EEGSym: Overcoming Inter-Subject Variability in Motor Imagery Based BCIs With Deep Learning

Abstract

In this study, we present a new Deep Learning (DL) architecture for Motor Imagery (MI) based Brain Computer Interfaces (BCIs) called EEGSym. Our implementation aims to improve previous state-of-the-art performances on MI classification by overcoming inter-subject variability and reducing BCI inefficiency, which has been estimated to affect 10-50% of the population. This convolutional neural network includes the use of inception modules, residual connections and a design that introduces the symmetry of the brain through the mid-sagittal plane into the network architecture. It is complemented with a data augmentation technique that improves the generalization of the model and with the use of transfer learning across different datasets. We compare EEGSym's performance on inter-subject MI classification with ShallowConvNet, Deep-ConvNet, EEGNet and EEG-Inception. This comparison is performed on 5 publicly available datasets that include left or right hand motor imagery of 280 subjects. This population is the largest that has been evaluated in similar studies to date. EEGSym significantly outperforms the baseline models reaching accuracies of 88.6±9.0 on Physionet, 83.3±9.3 on OpenBMI, 85.1±9.5 on Kaya2018, 87.4±8.0 on Meng2019 and 90.2±6.5 on Stieger2021. At the same time, it allows 95.7% of the tested population (268 out of 280 users) to reach BCI control (≥70% accuracy). Furthermore, these results are achieved using only 16 electrodes of the more than 60 available on some datasets. Our implementation of EEGSym, which includes new advances for EEG processing with DL, outperforms previous state-of-the-art approaches on inter-subject MI classification.

Spatial Variation Generation Algorithm for Motor Imagery Data Augmentation: Increasing the Density of Sample Vicinity

“…A comparison study is designed to evaluate the performance of the proposed SVG algorithm against six previous data augmentation algorithms: Hemisphere Perturbation, Patch Perturbation, and Random Shift, Mixup, Frequency Shift, FT surrogate [3], [21], [26]. The following sections detail each of these techniques:…”

“…• Hemisphere Perturbation randomly alters channel positions or replaces data with standard Gaussian noise, assuming that differences between classes can be distinguished regardless of hemisphere [3]; • Patch Perturbation randomly replaces several channels with either zeros or standard Gaussian noise, based on the hypothesis that the difference between classes can be distinguished from partially occluded objects [3], [56]; • Random Shift technique shifts the onset of the trial forward by up to half a second and replaces blank section with either zeros or Gaussian noise, based on the hypothesis that a robust classification model should be able to account for varied reaction times of users [3], [5]. • Mixup applies linear interpolations between two randomly selected pairs of data and labels, based on the hypothesis that the linear interpolations can broaden the support of the training distribution, eventually increasing model robustness and generalizability [21], [26].…”

“…From the perspective of model efficiency. the principal strategies to mitigate the overfitting issue can be categorized into three categories: 1) optimization of parameter numbers, such as incorporating depthwise or separable convolutions [4], [8], pooling layers [3], [9], and progressively reducing the number of layers [10]; 2) optimization of feature extraction efficiency, such as employing a multi-branch structure [11] and an inception block [10]); 3) dynamic modification of models, such as adding dropout regularization [3], [5], [10], [12]. Empirical evidence suggests that such a well-designed architecture can partially alleviate the issue of overfitting.…”

“…To address the challenge of data scarcity, researchers have focused on data augmentation as a potential solution, aiming to generate augmented data (satisfying the vicinal distribution) via the modification or recombination of existing data (satisfying the empirical distribution) according to specific inductive bias [21]- [23]. In the field of EEG, the data augmentation methods can be classified based on the information used for modification or recombination, including temporal [3], [5], [24]- [26], spatial [3], [27], spectral [26], [28], and component information [29], [30]. However, the existing work on data augmentation with EEG data is still limited.…”

Relation Learning Using Temporal Episodes for Motor Imagery Brain-Computer Interfaces

“…Alternatively, deep learning (DL) methods provide an endto-end framework to deal with the MI classification task and achieve a promising performance [10]. They reach the stateof-the-art (SOTA) classification accuracy [10], [11] in multiple public MI benchmarks such as Physionet [12] and those in BCI competitions [9], [13]. The convolutional neural network (CNN) is one of the popular choices in MI signal processing [10].…”