脑电(EEG)有监督建模:分类预测识别解码回归、传统机器学习、深度学习、特征提取、跨受试者建模、泛化、可解释性、数据泄漏

… Validating a model’s ability to solve the cross-subject … to the cross-subject generalization problem in EEG decoding … developed in the broader machine learning community. The most …

… This study introduces a hierarchical statistical validation … capabilities of machine learning (ML) and deep learning (DL) … for the fusion of multi-channel EEG data were evaluated on the …

Cross-subject variability problems hinder practical usages of Brain-Computer Interfaces. Recently, deep learning has been introduced into the BCI community due to its better generalization and feature representation abilities. However, most studies currently only have validated deep learning models for single datasets, and the generalization ability for other datasets still needs to be further verified. In this paper, we validated deep learning models for eight MI datasets and demonstrated that the cross-dataset variability problem weakened the generalization ability of models. To alleviate the impact of cross-dataset variability, we proposed an online pre-alignment strategy for aligning the EEG distributions of different subjects before training and inference processes. The results of this study show that deep learning models with online pre-alignment strategies could significantly improve the generalization ability across datasets without any additional calibration data.

Cognitive analytics using electroencephalography (EEG) can have a lot of potential in terms of intelligent learning systems, but assessment methods tend to ignore the fact that one subject may leak information to another, and thus, overestimate performance. This work explores the cognitive prediction of cognitive understanding through EEG approaches based on 68,831 samples that were recorded on eight subjects and eleven educational videos. It compared five baseline classifiers, which included Logistic Regression, Random Forest, Support Vector Machine, XGBoost, and LightGBM, on three validation strategies and they were on random row split, cross-subject GroupKFold and within-subject stratified modeling validation strategies. Near perfect results were obtained with random row splitting (LightGBM ROC-AUC = 1.0000), which is leakage-based upper-bound performance. Performance was much lower under subject-disjoint validation (optimal ROC-AUC = 0.6628 ± 0.1176 by XGBoost), demonstrating a poor cross-subject extrapolation. Conversely, within-subject modeling had close to perfect discrimination (ROC-AUC 0.999-1.000) among eligible subjects. The discussed leakage-conscious assessment system measures the generalization disjunctions and proves that custom modeling offers a more effective deployment planning. This work provides a methodological validation procedure of neuro-educational information systems and provides a practical implementation advice of EEG-based adaptive learning systems.

Highlights What are the main findings? The proposed Short-Term Second-Order Differential Entropy (ST-SODE) can capture fatigue from short-term band-power dynamics. ST-SODE improves the robustness of cross-domain EEG fatigue detection. What are the implications of the main findings? ST-SODE reduces calibration burden for real-world fatigue monitoring. ST-SODE enables lightweight cross-subject deployment. Abstract Background/Objectives: Electroencephalography (EEG) is a promising modality for fatigue detection because it directly reflects neural states; however, it is hindered by the need for subject-specific calibration and its reliance on unstable labeling. Moreover, classical EEG features are sensitive to intrinsic brain rhythm variations, causing pronounced domain shifts that degrade performance across sessions and subjects. Methods: Motivated by the biological fatigue rebound mechanism, we propose a robust cross-subject metric which we name Short-Term Second-Order Differential Entropy (ST-SODE). ST-SODE effectively suppresses the interference of background brain rhythms, enhancing robustness to cross-domain drift; consequently, its one-dimensional output can provide an indication of fatigue states without additional model training. Results: ST-SODE is validated on the public driving fatigue regression dataset SEED-VIG and on a private Vigilance classification dataset based on the N-Back task. ST-SODE achieves a correlation coefficient of 0.56 on SEED-VIG dataset (vs. 0.4 for differential entropy, DE) and a binary classification accuracy of 93.75% on the Vigilance dataset, outperforming other EEG-based fatigue metrics. Conclusions: ST-SODE offers a reliable solution for deployment in fields such as driving, manufacturing, and healthcare, where it could reduce safety incidents caused by fatigue.

Electroencephalography (EEG) signals are complex subject-dependent quantities, which leads to a continuing issue of cross-subject generalisation in the brain-computer interface (BCI) study. The purpose of the proposed research study is to construct a Cross-Subject EEG Signal Analysis Framework using the Self-Representation Learning (SRL) to further optimise the extraction of powerful, discriminatory, and transferable EEG features. The quantitative research methodology was employed through simulation-based EEG data of Python (Google Colab) convolutional and self-representation learning architecture to be used in a modified X-EEGNet-CNN architecture. The model was tested with Leave-One-Subject-Out (LOSO) cross-validation, and the perfect mean performance of the model was 0.99 with an accuracy score of 0.99, macro-F1 of 1.0, Cohen's kappa of 0.98 and an AUC score of 0.99 using ten subjects crossing subjects, and this is demonstrated in two plots of per-subject accuracy and macro-F1. The findings affirm the fact that SRL is useful in improving representational consistency since it models inter-sample dependencies in latent spaces. The paper has found that SRL, when combined with CNN architectures, is effective in establishing higher EEG generalisation in individuals. It is suggested that the framework should be broadened in future studies to realistic EEG records to check the applicability in the real world. The utility implication is that it would shorten calibration and enhance the transferability of clinical and neuroadaptive systems based on the EEG. The paper is restricted by the fact that simulated data has been used; this might not be a perfect method of modelling the variability of biological signals.

Traditional taste evaluation relies heavily on manual sensory analysis, which is highly subjective and inefficient with poor cross-individual generalization, limiting its application in industrial flavor detection. To achieve accurate cross-subject taste recognition, this paper proposes an electroencephalogram (EEG) classification method based on a meta-learning wavelet graph convolutional neural network (ML-WGCNet) under sweet- and bitter-taste stimuli. Sucrose (sweetness) and quinine (bitterness) were used as stimulation sources, each prepared at six concentration gradients, including a water control. EEG signals were detected from 20 subjects. First, the Morlet wavelet transform was applied to decompose the EEG signals in the time–frequency domain, extracting the maximum and average energy values from five frequency bands as core features. A graph structure was then constructed using electrodes as nodes and Pearson correlation coefficients between electrodes as edge weights. A lightweight graph convolutional neural network (GCN) is employed to model spatial correlations among brain regions. Finally, by integrating a meta-learning framework and adopting leave-one-subject-out cross-validation, the model can rapidly adapt to new subjects. The experimental results show that the proposed method achieves average accuracies of 76.03% and 77.01% in cross-subject classification of sweet and bitter tastes, respectively. The corresponding precision values are 79.94% and 79.53%, the recall values are 75.77% and 78.51%, and the F1-scores are 78.24% and 78.08%, respectively, demonstrating that the proposed model significantly outperforms existing mainstream EEG classification methods.

… The feature embeddings are used with conventional machine learning classifiers for binary classification of mental workload in two different sessions of same set of tasks. Experimental …

Electroencephalogram (EEG) signals are very weak and have low spatial resolution, which has led to less satisfactory accuracy in cross-subject EEG-based emotion classification studies. Microstate analyses of EEG sources can be performed to determine the important spatiotemporal characteristics of EEG signals. Such analyses can be used to cluster rapidly changing EEG signals into multiple brain prototype topographies, fully utilizing the spatial information contained in the EEG signals and providing a neural representation for emotional dynamics. To better utilize the spatial information of brain signals, source localization analysis on the EEG signals was first conducted. Then, a microstate analysis on the source-reconstructed EEG signals is conducted to extract the microstate features of the data. We conducted source microstate analysis on the participant data from the odor-video physiological signal database (OVPD-II) dataset. The experimental results show that the source microstate feature topologies of different participants under the same emotion exhibited a high degree of correlation, which was proven by the analysis of microstate feature topographic maps and the comparison of two-dimensional feature visualization maps of the differential entropy (DE) and power spectral density (PSD). The microstate features represent more abstract emotional information and are more robust. The extracted microstate features were then used with the style transfer mapping method to transfer the feature data from the source domain to the target domain and were then used in support vector machines (SVMs) and convolutional neural networks (CNNs) for emotion recognition. The experimental results show that the cross-subject classification accuracies of the microstate features in SVMs were 84.90 ± 8.24% and 87.43 ± 7.54%, which were 7.19 and 6.95% higher than those obtained with the PSD and 0.51 and 1.79% higher than those obtained with the DE features. In CNN, the average cross-subject classification accuracies of the microstate features were 86.44 and 91.49%, which were 7.71 and 19.41% higher than those obtained with the PSD and 2.7 and 11.76% higher than those obtained with the DE features.

… training and testing data, severely restricting the generalization capability of traditional supervised learning … that of most existing methods, and validated that the proposed method holds …

Electroencephalography (EEG) provides a low-cost and non-invasive way to monitor changes in mental state in real time. This paper studies whether simple spectral features can reliably separate relaxed and focused states across unseen individuals. Using the PhysioNet EEG Motor Movement/Imagery dataset, resting runs were labeled as relaxed and motor execution runs were labeled as focused [1], [2]. After standard filtering and re-referencing, windows of EEG were converted into power spectral density features using Welch's method, then summarized into band powers (delta through gamma) across channels. Two classical machine learning models were compared: a radial basis function support vector machine and a random forest classifier [3], [4]. Because relaxed segments were less frequent, Synthetic Minority Oversampling Technique was evaluated to reduce majority-class bias [5]. Performance was measured both with a pooled train-test split and with leave-one-subject-out cross-validation, which better reflects deployment on new users. Both models performed above chance, but the random forest produced more balanced detection of the relaxed class while maintaining high overall accuracy. Results were consistent with established physiology, showing stronger occipital alpha activity during relaxation and stronger frontal/frontoparietal beta activity during focused tasks. Overall, the study shows that a straightforward, interpretable pipeline can achieve strong subject-independent mental-state classification while highlighting the importance of class-imbalance handling and transparent, physiologically grounded interpretation.

