Dataset 2a Research Articles

Rethinking CNN Architecture for Enhancing Decoding Performance of Motor Imagery-Based EEG Signals

Brain–computer interface (BCI) is a technology that allows users to control computers by reflecting their intentions. Electroencephalogram (EEG)–based BCI has been developed because of its potential, however, its decoding performance is still insufficient to apply in the real–world environment. As deep learning methods achieve the significant performance in various domains, it has been applied in the EEG–based BCI domain. In particular, ShallowConvNet is one of the most widely used methods because of its robust decoding performance in multiple datasets. However, the model’s parameters have to be optimized to apply this model to various datasets each time, and we have also found some issues in architecture that disturb the stable training. In this paper, we highlight potential problems that might arise in ShallowConvNet and investigate the potential solutions. In addition, we propose a novel model, called M–ShallowConvNet, which solves the existing problems. The proposed model achieves the accuracies of 0.8164 and 0.8647 in datasets 2a and 2b of BCI Competition IV, respectively. Hence, we demonstrate that performance improvement can be achieved with only a few small modifications that resolve the problems of the conventional model.

EEG-ITNet: An Explainable Inception Temporal Convolutional Network for Motor Imagery Classification

In recent years, neural networks and especially deep architectures have received substantial attention for EEG signal analysis in the field of brain-computer interfaces (BCIs). In this ongoing research area, the end-to-end models are more favoured than traditional approaches requiring signal transformation pre-classification. They can eliminate the need for prior information from experts and the extraction of handcrafted features. However, although several deep learning algorithms have been already proposed in the literature, achieving high accuracies for classifying motor movements or mental tasks, they often face a lack of interpretability and therefore are not quite favoured by the neuroscience community. The reasons behind this issue can be the high number of parameters and the sensitivity of deep neural networks to capture tiny yet unrelated discriminative features. We propose an end-to-end deep learning architecture called EEG-ITNet and a more comprehensible method to visualise the network learned patterns. Using inception modules and causal convolutions with dilation, our model can extract rich spectral, spatial, and temporal information from multi-channel EEG signals with less complexity (in terms of the number of trainable parameters) than other existing end-to-end architectures, such as EEG-Inception and EEG-TCNet. By an exhaustive evaluation on dataset 2a from BCI competition IV and OpenBMI motor imagery dataset, EEG-ITNet shows up to 5.9% improvement in the classification accuracy in different scenarios with statistical significance compared to its competitors. We also comprehensively explain and support the validity of network illustration from a neuroscientific perspective. We have also made our code freely accessible at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/AbbasSalami/EEG-ITNet</uri> .

SincNet-Based Hybrid Neural Network for Motor Imagery EEG Decoding.

It is difficult to identify optimal cut-off frequencies for filters used with the common spatial pattern (CSP) method in motor imagery (MI)-based brain-computer interfaces (BCIs). Most current studies choose filter cut-frequencies based on experience or intuition, resulting in sub-optimal use of MI-related spectral information in the electroencephalography (EEG). To improve information utilization, we propose a SincNet-based hybrid neural network (SHNN) for MI-based BCIs. First, raw EEG is segmented into different time windows and mapped into the CSP feature space. Then, SincNets are used as filter bank band-pass filters to automatically filter the data. Next, we used squeeze-and-excitation modules to learn a sparse representation of the filtered data. The resulting sparse data were fed into convolutional neural networks to learn deep feature representations. Finally, these deep features were fed into a gated recurrent unit module to seek sequential relations, and a fully connected layer was used for classification. We used the BCI competition IV datasets 2a and 2b to verify the effectiveness of our SHNN method. The mean classification accuracies (kappa values) of our SHNN method are 0.7426 (0.6648) on dataset 2a and 0.8349 (0.6697) on dataset 2b, respectively. The statistical test results demonstrate that our SHNN can significantly outperform other state-of-the-art methods on these datasets.

