Regularized Linear Discriminant Analysis Research Articles

Fusing spectral and spatial features of hyperspectral reflectance imagery for differentiating between normal and defective blueberries

Effective defect detection of blueberries is important to ensuring supplies of high-quality products to the fresh market. In this study, hyperspectral reflectance imaging with machine learning was evaluated for discriminating between defective and normal blueberries. Fresh blueberries hand-harvested were scanned in the wavelength range of 400–1000 nm. An image analysis pipeline was developed to segment individual blueberries and extract mean spectra and spatial features. Defective blueberries were found to have lower near-infrared reflectance than sound samples, and spectral features produced a better separation between defective and sound samples than the spatial features in the scattering plots of their first two principal components. Nine types of machine learning models were built for classifying defective and sound samples using the spectral and spatial features separately as well as their concatenation. The regularized linear discriminant analysis (RLDA) model trained on the spectral features achieved the best overall accuracy of 95.7 %, as opposed to the best accuracy of 85.3 % based on spatial features, which was obtained by LDA. Simply concatenating spectral and spatial features did not improve over modeling using spectral or spatial features alone. A model ensemble strategy integrating the spectral features-based RLDA and the spatial features-based LDA resulted in a statistically significant improvement in the overall accuracy to 96.6 %. Model-level feature integration offers an effective means for improving the discrimination between defective and normal blueberries. Both the hyperspectral data11https://doi.org/10.5281/zenodo.11200576. and the software programs22https://github.com/vicdxxx/Blueberry-Defect-Detection-by-Hyperspectral-Imaging. of this study are made publicly available.

EEG-based hierarchical classification of level of demand and modality of auditory and visual sensory processing

Objective. To date, most research on electroencephalography (EEG)-based mental workload detection for passive brain-computer interface (pBCI) applications has focused on identifying the overall level of cognitive resources required, such as whether the workload is high or low. We propose, however, that being able to determine the specific type of cognitive resources being used, such as visual or auditory, would also be useful. This would enable the pBCI to take more appropriate action to reduce the overall level of cognitive demand on the user. For example, if a high level of workload was detected and it is determined that the user is primarily engaged in visual information processing, then the pBCI could cause some information to be presented aurally instead. In our previous work we showed that EEG could be used to differentiate visual from auditory processing tasks when the level of processing is high, but the two modalities could not be distinguished when the level of cognitive processing demand was very low. The current study aims to build on this work and move toward the overall objective of developing a pBCI that is capable of predicting both the level and the type of cognitive resources being used. Approach. Fifteen individuals undertook carefully designed visual and auditory tasks while their EEG data was being recorded. In this study, we incorporated a more diverse range of sensory processing conditions including not only single-modality conditions (i.e. those requiring one of either visual or auditory processing) as in our previous study, but also dual-modality conditions (i.e. those requiring both visual and auditory processing) and no-task/baseline conditions (i.e. when the individual is not engaged in either visual or auditory processing). Main results. Using regularized linear discriminant analysis within a hierarchical classification algorithm, the overall cognitive demand was predicted with an accuracy of more than 86%, while the presence or absence of visual and auditory sensory processing were each predicted with an accuracy of approximately 70%. Significance. The findings support the feasibility of establishing a pBCI that can determine both the level and type of attentional resources required by the user at any given moment. This pBCI could assist in enhancing safety in hazardous jobs by triggering the most effective and efficient adaptation strategies when high workload conditions are detected.

EEG-Based Emergency Braking Prediction Using Data Ablation and SVM Classification

Contemporary advanced driver assistance system (ADAS) features for semi-autonomous vehicles include braking assistance during collision avoidance. Although pre-collision detection typically relies on sensing systems to enable production vehicles to perceive oncoming road obstacles, the physiological state of the driver is not measured to predict emergency braking. On the other hand, previous driving simulation experiments have demonstrated the ability of regularized linear discriminant analysis (RLDA) to predict pre-collision braking using brain signals from multiple electroencephalogram (EEG) electrodes. In contrast, the current study used EEG data from these previous experiments to determine the quality of support vector machine (SVM) predictions as a first step towards realizing a brain-computer interface for emergency braking. Power spectral density (PSD) features were extracted from the EEG of one electrode to train and evaluate an SVM. Through a novel data ablation analysis, the optimal number of PSD components was determined to optimize model classification quality measured by the area under the curve (AUC). A comparison of the proposed model to the previous RLDA and other machine learning methods indicated that the SVM had a superior AU C. Thus, the proposed model is a candidate for assisting ADASs with pre-collision detection. Moreover, since the proposed model only utilized one electrode, our study potentially contributes to the facilitation of brain-computer interfaces for autonomous vehicles.

