A correntropy-based classifier for motor imagery brain-computer interfaces

Abstract

Objective. This work aims to present a deeper investigation of the classification performance achieved by a motor imagery (MI) EEG-based brain-computer interface (BCI) using functional connectivity (FC) measures as features. The analysis is performed for two different datasets and analytical setups, including an information-theoretic based FC estimator (correntropy). Approach. In the first setup, using data acquired by our group, correntropy was compared to Pearson and Spearman correlations for FC estimation followed by graph-based feature extraction and two different classification strategies -linear discriminant analysis (LDA) and extreme learning machines (ELMs) -coupled with a wrapper for feature selection in the mu (7-13 Hz) and beta (13-30 Hz) frequency bands. In the second setup, the BCI competition IV dataset 2a was considered for a broader comparison. Main results. For our own database the correntropy / degree centrality / ELM approach resulted in the most solid framework, with overall classification error as low as 5%. When using the BCI competition dataset, our best result provided a performance comparable to those of the top three competitors. Significance. Correntropy was shown to be the best FC estimator in all analyzed situations in the first experimental setup, capturing the signal temporal behavior and being less sensitive to outliers. The second experimental setup showed that the inclusion of different frequency bands can bring more information and improve the classification performance. Finally, our results pointed towards the importance of the joint use of different graph measures for the classification.

Reorganization of Resting-State EEG Functional Connectivity Patterns in Children with Cerebral Palsy Following a Motor Imagery Virtual-Reality Intervention

“…Hence, since the brain can be considered as a non-linear dynamic system [41], and since EEG exhibits important non-linear characteristics [42], we chose to use this measure for computing similarities between the recorded time series of each pair of electrodes. Moreover, a recent application of correntropy to study classification scenarios of hands MI tasks showed that this measure was able to capture the expected spatial patterns of MI, while also performing better in terms of classification, when compared to other types of correlation metrics (namely, Pearson and Spearman correlations) [43]. Furthermore, even though other nonlinear approaches have been proposed to deal with these issues in the EEG signal, correntropy can be directly applied to the recorded data and, thus, generally becomes simpler and easier to compute [44].…”

“…Consider the recorded EEG signals represented by X (w) (t) = {x 1 (t), x 2 (t), ..., x d (t)} (w) , with a total of d channels and N data points, for each sample t in time window w. Let us omit the indice w in the quantity X (w) (t), for notation simplicity, and define a correntropy matrix V w [τ] for a given lag τ as such [40,43]:…”

Single-Trial Kernel-Based Functional Connectivity for Enhanced Feature Extraction in Motor-Related Tasks

“…The authors plan to enhance the Gaussian functional connectivity that was developed for feature extraction as future work, allowing a better understanding of their impact and interaction on BCI-related tasks. To identify potential non-learners, the efforts can be directed toward a twofold aim: to enhance the feature extraction by profiting from more elaborate methods for measuring multivariate similarity, like centered kernel alignment [58,59], and to explore the robust estimation approaches based on information metrics (like correntropy) for dealing better with the variability [55,60,61]. Besides, modeling the temporal-dependencies within each trail to compute the FC is an exciting research line.…”

A Comprehensive Review on Critical Issues and Possible Solutions of Motor Imagery Based Electroencephalography Brain-Computer Interface

“…Similar to temporal filters, spatial filters extract the necessary spatial information associated with a motor-related task embedded in EEG signals. A common average reference (CAR) is a spatial filter that removes the common components from all channels, leaving channels with only channel specific signals [ 17 ]. This is done by removing the mean of all k channels from each channel : …”

“…In graph based methods, the EEG data are represented through graph adjacency matrices that correspond to temporal correlations (correlation approaches used like Pearson or Correntropy) between different brain regions (electrodes). Features are extracted from this graph in terms of the graph node’s importance, such as centrality measure [ 17 ].…”