A Skip Connection Generative Adversarial Network With EEG Background Noise for Motor Imagery Decoding

EEG decoding is a key for motor imagery-based brain-computer interfaces (MI-BCIs) to decode people’s motor intentions, holding a broad potential in medical rehabilitation. However, the amount of MI EEG data is often small because of some practical EEG collection limitations, which decreases the EEG decoding performance. This study proposes a skip connection generative adversarial network (GAN)-based data augmentation model to generate MI EEG samples, to expand the amount of the samples for the EEG decoding. The SGD utilizes independent component analysis (ICAs) to extract EEG background noise and to retain more biological information for the GAN and then employs a skip connection and multihead attention to integrate high- and low-level EEG features in the generator. Subsequently, the SGD adopts a temporal convolutional network (TCN) to discriminate between samples. The brain-computer interface (BCI) Competition IV-2a dataset is used to validate the SGD on four-class MI EEG decoding. The results show that the SGD can retain high accuracies (best accuracy: 86.15%, $p =0.034$ ) under different ratios of generated MI EEG samples, implying that the skip connection and multihead attention are helpful for reconstructing the temporal and spatial features of MI EEG by ablation experiment. This study demonstrates the feasibility of applying EEG background noise to generate MI EEG samples, providing a new way to solve the small sample problem faced by the MI-BCI.

I. Introduction

Brain-computer interfaces (BCIs) enable direct communication between the human brain and computers by decoding neural activity [1], highlighted as one of the seven technologies to watch in 2024 [2]. As a new type of human-machine interaction (HMI) technology, BCI has received increasing attention. It has been employed in the military, medical field, and smart HMI [3], [4], [5]. Motor imagery-based BCI (MI-BCI), a widely used BCI [6], requires the subject to imagine motor movements to produce desynchronization/synchronization (ERD/ERS) characteristics in the brain [7], [8] that are used to “decode” the subject’s motor intentions. The MI-BCIs have great prospects in medical rehabilitation and peripheral control [9], [10], while complex MI EEG features pose challenges to EEG decoding. With the rapid development of computer science, various machine learning models, especially deep learning (DL) models, have been progressively applied to MI EEG decoding, such as convolutional neural network (CNN), long short-term memory (LSTM) network, and hybrid CNNs [11], [12], [13]. Lawhern et al. [14] proposed a compact CNN, EEGNet, for EEG decoding. Li et al. [15] proposed a neural network feature fusion algorithm combining CNN and LSTM.