Adaptive transfer learning for EEG motor imagery classification with deep Convolutional Neural Network.

In recent years, deep learning has emerged as a powerful tool for developing Brain-Computer Interface (BCI) systems. However, for deep learning models trained entirely on the data from a specific individual, the performance increase has only been marginal owing to the limited availability of subject-specific data. To overcome this, many transfer-based approaches have been proposed, in which deep networks are trained using pre-existing data from other subjects and evaluated on new target subjects. This mode of transfer learning however faces the challenge of substantial inter-subject variability in brain data. Addressing this, in this paper, we propose 5 schemes for adaptation of a deep convolutional neural network (CNN) based electroencephalography (EEG)-BCI system for decoding hand motor imagery (MI). Each scheme fine-tunes an extensively trained, pre-trained model and adapt it to enhance the evaluation performance on a target subject. We report the highest subject-independent performance with an average (N=54) accuracy of 84.19% (±9.98%) for two-class motor imagery, while the best accuracy on this dataset is 74.15% (±15.83%) in the literature. Further, we obtain a statistically significant improvement (p=0.005) in classification using the proposed adaptation schemes compared to the baseline subject-independent model.

Classification of Individual Finger Movements Using Intracortical Recordings in Human Motor Cortex.

BACKGROUND: Intracortical microelectrode arrays have enabled people with tetraplegia to use a brain-computer interface for reaching and grasping. In order to restore dexterous movements, it will be necessary to control individual fingers. OBJECTIVE: To predict which finger a participant with hand paralysis was attempting to move using intracortical data recorded from the motor cortex. METHODS: A 31-yr-old man with a C5/6 ASIA B spinal cord injury was implanted with 2 88-channel microelectrode arrays in left motor cortex. Across 3 d, the participant observed a virtual hand flex in each finger while neural firing rates were recorded. A 6-class linear discriminant analysis (LDA) classifier, with 10 × 10-fold cross-validation, was used to predict which finger movement was being performed (flexion/extension of all 5 digits and adduction/abduction of the thumb). RESULTS: The mean overall classification accuracy was 67% (range: 65%-76%, chance: 17%), which occurred at an average of 560 ms (range: 420-780 ms) after movement onset. Individually, thumb flexion and thumb adduction were classified with the highest accuracies at 92% and 93%, respectively. The index, middle, ring, and little achieved an accuracy of 65%, 59%, 43%, and 56%, respectively, and, when incorrectly classified, were typically marked as an adjacent finger. The classification accuracies were reflected in a low-dimensional projection of the neural data into LDA space, where the thumb-related movements were most separable from the finger movements. CONCLUSION: Classification of intention to move individual fingers was accurately predicted by intracortical recordings from a human participant with the thumb being particularly independent.

Data augmentation for self-paced motor imagery classification with C-LSTM.

OBJECTIVE: Brain-computer interfaces (BCI) are becoming important tools for assistive technology, particularly through the use of motor imagery (MI) for aiding task completion. However, most existing methods of MI classification have been applied in a trial-wise fashion, with window sizes of approximately 2 s or more. Application of this type of classifier could cause a delay when switching between MI events. APPROACH: In this study, state-of-the-art classification methods for motor imagery are assessed offline with considerations for real-time and self-paced control, and a convolutional long-short term memory (C-LSTM) network based on filter bank common spatial patterns (FBCSP) is proposed. In addition, the effects of several methods of data augmentation on different classifiers are explored. MAIN RESULTS: The results of this study show that the proposed network achieves adequate results in distinguishing between different control classes, but both considered deep learning models are still less reliable than a Riemannian MDM classifier. In addition, controlled skewing of the data and the explored data augmentation methods improved the average overall accuracy of the classifiers by 14.0% and 5.3%, respectively. SIGNIFICANCE: This manuscript is among the first to attempt combining convolutional and recurrent neural network layers for the purpose of MI classification, and is also one of the first to provide an in-depth comparison of various data augmentation methods for MI classification. In addition, all of these methods are applied on smaller windows of data and with consideration to ambient data, which provides a more realistic test bed for real-time and self-paced control.

Towards error categorisation in BCI: single-trial EEG classification between different errors.

