Feature extraction for BCIs based on electromagnetic source localization and multiclass Filter Bank Common Spatial Patterns

Abstract

Brain-Computer Interfaces (BCIs) provide means for communication and control without muscular movement and, therefore, can offer significant clinical benefits. Electrical brain activity recorded by electroencephalography (EEG) can be interpreted into software commands by various classification algorithms according to the descriptive features of the signal. In this paper we propose a novel EEG BCI feature extraction method employing EEG source reconstruction and Filter Bank Common Spatial Patterns (FBCSP) based on Joint Approximate Diagonalization (JAD). The proposed method is evaluated by the commonly used reference EEG dataset yielding an average classification accuracy of 77.1 ± 10.1 %. It is shown that FBCSP feature extraction applied to reconstructed source components outperforms conventional CSP and FBCSP feature extraction methods applied to signals in the sensor domain.

A Deep Learning Approach for Motor Imagery EEG Signal Classification

“…This prevents the decoding system to correctly decode the user thoughts. Various techniques have been proposed by the scientific community aiming to improve the temporal filtering methods [11,12], spatial filtering [13][14][15], feature extraction [16,17] and feature selection [18][19][20][21] techniques, dimensionality reduction techniques [22][23][24] and classification algorithms [25][26][27][28]. Several feature extraction techniques such as power spectral density (PSD), common spatial pattern (CSP) [29,30], statistical features, selforganizing maps (SOM), correlation, spectral coherence [31] and information entropy [17,32] have been studied.…”

As above, so below? Towards understanding inverse models in BCI

“…Many volume BCI solutions use spherical, off-the-shelf leadfield models [12,18,21,28]. Other studies have used average brain models [30,31], with a few studies relying on user-specific head models [7,23,24]. In source localization in general, the effects of improving the forward model A are well known (e.g.…”

“…In inverse-based BCI studies, the ratio m/n is usually between 20 and 100, see e.g. [24,28,30]. As a consequence, many different vectors s can result in the same measurement x (plus noise).…”

“…Despite this work, the evidence that volume reconstruction would lead to better classification accuracies than the surface approaches remains inconclusive. Although some empirical studies claim benefits for the volume approach [7,21,24,25,30,31], yet other reports present accuracies that do not much differ from those obtained by surface methods [12,18,23,26,28]. In addition, some research groups have changed their opinion over time, for example the initial supportive claims of Goel et al [25] about the volume approach were later replaced by a much more guarded statement [26].…”

“…A nonnegligible amount of research work has gone into empirical EEG studies where inverse modeling is used in the BCI classification pipeline [7,12,[16][17][18][19][20][21][22][23][24][25][26][27][28][29][30][31]. Despite this work, the evidence that volume reconstruction would lead to better classification accuracies than the surface approaches remains inconclusive.…”