Classifying Children with ADHD Based on Prefrontal Functional Near-infrared Spectroscopy Using Machine Learning.

: Attention deficit hyperactivity disorder (ADHD) is a common neurodevelopmental disorder in children and adults characterized by cognitive and emotional self-control deficiencies. Previous functional near-infrared spectroscopy (fNIRS) studies found significant group differences between ADHD children and healthy controls during cognitive flexibility tasks in several brain regions. This study aims to apply a machine learning approach to identify medication-naive ADHD patients and healthy control (HC) groups using task-based fNIRS data. : fNIRS signals from 33 ADHD children and 39 HC during the Stroop task were analyzed. In addition, regularized linear discriminant analysis (RLDA) was used to identify ADHD individuals from healthy controls, and classification performance was evaluated. : We found that participants can be correctly classified in RLDA leave-one-out cross validation, with a sensitivity of 0.67, specificity of 0.93, and accuracy of 0.82. : RLDA using only fNIRS data can effectively discriminate children with ADHD from HC. This study suggests the potential utility of the fNIRS signal as a diagnostic biomarker for ADHD children.

Feasibility of imaging under structured illumination for evaluation of white striping in broiler breast fillets

Muscle myopathies or defects, such as white striping (WS), downgrade the quality of broiler meat, leading to the loss of millions of dollars annually for the U.S. poultry industry. Visual inspection is currently practiced for assessing myopathies in broiler meat, which is however prone to human evaluation error, labor-intensive, and costly. Imaging under uniform or diffuse illumination has been investigated for assessing broiler meat quality, but their performance is not always satisfactory, especially for detecting subtle defects with few visual symptoms. This study presents the first proof-of-concept evaluation of the utility of an innovative technique for improved imaging of structures of biological materials, i.e., structured-illumination reflectance imaging (SIRI), for detecting WS conditions in broiler breast meat. Broiler breast fillets with varying degrees of WS severity were imaged by an in-house assembled SIRI system under sinusoidal illumination at different spatial frequencies (0.05–0.40 cycles/mm). The acquired pattern images were demodulated into direct components (DC) and amplitude components (AC) at each spatial frequency. Texture features were extracted from DC and AC images for building classification models using regularized linear discriminant analysis, which were to classify meat samples into 2, 3, and 4 classes, respectively, according to WS conditions. AC images were found by visual inspection more effective than DC in resolving WS characteristics, depending on the spatial frequency of illumination patterns. Compared to DC, AC yielded the maximum accuracy improvements of 8.33%, 7.93%, and 12.3% in 2-, 3- and 4-class classifications, respectively. Models using ranked features yielded similar improvements of up to 9.53% by AC over DC. The SIRI technique is promising for enhanced detection of WS and potentially other defects of broiler breast meat. • Structured-illumination reflectance imaging (SIRI) was used for poultry quality detection. • Amplitude component (AC) images enabled resolving white striping (WS) of broiler breast meat. • Broiler breast fillets were classified by models into 2, 3 and 4 classes according to WS conditions. • SIRI offered substantial improvements of up to 12% in classifying WS-affected poultry samples.

Early Detection of Alzheimer’s Using EEG

Alzheimer’s disease (AD) is a brain disorder which is irreversible as well as progressive. AD slowly destroys memory and thinking skills. The most common cause of Dementia is AD. Dementia that consists of (loss of Cognitive functioning – thinking, remembering and reasoning – and behavioural ability) is caused due to death of neurons due to AD, it has serious interference with Daily life. The chronicity of the disease is increasing rapidly in the world. The advancements in medical have initiated many research studies to automatically detect the disease by capturing the brain signals. The main aim of this paper is to help the doctors detect the Alzheimer’s disease at earliest and more number of patients can be prevented before irreversible changes occur in brain. In this paper we present some common features that will improve the detection of AD accurately from EEG signals in early stage of the disease. Apart from this, different signal decomposition methods including filtering into brain frequency bands, discrete wavelet transform (DWT) and empirical mode decomposition (EMD), and various classification algorithms including support vector machine (SVM), K-nearest neighbours (KNN) and regularized linear discriminant analysis (RLDA) are evaluated.

