Enhanced BCI Performance using Diffusion Model for EEG Generation.

In the realm of Motor Imagery (MI)-based Brain-Computer Interface (BCI), the widespread adoption of deep learning-based algorithms has resulted in an increased demand for a larger training sample size, thereby placing a heightened burden on users. This study advocates the utilization of one of the most advanced generative models, the denoising diffusion probabilistic model (DDPM), for the artificial synthesis of Electroencephalogram (EEG) raw signals. The quality of the generated EEG signals is evaluated through both qualitative and quantitative analyses. Through dimensionality reduction projection, we observed a notable similarity in the data distributions between the generated EEG signals and real EEG signals. Additionally, spectral analysis indicates a striking similarity in energy distribution between the two, accompanied by the presence of an event-related synchronization (ERS) phenomenon in the generated EEG signals. Quantitative analysis reveals that the accuracy of generated EEG signals for left and right-hand motor imagery tasks is 89.81 ± 2.11%, with discriminative information related to classes predominantly concentrated in the motor-sensory cortex area and alpha-beta frequency band. Furthermore, the integration of generated EEG samples contributes to a 3.17% improvement in the classification performance of BCI-deficiency subjects. These artificially generated EEG signals exhibit promising potential for application in calibrating MI-BCI deep learning models, thereby alleviating the burden on participants.