OBJECTIVE: Error-related potentials (ErrP) are generated in the brain when humans perceive errors. These ErrP signals can be used to classify actions as erroneous or non-erroneous, using single-trial electroencephalography (EEG). A small number of studies have demonstrated the feasibility of using ErrP detection as feedback for reinforcement-learning-based brain-computer interfaces (BCI), confirming the possibility of developing more autonomous BCI. These systems could be made more efficient with specific information about the type of error that occurred. A few studies differentiated the ErrP of different errors from each other, based on direction or severity. However, errors cannot always be categorised in these ways. We aimed to investigate the feasibility of differentiating very similar error conditions from each other, in the absence of previously explored metrics. APPROACH: In this study, we used two data sets with 25 and 14 participants to investigate the differences between errors. The two error conditions in each task were similar in terms of severity, direction and visual processing. The only notable differences between them were the varying cognitive processes involved in perceiving the errors, and differing contexts in which the errors occurred. We used a linear classifier with a small feature set to differentiate the errors on a single-trial basis. MAIN RESULTS: For both data sets, we observed neurophysiological distinctions between the ErrPs related to each error type. We found further distinctions between age groups. Furthermore, we achieved statistically significant single-trial classification rates for most participants included in the classification phase, with mean overall accuracy of 65.2% and 65.6% for the two tasks. SIGNIFICANCE: As a proof of concept our results showed that it is feasible, using single-trial EEG, to classify these similar error types against each other. This study paves the way for more detailed and efficient learning in BCI, and thus for a more autonomous human-machine interaction.

A BCI based visual-haptic neurofeedback training improves cortical activations and classification performance during motor imagery.

OBJECTIVE: We proposed a brain-computer interface (BCI) based visual-haptic neurofeedback training (NFT) by incorporating synchronous visual scene and proprioceptive electrical stimulation feedback. The goal of this work was to improve sensorimotor cortical activations and classification performance during motor imagery (MI). In addition, their correlations and brain network patterns were also investigated respectively. APPROACH: 64-channel electroencephalographic (EEG) data were recorded in nineteen healthy subjects during MI before and after NFT. During NFT sessions, the synchronous visual-haptic feedbacks were driven by real-time lateralized relative event-related desynchronization (lrERD). MAIN RESULTS: By comparison between previous and posterior control sessions, the cortical activations measured by multi-band (i.e. alpha_1: 8-10 Hz, alpha_2: 11-13 Hz, beta_1: 15-20 Hz and beta_2: 22-28 Hz) absolute ERD powers and lrERD patterns were significantly enhanced after the NFT. The classification performance was also significantly improved, achieving a ~9% improvement and reaching ~85% in mean classification accuracy from a relatively poor performance. Additionally, there were significant correlations between lrERD patterns and classification accuracies. The partial directed coherence based functional connectivity (FC) networks covering the sensorimotor area also showed an increase after the NFT. SIGNIFICANCE: These findings validate the feasibility of our proposed NFT to improve sensorimotor cortical activations and BCI performance during motor imagery. And it is promising to optimize conventional NFT manner and evaluate the effectiveness of motor training.

Deep learning for electroencephalogram (EEG) classification tasks: a review.

OBJECTIVE: Electroencephalography (EEG) analysis has been an important tool in neuroscience with applications in neuroscience, neural engineering (e.g. Brain-computer interfaces, BCI's), and even commercial applications. Many of the analytical tools used in EEG studies have used machine learning to uncover relevant information for neural classification and neuroimaging. Recently, the availability of large EEG data sets and advances in machine learning have both led to the deployment of deep learning architectures, especially in the analysis of EEG signals and in understanding the information it may contain for brain functionality. The robust automatic classification of these signals is an important step towards making the use of EEG more practical in many applications and less reliant on trained professionals. Towards this goal, a systematic review of the literature on deep learning applications to EEG classification was performed to address the following critical questions: (1) Which EEG classification tasks have been explored with deep learning? (2) What input formulations have been used for training the deep networks? (3) Are there specific deep learning network structures suitable for specific types of tasks? APPROACH: A systematic literature review of EEG classification using deep learning was performed on Web of Science and PubMed databases, resulting in 90 identified studies. Those studies were analyzed based on type of task, EEG preprocessing methods, input type, and deep learning architecture. MAIN RESULTS: For EEG classification tasks, convolutional neural networks, recurrent neural networks, deep belief networks outperform stacked auto-encoders and multi-layer perceptron neural networks in classification accuracy. The tasks that used deep learning fell into five general groups: emotion recognition, motor imagery, mental workload, seizure detection, event related potential detection, and sleep scoring. For each type of task, we describe the specific input formulation, major characteristics, and end classifier recommendations found through this review. SIGNIFICANCE: This review summarizes the current practices and performance outcomes in the use of deep learning for EEG classification. Practical suggestions on the selection of many hyperparameters are provided in the hope that they will promote or guide the deployment of deep learning to EEG datasets in future research.