Improving Intention Detection in Single-Trial Classification Through Fusion of EEG and Eye-Tracker Data

Intention decoding is an indispensable procedure in hands-free human–computer interaction (HCI). A conventional eye-tracker system using a single-model fixation duration may issue commands that ignore users' real expectations. Here, an eye-brain hybrid brain–computer interface (BCI) interaction system was introduced for intention detection through the fusion of multimodal eye-tracker and event-related potential (ERP) [a measurement derived from electroencephalography (EEG)] features. Eye-tracking and EEG data were recorded from 64 healthy participants as they performed a 40-min customized free search task of a fixed target icon among 25 icons. The corresponding fixation duration of eye tracking and ERP were extracted. Five previously-validated linear discriminant analysis (LDA)-based classifiers [including regularized LDA, stepwise LDA, Bayesian LDA, shrinkage linear discriminant analysis (SKLDA), and spatial-temporal discriminant analysis] and the widely-used convolutional neural network (CNN) method were adopted to verify the efficacy of feature fusion from both offline and pseudo-online analysis, and the optimal approach was evaluated by modulating the training set and system response duration. Our study demonstrated that the input of multimodal eye tracking and ERP features achieved a superior performance of intention detection in the single-trial classification of active search tasks. Compared with the single-model ERP feature, this new strategy also induced congruent accuracy across classifiers. Moreover, in comparison with other classification methods, we found that SKLDA exhibited a superior performance when fusing features in offline tests (ACC = 0.8783, AUC = 0.9004) and online simulations with various sample amounts and duration lengths. In summary, this study revealed a novel and effective approach for intention classification using an eye-brain hybrid BCI and further supported the real-life application of hands-free HCI in a more precise and stable manner.

Decoding EEG rhythms offline and online during motor imagery for standing and sitting based on a brain-computer interface.

Motor imagery (MI)-based brain-computer interface (BCI) systems have shown promising advances for lower limb motor rehabilitation. The purpose of this study was to develop an MI-based BCI for the actions of standing and sitting. Thirty-two healthy subjects participated in the study using 17 active EEG electrodes. We used a combination of the filter bank common spatial pattern (FBCSP) method and the regularized linear discriminant analysis (RLDA) technique for decoding EEG rhythms offline and online during motor imagery for standing and sitting. The offline analysis indicated the classification of motor imagery and idle state provided a mean accuracy of 88.51 ± 1.43% and 85.29 ± 1.83% for the sit-to-stand and stand-to-sit transitions, respectively. The mean accuracies of the sit-to-stand and stand-to-sit online experiments were 94.69 ± 1.29% and 96.56 ± 0.83%, respectively. From these results, we believe that the MI-based BCI may be useful to future brain-controlled standing systems.

RLDA: Valid test pattern identification by machine learning classification method for VLSI test

Testing of integrated circuits (ICs) is essential for weeding out defects before the products are shipped to customers. However, as circuitry shrinks in scale, the complexity of the circuit increases, leading to test cost increases. Considering test time (TT) taking too long due to pattern increases in digital circuits, a compact test set was selected to save TT without increasing test escape (when faulty circuits are undetected). Therefore, a machine learning (ML) classification method called regularized linear discriminant analysis (RLDA) algorithm was proposed. In this way, valid (test failed) patterns can be selected and invalid (test pass) patterns can be dropped to shorten TT. Results show that the least DPPM (97) for the same TT level, or the least TT (1.2s) for the same DPPM can be obtained.