Cross-subject emotion recognition based on electroencephalogram (EEG) signals faces significant challenges, mainly because EEG data are highly non-stationary and easily influenced by time, environment, and individual physiological states. Meanwhile, substantial inter-subject variability leads to obvious differences in signal patterns across different people, which makes it difficult for a single model to learn stable and transferable emotional features. As a result, these factors severely hinder model generalization and reduce recognition performance in real-world applications. Unlike previous reviews that categorize methods based on network architectures, this paper proposes a novel taxonomy grounded in the “generalization hypothesis,” synthesizing existing approaches into five major paradigms: statistical and adversarial distribution alignment, topological and structural modeling, advanced representation learning, generative modeling and style reconstruction, and multimodal complementary fusion. Our analysis reveals that the core conflict lies in the trade-off between alignment intensity and semantic integrity. Future research should integrate causal representation learning with source-free domain adaptation to realize truly plug-and-play affective brain–computer interfaces (aBCIs).

… the proposed method obtains better performance on various machine learning methods. … calibrated EEG is insufficient. To address these issues, we proposed a novel cross-subject …

… -time physiological assessment when coupled with machine learning methods. These findings reinforce … [6] highlighted the benefit of integrating ECG, GSR, and EEG signals to capture …

Electroencephalography (EEG) is a non-invasive technique used to record the brain’s evoked and induced electrical activity from the scalp. Artificial intelligence, particularly machine learning (ML) and deep learning (DL) algorithms, are increasingly being applied to EEG data for pattern analysis, group membership classification, and brain-computer interface purposes. This study aimed to systematically review recent advances in ML and DL supervised models for decoding and classifying EEG signals. Moreover, this article provides a comprehensive review of the state-of-the-art techniques used for EEG signal preprocessing and feature extraction. To this end, several academic databases were searched to explore relevant studies from the year 2000 to the present. Our results showed that the application of ML and DL in both mental workload and motor imagery tasks has received substantial attention in recent years. A total of 75% of DL studies applied convolutional neural networks with various learning algorithms, and 36% of ML studies achieved competitive accuracy by using a support vector machine algorithm. Wavelet transform was found to be the most common feature extraction method used for all types of tasks. We further examined the specific feature extraction methods and end classifier recommendations discovered in this systematic review.

Machine-learning (ML) decoding methods have become a valuable tool for analyzing information represented in electroencephalogram (EEG) data. However, a systematic quantitative comparison of the performance of major ML classifiers for the decoding of EEG data in neuroscience studies of cognition is lacking. Using EEG data from two visual word-priming experiments examining well-established N400 effects of prediction and semantic relatedness, we compared the performance of three major ML classifiers that each use different algorithms: support vector machine (SVM), linear discriminant analysis (LDA), and random forest (RF). We separately assessed the performance of each classifier in each experiment using EEG data averaged over cross-validation blocks and using single-trial EEG data by comparing them with analyses of raw decoding accuracy, effect size, and feature importance weights. The results of these analyses demonstrated that SVM outperformed the other ML methods on all measures and in both experiments.

Advances in intracranial electroencephalography (iEEG) and neurophysiology have enabled the study of previously inaccessible brain regions with high fidelity temporal and spatial resolution. Studies of iEEG have revealed a rich neural code subserving healthy brain function and which fails in disease states. Machine learning (ML), a form of artificial intelligence, is a modern tool that may be able to better decode complex neural signals and enhance interpretation of these data. To date, a number of publications have applied ML to iEEG, but clinician awareness of these techniques and their relevance to neurosurgery, has been limited. The present work presents a review of existing applications of ML techniques in iEEG data, discusses the relative merits and limitations of the various approaches, and examines potential avenues for clinical translation in neurosurgery. One-hundred-seven articles examining artificial intelligence applications to iEEG were identified from 3 databases. Clinical applications of ML from these articles were categorized into 4 domains: i) seizure analysis, ii) motor tasks, iii) cognitive assessment, and iv) sleep staging. The review revealed that supervised algorithms were most commonly used across studies and often leveraged publicly available timeseries datasets. We conclude with recommendations for future work and potential clinical applications.

Electroencephalography (EEG) signal analysis is a fast, inexpensive, and accessible technique to detect the early stages of dementia, such as Mild Cognitive Impairment (MCI) and Alzheimer’s disease (AD). In the last years, EEG signal analysis has become an important topic of research to extract suitable biomarkers to determine the subject’s cognitive impairment. In this work, we propose a novel simple and efficient method able to extract features with a finite response filter (FIR) in the double time domain in order to discriminate among patients affected by AD, MCI, and healthy controls (HC). Notably, we compute the power intensity for each high- and low-frequency band, using their absolute differences to distinguish among the three classes of subjects by means of different supervised machine learning methods. We use EEG recordings from a cohort of 105 subjects (48 AD, 37 MCI, and 20 HC) referred for dementia to the IRCCS Centro Neurolesi “Bonino-Pulejo” of Messina, Italy. The findings show that this method reaches 97%, 95%, and 83% accuracy when considering binary classifications (HC vs. AD, HC vs. MCI, and MCI vs. AD) and an accuracy of 75% when dealing with the three classes (HC vs. AD vs. MCI). These results improve upon those obtained in previous studies and demonstrate the validity of our approach. Finally, the efficiency of the proposed method might allow its future development on embedded devices for low-cost real-time diagnosis.

Amyotrophic lateral sclerosis (ALS) is a neurodegenerative disease that affects the nerve cells in the brain and spinal cord. This condition leads to the loss of motor skills and, in many cases, the inability to speak. Decoding spoken words from electroencephalography (EEG) signals emerges as an essential tool to enhance the quality of life for these patients. This study compares two classification techniques: (1) the extraction of spectral power features across various frequency bands combined with support vector machines (PSD + SVM) and (2) EEGNet, a convolutional neural network specifically designed for EEG-based brain–computer interfaces. An EEG dataset was acquired from 32 electrodes in 28 healthy participants pronouncing five words in Spanish. Average accuracy rates of 91.04 ± 5.82% for Attention vs. Pronunciation, 73.91 ± 10.04% for Short words vs. Long words, 81.23 ± 10.47% for Word vs. Word, and 54.87 ± 14.51% in the multiclass scenario (All words) were achieved. EEGNet outperformed the PSD + SVM method in three of the four classification scenarios. These findings demonstrate the potential of EEGNet for decoding words from EEG signals, laying the groundwork for future research in ALS patients using non-invasive methods.

Emotional decoding and automatic identification of major depressive disorder (MDD) are helpful for the timely diagnosis of the disease. Electroencephalography (EEG) is sensitive to changes in the functional state of the human brain, showing its potential to help doctors diagnose MDD. In this paper, an approach for identifying MDD by fusing interhemispheric asymmetry and cross-correlation with EEG signals is proposed and tested on 32 subjects [16 patients with MDD and 16 healthy controls (HCs)]. First, the structural features and connectivity features of the θ-, α-, and β-frequency bands are extracted on the preprocessed and segmented EEG signals. Second, the structural feature matrix of the θ-, α-, and β-frequency bands are added to and subtracted from the connectivity feature matrix to obtain mixed features. Finally, the structural features, connectivity features, and the mixed features are fed to three classifiers to select suitable features for the classification, and it is found that our mode achieves the best classification results using the mixed features. The results are also compared with those from some state-of-the-art methods, and we achieved an accuracy of 94.13%, a sensitivity of 95.74%, a specificity of 93.52%, and an F1-score (f1) of 95.62% on the data from Beijing Anding Hospital, Capital Medical University. The study could be generalized to develop a system that may be helpful in clinical purposes.

In recent years, deep learning algorithms have been developed rapidly, and they are becoming a powerful tool in biomedical engineering. Especially, there has been an increasing focus on the use of deep learning algorithms for decoding physiological or pathological status of the brain from electroencephalographic (EEG). This paper overviews current application of deep learning algorithms in various EEG decoding tasks, and introduces commonly used algorithms, typical application scenarios, important progresses and existing problems. Firstly, the basic principles of deep learning algorithms used in EEG decoding is briefly described, including convolutional neural network, deep belief network, auto-encoder and recurrent neural network. In this paper, existing applications of deep learning on EEG is discussed, including brain-computer interfaces, cognitive neuroscience and diagnosis of brain disorders. Finally, this paper outlines some key problems that will be addressed in future applications of deep learning for EEG decoding, such as parameter selection, computational complexity, and the capability of generalization.

Autism Spectrum Disorder is a neurological and developmental disorder that starts early in adolescence and lasts throughout a person's life affecting information flow in the brain leading to secondary problems for the patient [1],[2]. Current diagnostic approaches for autism are time-consuming and are mainly based on clinical interviews, to accelerate this process of diagnosing the disease as early as possible with fewer efforts and better accuracy machine learning methods have been proposed recently [3],[4]. Early detection of ASD is vital in enhancing the efficiency of the treatment [5]. The motivation behind this study is the absence of well-defined automated diagnostic procedures for ASD. The objective of this study is to explore and analyze the techniques for EEG pre-processing, feature extraction, classification and identify the abnormal activity for the diagnosis of ASD based on the power spectral density of EEG signals applying machine learning models.

