On robust spatial filtering of EEG in nonstationary environments

Abstract

bibliography 119 I N T R O D U C T I O NVariability is the law of life, and as no two faces are the same, so no two bodies are alike, and no two individuals react alike and behave alike under the abnormal conditions which we know as disease. (Sir William Osler, 1903) T he Canadian physician Sir William Osler (1849 -1919 revolutionized the medical world of the 19th century by advocating an approach to therapy which focuses on the needs of individual patients. He recognized the great variability among individuals with respect to their physical conditions, their mental status and their responses to drugs and concluded that focusing solely on clinical signs and symptoms does not result in the most effective therapy. The development of this novel patient-centered paradigm made him the father of personalized medicine; a branch of medicine that is also very popular today due to a better understanding of the genetic influences on a person's risk of acquiring certain diseases.In the last two decades personalized approaches have also revolutionized the field of Brain-Computer Interfacing (BCI). A BCI system (see e.g. Wolpaw et al., 2002;Dornhege et al., 2007;Wolpaw and Wolpaw, 2012) aims to decode the intention of a user from recordings of brain activity and to use this information for controlling a computer application or a robotic device. Today's BCIs extract user-specific features from electroencelographic (EEG) recordings and adapt to the user's signal characteristics. This subject-centered or machine learning approach not only largely reduces the calibration time in comparison to classical BCI systems based on neurofeedback training (Kamiya et al., 1969) but also increases classification accuracy. Unfortunately, EEG responses to a stimulus or a task not only differ from subject to subject but also from trial to trial and from day to day. This change in the signal properties over time is termed the intrinsic nonstationarity of EEG (Kaplan et al., 2005) and constitutes a major challenge for data analysis and classification by violating a basic assumption of many machine learning methods, namely that data are sampled from a fixed (but unknown) distribution (Vapnik, 1998;Hastie et al., 2001). The analysis of EEG is further aggravated by the presence of artifacts in the data, e.g., eye movements, muscular activity or loose electrodes. The lack of robustness to unexpected events and the nonstationary nature of EEG are major reasons in explaining why current BCI technology is seldom used in out-of-lab scenarios and clinical practice. Since artifacts and nonstationarity can not be fully eliminated, even with the best experimental protocol, robust and invariant feature 1 2 introduction representations are required for optimal signal analysis and high classification accuracy. structure of the thesisThis thesis is divided into three parts. The first part introduces the basic concepts of motor imagery-based Brain-Computer Interfacing and discusses the advantages and limitations of several state-of-the-art spatial filtering al...

Deep learning with convolutional neural networks for EEG decoding and visualization

“…While we have shown deep ConvNets to be competitive with standard FBCSP, a lot of variants of FBCSP exist. For example, many regularized variants of CSP exist that can be used inside FBCSP [Lotte and Guan, ; Samek, ]; a comparison to these could further show the exact tradeoff between the more generic ConvNets and the more domain‐specific FBCSP.…”

Formulating Divergence Framework for Multiclass Motor Imagery EEG Brain Computer Interface

“…Similar to CSP algorithm, the OVR approach computes CSP filters that discriminate each class from all the other classes. For each binary classification, generally two spatial filters are selected according to α sorting criteria (sorting according to discriminativity value, check [6] for detail) Assume that the class covariance matrix for class c i is denoted by Σ ci then Σ ovri is defined as follows:…”

“…where K is number of different motor imagery classes (the number of different matrices to be jointly diagonalized). p c is the prior probability corresponding to class c. The transform V can be decomposed as a product of orthogonal matrix and a whitening transform V T = RW [6]. We can write (5) in terms of orthogonal matrix R as follows:…”

“…Authors in [5] also proposed to optimize the discriminativity and stationarity simultaneously. Later, Samek et al [6][7] proposed a composite divergence based framework to optimize various types of non-stationarities for CSP to incorporate the stationarity. Horev et al used a riemannian geometry framework to optimise stationarity in binary class motor imagery BCI [8].…”

“…Later different strategies such as One-Versus-Rest CSP [9], pairwise binary classification followed by voting [10], Information theoretic filter selection criteria [11] and Riemannian geometry based techniques [12] [13] were proposed for feature extraction and classification in multiclass motor imagery based BCIs. The proposed information theoretic approaches to optimise stationarities in earlier research works [6][7] [14] can not be directly used in multiclass BCI settings. The primary goal of this letter is to develop an information theoretic divergence based framework for multiclass motor imagery BCI.…”