Deep Adversarial Domain Adaptation With Few-Shot Learning for Motor-Imagery Brain-Computer Interface

Electroencephalography (EEG) is the most prevalent signal acquisition technique for brain-computer interface (BCI). However, the statistical distribution of EEG data varies across subjects and sessions, resulting in poor generalization of the domain-specific classifier. Although the collection of a large number of recordings may alleviate this issue, it is often impractical and not user-friendly. This study proposes the integration of deep domain adaptation with few-shot learning to address the challenge by leveraging the knowledge from multiple source subjects to enhance the performance of a single target subject. The framework incorporated 3 modules: a feature extractor, domain discriminator, and classifier. The feature extractor utilized the available labeled samples with supervised contrastive loss to map the discriminate features onto a deep representation space, where the features from the same class were more similar than those from different classes. The domain discriminator was used to reduce domain drift, through adversarial training. The classifier predicted the user motor intention, based on EEG features. The framework was extensively evaluated through the BCI Competition IV Datasets 2a and 2b. The results of this study indicate that the framework is capable of enhancing the BCI performance and potentially decreases the calibration effort compared to the traditional approach, but the major limitation of this method is that it requires meticulous selection of source subjects.

A Self-Supervised Learning Based Channel Attention MLP-Mixer Network for Motor Imagery Decoding.

Convolutional Neural Network (CNN) is commonly used for the Electroencephalogram (EEG) based motor-imagery (MI) decoding. However, its performance is generally limited due to the small size sample problem. An alternative way to address such issue is to segment EEG trials into small slices for data augmentation, but this approach usually inevitably loses the valuable long-range dependencies of temporal information in EEG signals. To this end, we propose a novel self-supervised learning (SSL) based channel attention MLP-Mixer network (S-CAMLP-Net) for MI decoding with EEG. Specifically, a new EEG slice prediction task is designed as the pretext task to capture the long-range information of EEG trials in the time domain. In the downstream task, a newly proposed MLP-Mixer is applied to the classification task for signals rather than for images. Moreover, in order to effectively learn the discriminative spatial representations in EEG slices, an attention mechanism is integrated into MLP-Mixer to adaptively estimate the importance of each EEG channel without any prior information. Thus, the proposed S-CAMLP-Net can effectively learn more long-range temporal information and global spatial features of EEG signals. Extensive experiments are conducted on the public MI-2 dataset and the BCI Competition IV Dataset 2A. The experimental results indicate that our proposed S-CAMLP-Net achieves superior classification performance over all the compared algorithms.

Electroencephalography-Based Brain-Computer Interface Motor Imagery Classification.

Background:Advances in the medical applications of brain–computer interface, like the motor imagery systems, are highly contributed to making the disabled live better. One of the challenges with such systems is to achieve high classification accuracy.Methods:A highly accurate classification algorithm with low computational complexity is proposed here to classify different motor imageries and execution tasks. An experimental study is performed on two electroencephalography datasets (Iranian Brain–Computer Interface competition [iBCIC] dataset and the world BCI Competition IV dataset 2a) to validate the effectiveness of the proposed method. For lower complexity, the common spatial pattern is applied to decrease the 64 channel signal to four components, in addition to increase the class separability. From these components, first, some features are extracted in the time and time–frequency domains, and next, the best linear combination of these is selected by adopting the stepwise linear discriminant analysis (LDA) method, which are then applied in training and testing the classifiers including LDA, random forest, support vector machine, and K nearest neighbors. The classification strategy is of majority voting among the results of the binary classifiers.Results:The experimental results indicate that the proposed algorithm accuracy is much higher than that of the winner of the first iBCIC. As to dataset 2a of the world BCI competition IV, the obtained results for subjects 6 and 9 outperform their counterparts. Moreover, this algorithm yields a mean kappa value of 0.53, which is higher than that of the second winner of the competition.Conclusion:The results indicate that this method is able to classify motor imagery and execution tasks in both effective and automatic manners.

EEG representation approach based on Kernel Canonical Correlation Analysis highlighting discriminative patterns for BCI applications.

Brain-Computer Interface (BCI) is applied in the study of different cognitive processes or clinical conditions as enhancing cognitive skills, motor rehabilitation, and control. However, many approaches focus on using a robust classifier instead of providing a better feature space. This work develops a feature representation methodology through the kernel canonical correlation analysis to reveal nonlinear relations between filter-banked common spatial patterns (CSP) extracted. Our approach reveals nonlinear relations between ranked filter-banked multi-class CSP features and the labels in a finite-dimensional canonical space. We tested the performance of our methodology on the BCI Competition IV dataset 2a. The introduced feature representation using a classic linear SVM achieves accuracy rates competitive with the state-of-the-art BCI strategies. Besides, the processing pipeline allows identifying the spatial and spectral features driven by the underlying brain activity and best modeling the motor imagery intentions.Clinical relevance- This BCI strategy assesses the nonlinear relationships between time series to improve the interpretation of brain electrical activity, taking into account the spatial and spectral features driven by the underlying brain dynamic.