Deep learning with convolutional neural networks (deep ConvNets) has revolutionized computer vision through end-to-end learning, that is, learning from the raw data. There is increasing interest in using deep ConvNets for end-to-end EEG analysis, but a better understanding of how to design and train ConvNets for end-to-end EEG decoding and how to visualize the informative EEG features the ConvNets learn is still needed. Here, we studied deep ConvNets with a range of different architectures, designed for decoding imagined or executed tasks from raw EEG. Our results show that recent advances from the machine learning field, including batch normalization and exponential linear units, together with a cropped training strategy, boosted the deep ConvNets decoding performance, reaching at least as good performance as the widely used filter bank common spatial patterns (FBCSP) algorithm (mean decoding accuracies 82.1% FBCSP, 84.0% deep ConvNets). While FBCSP is designed to use spectral power modulations, the features used by ConvNets are not fixed a priori. Our novel methods for visualizing the learned features demonstrated that ConvNets indeed learned to use spectral power modulations in the alpha, beta, and high gamma frequencies, and proved useful for spatially mapping the learned features by revealing the topography of the causal contributions of features in different frequency bands to the decoding decision. Our study thus shows how to design and train ConvNets to decode task-related information from the raw EEG without handcrafted features and highlights the potential of deep ConvNets combined with advanced visualization techniques for EEG-based brain mapping. Hum Brain Mapp 38:5391-5420, 2017. © 2017 Wiley Periodicals, Inc.

Background: Brain traumas, mental disorders, and vocal abuse can result in permanent or temporary speech impairment, significantly impairing one’s quality of life and occasionally resulting in social isolation. Brain–computer interfaces (BCI) can support people who have issues with their speech or who have been paralyzed to communicate with their surroundings via brain signals. Therefore, EEG signal-based BCI has received significant attention in the last two decades for multiple reasons: (i) clinical research has capitulated detailed knowledge of EEG signals, (ii) inexpensive EEG devices, and (iii) its application in medical and social fields. Objective: This study explores the existing literature and summarizes EEG data acquisition, feature extraction, and artificial intelligence (AI) techniques for decoding speech from brain signals. Method: We followed the PRISMA-ScR guidelines to conduct this scoping review. We searched six electronic databases: PubMed, IEEE Xplore, the ACM Digital Library, Scopus, arXiv, and Google Scholar. We carefully selected search terms based on target intervention (i.e., imagined speech and AI) and target data (EEG signals), and some of the search terms were derived from previous reviews. The study selection process was carried out in three phases: study identification, study selection, and data extraction. Two reviewers independently carried out study selection and data extraction. A narrative approach was adopted to synthesize the extracted data. Results: A total of 263 studies were evaluated; however, 34 met the eligibility criteria for inclusion in this review. We found 64-electrode EEG signal devices to be the most widely used in the included studies. The most common signal normalization and feature extractions in the included studies were the bandpass filter and wavelet-based feature extraction. We categorized the studies based on AI techniques, such as machine learning and deep learning. The most prominent ML algorithm was a support vector machine, and the DL algorithm was a convolutional neural network. Conclusions: EEG signal-based BCI is a viable technology that can enable people with severe or temporal voice impairment to communicate to the world directly from their brain. However, the development of BCI technology is still in its infancy.

Electroencephalography (EEG)-based motor imagery (MI) has potential applications in diverse fields including rehabilitation, drone control, and virtual reality. However, its practical use is hindered by low generalization performance in decoding brain signals, primarily due to the subject-dependency of EEG signals. Although multitask autoencoder (MTAE) techniques have recently been used to mitigate this issue, these approaches encounter an imbalance problem between loss functions with different objectives, particularly between reconstruction loss and cross-entropy. To address this, we propose a novel two-phase multitask autoencoder (2PMTAE) framework that not only rectifies the imbalance issue but also ensures stable training of the MTAE. Our framework comprises two phases: first, the generation of a class-specific target signal, and second, the calculation of the reconstruction loss based on the generated target signals, effectively aligning the objectives of the two loss functions. In subject-independent experiments, our proposed method significantly outperformed state-of-the-art techniques, achieving accuracies of 71.68% and 75.78% on the BCI competition IV-2a and OpenBMI datasets, respectively. We also show that 2PMTAE is a generic framework for MI applications that can accept any encoder the practitioner wishes to employ. These results highlight the efficacy of our approach in enhancing the generalization performance of MI-EEG decoding.

Deep learning has significantly enhanced the research on the emerging issue of Electroencephalogram (EEG)-based visual classification and reconstruction, which has gained a growth of attention and concern recently. To promote the research progress, at this critical moment, a review work on the deep learning methodology for the issue becomes necessary and important. However, such a work seems absent in the literature. This paper provides the first review on EEG-based visual classification and reconstruction, whose contents can be categorized into the following four main parts: 1) comprehensively summarizing and systematically analyzing the representative deep learning methods from both feature encoding and decoding perspectives; 2) introducing the available benchmark datasets, describing the experimental paradigms, and displaying the method performances; 3) proposing the methodological essences and neuroscientific insights as well as the dynamic closed-loop interaction and promotion between them, which are potentially beneficial for technological innovations and academic progress; 4) discussing the potential challenges of current research and the prospective opportunities in future trends. We expect that this work can shed light on the technological directions and also enlighten the academic breakthroughs for the issue in the not-so-far future.

Decoding of continuous speech from electroencephalography (EEG) presents a promising avenue for understanding neural mechanisms of auditory processing and developing applications in hearing diagnostics. Recent advances in deep learning have improved decoding accuracy. However, challenges remain due to the low signal-to-noise ratio of the recorded brain signals. This study explores the application of contrastive learning, a self-supervised learning technique, to learn robust latent representations of EEG signals. We introduce a novel model architecture that leverages contrastive learning and transformer networks to capture relationships between auditory stimuli and EEG responses. Our model is evaluated on two tasks from the ICASSP 2023 Auditory EEG Decoding Challenge: a binary stimulus classification task (match-mismatch) and stimulus envelope decoding. We achieve state-of-the-art performance on both tasks, significantly outperforming previous winners with 87% accuracy in match-mismatch classification and a 0.176 Pearson correlation in envelope regression. Furthermore, we investigate the impact of model architecture, training set size, and finetuning on decoding performance, providing insights into the factors influencing model generalizability and accuracy. Our findings underscore the potential of contrastive learning for advancing the field of auditory EEG decoding and its potential applications in clinical settings.

To decode the pilot’s behavioral awareness, an experiment is designed to use an aircraft simulator obtaining the pilot’s physiological behavior data. Existing pilot behavior studies such as behavior modeling methods based on domain experts and behavior modeling methods based on knowledge discovery do not proceed from the characteristics of the pilots themselves. The experiment starts directly from the multimodal physiological characteristics to explore pilots’ behavior. Electroencephalography, electrocardiogram, and eye movement were recorded simultaneously. Extracted multimodal features of ground missions, air missions, and cruise mission were trained to generate support vector machine behavior model based on supervised learning. The results showed that different behaviors affects different multiple rhythm features, which are power spectra of the [Formula: see text] waves of EEG, standard deviation of normal to normal, root mean square of standard deviation and average gaze duration. The different physiological characteristics of the pilots could also be distinguished using an SVM model. Therefore, the multimodal physiological data can contribute to future research on the behavior activities of pilots. The result can be used to design and improve pilot training programs and automation interfaces.

Despite numerous successful applications of Deep Learning (DL) to large-scale image, video, speech and text data, they remain relatively unexplored in brain imaging field. In this paper, we make an overview of recent DL architectures for recognizing cognitive brain activities from Electroencephalogram (EEG) data with particular emphasis on Brain Computer Interface(BCI) technologies and Affective Neurocomputing. We discuss the use of convolutional, recurrent neural nets, as well as deep belief networks, echo-state networks, reservoir computing, and denoising auto encoder models. A major challenge in modeling brain cognitive activity from EEG data is finding representations that are invariant to inter- and intra-subject differences, as well as the inherent noise in the EEG recordings. The reviewed studies reveal the great potential of DL to decode human intentions in BCI applications and to find the invariant descriptors of human emotions across subjects in Affective Neurocomputing applications. Many of the DL models prove to be more accurate and efficient than traditional machine learning models.

… This study explores the use of machine learning algorithms and EEG data to predict … Machine learning methods for EEG data are categorized into supervised and unsupervised …

Various convolutional neural network (CNN)-based approaches have been recently proposed to improve the performance of motor imagery based-brain-computer interfaces (BCIs). However, the classification accuracy of CNNs is compromised when target data are distorted. Specifically for motor imagery electroencephalogram (EEG), the measured signals, even from the same person, are not consistent and can be significantly distorted. To overcome these limitations, we propose to apply a capsule network (CapsNet) for learning various properties of EEG signals, thereby achieving better and more robust performance than previous CNN methods. The proposed CapsNet-based framework classifies the two-class motor imagery, namely right-hand and left-hand movements. The motor imagery EEG signals are first transformed into 2D images using the short-time Fourier transform (STFT) algorithm and then used for training and testing the capsule network. The performance of the proposed framework was evaluated on the BCI competition IV 2b dataset. The proposed framework outperformed state-of-the-art CNN-based methods and various conventional machine learning approaches. The experimental results demonstrate the feasibility of the proposed approach for classification of motor imagery EEG signals.

Deep learning holds considerable promise for electroencephalography (EEG) analysis but faces challenges due to scarce and noisy EEG data, and the limited generality of existing data augmentation techniques. To address these issues, we propose an end-to-end EEG augmentation framework with an adaptive mechanism. This approach utilizes contrastive learning to mitigate representational distortions caused by augmentation, thereby strengthening the encoder’s feature learning. A selective augmentation strategy is further incorporated to dynamically determine optimal augmentation combinations based on performance. We also introduce NeuroBrain, a novel neural architecture specifically designed for auditory EEG decoding. It effectively captures both local and global dependencies within EEG signals. Comprehensive evaluations on the SparrKULee and WithMe datasets confirm the superiority of our proposed framework and architecture, demonstrating a 29.42% performance gain over HappyQuokka and a 5.45% accuracy improvement compared to EEGNet. These results validate our method’s efficacy in tackling key challenges in EEG analysis and advancing the state of the art.