Identifying chronic rhinosinusitis without nasal polyps by analyzing aspirated nasal air with an electronic nose based on differential mobility spectrometry

Background The diagnosis of chronic rhinosinusitis (CRS) is a complicated procedure. An electronic nose (eNose) is a novel method that detects disease from gas-phase mixtures, such as human breath. Aims/Objectives To determine whether an eNose based on differential mobility spectrometry (DMS) can detect chronic rhinosinusitis without nasal polyps (CRSsNP) by analyzing aspirated nasal air. Materials and methods Adult patients with CRSsNP were examined. The control group consisted of patients with septal deviation. Nasal air was aspirated into a collection bag and analyzed with DMS. The DMS data were classified using regularized linear discriminant analysis (LDA) models with 10-fold cross-validation. Results The accuracy of the DMS to distinguish CRSsNP from patients with septal deviation was 69%. Sensitivity and specificity were 67 and 70%, respectively. Bonferroni-corrected statistical differences were clearly noted. When a subgroup with more severe inflammatory disease was compared to controls, the classification accuracy increased to 82%. Conclusions The results of this feasibility study demonstrate that CRSsNP can potentially be differentiated distinguished from patients with similar nasal symptoms by analyzing the aspirated nasal air using DMS. Further research is warranted to evaluate the ability of this novel method in the differential diagnostics of CRS.

Multiscale principle of relevant information for hyperspectral image classification

This paper proposes a novel architecture, termed multiscale principle of relevant information (MPRI), to learn discriminative spectral-spatial features for hyperspectral image classification. MPRI inherits the merits of the principle of relevant information (PRI) to effectively extract multiscale information embedded in the given data, and also takes advantage of the multilayer structure to learn representations in a coarse-to-fine manner. Specifically, MPRI performs spectral-spatial pixel characterization (using PRI) and feature dimensionality reduction (using regularized linear discriminant analysis) iteratively and successively. Extensive experiments on three benchmark data sets demonstrate that MPRI outperforms existing state-of-the-art methods (including deep learning based ones) qualitatively and quantitatively, especially in the scenario of limited training samples. Code of MPRI is available at http://bit.ly/MPRI_HSI.

A Doubly Regularized Linear Discriminant Analysis Classifier With Automatic Parameter Selection

Linear discriminant analysis (LDA) based classifiers tend to falter in many practical settings where the training data size is smaller than, or comparable to, the number of features. As a remedy, different regularized LDA (RLDA) methods have been proposed. These methods may still perform poorly depending on the size and quality of the available training data. In particular, the test data deviation from the training data model, for example, due to noise contamination, can cause severe performance degradation. Moreover, these methods commit further to the Gaussian assumption (upon which LDA is established) to tune their regularization parameters, which may compromise accuracy when dealing with real data. To address these issues, we propose a doubly regularized LDA classifier that we denote as R2LDA. In the proposed R2LDA approach, the RLDA score function is converted into an inner product of two vectors. By substituting the expressions of the regularized estimators of these vectors, we obtain the R2LDA score function that involves two regularization parameters. To set the values of these parameters, we adopt three existing regularization techniques; the constrained perturbation regularization approach (COPRA), the bounded perturbation regularization (BPR) algorithm, and the generalized cross-validation (GCV) method. These methods are used to tune the regularization parameters based on linear estimation models, with the sample covariance matrix's square root being the linear operator. Results obtained from both synthetic and real data demonstrate the consistency and effectiveness of the proposed R2LDA approach, especially in scenarios involving test data contaminated with noise that is not observed during the training phase.

Early detection of Alzheimer’s disease from EEG signals using Hjorth parameters

BackgroundAlzheimer's disease (AD) is a progressive neurodegenerative disorder of the brain that ultimately results in the death of neurons and dementia. The prevalence of the disease in the world is increasing rapidly. In recent years, many studies have been done to automatically detect this disease from brain signals. MethodIn this paper, the Hjorth parameters are used along with other common features to improve the AD detection accuracy from EEG signals in early stages. Also different signal decomposition methods including filtering into brain frequency bands, discrete wavelet transform (DWT) and empirical mode decomposition (EMD), and various classification algorithms including support vector machine (SVM), K-nearest neighbors (KNN) and regularized linear discriminant analysis (RLDA) are evaluated. ResultsAfter preprocessing and extracting the discriminative features from EEG signals for 35 healthy, 31 mild AD, and 20 moderate AD subjects, the performance of different decomposition methods and different classifiers was evaluated before and after combining Hjorth parameters. ConclusionsIt was shown that combining Hjorth parameters to the common features improved the accuracy of detection and by using DWT method for signal decomposition and the KNN algorithm for classification the highest accuracy is obtained as 97.64%.