Double ErrP Detection for Automatic Error Correction in an ERP-Based BCI Speller.

Brain-computer interface (BCI) is a useful device for people with severe motor disabilities. However, due to its low speed and low reliability, BCI still has a very limited application in daily real-world tasks. This paper proposes a P300-based BCI speller combined with a double error-related potential (ErrP) detection to automatically correct erroneous decisions. This novel approach introduces a second error detection to infer whether wrong automatic correction also elicits a second ErrP. Thus, two single-trial responses, instead of one, contribute to the final selection, improving the reliability of error detection. Moreover, to increase error detection, the evoked potential detected as target by the P300 classifier is combined with the evoked error potential at a feature-level. Discriminable error and positive potentials (response to correct feedback) were clearly identified. The proposed approach was tested on nine healthy participants and one tetraplegic participant. The online average accuracy for the first and second ErrPs were 88.4% and 84.8%, respectively. With automatic correction, we achieved an improvement around 5% achieving 89.9% in spelling accuracy for an effective 2.92 symbols/min. The proposed approach revealed that double ErrP detection can improve the reliability and speed of BCI systems.

Covert Verb Reading Contributes to Signal Classification of Motor Imagery in BCI.

Motor imagery is widely used in the brain-computer interface (BCI) systems that can help people actively control devices to directly communicate with the external world, but its training and performance effect is usually poor for normal people. To improve operators' BCI performances, here we proposed a novel paradigm, which combined the covert verb reading in the traditional motor imagery paradigm. In our proposed paradigm, participants were asked to covertly read the presented verbs during imagining right hand or foot movements referred by those verbs. EEG signals were recorded with both our proposed paradigm and the traditional paradigm. By the common spatial pattern method, we, respectively, decomposed these signals into spatial patterns and extracted their features used in the following classification of support vector machine. Compared with the traditional paradigm, our proposed paradigm could generate clearer spatial patterns following a somatotopic distribution, which led to more distinguishable features and higher classification accuracies than those in the traditional paradigm. These results suggested that semantic processing of verbs can influence the brain activity of motor imagery and enhance the mu event-related desynchronisation. The combination of semantic processing with motor imagery is therefore a promising method for the improvement of operators' BCI performances.

A Multi-Class Tactile Brain-Computer Interface Based on Stimulus-Induced Oscillatory Dynamics.

We proposed a multi-class tactile brain-computer interface that utilizes stimulus-induced oscillatory dynamics. It was hypothesized that somatosensory attention can modulate tactile-induced oscillation changes, which can decode different sensation attention tasks. Subjects performed four tactile attention tasks, prompted by cues presented in random order and while both wrists were simultaneously stimulated: 1) selective sensation on left hand (SS-L); 2) selective sensation on right hand (SS-R); 3) bilateral selective sensation; and 4) selective sensation suppressed or idle state (SS-S). The classification accuracy between SS-L and SS-R (79.9 ± 8.7%) was comparable with that of a previous tactile BCI system based on selective sensation. Moreover, the accuracy could be improved to an average of 90.3 ± 4.9% by optimal class-pair and frequency-band selection. Three-class discrimination had an accuracy of 75.2 ± 8.3%, with the best discrimination reached for the classes SS-L, SS-R, and SS-S. Finally, four classes were classified with an accuracy of 59.4 ± 7.3%. These results show that the proposed system is a promising new paradigm for multi-class BCI.

Inferring imagined speech using EEG signals: a new approach using Riemannian manifold features.

OBJECTIVE: In this paper, we investigate the suitability of imagined speech for brain-computer interface (BCI) applications. APPROACH: A novel method based on covariance matrix descriptors, which lie in Riemannian manifold, and the relevance vector machines classifier is proposed. The method is applied on electroencephalographic (EEG) signals and tested in multiple subjects. MAIN RESULTS: The method is shown to outperform other approaches in the field with respect to accuracy and robustness. The algorithm is validated on various categories of speech, such as imagined pronunciation of vowels, short words and long words. The classification accuracy of our methodology is in all cases significantly above chance level, reaching a maximum of 70% for cases where we classify three words and 95% for cases of two words. SIGNIFICANCE: The results reveal certain aspects that may affect the success of speech imagery classification from EEG signals, such as sound, meaning and word complexity. This can potentially extend the capability of utilizing speech imagery in future BCI applications. The dataset of speech imagery collected from total 15 subjects is also published.

