Motor imagery EEG signal classification using SAGAN-based data augmentation and ResNet-18 for brain–computer interface applications

Electroencephalogram signals are an important basis for motor skill rehabilitation and motor control in brain–computer interface (BCI) research. However, the limited labelled EEG data may limit the development of effective motor imagery (MI) classifiers. To address this limitation, this study presents a data augmentation framework using an improved Synaptic Augmented Generative Adversarial Network (SA-GAN). The EEG signals of the BCI Competition IV-2a dataset were transformed into two-dimensional time–frequency representations from one-dimensional time-series signals using the Short-Time Fourier Transform (STFT). The resulting representations were augmented using SA-GAN to generate synthetic MI samples and used to train ResNet-18 under augmented and non-augmented conditions. The proposed SAGAN-ResNet18 framework achieved better classification performance than the considered GAN-based augmentation approaches for all the evaluated subjects. The results show that the generated samples can provide valuable extra training information and reduce the negative impact of limited training data in within-subject MI classification. The results show that the proposed augmentation strategy is effective for the conditions evaluated in the BCI Competition IV-2a experiment.