Classification of Selective Attention Within Steady-State Somatosensory Evoked Potentials From Dry Electrodes Using Mutual Information-Based Spatio-Spectral Feature Selection

Nowadays, the steady-state somatosensory evoked potential (SSSEP)-based brain-computer interfaces (BCIs) has been developed for improving the quality of daily life for people with physical disabilities. However, due to its poor performance of recognizing selective attention tasks and inattention(rest)-state, the SSSEP-based BCI has not been widely used for practical interfaces. In this paper, we propose a mutual information-based spatio-spectral feature selection method for recognizing selective attention tasks and inattention(rest)-state using dry electrodes considering a real-life application, when vibration stimuli were applied to both index fingers. In our methods, the filter-bank common spatial pattern (FBCSP) was used for extracting spatio-spectral features of the SSSEP. Then, discriminative features were selected using a mutual information-based best individual feature (MIBIF) algorithm. The regularized linear discriminant analysis (RLDA) used as the classifier. The feasibility of the proposed method was demonstrated through eight healthy subjects using the vibration stimuli induced SSSEP with spatially clear and distinguishable patterns for SSSEP-based BCI. From our study, the proposed method showed the best classification accuracy with a kappa value of 0.35±0.17. Furthermore, based on the ANOVA with posthoc tests, the proposed method showed significantly higher accuracy as 57.9% in decoding three classes (p -value <; 0.01) compared to the fast Fourier transform (FFT) and common spatial pattern (CSP)-based previous feature extraction methods. Consequently, the proposed FBCSP and MIBIF-based methods and findings can further help to improve decoding performance and develop the SSSEP-based BCI systems for real-world applications.

Asymptotically Bias-Corrected Regularized Linear Discriminant Analysis for Cost-Sensitive Binary Classification

In this letter, the theory of random matrices of increasing dimension is used to construct a form of regularized linear discriminant analysis (RLDA) that asymptotically yields the lowest overall risk with respect to the bias of the discriminant in cost-sensitive classification of two multivariate Gaussian distributions. Numerical experiments using both synthetic and real data show that even in finite-sample settings, the proposed classifier can uniformly outperform RLDA in terms of achieving a lower risk as a function of regularization parameter and misclassification costs.

Robust Face Recognition Method Based on Kernel Regularized Relevance Weighted Discriminant Analysis and Deterministic Approach

A novel feature dimensionality reduction strategy based on kernel regularized relevance weighted discriminant analysis is proposed in this paper with some interesting characteristics. First, the proposed method has shown its effectiveness in dealing with a small sample size problem when using the regularized linear discriminant analysis (RLDA) technique and Kernel theory. Second, a new computation method is proposed to solve the complicated and inefficient computation procedure problem in the traditional RLDA technique while using the cross-validation method. The experimental results indicate that the proposed algorithm shows better performance than the other methods.

Abstract 3516: A genomics model to predict immune-related adverse events in cancer patients treated with checkpoint inhibitors

Abstract The overall objective of this study is to use next-generation sequencing technology and bioinformatics to better inform the safety of immunotherapy treatment for cancer. Cancer patients commonly develop immune-related adverse events (irAEs) during and after treatment with checkpoint inhibitors. These irAEs can be serious or even fatal. Therefore, a biomarker for prediction of irAE development could have utility for heightened surveillance, personalized therapy decisions, and regulatory evaluation of drugs. Due to the similarity between irAEs and autoimmune diseases and the high heritability of autoimmune diseases, we hypothesized that certain patients could have a genetic predisposition for developing irAEs. To test this hypothesis, we conducted whole exome sequencing on an Illumina NextSeq to interrogate germline genomes of solid tumor patients (n=50) treated with an anti-PD-L1 antibody (NCT01772004). Relevant clinical data, such as adverse events and irAE classification, were provided for this retrospective analysis; twenty percent of patients (10/50) had irAEs. A preliminary germline genetic model of irAEs was constructed using short variant calls from this initial training set. This was generated using a proprietary algorithm that implements a Monte-Carlo simulation expansion of Fisher’s regularized linear discriminant analysis (RLDA) in a multidimensional measurement system to create a model that maximizes separation between two groups. This model consists of 131 genes, each of which make a relatively small contribution to the overall signature. The ten genes with the highest contribution coefficients together account for 21% of the signature. Ingenuity Pathway Analysis (IPA) identified a network associated with infectious diseases, antimicrobial response, and inflammatory response that contains 21 interconnected genes from the signature. IPA also revealed that genes in the signature have a variety of molecular and cellular functions, the most significant of which are cell death and survival, cellular movement, and cell-to-cell signaling and interaction. This model has 100% sensitivity, specificity, and accuracy on the training set. Future directions will test the performance of this putative genomic model on a new dataset to assess the validity and utility of the model as a predictive biomarker to identify patients at risk for developing irAEs in response to checkpoint inhibition. Citation Format: Emma C. Scott, Dickran Kazandjian, Luis Santana-Quintero, Tigran Ghazanchyan, Svetlana Petrovskaya, Yong Zhang, Amy Rosenberg, V. Ashutosh Rao, Jennifer L. Marte, Gideon M. Blumenthal, Marc R. Theoret, Richard Pazdur, James L. Gulley, Julia A. Beaver. A genomics model to predict immune-related adverse events in cancer patients treated with checkpoint inhibitors [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 3516.