A review and experimental study on the application of classifiers and evolutionary algorithms in EEG-based brain-machine interface systems.

OBJECTIVE: Considering the importance and the near-future development of noninvasive brain-machine interface (BMI) systems, this paper presents a comprehensive theoretical-experimental survey on the classification and evolutionary methods for BMI-based systems in which EEG signals are used. APPROACH: The paper is divided into two main parts. In the first part, a wide range of different types of the base and combinatorial classifiers including boosting and bagging classifiers and evolutionary algorithms are reviewed and investigated. In the second part, these classifiers and evolutionary algorithms are assessed and compared based on two types of relatively widely used BMI systems, sensory motor rhythm-BMI and event-related potentials-BMI. Moreover, in the second part, some of the improved evolutionary algorithms as well as bi-objective algorithms are experimentally assessed and compared. MAIN RESULTS: In this study two databases are used, and cross-validation accuracy (CVA) and stability to data volume (SDV) are considered as the evaluation criteria for the classifiers. According to the experimental results on both databases, regarding the base classifiers, linear discriminant analysis and support vector machines with respect to CVA evaluation metric, and naive Bayes with respect to SDV demonstrated the best performances. Among the combinatorial classifiers, four classifiers, Bagg-DT (bagging decision tree), LogitBoost, and GentleBoost with respect to CVA, and Bagging-LR (bagging logistic regression) and AdaBoost (adaptive boosting) with respect to SDV had the best performances. Finally, regarding the evolutionary algorithms, single-objective invasive weed optimization (IWO) and bi-objective nondominated sorting IWO algorithms demonstrated the best performances. SIGNIFICANCE: We present a general survey on the base and the combinatorial classification methods for EEG signals (sensory motor rhythm and event-related potentials) as well as their optimization methods through the evolutionary algorithms. In addition, experimental and statistical significance tests are carried out to study the applicability and effectiveness of the reviewed methods.

Invariance and variability in interaction error-related potentials and their consequences for classification.

OBJECTIVE: This paper discusses the invariance and variability in interaction error-related potentials (ErrPs), where a special focus is laid upon the factors of (1) the human mental processing required to assess interface actions (2) time (3) subjects. APPROACH: Three different experiments were designed as to vary primarily with respect to the mental processes that are necessary to assess whether an interface error has occurred or not. The three experiments were carried out with 11 subjects in a repeated-measures experimental design. To study the effect of time, a subset of the recruited subjects additionally performed the same experiments on different days. MAIN RESULTS: The ErrP variability across the different experiments for the same subjects was found largely attributable to the different mental processing required to assess interface actions. Nonetheless, we found that interaction ErrPs are empirically invariant over time (for the same subject and same interface) and to a lesser extent across subjects (for the same interface). SIGNIFICANCE: The obtained results may be used to explain across-study variability of ErrPs, as well as to define guidelines for approaches to the ErrP classifier transferability problem.

Classification of different reaching movements from the same limb using EEG.

OBJECTIVE: Brain-computer-interfaces (BCIs) have been proposed not only as assistive technologies but also as rehabilitation tools for lost functions. However, due to the stochastic nature, poor spatial resolution and signal to noise ratio from electroencephalography (EEG), multidimensional decoding has been the main obstacle to implement non-invasive BCIs in real-live rehabilitation scenarios. This study explores the classification of several functional reaching movements from the same limb using EEG oscillations in order to create a more versatile BCI for rehabilitation. APPROACH: Nine healthy participants performed four 3D center-out reaching tasks in four different sessions while wearing a passive robotic exoskeleton at their right upper limb. Kinematics data were acquired from the robotic exoskeleton. Multiclass extensions of Filter Bank Common Spatial Patterns (FBCSP) and a linear discriminant analysis (LDA) classifier were used to classify the EEG activity into four forward reaching movements (from a starting position towards four target positions), a backward movement (from any of the targets to the starting position and rest). Recalibrating the classifier using data from previous or the same session was also investigated and compared. MAIN RESULTS: Average EEG decoding accuracy were significantly above chance with 67%, 62.75%, and 50.3% when decoding three, four and six tasks from the same limb, respectively. Furthermore, classification accuracy could be increased when using data from the beginning of each session as training data to recalibrate the classifier. SIGNIFICANCE: Our results demonstrate that classification from several functional movements performed by the same limb is possible with acceptable accuracy using EEG oscillations, especially if data from the same session are used to recalibrate the classifier. Therefore, an ecologically valid decoding could be used to control assistive or rehabilitation mutli-degrees of freedom (DoF) robotic devices using EEG data. These results have important implications towards assistive and rehabilitative neuroprostheses control in paralyzed patients.