While deep learning techniques are nowadays a powerful field to automatically learn and perform accurate EEG data classification in the clinical context, they still lack wide acceptance within the medical and health research community. This lack of trust is associated with the high complexity deep learning algorithms typically have, which contributes to a low level of interpretability of the outcome and predictions. This survey aims to provide a comprehensive discussion of the latest advancements in deep learning models applied to EEG analysis, while also emphasizing research in explainable AI within this domain. It explores commonly used algorithms in EEG analysis, their main application areas, and the insights provided by XAI. Moreover, the survey addresses current limitations, such as the evaluation of XAI methods and the need for clinical validation, as well as ongoing challenges in this field. One critical insight from this review is the relative paucity of clinical evaluations of the rich stock of proposed techniques and methods for AI explainability. By showcasing various applications and breakthroughs in EEG analysis facilitated by XAI, the survey underscores the potential of these technologies to revolutionize neurological diagnosis and treatment, paving the way for wider acceptance and implementation in clinical settings.

While deep learning has drastically improved the performance of electroencephalography (EEG) analysis, it remains unclear what these models, such as EEGNet, learn from the data and how their learned features relate to neuroscientific concepts. In this work, we introduce a comprehensive interpretability framework for deep learning models of neural data based on Concept Relevance Propagation (CRP), an extension of layer‐wise relevance propagation that enables the analysis of abstract concepts encoded by individual neurons and filters. We apply CRP to individual filters of convolutional neural networks (EEGNet) trained using leave‐one‐out cross‐validation. To identify common classification strategies across models, we guide the selection of representative data for individual filters using relevance maximization, reduce dimensionality via UMAP, and identify clusters of filters encoding similar concepts through density‐based clustering. To gain insight into the neural correlates of these tasks, we analyze the learned features across multiple data domains without requiring model retraining. We integrate a virtual inspection layer to project explanations into the frequency domain, enabling the simultaneous analysis of spatial, temporal, and spectral aspects using topographic maps, functional grouping, and independent component analysis (ICA). Using three EEG classification tasks—auditory attention, internal/external attention, and motor imagery—we demonstrate that our approach reveals interpretable, task‐relevant neural patterns that generalize across participants. Overall, this framework provides a step toward understanding the models itself and gaining insights into the tasks in terms of neuroscience.

Convolutional neural networks (CNNs) achieve high performance in electroencephalographic (EEG) classification tasks; however, their decision-making mechanisms remain difficult to interpret. Explainable artificial intelligence (XAI) methods are typically applied to provide insight into individual model decisions, yet such explanations do not reveal the overall structure of the patterns learned by the network. In this study, we hypothesize that, in EEG analysis, XAI can serve a deeper role: when appropriately applied, it can expose the general patterns learned by a trained CNN, thereby transforming it from a purely predictive model into a framework capable of revealing candidate discriminative structures that may form the basis for future neuroscientific hypotheses. This shift is enabled by structured aggregation of local explanations, through which instance-level insights are consolidated into cross-subject patterns. To implement this strategy, we employ averagedLIME, an extension of Local Interpretable Model-agnostic Explanations, which aggregates sample-level explanations into global class-level saliency maps. In trial-based EEG paradigms, such aggregation reinforces consistent patterns across samples analogously to event-related potentials averaging. In this study, we examine whether this strategy remains effective in a temporally unaligned resting-state setting, taking subject-independent classification of Alzheimer's disease and cognitively normal controls as a case study. Four neural architectures are compared, with spectral CNN yielding the most robust performance (95.81% test accuracy). The best-performing model is subsequently analyzed using averagedLIME, while SHapley Additive exPlanations (SHAP) and Grad-CAM are employed as complementary explanation techniques. Quantitative analyses demonstrated that the averagedLIME patterns were stable across cross-validation folds, robust to perturbation parameter selection, reproducible on unseen test data, and strongly supported by independent SHAP explanations. When interpreted in conjunction with the averaged input representations, the resulting saliency maps reproduced established EEG slowing patterns in Alzheimer's disease, while also highlighting localized spatial-spectral structures that were less apparent in the averaged EEG representations alone. These findings demonstrate that structured aggregation of CNN explanations enables extraction of stable cross-subject patterns from resting-state EEG and may reveal candidate discriminative structures that can motivate future neuroscientific hypotheses.

Recent advancements in XAI have radically changed the way that AI systems are evaluated, as transparency and trustworthiness are now valued as highly as performance. This is especially true in medical applications, as, in order for such tools to be used in practical applications, interpretability is a key requirement for clinical adoption. Electroencephalography (EEG) analysis, in particular, has seen a significant rise in research, as the difficult and complex nature of EEG signals benefits from these methods, enabling researchers and practitioners to gain new insights from the vast amount of data that is now available. This survey presents a comprehensive analysis of the latest trends and advancements in XAI for EEG analysis. First, we provide a brief overview of fundamental EEG tasks, available datasets, and AI model approaches used for analysis. Then, we classify XAI methods using well-established taxonomies in XAI research, such as locality and generalization of explanations. By exploring all relevant XAI techniques in EEG analysis, our study offers researchers a clear perspective on the current state of the field and identifies potential research gaps. Our review indicates that current XAI approaches for EEG often face limitations in robustness, consistency, and neuroscientific grounding. These findings highlight the need for more reliable and domain-informed explainability methods to support trustworthy EEG analysis in research and clinical practice.

Decoding reaching movements from non-invasive brain signals is a key challenge for the development of naturalistic brain–computer interfaces (BCIs). While this decoding problem has been addressed via traditional machine learning, the exploitation of deep learning is still limited. Here, we evaluate a convolutional neural network (CNN) for decoding movement direction during a delayed center-out reaching task from the EEG. Signals were collected from twenty healthy participants and analyzed using EEGNet to discriminate reaching endpoints in three scenarios: fine-direction (five endpoints), coarse-direction (three endpoints), and proximity (two endpoints) classifications. To interpret the decoding process, the CNN was coupled with explanation techniques, including DeepLIFT and occlusion tests, enabling a data-driven analysis of spatio-temporal EEG features. The proposed approach achieved accuracies well above chance, with accuracies of 0.45 (five endpoints), 0.64 (three endpoints) and 0.70 (two endpoints) on average across subjects. Explainability analyses revealed that directional information is predominantly encoded during movement preparation, particularly in parietal and parietal–occipital regions, consistent with known visuomotor planning mechanisms and with EEG analysis based on event-related spectral perturbations. These results demonstrate the feasibility and interpretability of CNN-based EEG decoding for reaching movements, providing insights relevant for both neuroscience and the prospective development of non-invasive BCIs.

Accurate classification among Alzheimer’s Disease (AD), Fronto Temporal Dementia (FTD), and Cognitively Normal (CN) adults from EEG remains challenging. We propose a multi-class classification method that fuses interpretable spectral/connectivity biomarkers (band power, spectral entropy, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}-coherence) with compact temporal embeddings from a customized lightweight one Dimensional Convolutional Neural Network (1D-CNN). The fused features are reduced by Principal Component Analysis (PCA) and classified with Support Vector Machine (SVM). All data-dependent steps like Synthetic Minority Over Sampling Technique (SMOTE), z-scoring and PCA are fit strictly on training folds to prevent leakage. Hyperparameters including PCA dimensionality and SMOTE neighbors were selected via inner-loop grid sweeps that maximized macro-F1; full grids and class distributions before/after SMOTE (inner-train only) are reported in the Supplement. On OpenNeuro ds004504 (eyes-closed; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N{=}88$$\end{document}; 36 AD/23 FTD/29 CN) the model achieved 94.5% accuracy, macro-F1 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$=0.95$$\end{document}, and AUC \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$=0.96$$\end{document}. Robustness was examined on two public datasets using the identical preprocessing/segmentation: (i) cross-condition testing on ds006036 (eyes-open recordings from the same participants) and (ii) zero-shot transfer to an independent OSF dataset with different instrumentation and demographics. SHapley Additive exPlanations (SHAP) analyses provided global, class-wise, and subject-level attributions aligned with known electrophysiology (e.g., reduced \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}-coherence, elevated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}), supporting clinical interpretability. These results indicate that a transparent hybrid feature design can deliver accurate, leakage-safe, and explainable EEG-based differentiation of AD, FTD, and CN with preliminary external checks.

Motor imagery (MI)-based brain–computer interfaces (BCIs) enable users to control external devices using EEG signals, offering great potential in assistive and rehabilitation technologies. However, MI recognition remains challenging due to EEG’s low signal-to-noise ratio (SNR), inter-subject variability, and complex spatiotemporal patterns. Existing approaches often suffer from limited accuracy, high computational cost, and poor interpretability. In response to these challenges, we present the first comprehensive benchmarking of the publicly available EEG-hand movement (EEG-HM) dataset. Our study aims to establish a standardized performance baseline, guide the selection of optimal models by jointly considering accuracy, prediction time, and explainability, and ultimately accelerate progress in MI-BCI development. We have proposed a two-stage optimization of machine learning models that employs both feature selection and hyperparameter tuning. We exploit five feature selection algorithms for selecting the best set of EEG electrodes and frequency bands, while Bayesian optimization is exploited for machine learning model optimization through hyperparameter tuning. Furthermore, to validate the neurophysiological basis of our model’s decisions, we leverage explainable AI (XAI) algorithms—LIME and SHAP—quantifying the contributions of specific EEG electrodes and frequency bands to interpret its decision-making process. Through extensive simulations, the proposed two-stage optimization of the machine learning model demonstrates a superior performance in terms of accuracy, precision, and recall. This method outperforms the existing methods by 21.47% in accuracy with competitive prediction time. Its performance is further evaluated on the PhysioNet MI dataset, achieving a 4.67% accuracy improvement over state-of-the-art methods. Through LIME and SHAP, we provide the local and global explanations for no activity, left-hand, and right-hand imagery movements. Additionally, we analyze how various EEG frequency bands and electrode locations interact during the performance of different motor imagery hand movements.

Depression is a major mental health disorder, and EEG‐based automated detection is emerging as a potential objective diagnostic tool. However, achieving both high accuracy and interpretability remains challenging due to the complex spatiotemporal structure of EEG signals. This study proposes an explainable deep learning framework for depression detection using resting‐state EEG data.