Improvement Motor Imagery EEG Classification Based on Regularized Linear Discriminant Analysis.

Mental tasks classification such as motor imagery, based on EEG signals is an important problem in brain computer interface systems (BCI). One of the major concerns in BCI is to have a high classification accuracy. The other concerning one is with the favorable result is guaranteed how to improve the computational efficiency. In this paper, Mu/Beta rhythm was obtained by bandpass filter from EEG signal. And the classical linear discriminant analysis (LDA) was used for deciding which rhythm can give the better classification performance. During this, the common spatial pattern (CSP) was used to project data subject to the ratio of projected energy of one class to that of the other class was maximized. The optimal projection dimension was determined corresponding to the maximum of area under the curve (AUC) for each participant. Eventually, regularized linear discriminant analysis (RLDA) is possible to decode the imagined motor sensed using electroencephalogram (EEG). Results show that higher classification accuracy can be provided by RLDA. And optimal projection dimensions determined by LDA and RLDA are of consistent solution, this improves computational efficiency of CSP-RLDA method without computation of projection dimension.

Optimized regularized linear discriminant analysis for feature extraction in face recognition

In a reduced dimensional space, linear discriminant analysis looks for a projective transformation that can maximizes separability among classes. Since linear discriminant analysis demands the within-class scatter matrix appear to non-singular, which cannot directly used in condition of small sample size (SSS) issues in which the dimension of image is much higher, while the number of samples isn’t unlimited. Both the between-class and within-class scatter matrices are always exceedingly ill-posed in SSS problems. And many algorithms are suffered from small sample size issues still. To solve SSS problems, many methods including regularized linear discriminant analysis were proposed. In this article, a way was presented by optimized regularized linear discriminant analysis for feature extraction in FR which can not only fix the singularity problem existing in scatter matrix but also the problem of parameter estimation. The experiment is conducted on several databases and promising results are obtained compared to some state-of-the-art methods to demonstrate the effectiveness of the proposed approach.

Online classification of imagined speech using functional near-infrared spectroscopy signals

Objective. Most brain–computer interfaces (BCIs) based on functional near-infrared spectroscopy (fNIRS) require that users perform mental tasks such as motor imagery, mental arithmetic, or music imagery to convey a message or to answer simple yes or no questions. These cognitive tasks usually have no direct association with the communicative intent, which makes them difficult for users to perform. Approach. In this paper, a 3-class intuitive BCI is presented which enables users to directly answer yes or no questions by covertly rehearsing the word ‘yes’ or ‘no’ for 15 s. The BCI also admits an equivalent duration of unconstrained rest which constitutes the third discernable task. Twelve participants each completed one offline block and six online blocks over the course of two sessions. The mean value of the change in oxygenated hemoglobin concentration during a trial was calculated for each channel and used to train a regularized linear discriminant analysis (RLDA) classifier. Main results. By the final online block, nine out of 12 participants were performing above chance (p < 0.001 using the binomial cumulative distribution), with a 3-class accuracy of 83.8% ± 9.4%. Even when considering all participants, the average online 3-class accuracy over the last three blocks was 64.1 % ± 20.6%, with only three participants scoring below chance (p < 0.001). For most participants, channels in the left temporal and temporoparietal cortex provided the most discriminative information. Significance. To our knowledge, this is the first report of an online 3-class imagined speech BCI. Our findings suggest that imagined speech can be used as a reliable activation task for selected users for development of more intuitive BCIs for communication.