Pure visual imagery as a potential approach to achieve three classes of control for implementation of BCI in non-motor disorders.

OBJECTIVE: The achievement of multiple instances of control with the same type of mental strategy represents a way to improve flexibility of brain-computer interface (BCI) systems. Here we test the hypothesis that pure visual motion imagery of an external actuator can be used as a tool to achieve three classes of electroencephalographic (EEG) based control, which might be useful in attention disorders. APPROACH: We hypothesize that different numbers of imagined motion alternations lead to distinctive signals, as predicted by distinct motion patterns. Accordingly, a distinct number of alternating sensory/perceptual signals would lead to distinct neural responses as previously demonstrated using functional magnetic resonance imaging (fMRI). We anticipate that differential modulations should also be observed in the EEG domain. EEG recordings were obtained from twelve participants using three imagery tasks: imagery of a static dot, imagery of a dot with two opposing motions in the vertical axis (two motion directions) and imagery of a dot with four opposing motions in vertical or horizontal axes (four directions). The data were analysed offline. MAIN RESULTS: An increase of alpha-band power was found in frontal and central channels as a result of visual motion imagery tasks when compared with static dot imagery, in contrast with the expected posterior alpha decreases found during simple visual stimulation. The successful classification and discrimination between the three imagery tasks confirmed that three different classes of control based on visual motion imagery can be achieved. The classification approach was based on a support vector machine (SVM) and on the alpha-band relative spectral power of a small group of six frontal and central channels. Patterns of alpha activity, as captured by single-trial SVM closely reflected imagery properties, in particular the number of imagined motion alternations. SIGNIFICANCE: We found a new mental task based on visual motion imagery with potential for the implementation of multiclass (3) BCIs. Our results are consistent with the notion that frontal alpha synchronization is related with high internal processing demands, changing with the number of alternation levels during imagery. Together, these findings suggest the feasibility of pure visual motion imagery tasks as a strategy to achieve multiclass control systems with potential for BCI and in particular, neurofeedback applications in non-motor (attentional) disorders.

A systematic review of hybrid brain-computer interfaces: Taxonomy and usability perspectives.

A new Brain-Computer Interface (BCI) technique, which is called a hybrid BCI, has recently been proposed to address the limitations of conventional single BCI system. Although some hybrid BCI studies have shown promising results, the field of hybrid BCI is still in its infancy and there is much to be done. Especially, since the hybrid BCI systems are so complicated and complex, it is difficult to understand the constituent and role of a hybrid BCI system at a glance. Also, the complicated and complex systems make it difficult to evaluate the usability of the systems. We systematically reviewed and analyzed the current state-of-the-art hybrid BCI studies, and proposed a systematic taxonomy for classifying the types of hybrid BCIs with multiple taxonomic criteria. After reviewing 74 journal articles, hybrid BCIs could be categorized with respect to 1) the source of brain signals, 2) the characteristics of the brain signal, and 3) the characteristics of operation in each system. In addition, we exhaustively reviewed recent literature on usability of BCIs. To identify the key evaluation dimensions of usability, we focused on task and measurement characteristics of BCI usability. We classified and summarized 31 BCI usability journal articles according to task characteristics (type and description of task) and measurement characteristics (subjective and objective measures). Afterwards, we proposed usability dimensions for BCI and hybrid BCI systems according to three core-constructs: Satisfaction, effectiveness, and efficiency with recommendations for further research. This paper can help BCI researchers, even those who are new to the field, can easily understand the complex structure of the hybrid systems at a glance. Recommendations for future research can also be helpful in establishing research directions and gaining insight in how to solve ergonomics and HCI design issues surrounding BCI and hybrid BCI systems by usability evaluation.

Weighted spatial based geometric scheme as an efficient algorithm for analyzing single-trial EEGS to improve cue-based BCI classification.