OBJECTIVE Detecting Alzheimer's disease (AD) at an early stage is essential for administering effective treatments and preventing neuronal damage. Unfortunately, current diagnostic techniques are often invasive and expensive. Our research focuses on creating a cost-effective and non-invasive method for the early detection of cognitive decline. METHODS Using a publicly available dataset of resting state electroencephalographic (EEG) data on healthy controls and patients with Mild Cognitive Impairment (MCI), two novel deep learning (DL) algorithms with self-attention mechanisms were developed and evaluated for their performance in predicting MCI and cognitive decline. RESULTS Both proposed DL algorithms outperformed a traditional convolutional neural network (CNN) model in predicting MCI, achieving test accuracy improvements of 8.5% and 10%, respectively, while utilizing significantly fewer trainable parameters. An ablation study highlighted the attention layer as a key feature, enhancing model accuracy by 8.5%. Analysis of the attention layers indicated that beta band frequencies (13-30 Hz) were essential for distinguishing MCI from control subjects, highlighting the role of high EEG frequencies in early cognitive deficits. Predicting pre-clinical cognitive decline in healthy subjects proved more challenging than predicting diagnosed MCI. However, using transfer-learning methods, we achieved a test accuracy of 56.08%. CONCLUSION Our models achieved state-of-the-art results in the MCI classification task, and demonstrated learning progress in predicting cognitive decline in the preclinical stage. As this is the first time DL models have been evaluated to classify healthy subjects based on cognitive scores, where brain changes are minimal and difficult to detect, this study opens new avenues for discovering biomarkers in early AD diagnosis and facilitating early interventions. Interpretation of the trained DL attention models provided valuable insights that aligned with the existing brain research, serving as a helpful tool for validating AI in healthcare applications.

Epilepsy is a condition that affects 50 million individuals globally, significantly impacting their quality of life. Epileptic seizures, a transient occurrence, are characterized by a spectrum of manifestations, including alterations in motor function and consciousness. These events impose restrictions on the daily lives of those affected, frequently resulting in social isolation and psychological distress. In response, numerous efforts have been directed towards the detection and prevention of epileptic seizures through EEG signal analysis, employing machine learning and deep learning methodologies. This study presents a methodology that reduces the number of features and channels required by simpler classifiers, leveraging Explainable Artificial Intelligence (XAI) for the detection of epileptic seizures. The proposed approach achieves performance metrics exceeding 95% in accuracy, precision, recall, and F1-score by utilizing merely six features and five channels in a temporal domain analysis, with a time window of 1 s. The model demonstrates robust generalization across the patient cohort included in the database, suggesting that feature reduction in simpler models—without resorting to deep learning—is adequate for seizure detection. The research underscores the potential for substantial reductions in the number of attributes and channels, advocating for the training of models with strategically selected electrodes, and thereby supporting the development of effective mobile applications for epileptic seizure detection.

… from EEG signals for accurate classification. The system preprocesses the data to optimize it for model training, and multiple deep learning models … EEG analysis with deep learning, it …

Biomedical signal analysis underpins modern healthcare by enabling accurate diagnosis, continuous physiological monitoring, and informed patient management. While deep learning excels at automated feature extraction and end-to-end modeling, classical ML remains essential for tasks requiring interpretability, data efficiency, and clinical transparency. This review synthesizes advances in ML methods including Support Vector Machines, Random Forests, and Decision Trees focusing on physiologically informed feature engineering, robust feature selection, and meaningful model interpretation. We provide guidelines for signal preprocessing, domain-specific feature extraction, and selection strategies across standard biomedical signals such as electrocardiograms (ECGs), electromyograms (EMGs), electroencephalograms (EEGs), Electrovestibulography (EVestG), and tracheal breathing sounds (TBSs). Reviewing TBS studies illustrates an end-to-end workflow highlighting common features and classifiers alongside practical challenges and solutions. Reported ML application performance ranges from 85 to 94% accuracy for EEG, ECG, and EMG, to 82% specificity for TBSs, emphasizing the trade-off between interpretability and predictive performance. Marginal accuracy gains alone do not constitute meaningful progress unless they enhance clinical insight, actionable decision-making, or model transparency. Finally, we compare ML with DL, discuss strengths and limitations, and provide recommendations and future directions for developing robust, interpretable, and clinically relevant biomedical ML.

Many state‐of‐the‐art methods for seizure prediction, using the electroencephalogram, are based on machine learning models that are black boxes, weakening the trust of clinicians in them for high‐risk decisions. Seizure prediction concerns a multidimensional time‐series problem that performs continuous sliding window analysis and classification. In this work, we make a critical review of which explanations increase trust in models' decisions for predicting seizures. We developed three machine learning methodologies to explore their explainability potential. These contain different levels of model transparency: a logistic regression, an ensemble of 15 support vector machines, and an ensemble of three convolutional neural networks. For each methodology, we evaluated quasi‐prospectively the performance in 40 patients (testing data comprised 2055 hours and 104 seizures). We selected patients with good and poor performance to explain the models' decisions. Then, with grounded theory, we evaluated how these explanations helped specialists (data scientists and clinicians working in epilepsy) to understand the obtained model dynamics. We obtained four lessons for better communication between data scientists and clinicians. We found that the goal of explainability is not to explain the system's decisions but to improve the system itself. Model transparency is not the most significant factor in explaining a model decision for seizure prediction. Even when using intuitive and state‐of‐the‐art features, it is hard to understand brain dynamics and their relationship with the developed models. We achieve an increase in understanding by developing, in parallel, several systems that explicitly deal with signal dynamics changes that help develop a complete problem formulation.

Epilepsy is a complex neurological disorder characterized by recurrent seizures, requiring accurate and timely diagnosis to ensure effective clinical management. This is particularly important in Brain-Computer Interfaces (BCIs), where neurological monitoring plays a key role in assistive and diagnostic applications. In this context, Electroencephalography (EEG) is the primary diagnostic modality; however, traditional manual analysis is labor-intensive, subjective, and prone to inter-expert variability. Automated machine learning techniques offer a promising alternative to manual EEG analysis. The opaque, “black-box” nature of many Machine Learning (ML) models often raise concerns about their reliability in clinical settings, which can hinder their practical implementation. To tackle this issue, Explainable Artificial Intelligence (XAI) has emerged as a transformative solution, providing greater transparency and shedding light on the critical features driving model predictions. In this study, we introduce Explainable LIMEbased Selection of EEG Electrodes (XLISEE), a novel approach that not only enhances EEG channel selection but also strengthens interpretability by leveraging XAI methodologies. The suggested approach is evaluated using the dataset provided by Guinea-Bissau Hospital. The most important EEG channels are identified by XLISEE to differentiate between epileptics and healthy subjects, reducing data complexity without sacrificing the precision of the diagnosis. With accuracy rates of 99.76% for both the Support Vector Machine (SVM) and k-Nearest Neighbor (kNN) classifiers, our experimental results show that the suggested method delivers remarkable classification performance. It also attains 100% accuracy, specificity, sensitivity, and F1-score, as well as a Matthews Correlation Coefficient (MCC) of 1.0. Moreover, lower computational demands improve efficiency, making real-time epilepsy detection more practical and feasible for clinical use. Beyond performance gains, integrating XAI enhances clinician confidence by offering interpretable, evidence-based insights, promoting wider adoption of AI-driven diagnostics.

Early and accurate prediction of neurological outcomes in comatose patients following cardiac arrest is critical for informed clinical decision-making. Existing studies have predominantly focused on EEG for assessing brain injury, with some exploring ECG data. However, the integration of EEG, ECG, and clinical features remains insufficiently investigated, and its potential to enhance predictive accuracy has not been fully established. Moreover, the limited interpretability of current models poses significant barriers to clinical application. Using the I-CARE database, we analyzed EEG, ECG, and clinical data from comatose cardiac arrest patients. After rigorous preprocessing and feature engineering, machine learning models (Logistic Regression, SVM, Random Forest, and Gradient Boosting) were developed. Performance was evaluated through AUC-ROC, accuracy, sensitivity, and specificity, with SHAP applied to interpret feature contributions. Our multi-modal model outperformed single-modality models, achieving AUC values from 0.75 to 1.0. Notably, the model’s accuracy peaked at a critical point within the 12–24 h window (e.g., 18 h, AUC = 1.0), surpassing EEG-only (AUC 0.7–0.8) and ECG-only (AUC < 0.6) models. SHAP identified Shockable Rhythm as the most influential feature (mean SHAP value 0.17), emphasizing its role in predictive accuracy. This study presents a novel multi-modal approach that significantly enhances early neurological outcome prediction in critical care. SHAP-based interpretability further supports clinical applicability, paving the way for more personalized patient management post-cardiac arrest.

… to other machine learning (ML) models using well-known Bonn and UCI-EEG benchmark datasets. Finally, SHapley additive exPlanation (SHAP) was used to interpret and explain the …

Deep learning approaches have been widely adopted in the medical image analysis field. However, a most of existing deep learning approaches focus on achieving promising performances such as classification, detection, and segmentation, and much less effort is devoted to the explanation of the designed models. Similarly, in the brain imaging field, many deep learning approaches have been designed and applied to characterize and predict human brain states. However, these models lack interpretation. In response, we propose a novel domain knowledge informed self-attention graph pooling-based (SAGPool) graph convolutional neural network to study human brain states. Specifically, the dense individualized and common connectivity-based cortical landmarks system (DICCCOL, structural brain connectivity profiles) and holistic atlases of functional networks and interactions system (HAFNI, functional brain connectivity profiles) are integrated with the SAGPool model to better characterize and interpret the brain states. Extensive experiments are designed and carried out on the large-scale human connectome project (HCP) Q1 and S1200 dataset. Promising brain state classification performances are observed (e.g., an average of 93.7% for seven-task classification and 100% for binary classification). In addition, the importance of the brain regions, which contributes most to the accurate classification, is successfully quantified and visualized. A thorough neuroscientific interpretation suggests that these extracted brain regions and their importance calculated from self-attention graph pooling layer offer substantial explainability.

