation technique for SSVEP-BCI.
• To mitigate overfitting in DL models by enhancing feature learning.
• To improve the generalization capabilities of SSVEP-BCI systems.

Main Methods:

• Proposed EEG mask encoding (EEG-ME) by masking partial EEG data.
• Validated EEG-ME on three DL architectures: CNN-Former, tCNN, and EEGNet.
• Evaluated performance on benchmark and BETA datasets with varying time window lengths.

Main Results:

• EEG-ME significantly improved average classification accuracy across different DL methods and time windows.
• CNN-Former, tCNN, and EEGNet showed accuracy increases of up to 3.18% and 11.09% on benchmark and BETA datasets, respectively.
• The 1-second time window demonstrated notable performance enhancements.

Conclusions:

• EEG-ME effectively enhances the robustness and generalization of DL models in SSVEP-BCI.
• The proposed method promotes the practical implementation of asynchronous SSVEP-BCI systems.
• Improved BCI performance leads to more flexible and reliable human-machine interaction.