Compact convolutional neural network (CNN) based on SincNet for end-to-end motor imagery decoding and analysis

In the field of human-computer interaction, the detection, extraction and classification of the electroencephalogram (EEG) spectral and spatial features are crucial towards developing a practical and robust non-invasive EEG-based brain-computer interface. Recently, due to the popularity of end-to-end deep learning, the applicability of algorithms such as convolutional neural networks (CNN) has been explored to achieve the mentioned tasks. This paper presents an improved and compact CNN algorithm for motor imagery decoding based on the adaptation of SincNet, which was initially developed for speaker recognition task from the raw audio input. Such adaptation allows for a compact end-to-end neural network with state-of-the-art (SOTA) performances and enables network interpretability for neurophysiological validation in cortical rhythms and spatial analysis. In order to validate the performance of proposed algorithms, two datasets were used; the first is the publicly available BCI Competition IV dataset 2a, which was often used as a benchmark in validating motor imagery classification algorithms, and the second is a dataset consists of primary data initially collected to study the difference between motor imagery and mental-task associated motor imagery BCI and was used to test the plausibility of the proposed algorithm in highlighting the differences in terms of cortical rhythms. Competitive decoding performance was achieved in both datasets in comparisons with SOTA CNN models, albeit with the lowest number of trainable parameters. In addition, it was shown that the proposed architecture performs a cleaner band-pass, highlighting the necessary frequency bands that were crucial and neurophysiologically plausible in solving the classification tasks.

A novel method for classification of multi-class motor imagery tasks based on feature fusion

Motor imagery based brain computer interface (MI-BCI) has the advantage of strong independence that can rely on the spontaneous brain activity of the user to operate external devices. However, MI-BCI still has the problem of poor control effect, which requires more effective feature extraction algorithms and classification methods to extract distinctly separable features from electroencephalogram (EEG) signals. This paper proposes a novel framework based on Bispectrum, Entropy and common spatial pattern (BECSP). Here we use three methods of bispectrum in higher order spectra, entropy and CSP to extract MI-EEG signal features, and then select the most contributing features through tree-based feature selection algorithm. By comparing the classification results of SVM, Random Forest, Naive Bayes, LDA, KNN, Xgboost and Adaboost, we finally decide to use the SVM algorithm based on RBF kernel function which obtained the best result among them for classification. The proposed method is applied to the BCI competition IV data set 2a and BCI competition III data set IVa. On data set 2a, the highest accuracy on the evaluation data set reaches 85%. The experiment on data set IVa can also achieve good results. Compared with other algorithms that use the same data set, the performance of our algorithm has also been improved.

Distinguishable spatial-spectral feature learning neural network framework for motor imagery-based brain–computer interface

Objective. Spatial and spectral features extracted from electroencephalogram (EEG) are critical for the classification of motor imagery (MI) tasks. As prevalently used methods, the common spatial pattern (CSP) and filter bank CSP (FBCSP) can effectively extract spatial-spectral features from MI-related EEG. To further improve the separability of the CSP features, we proposed a distinguishable spatial-spectral feature learning neural network (DSSFLNN) framework for MI-based brain–computer interfaces (BCIs) in this study. Approach. The first step of the DSSFLNN framework was to extract FBCSP features from raw EEG signals. Then two squeeze-and-excitation modules were used to re-calibrate CSP features along the band-wise axis and the class-wise axis, respectively. Next, we used a parallel convolutional neural network module to learn distinguishable spatial-spectral features. Finally, the distinguishable spatial-spectral features were fed to a fully connected layer for classification. To verify the effectiveness of the proposed framework, we compared it with the state-of-the-art methods on BCI competition IV datasets 2a and 2b. Main results. The results showed that the DSSFLNN framework can achieve a mean Cohen’s kappa value of 0.7 on two datasets, which outperformed the state-of-the-art methods. Moreover, two additional experiments were conducted and they proved that the combination of band-wise feature learning and class-wise feature learning can achieve significantly better performance than only using either one of them. Significance. The proposed DSSFLNN can effectively improve the decoding performance of MI-based BCIs.