Autism spectrum disorder (ASD) is a complex neurodevelopmental condition that remains challenging to diagnose due to its reliance on subjective clinical assessments. To address this limitation, an automated electroencephalography (EEG)-based ASD identification framework is proposed, aiming to enhance identification accuracy. Temporal dynamics of EEG signals analysed through three-phase feature selection approach involving independent samples t-test elimination, followed by a novel binary search-driven mutual information (BiS-MI) and recursive feature elimination (BiS-RFE). Six ensemble models trained on the selected features, and feature importance and classification predictions interpreted using explainable artificial intelligence (XAI). CatBoost achieved the highest performance, with an accuracy of 0.9923 and recall of 0.9864 using BiS-MI, and an accuracy of 0.9936 and recall of 0.9911 using BiS-RFE. SHAP analysis identified features from frontal and central EEG electrodes as the most significant contributors. These results highlight the potential for developing interpretable, non-invasive and improved diagnostic tools for ASD identification.

In order to improve the performance of binary motor imagery (MI) - based brain-computer interfaces (BCIs) using electroencephalography (EEG), a novel method (PSS-CSP) is proposed, which combines spectral subtraction with common spatial pattern. Spectral subtraction is an effective denoising method which is initially adopted to process MI-based EEG signals for binary BCIs in this work. On this basis, we proposed a novel feature extraction method called power spectral subtraction-based common spatial pattern (PSS-CSP) , which calculates the differences in power spectrum between binary classes of EEG signals and uses the differences in the feature extraction process. Additionally, support vector machine (SVM) algorithm is used for signal classification. Results show the proposed method (PSS-CSP) outperforms certain existing methods, achieving a classification accuracy of 76.8% on the BCIIV dataset 2b, and 76.25% and 77.38% on the OpenBMI dataset session 1 and session 2, respectively.

… of two classes of EEG related motor imagery [14]. In … feature extraction techniques on the EEG data using common spatial pattern based on entropy, energy and logarithmic band power. …

… Feature extraction of electroencephalogram (EEG) plays a vital role … features that are fitted for the specific subject. So a subject-based feature extraction method using fisher WPD-CSP is …

… (1) Can CSP-based methods identify AM pedaling using EEG … (2) What is an adequate time window for feature extraction … task by relating the EEG spectral power during pedaling with …

A brain-computer interface (BCI) acquires brain signals, extracts informative features, and translates these features to commands to control an external device. This paper investigates the application of a noninvasive electroencephalography (EEG)based BCI to identify brain signal features in regard to actual hand movement speed. This provides a more refined control for a BCI system in terms of movement parameters. An experiment was performed to collect EEG data from subjects while they performed right-hand movement at two different speeds, namely fast and slow, in four different directions. The informative features from the data were obtained using the Wavelet-Common Spatial Pattern (W-CSP) algorithm that provided high-temporal-spatial-spectral resolution. The applicability of these features to classify the two speeds and to reconstruct the speed profile was studied. The results for classifying speed across seven subjects yielded a mean accuracy of 83.71% using a Fisher Linear Discriminant (FLD) classifier. The speed components were reconstructed using multiple linear regression and significant correlation of 0.52 (Pearson's linear correlation coefficient) was obtained between recorded and reconstructed velocities on an average. The spatial patterns of the W-CSP features obtained showed activations in parietal and motor areas of the brain. The results achieved promises to provide a more refined control in BCI by including control of movement speed.

This paper presents a novel approach of spectral feature selection using spatial filters for the classification of four cognitive imagery tasks. The input dataset consists of electroencephalogram (EEG) signals acquired through a commercial wireless headset. The spectral features included mel frequency (MF) components extracted from the low frequency bands of EEG signal. A spatial projection filter was used for the selection of the most relevant features before classification. The popular method of multiclass common spatial pattern (CSP) and regularized CSP (RCSP) are investigated for a subject dependent (intra) and subject independent (inter) generation of spatial projection filter, respectively. Based upon this, present study used two different algorithmic approaches namely MF-CSP and MF-RCSP. The developed algorithm successfully classified four imagery actions with the reported prediction accuracy of 46.23% and 64.01% and standard deviation of 11.60% and 8.67% for MF-CSP and MF-RCSP, respectively.

The common spatial pattern (CSP) is a very effective feature extraction method in motor imagery based brain computer interface (BCI), but its performance depends on the selection of the optimal frequency band. Although a lot of research works have been proposed to improve CSP, most of these works have the problems of large computation costs and long feature extraction time. To this end, three new feature extraction methods based on CSP and a new feature selection method based on non-convex log regularization are proposed in this paper. Firstly, EEG signals are spatially filtered by CSP, and then three new feature extraction methods are proposed. We called them CSP-wavelet, CSP-WPD and CSP-FB, respectively. For CSP-Wavelet and CSP-WPD, the discrete wavelet transform (DWT) or wavelet packet decomposition (WPD) is used to decompose the spatially filtered signals, and then the energy and standard deviation of the wavelet coefficients are extracted as features. For CSP-FB, the spatially filtered signals are filtered into multiple bands by a filter bank (FB), and then the logarithm of variances of each band are extracted as features. Secondly, a sparse optimization method regularized with a non-convex log function is proposed for the feature selection, which we called LOG, and an optimization algorithm for LOG is given. Finally, ensemble learning is used for secondary feature selection and classification model construction. Combing feature extraction and feature selection methods, a total of three new EEG decoding methods are obtained, namely CSP-Wavelet+LOG, CSP-WPD+LOG, and CSP-FB+LOG. Four public motor imagery datasets are used to verify the performance of the proposed methods. Compared to existing methods, the proposed methods achieved the highest average classification accuracy of 88.86, 83.40, 81.53, and 80.83 in datasets 1–4, respectively. The feature extraction time of CSP-FB is the shortest. The experimental results show that the proposed methods can effectively improve the classification accuracy and reduce the feature extraction time. With comprehensive consideration of classification accuracy and feature extraction time, CSP-FB+LOG has the best performance and can be used for the real-time BCI system.

Brain-computer interface (BCI) provides a new way for people who are unable to communicate with each other. The traditional EEG signal feature extraction method based on frequency characteristics only extracts the energy features of each channel, but ignores the correlation information between different channels. In order to obtain better feature extraction results, the method of EEG signal feature extraction based on wavelet packet and Common Space Pattern (CSP) is adopted in this paper. Firstly, on the basis of analyzing channels and frequency bands closely related to event desynchronization, wavelet packet decomposition was carried out for EEG signals to extract the activity imagination EEG co-rhythms and beta rhythms. Spatial filtering was carried out to extract features through the CSP algorithm, and then the related nodes were selected to calculate the wavelet packet energy. Combining the advantages of wavelet packet and CSP method, the correlation information between different channels can be fully utilized, and the Support Vector Machine (SVM) can be used to classify the two kinds of EEG signals. Corresponding experiments were conducted on BCI competition data sets, the classify results show that the proposed feature extraction algorithm can extract useful features for motor imagery EEG signals and get high classify accuracy.

The effective decoding of motor imagination EEG signals depends on significant temporal, spatial, and frequency features. For example, the motor imagination of the single limbs is embodied in the μ (8–13 Hz) rhythm and β (13–30 Hz) rhythm in frequency features. However, the significant temporal features are not necessarily manifested in the whole motor imagination process. This paper proposes a Multi-Time and Frequency band Common Space Pattern (MTF-CSP)-based feature extraction and EEG decoding method. The MTF-CSP learns effective motor imagination features from a weak Electroencephalogram (EEG), extracts the most effective time and frequency features, and identifies the motor imagination patterns. Specifically, multiple sliding window signals are cropped from the original signals. The multi-frequency band Common Space Pattern (CSP) features extracted from each sliding window signal are fed into multiple Support Vector Machine (SVM) classifiers with the same parameters. The Effective Duration (ED) algorithm and the Average Score (AS) algorithm are proposed to identify the recognition results of multiple time windows. The proposed method is trained and evaluated on the EEG data of nine subjects in the 2008 BCI-2a competition dataset, including a train dataset and a test dataset collected in other sessions. As a result, the average cross-session recognition accuracy of 78.7% was obtained on nine subjects, with a sliding window length of 1 s, a step length of 0.4 s, and the six windows. Experimental results showed the proposed MTF-CSP outperforming the compared machine learning and CSP-based methods using the original signals or other features such as time-frequency picture features in terms of accuracy. Further, it is shown that the performance of the AS algorithm is significantly better than that of the Max Voting algorithm adopted in other studies.

… [1] introduced EEGNet, utilizing depthwise and separable … verify the performance and generalization ability of the … shift phenomenon often encountered in cross-subject EEG analysis, …

… demonstrates strong generalization capability in both within-subject and cross-subject … Owing to its compact and interpretable architecture, EEGNet enables practical deployment …

Motor imagery electroencephalogram (MI-EEG) analysis is essential for natural interaction and autonomous control in brain-computer interfaces (BCIs). However, deep learning models often struggle with inter-subject variability, which limits their ability to generalize across subjects. This study proposes RMETNet, a novel framework that integrates TSLANet, a spatio-temporal convolution module, and a multi-scale Riemannian geometry feature module. TSLANet suppresses noise and captures complex temporal patterns for preliminary signal decoding, while the spatio-temporal convolution module extracts higher-order representations. The Riemannian branch learns geometry-based distribution features across subjects, and the fused features are used for classification. To address inter-subject distribution shifts, RMETNet incorporates Maximum Mean Discrepancy (MMD) loss for domain adaptation, aligning feature distributions between source and target domains. Experiments show that on the four-class BCI Competition IV 2a (BCICIV2a) dataset, RMETNet achieved accuracies of 71.39% in the cross-subject setting and 80.71% in the subject-dependent setting; on the two-class BCI Competition IV 2b (BCICIV2b) dataset, it achieved 80.93% and 86.76%, respectively. The model consistently outperformed baseline algorithms. Ablation and visualization analyses further validated its effectiveness in reducing inter-subject feature distribution disparities and enhancing MI-EEG decoding. The code is available at: https://github.com/rokanfeermecer486/RMETNet.