There is a growing interest in analyzing the geometrical behavior of electroencephalogram (EEG) covariance matrix in the context of brain computer interface (BCI). The bottleneck of the current Riemannian framework is the bias of the mean vector of EEG signals to the noisy trials, which deteriorates the covariance matrix in the manifold space. This study presents a spatial weighting scheme to reduce the effect of noisy trials on the mean vector. To assess the proposed method, dataset IIa from BCI competition IV, containing the EEG trials of 9 subjects performing four mental tasks, was utilized. The performance of the proposed method is compared to the classical Riemannian method along with Common Spatial Pattern (CSP) on the dataset. The results show that when considering just two imagery classes, the proposed method performs on par with CSP method, whereas in the multi class scenario, the proposed algorithm outperforms the CSP approach on seven out of nine subjects. Incidentally, the proposed method obtains better accuracy for the majority of subjects compared to the classical Riemannian method.

A hybrid three-class brain-computer interface system utilizing SSSEPs and transient ERPs.

OBJECTIVE: This paper investigates the fusion of steady-state somatosensory evoked potentials (SSSEPs) and transient event-related potentials (tERPs), evoked through tactile simulation on the left and right-hand fingertips, in a three-class EEG based hybrid brain-computer interface. It was hypothesized, that fusing the input signals leads to higher classification rates than classifying tERP and SSSEP individually. APPROACH: Fourteen subjects participated in the studies, consisting of a screening paradigm to determine person dependent resonance-like frequencies and a subsequent online paradigm. The whole setup of the BCI system was based on open interfaces, following suggestions for a common implementation platform. During the online experiment, subjects were instructed to focus their attention on the stimulated fingertips as indicated by a visual cue. The recorded data were classified during runtime using a multi-class shrinkage LDA classifier and the outputs were fused together applying a posterior probability based fusion. Data were further analyzed offline, involving a combined classification of SSSEP and tERP features as a second fusion principle. The final results were tested for statistical significance applying a repeated measures ANOVA. MAIN RESULTS: A significant classification increase was achieved when fusing the results with a combined classification compared to performing an individual classification. Furthermore, the SSSEP classifier was significantly better in detecting a non-control state, whereas the tERP classifier was significantly better in detecting control states. Subjects who had a higher relative band power increase during the screening session also achieved significantly higher classification results than subjects with lower relative band power increase. SIGNIFICANCE: It could be shown that utilizing SSSEP and tERP for hBCIs increases the classification accuracy and also that tERP and SSSEP are not classifying control- and non-control states with the same level of accuracy.

Integrating language models into classifiers for BCI communication: a review.

OBJECTIVE: The present review systematically examines the integration of language models to improve classifier performance in brain-computer interface (BCI) communication systems. APPROACH: The domain of natural language has been studied extensively in linguistics and has been used in the natural language processing field in applications including information extraction, machine translation, and speech recognition. While these methods have been used for years in traditional augmentative and assistive communication devices, information about the output domain has largely been ignored in BCI communication systems. Over the last few years, BCI communication systems have started to leverage this information through the inclusion of language models. MAIN RESULTS: Although this movement began only recently, studies have already shown the potential of language integration in BCI communication and it has become a growing field in BCI research. BCI communication systems using language models in their classifiers have progressed down several parallel paths, including: word completion; signal classification; integration of process models; dynamic stopping; unsupervised learning; error correction; and evaluation. SIGNIFICANCE: Each of these methods have shown significant progress, but have largely been addressed separately. Combining these methods could use the full potential of language model, yielding further performance improvements. This integration should be a priority as the field works to create a BCI system that meets the needs of the amyotrophic lateral sclerosis population.

Tensor-based classification of an auditory mobile BCI without a subject-specific calibration phase.

OBJECTIVE: One of the major drawbacks in EEG brain-computer interfaces (BCI) is the need for subject-specific training of the classifier. By removing the need for a supervised calibration phase, new users could potentially explore a BCI faster. In this work we aim to remove this subject-specific calibration phase and allow direct classification. APPROACH: We explore canonical polyadic decompositions and block term decompositions of the EEG. These methods exploit structure in higher dimensional data arrays called tensors. The BCI tensors are constructed by concatenating ERP templates from other subjects to a target and non-target trial and the inherent structure guides a decomposition that allows accurate classification. We illustrate the new method on data from a three-class auditory oddball paradigm. MAIN RESULTS: The presented approach leads to a fast and intuitive classification with accuracies competitive with a supervised and cross-validated LDA approach. SIGNIFICANCE: The described methods are a promising new way of classifying BCI data with a forthright link to the original P300 ERP signal over the conventional and widely used supervised approaches.