Relevance-based channel selection in motor imagery brain–computer interface

Abstract

Objective: Channel selection in electroencephalogram (EEG)-based brain-computer interface (BCI) has been extensively studied for over two decades, with the goal to select optimal subject-specific channels that can enhance the overall decoding efficacy of BCI. With the emergence of deep learning (DL) based BCI models, there arises a need for fresh perspectives and novel techniques to conduct channel selection. In this regard, subject-independent channel selection is relevant, since DL models trained using cross-subject data offer superior performance, and the impact of inherent inter-subject variability of EEG characteristics on subject-independent DL training is not yet fully understood. Approach: Here, we propose a novel methodology for implementing subject-independent channel selection in DL based motor imagery (MI)-BCI, using layer-wise relevance propagation (LRP) and neural network pruning. Experiments were conducted using Deep ConvNet and 62-channel MI data from Korea University (KU) EEG dataset. Main Results: Using our proposed methodology, we achieved a 61% reduction in the number of channels without any significant drop (p=0.09) in subject-independent classification accuracy, due to the selection of highly relevant channels by LRP. LRP relevance based channel selections provide significantly better accuracies compared to conventional weight based selections while using less than 40% of the total number of channels, with differences in accuracies ranging from 5.96% to 1.72%. The performance of the adapted sparse-LRP model using only 16% of the total number of channels is similar to that of the adapted baseline model (p=0.13). Furthermore, the accuracy of the adapted sparse-LRP model using only 35% of the total number of channels exceeded that of the adapted baseline model by 0.53% (p=0.81). Analyses of channels chosen by LRP confirm the neurophysiological plausibility of selection, and emphasize the influence of motor, parietal, and occipital channels in MI-EEG classification. Significance: The proposed method addresses a traditional issue in EEG-BCI decoding, while being relevant and applicable to the latest developments in the field of BCI. We believe that our work brings forth an interesting and important application of model interpretability as a problem-solving technique.

Transferring a deep learning model from healthy subjects to stroke patients in a motor imagery brain–computer interface

“…The influence of the middle parietal region observed in H-to-S transfer contrasts with the middle frontal influence seen during S-to-S transfer. The parietal influence specific to transfer from models pre-trained using data from healthy subjects is in line with findings from the literature reported for healthy subjects during MI performance [28,81].…”

“…The identified associative patterns also underlie the good MI detection performance in stroke patients achieved through our proposed method. Our analysis of both types of transfer suggests that the bilateral motor, middle frontal, and middle parietal regions are highly relevant for MI detection in stroke patients [28,46,78,81].…”

“…In the study by Lee et al [81], the authors observed a pattern of neural activity over the frontal, motor, and parietal areas of the cortex in healthy controls during MI performance. More recently, in a cross-subject channel selection study using the OpenBMI dataset, Nagarajan et al [28] identified the influence of motor, parietal, and occipital channels in MI-EEG classification among healthy subjects. However, the bilateral nature of MI relevance found in both S-to-S and H-to-S transfer models raised questions regarding how these average patterns are influenced by the respective relevances pertaining to stroke MI of the LAL versus the RAL.…”

“…In contrast, other state-ofthe-art (SOA) models, such as Deep ConvNet [25] and EEGNet [26], utilize raw EEG to classify MI, thus removing the need for feature engineering. The advent of deep neural networks (DNNs) in MI-BCI has piqued the interest of BCI researchers, leading to several investigations into their functioning [27,28]. These DL algorithms have demonstrated promising results in classifying MI using EEG data, more so with subject-independent modeling [28,29], where networks pre-trained using numerous multi-subject EEG trials are evaluated on a new target subject.…”

“…Their results are based on data collected from five stroke patients over two months, and are subject-specific. Previous research has examined SOA MI classifiers based on DL, such as the Deep ConvNet, for subject-independent classification using data from healthy individuals, and has demonstrated promising outcomes [28,29]. However, Raza et al recently evaluated the efficacy of EEGNet for cross-subject classification using data from 10 hemiparetic stroke patients and found no significant difference between cross-subject and withinsubject classification [43].…”