Early and accurate detection of Mild Cognitive Impairment (MCI) is essential for preventing progression toward Alzheimer’s disease (AD). In this cross-subject study, we investigate the effectiveness of entropy- and graph-based EEG features for distinguishing MCI from healthy controls (HC), using two modeling approaches: (1) a Transformer network applied to the engineered feature set, and (2) an EEGNet model trained on the same feature representation for comparison. The dataset consists of resting-state, eyes-closed EEG recordings from 183 participants (127 HC, 56 MCI), collected using a 20-channel STAT™ X24 wireless system and segmented into 3-second epochs. EEG data underwent standard preprocessing, including band-pass filtering, downsampling, normalization, and class-balancing augmentation applied to the minority class. From each channel, nonlinear dynamical measures (e.g., sample and fuzzy entropy, Higuchi fractal dimension, Lyapunov exponent) and graph-theoretic connectivity descriptors derived from coherence matrices across five frequency bands were extracted, yielding a structured 19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}77 feature representation. The feature-based Transformer achieved the best performance (97.04% ± 0.72), outperforming the feature-based EEGNet baseline and highlighting the benefits of combining rich handcrafted features with attention-based modeling. SHAP (SHapley Additive exPlanations) analysis provided global and local interpretability, revealing the most influential nonlinear and connectivity features as well as the EEG channels contributing most to classification. Overall, these results demonstrate the effectiveness of feature-Transformer integration and support the potential of interpretable feature-driven deep learning models for early MCI detection.

Electroencephalography (EEG) has emerged as a promising non-invasive tool for the diagnosis of neurodegenerative disorders, and artificial intelligence (AI) has shown significant potential in this domain, as demonstrated by recent studies. However, strong inter-subject variability remains a major challenge, limiting the ability of AI-based models to learn disease-specific features that generalize across individuals, thereby hindering the development of clinically deployable subject-independent systems. In this work, we propose a cross-subject, AI-based EEG classification framework to distinguish between Alzheimer’s disease (AD), Creutzfeldt–Jakob disease (CJD), and healthy control subjects using clinical EEG data collected from a local hospital. A lightweight hybrid deep learning model is developed, combining a two-layer one-dimensional convolutional neural network with a two-layer Transformer encoder to capture both local temporal patterns and long-range dependencies in EEG signals. The proposed model achieves an average classification accuracy of 97%, representing a 3% improvement over a baseline model evaluated on a cohort of 36 subjects. To assess deployment feasibility in real-time clinical settings, the trained model is implemented and evaluated on an edge-AI platform (NVIDIA Jetson AGX Orin), demonstrating energy efficiency for the inference with a compact model footprint. These results indicate that the proposed approach provides an accurate, efficient, and practically deployable solution for subject-independent EEG-based classification of neurological disorders.

BACKGROUND Alzheimer's disease (AD) and mild cognitive impairment (MCI) are progressive neurodegenerative disorders with no effective treatments currently, underscoring the urgent need for early diagnosis. Electroencephalography and event-related potentials (ERP) provide noninvasive, cost-effective methods with high temporal resolution for detecting cognitive decline, while traditional Chinese medicine (TCM) features such as body constitutions have been identified as risk factors for MCI. Recent developments in artificial intelligence (AI) especially deep learning architectures have further improved the diagnostic accuracies of AD and MCI. This study aimed to assess the efficacy of deep learning models based on fused ERP and TCM features in the cross-subject classification of cognitive impairment. METHODS Visual oddball ERP tasks under Neutral, Happiness, or Sadness stimulus were conducted among 30 healthy controls (HC, 12 males and 18 females), 30 MCI (10 males and 20 females), and 30 AD (10 males and 20 females) patients. Deep learning models, including EEGNet, Convolutional Neural Network - Long Short-Term Memory, Graph Convolutional Network, (GCN), and multi-scale feature reconstruction GCN, were employed to extract differential entropy features from ERP data, and multilayer perceptron was utilized to extract features from TCM questionnaires. After feature fusion, 10-fold cross-subject binary (HC vs. MCI+AD; MCI vs. AD) and ternary (HC, MCI, AD) classification tasks were performed subsequently. RESULTS GCN significantly outperformed other models in all three cross-subject classification tasks. In binary classification tasks distinguishing HC from MCI and AD, GCN achieved accuracies of 90.17 ± 5.58 %, 86.73 ± 2.34 %, and 84.73 ± 4.28 % under Neutral, Happiness, and Sadness, respectively. Similarly, in ternary classification of HC, MCI, and AD, GCN reached the highest accuracy of 72.67 ± 1.89 % under Neutral stimulus. CONCLUSIONS Leveraging fused ERP and TCM features, deep learning models have demonstrated robust cross-subject efficacy in the early diagnosis of cognitive decline. Particularly in distinguishing HC from MCI and AD, the performance of GCN was comparable to that of hematological biomarkers. Our study, therefore, highlights a reliable and effective AI-driven methodology for the early diagnosis of cognitive impairment in clinical settings.

Background Inner speech—the covert articulation of words in one’s mind—is a fundamental phenomenon in human cognition with growing interest across BCI. This pilot study evaluates and compares deep learning models for inner-speech classification using non-invasive EEG derived from a bimodal EEG-fMRI dataset (4 participants, 8 words). The study assesses a compact CNN (EEGNet) and a spectro-temporal Transformer using leave-one-subject-out validation, reporting accuracy. Macro-F1, precision, and recall. Objective This study aims to evaluate and compare deep learning models for inner speech classification using non-invasive electroencephalography (EEG) data, derived from a bimodal EEG-fMRI dataset. The goal is to assess the performance and generalizability of two architectures: the compact convolutional EEGNet and a novel spectro-temporal Transformer. Methods Data were obtained from four healthy participants who performed structured inner speech tasks involving eight target words. EEG signals were preprocessed and segmented into epochs for each imagined word. EEGNet and Transformer models were trained using a leave-one-subject-out (LOSO) cross-validation strategy. Performance metrics included accuracy, macro-averaged F1 score, precision, and recall. An ablation study examined the contribution of Transformer components, including wavelet decomposition and self-attention mechanisms. Results The spectro-temporal Transformer achieved the highest classification accuracy (82.4%) and macro-F1 score (0.70), outperforming both the standard and improved EEGNet models. Discriminative power was also substantially improved by using wavelet-based time-frequency features and attention mechanisms. Results showed that confusion patterns of social word categories outperformed those of number concepts, corresponding to different mental processing strategies. Conclusion Deep learning models, in particular attention-based Transformers, demonstrate great promise in decoding internal speech from EEG. These findings lay the groundwork for non-invasive, real-time BCIs for communication rehabilitation in severely disabled patients. Future work will take into account vocabulary expansion, wider participant variety, and real-time validation in clinical settings.

Intraoperative awareness due to inappropriate depth of anesthesia remains a critical concern in clinical practice. Traditional binary classifications of conscious ness and unconsciousness may fail to capture the gradual and variable transitions that occur during the induction and emergencephasesofanesthesia. These transitions can differ significantly across individuals and anesthetic agents. This study aims to classify three conscious states, such as consciousness, transitions, and unresponsiveness under sedation with propofol and midazolam, based on the electroencephalogram (EEG) signals. Using patient-controlled sedation paradigms, we identified transitions through be havioral responsiveness and proposeanoveldeeplearning framework, Deep-ConTrans, incorporating common spatial pattern-based spatial filtering, multi-domain feature extrac tion, attention-based fusion, and domain-adversarial training for robust classification. The average classification accuracies were improved, achieving 93.93 (±3.32)% for propofol and 97.42 (±1.68)% for midazolam, respectively, demonstrating superior performance over conventional methods. The model also demonstrated strong cross anesthetic generalizability, maintaining high performance when evaluated across propofol and midazolam in external validation. Specifically, transitions toward unresponsive ness were marked by increased delta power in frontal regions and increased alpha power in parietal regions. These spectral changes are consistent with cortical bistability and disruption in the posterior hot zone, both of which have been linked to alterations in consciousness. By identifying robust and drug-independent EEG signatures of transitions, this study highlights the potential for more granular and reliable intraoperative monitoring beyond binary assessment. The enhanced generalizability and sensitivity of Deep-ConTrans enable anesthesiologists to precisely identify critical transitions, thus improving anesthetic man agement, minimizing the risk of intraoperative awareness, and facilitating personalized sedation protocols based on real-time EEG dynamics.

… These results highlight the varying degrees of cross-subject generalization, with EEGNet v4 emerging as the most robust model for cross-subject motor imagery classification in this …

… and demographic stratification to serve as digital biomarkers. Methodological advances continue, … GNNs, provide valuable constraints on generalisation but often neglect EEG temporal …

Electroencephalography (EEG)-based four-level fear classification is important for mental-health assessment and emotion-aware systems, yet robust generalization to unseen subjects remains challenging. This study aims to improve subject-independent performance and analyze factors that limit it. Using the DEAP dataset, we propose a GIN–LSTM model that combines a Graph Isomorphism Network (GIN) and Long Short-Term Memory (LSTM) for spatio-temporal EEG modeling. We compare raw EEG and differential-entropy (DE) features under 10-fold and leave-one-subject-out (LOSO) cross-validation (CV) with domain generalization. We further analyze calibration results, feature distributions, and channel importance to investigate low LOSO performance. In 10-fold CV, the model achieves over 98% accuracy, whereas in LOSO CV accuracy is about 37%, showing limited generalization. Calibration experiments suggest that the high 10-fold accuracy is largely due to data leakage between windows from the same trials, while distributional analyses indicate that data scarcity and trial-to-trial variability are major limiting factors for subject-independent performance. They also show that raw EEG preserves more consistent spatial patterns across subjects than DE features. These findings highlight the risk of overestimating performance with conventional CV and suggest data augmentation, stronger domain generalization, and attention mechanisms as promising directions for more robust EEG-based fear recognition.

