SAD-VER: A Self-supervised, Diffusion probabilistic model-based data augmentation framework for Visual-stimulus EEG Recognition

Abstract

The decoding of EEG-based visual stimuli has become a major and important topic in the field of Brain–Computer Interfaces (BCI) research. However, there is a problem of EEG data scarcity in visual stimulus EEG decoding research, making it difficult to establish effective and stable deep learning models. Therefore, in this paper we propose a novel data augmentation framework: the Self-supervised, Adaptive variance Diffusion probabilistic model-based Visual-stimulus EEG Augmentation Framework (SAD-VER), for enhancing and recognizing visual stimulus EEG data. As the first to introduce diffusion model to EEG-based visual stimulus research, the generating process of SAD-VER is composed of a well-designed diffusion model to generate high-quality and diverse EEG samples. Additionally, this process is self-optimized with a Bayesian method-based hyperparameter optimizer to maximize the quality of the generated EEG samples in a self-supervised manner. A modified convolutional network is also utilized for quality analysis and decoding of augmented EEG. Experimental results demonstrate that the proposed SAD-VER can improve the decoding accuracy of existing models by generating high-quality EEG samples, and achieve the state-of-the-art performance in various visual stimulus EEG decoding tasks. Further analysis indicates that EEG generated by SAD-VER enhances the separability of features between different categories, and contributes to locating crucial brain region information. Code of this research is available at: https://github.com/yellow006/SAD-VER.