Classification of the four‐class motor imagery signals using continuous wavelet transform filter bank‐based two‐dimensional images

AbstractThe feature extraction technique plays a vital role in obtaining better classification accuracy. In this paper, a novel framework is proposed, which develops two‐dimensional (2D) images for convolutional neural network (CNN) to classify four (left hand, right hand, feet, and tongue) MI tasks. 2D image is formed by decomposing each trial using continuous wavelet transform (CWT) filter bank after pre‐processing the MI‐based EEG data by multi‐class common spatial pattern (CSP) method. Obtained images are used to train the CNN model for classification. The proposed framework is evaluated using publicly available BCI competition IV dataset 2a by calculating the classification accuracy for all subjects. Results show that the proposed framework has been giving better classification accuracy than some existing CNN‐based and conventional machine learning‐based approaches compared in this paper. The average time required to train CNN using the proposed framework is 12.67 s, acceptable for online MI‐based BCI applications.

Classification of Motor Imagery EEG Based on Time-Domain and Frequency-Domain Dual-Stream Convolutional Neural Network

Background and objectiveAn important task of the brain-computer interface (BCI) of motor imagery is to extract effective time-domain features, frequency-domain features or time-frequency domain features from the raw electroencephalogram (EEG) signals for classification of motor imagery. However, choosing an appropriate method to combine time domain and frequency domain features to improve the performance of motor imagery recognition is still a research hotspot. MethodsIn order to fully extract and utilize the time-domain and frequency-domain features of EEG in classification tasks, this paper proposed a novel dual-stream convolutional neural network (DCNN), which can use time domain signal and frequency domain signal as the inputs, and the extracted time-domain features and frequency-domain features are fused by linear weighting for classification training. Furthermore, the weight can be learned by the DCNN automatically. ResultsThe experiments based on BCI competition II dataset III and BCI competition IV dataset 2a showed that the model proposed by this study has better performance than other conventional methods. The model used time-frequency signal as the inputs had better performance than the model only used time-domain signals or frequency-domain signals. The accuracy of classification was improved for each subject compared with the models only used one signals as the inputs. ConclusionsFurther analysis shown that the fusion weight of different subject is specifically, adjusting the weight coefficient automatically is helpful to improve the classification accuracy.

Combining generative adversarial networks and multi-output CNN for motor imagery classification

Objective. Motor imagery (MI) classification is an important task in the brain–computer interface (BCI) field. MI data exhibit highly dynamic characteristics and are difficult to obtain. Therefore, the performance of the classification model will be challenged. Recently, convolutional neural networks (CNNs) have been employed for MI classification and have demonstrated favorable performances. However, the traditional CNN model uses an end-to-end output method, and part of the feature information is discarded during the transmission process. Approach. Herein, we propose a novel algorithm, that is, a combined long short-term memory generative adversarial networks (LGANs) and multi-output convolutional neural network (MoCNN) for MI classification, and an attention network for improving model performance. Specifically, the proposed method comprises three steps. First, MI data are obtained, and preprocessing is performed. Second, additional data are generated for training. Here, a data augmentation method based on a LGAN is designed. Last, the MoCNN is proposed to improve the classification performance. Main results. The BCI competition IV datasets 2a and 2b are employed to evaluate the performance of the proposed method. With multiple evaluation indicators, the proposed generative model can generate more realistic data. The expanded training set improves the performance of the classification model. Significance. The results show that the proposed method improves the classification of MI data, which facilitates motion imagination.

Improving the performance of multisubject motor imagery-based BCIs using twin cascaded softmax CNNs

Objective. Motor imagery (MI) EEG signals vary greatly among subjects, so scholarly research on motor imagery-based brain–computer interfaces (BCIs) has mainly focused on single-subject systems or subject-dependent systems. However, the single-subject model is applicable only to the target subject, and the small sample number greatly limits the performance of the model. This paper aims to study a convolutional neural network to achieve an adaptable MI-BCI that is applicable to multiple subjects. Approach. In this paper, a twin cascaded softmax convolutional neural network (TCSCNN) is proposed for multisubject MI-BCIs. The proposed TCSCNN is independent and can be applied to any single-subject MI classification convolutional neural network (CNN) model. First, to reduce the influence of individual differences, subject recognition and MI recognition are accomplished simultaneously. A cascaded softmax structure consisting of two softmax layers, related to subject recognition and MI recognition, is subsequently applied. Second, to improve the MI classification precision, a twin network structure is proposed on the basis of ensemble learning. TCSCNN is built by combining a cascaded softmax structure and twin network structure. Main results. Experiments were conducted on three popular CNN models (EEGNet and Shallow ConvNet and Deep ConvNet from EEGDecoding) and three public datasets (BCI Competition IV datasets 2a and 2b and the high-gamma dataset) to verify the performance of the proposed TCSCNN. The results show that compared with the state-of-the-art CNN model, the proposed TCSCNN obviously improves the precision and convergence of multisubject MI recognition. Significance. This study provides a promising scheme for multisubject MI-BCI, reflecting the progress made in the development and application of MI-BCIs.

