Multilinear Discriminative Spatial Patterns for Movement-Related Cortical Potential Based on EEG Classification with Tensor Representation.

Qian Cai,Jianfeng Yan,Hongfang Han,Haixian Wang,Weiqiang Gong,Anastasios D Doulamis

doi:10.1155/2021/6634672

Abstract

The discriminative spatial patterns (DSP) algorithm is a classical and effective feature extraction technique for decoding of voluntary finger premovements from electroencephalography (EEG). As a purely data-driven subspace learning algorithm, DSP essentially is a spatial-domain filter and does not make full use of the information in frequency domain. The paper presents multilinear discriminative spatial patterns (MDSP) to derive multiple interrelated lower dimensional discriminative subspaces of low frequency movement-related cortical potential (MRCP). Experimental results on two finger movement tasks' EEG datasets demonstrate the effectiveness of the proposed MDSP method.

Highlights

Computational Intelligence and Neuroscience encoded as a tensor with second or higher order by continuous wavelet transform (CWT). en, tensor-based discriminant analysis theory is explored to optimize subspaces algorithm
To uncover the underlying structures in these problems for EEG analysis, this paper proposes the multilinear discriminative spatial patterns (MDSP) as a subspace learning method that includes classification
Against the special form of features extracted by MDSP, we propose a new tensor classification method based on nearest neighbors. e advantages of our MDSP algorithm are as follows: (1) MDSP is a general multidimensional dimensionality reduction method

Summary

Methods

It implies that the subspaces derived from frequency-domain have additional discrimination capability compared with time domain and spatial domain. For original EEG signal X ∈ Rc×t, the time complexity of the MDSP is O(c3 + t3) for each loop. R is the number of iterations that makes the MDSP optimization procedure converge and f is the dimensionality of frequency after CWT. For any an nth-order tensor X ∈ Rm1×m2×···×mk×···×mh , the time complexity of the MDSP is O(r(􏽐ni i m3i )) and the space complexity O(􏽐ni i m2i ). The MDSP training procedure requires many loops to converge, it is acceptable for ordinary computer. Compared with traditional subspace methods, for example, LDA with the time complexity of O(􏽑hi 1 m3i ) and the space complexity of O(􏽑hi 1 m2i )

Experiments

Conclusions