Time window and frequency band optimization using regularized neighbourhood component analysis for Multi-View Motor Imagery EEG classification

Spatial features optimized at frequency bands have been widely used in motor imagery (MI) based brain-computer interface (BCI) systems. However, using a fixed time window of electroencephalogram (EEG) to extract discriminatory features results in suboptimal MI classification performance because time latency during MI tasks is inconsistent between different subjects. Thus, apart from frequency band optimization, time window optimization is equally important to develop a subject-specific MI-BCI. With time windows, extracted feature space becomes a higher-order tensor problem that requires multi-view learning approaches to optimize features. This study proposes a novel multi-view feature selection method based on regularized neighbourhood component analysis to simultaneously optimize time windows and frequency bands. In the experiment, we extracted spatial features using common spatial patterns (CSP) from MI related EEG data at multiple time windows and frequency bands and optimized them using the proposed feature selection method. A support vector machine is trained to classify optimized CSP features to identify MI tasks. The proposed method achieved classification accuracies on three public BCI datasets (BCI competition IV dataset 2a, BCI competition III dataset IIIa, and BCI competition IV dataset 2b), which are 82.1 %, 91.7 %, and 84.5 %, respectively. Obtained results are superior to those obtained using standard competing algorithms. Hence, the proposed multi-view learning approach for simultaneous optimization of time windows and frequency bands of MI signals shows the potential to enhance a practical MI BCI device's performance.

EEG Channel Selection Using A Modified Grey Wolf Optimizer

Consider an increasingly growing field of research, Brain-Computer Interface (BCI) is to form a direct channel of communication between a computer and the brain. However, extracting features of random time-varying EEG signals and their classification is a major challenge that faces current BCI. This paper proposes a modified grey wolf optimizer (MGWO) that can select optimal EEG channels to be used in (BCIs), the way that identifies main features and the immaterial ones from that dataset and the complexity to be removed. This allows (MGWO) to opt for optimal EEG channels as well as helping machine learning classification in its tasks when doing training to the classifier with the dataset. (MGWO), which imitates the grey wolves leadership and hunting manner nature and which consider metaheuristics swarm intelligence algorithms, is an integration with two modification to achieve the balance between exploration and exploitation the first modification applies exponential change for the number of iterations to increase search space accordingly exploitation, the second modification is the crossover operation that is used to increase the diversity of the population and enhance exploitation capability. Experimental results use four different EEG datasets BCI Competition IV- dataset 2a, BCI Competition IV- data set III, BCI Competition II data set III, and EEG Eye State from UCI Machine Learning Repository to evaluate the quality and effectiveness of the (MGWO). A cross-validation method is used to measure the stability of the (MGWO).

EEG Signal Classification Using Manifold Learning and Matrix‐Variate Gaussian Model

In brain-computer interface (BCI), feature extraction is the key to the accuracy of recognition. There is important local structural information in the EEG signals, which is effective for classification; and this locality of EEG features not only exists in the spatial channel position but also exists in the frequency domain. In order to retain sufficient spatial structure and frequency information, we use one-versus-rest filter bank common spatial patterns (OVR-FBCSP) to preprocess the data and extract preliminary features. On this basis, we conduct research and discussion on feature extraction methods. One-dimensional feature extraction methods like linear discriminant analysis (LDA) may destroy this kind of structural information. Traditional manifold learning methods or two-dimensional feature extraction methods cannot extract both types of information at the same time. We introduced the bilinear structure and matrix-variate Gaussian model into two-dimensional discriminant locality preserving projection (2DDLPP) algorithm and decompose EEG signals into spatial and spectral parts. Afterwards, the most discriminative features were selected through a weight calculation method. We tested the method on BCI competition data sets 2a, data sets IIIa, and data sets collected by our laboratory, and the results were expressed in terms of recognition accuracy. The cross-validation results were 75.69%, 70.46%, and 54.49%, respectively. The average recognition accuracy of new method is improved by 7.14%, 7.38%, 4.86%, and 3.8% compared to those of LDA, two-dimensional linear discriminant analysis (2DLDA), discriminant locality property projections (DLPP), and 2DDLPP, respectively. Therefore, we consider that the proposed method is effective for EEG classification.

Channel Selection for Optimal EEG Measurement in Motor Imagery-Based Brain-Computer Interfaces.

A method for selecting electroencephalographic (EEG) signals in motor imagery-based brain-computer interfaces (MI-BCI) is proposed for enhancing the online interoperability and portability of BCI systems, as well as user comfort. The attempt is also to reduce variability and noise of MI-BCI, which could be affected by a large number of EEG channels. The relation between selected channels and MI-BCI performance is therefore analyzed. The proposed method is able to select acquisition channels common to all subjects, while achieving a performance compatible with the use of all the channels. Results are reported with reference to a standard benchmark dataset, the BCI competition IV dataset 2a. They prove that a performance compatible with the best state-of-the-art approaches can be achieved, while adopting a significantly smaller number of channels, both in two and in four tasks classification. In particular, classification accuracy is about 77-83% in binary classification with down to 6 EEG channels, and above 60% for the four-classes case when 10 channels are employed. This gives a contribution in optimizing the EEG measurement while developing non-invasive and wearable MI-based brain-computer interfaces.

A Multifrequency Brain Network-Based Deep Learning Framework for Motor Imagery Decoding.

Motor imagery (MI) is an important part of brain-computer interface (BCI) research, which could decode the subject's intention and help remodel the neural system of stroke patients. Therefore, accurate decoding of electroencephalography- (EEG-) based motion imagination has received a lot of attention, especially in the research of rehabilitation training. We propose a novel multifrequency brain network-based deep learning framework for motor imagery decoding. Firstly, a multifrequency brain network is constructed from the multichannel MI-related EEG signals, and each layer corresponds to a specific brain frequency band. The structure of the multifrequency brain network matches the activity profile of the brain properly, which combines the information of channel and multifrequency. The filter bank common spatial pattern (FBCSP) algorithm filters the MI-based EEG signals in the spatial domain to extract features. Further, a multilayer convolutional network model is designed to distinguish different MI tasks accurately, which allows extracting and exploiting the topology in the multifrequency brain network. We use the public BCI competition IV dataset 2a and the public BCI competition III dataset IIIa to evaluate our framework and get state-of-the-art results in the first dataset, i.e., the average accuracy is 83.83% and the value of kappa is 0.784 for the BCI competition IV dataset 2a, and the accuracy is 89.45% and the value of kappa is 0.859 for the BCI competition III dataset IIIa. All these results demonstrate that our framework can classify different MI tasks from multichannel EEG signals effectively and show great potential in the study of remodelling the neural system of stroke patients.

Multiple Kernel Stein Spatial Patterns for the Multiclass Discrimination of Motor Imagery Tasks

Brain–computer interface (BCI) systems communicate the human brain and computers by converting electrical activity into commands to use external devices. Such kind of system has become an alternative for interaction with the environment for people suffering from motor disabilities through the motor imagery (MI) paradigm. Despite being the most widespread, electroencephalography (EEG)-based MI systems are highly sensitive to noise and artifacts. Further, spatially close brain activity sources and variability among subjects hampers the system performance. This work proposes a methodology for the classification of EEG signals, termed Multiple Kernel Stein Spatial Patterns (MKSSP) dealing with noise, raveled brain activity, and subject variability issues. Firstly, a bank of bandpass filters decomposes brain activity into spectrally independent multichannel signals. Then, Multi-Kernel Stein Spatial Patterns (MKSSP) maps each signal into low-dimensional covariance matrices preserving the nonlinear channel relationships. The Stein kernel provides a parameterized similarity metric for covariance matrices that belong to a Riemannian manifold. Lastly, the multiple kernel learning assembles the similarities from each spectral decomposition into a single kernel matrix that feeds the classifier. Experimental evaluations in the well-known four-class MI dataset 2a BCI competition IV proves that the methodology significantly improves state-of-the-art approaches. Further, the proposal is interpretable in terms of data distribution, spectral relevance, and spatial patterns. Such interpretability demonstrates that MKSSP encodes features from different spectral bands into a single representation improving the discrimination of mental tasks.