We report a deep learning-based emotion recognition method using EEG data collected while applying cosmetic creams. Four creams with different textures were randomly applied, and they were divided into two classes, “like (positive)” and “dislike (negative)”, according to the preference score given by the subject. We extracted frequency features using well-known frequency bands, i.e., alpha, beta and low and high gamma bands, and then we created a matrix including frequency and spatial information of the EEG data. We developed seven CNN-based models: (1) inception-like CNN with four-band merged input, (2) stacked CNN with four-band merged input, (3) stacked CNN with four-band parallel input, and stacked CNN with single-band input of (4) alpha, (5) beta, (6) low gamma, and (7) high gamma. The models were evaluated by the Leave-One-Subject-Out Cross-Validation method. In like/dislike two-class classification, the average accuracies of all subjects were 73.2%, 75.4%, 73.9%, 68.8%, 68.0%, 70.7%, and 69.7%, respectively. We found that the classification performance is higher when using multi-band features than when using single-band feature. This is the first study to apply a CNN-based deep learning method based on EEG data to evaluate preference for cosmetic creams.

BACKGROUND Electroencephalography (EEG) signals play a crucial role in understanding brain activity because they provide useful information about real emotions and intentions. Many machine learning models have been used for automatic EEG-based emotion classification. However, previous studies remain limited by restricted feature representations and insufficient subject-independent validation. METHODS In this work, we developed a novel EEG emotion dataset from 22 healthy participants. The dataset includes 14-channel EEG recordings with arousal and valence labels. We also proposed a new feature-extraction function named multiple attention local binary pattern (MATLBP), which generates five feature vectors from EEG signals. To improve the feature-engineering process, we designed an architecture with multi-level MATLBP-based feature extraction, multiple feature-selection methods, and a multi-classifier classification phase. We also used a channel-wise evaluation approach for all EEG channels. Then, iterative majority voting (IMV) was applied during classification to generate predicted vectors. Finally, the MATLBP-based feature-engineering architecture selected the best prediction vector as the final outcome. Thus, the proposed model can automatically select the most accurate result. RESULTS The proposed framework was assessed on both the self-collected EEG dataset and the publicly available DREAMER dataset using leave-one-subject-out cross-validation. The model achieved accuracies of 93.38% for arousal and 88.64% for valence on the self-collected dataset. On the DREAMER dataset, it achieved accuracies of 93.56% for arousal, 97.22% for valence, and 86.73% for dominance. Relative to previously reported methods, the proposed approach demonstrated competitive performance under subject-independent validation conditions, with the additional advantage of reduced computational complexity. CONCLUSIONS The proposed MATLBP-based framework provides an accurate and computationally efficient approach for subject-independent EEG emotion classification across the self-collected and DREAMER datasets. Its strong cross-subject performance and lightweight architecture indicate its potential use in affect-aware clinical decision-support systems, neurofeedback applications, and assistive technologies for individuals with impaired emotional expression. The primary limitation is potential optimistic bias due to model selection on the same dataset. Independent validation in larger, multicenter cohorts is needed to confirm generalizability.

… We propose a lightweight feature-engineered pipeline for binary MI open–close hand classification using a low-cost EEG system and evaluate it under a strict leave-one-subject-out (…

… with leave-one-subject-out cross-validation to ensure that our results reflect genuine subject-… classification capability. This approach builds upon our previous work characterizing EEG …

Detecting Alzheimer’s Disease (AD), Frontotemporal Dementia (FTD), and Healthy Control (Control) remains a critical challenge in Clinical Neurophysiology, particularly in terms of achieving timely and accurate diagnosis. Electroencephalography (EEG) is a low-cost and non-invasive method of measuring functional alterations in brain activities. The various handcrafted features calculated in this research include logarithmic band power (LBP), energy (Energy), root mean square (RMS), and the signal norm (Norm), which were then used to develop a set of robust feature vectors. We leveraged Leave-One-Subject-Out (LOSO) cross-validation to ensure subject independence during evaluation. Additionally, we evaluated four distinct machine learning classifiers: Random Forest (RF), Support Vector Machine (SVM), Decision Tree (DT), and Linear Discriminant Analysis (LDA). Our findings indicate that RF using norm features performed best, producing an accuracy of $\mathbf{8 4. 6 2 \%}$ when differentiating AD patients from Control and exhibiting robust generalizability across multiple subjects.

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition where scalable biomarkers remain limited. Electroencephalography (EEG), with its high temporal resolution and portability, offers a promising foundation for computational analysis. This work benchmarks deep and handcrafted EEG pipelines on paediatric data from 34 children, using strict leave-one-subject-out (LOSO) validation to prevent data leakage. While many models achieved inflated epoch- level accuracies of 80-90%, subject-level performance dropped sharply, often nearing chance. Supervised Transformers reached 60-78% accuracy across sites but underperformed Random Forests, which paired with spectral and entropy features achieved the most consistent subject-level accuracy 71%. Self-supervised pretraining partially closed the transformer gap, improving calibration but not surpassing RF. These findings highlight the importance of rigorous validation and suggest that physiologically informed, interpretable features may currently offer more robust support for ASD assessment than purely deep models, while also clarifying both the potential and current limitations of EEG- based pipelines.

Deep learning for motor imagery electroencephalography (MI-EEG) has repeatedly reported near-ceiling accuracy on public benchmarks. However, many studies use data partitioning strategies in which windows, images, or pooled samples from the same subject can influence both model development and evaluation, so the resulting numbers do not answer the stronger question of cross-subject generalization. This study presents a subject-wise leave-one-subject-out cross-validation (LOSOCV) analysis of a Gramian Angular Field (GAF) and Phase-Locking Value (PLV) parallel convolutional neural network developed for MI-EEG representation learning. Under segment-level splitting, the framework achieved 99.73% binary accuracy in proof-of-concept benchmarking. Here, the same feature-construction logic is examined under LOSOCV on the 105 retained PhysioNet subjects for which subject-wise rerun outputs were available in this study. Under this subject-wise setting, the model achieves 58.07% ± 8.27% mean accuracy, 53.48% ± 11.19% macro-F1, and 0.1615 ± 0.1654 Cohen’s kappa, with held-out subject accuracy ranging from 38.10% to 78.57%. Relative to the earlier segment-wise benchmark, the mean generalization gap is 41.66 points, while held-out-subject gaps span 21.16–61.63 points. By combining complete held-out-fold disclosure, retained-cohort accounting, bootstrap confidence intervals, and explicit protocol-sensitive comparison, the study provides a stronger subject-wise reference point for future MI-EEG evaluation and a more defensible basis for interpreting translational claims. The significance of the work lies in showing how a strong proof-of-concept benchmark behaves under a subject-wise inference target and in providing a clearer field reference for subject-independent MI-EEG research.

To address domain shift in motor imagery electroencephalogram (MI-EEG) caused by individual differences and temporal variations, this paper proposes a Subject Selection-based Unsupervised Domain Adaptation model (SUDA). The method begins by applying K-Means clustering to group subjects in a balanced manner. It then incorporates a multi-scale feature extraction module to capture temporal discriminative features. Using the pseudo-label generation strategy, it generates target domain pseudo-labels for adaptive alignment. Finally, model optimization is performed through a hierarchical loss function. Using leave-one-subject-out cross-validation on the OpenBMI dataset, SUDA achieved accuracy of 81.0% and 83.4% across two independent sessions. These results surpass both traditional CSP and various deep learning benchmarks, confirming the model's effectiveness and generalization capability in cross-subject motor imagery classification.

This study presents a validation-aware EEG framework based on Chaotic Pattern of Prime Numbers (CPPN) features for depression treatment-response modelling across one SSRI cohort and two rTMS cohorts. CPPN features were evaluated through a seven-protocol validation hierarchy spanning random segment splitting, segment-level cross-validation, nested segment-level cross-validation, leave-N-subjects-out, fixed-feature leave-one-subject-out (LOSO), nested leave-N-subjects-out, and nested LOSO, with normalisation, NCA ranking, feature-count selection where applicable, and model fitting confined to the appropriate training partitions. In the representative K-nearest neighbour (KNN) comparison, segment-level 10-fold CV achieved accuracies of 98.79% for Mumtaz SSRI, 99.32% for small Atieh rTMS, and 99.42% for big Atieh rTMS, demonstrating strong discriminative structure in the CPPN feature space. In the available segment-level KNN comparison, CPPN features with fold-internal NCA-selected feature sets exceeded conventional statistical EEG features by 29.53, 16.33, and 11.45 percentage points across the three cohorts. Subject-wise validation produced lower and more cohort-dependent estimates, with the best fixed-feature LOSO accuracy of 80.00% and the best nested LOSO accuracy of 73.33% in the small Atieh rTMS cohort. These results show that CPPN provides a compact, inspectable and computationally accessible EEG feature representation, while the validation hierarchy gives a transparent account of how performance changes from segment-level separability to held-out-subject evaluation. The main contribution is methodological: this study combines an original CPPN feature representation with explicit validation-depth analysis, leakage-aware feature selection, and interpretable channel/bin inspection. It therefore provides a rigorous basis for future externally validated EEG treatment-response studies without claiming prospective clinical deployment from the present retrospective cohorts.

Decoding visual information from electroencephalography (EEG) has advanced with contrastive learning, but cross-subject generalization remains difficult due to strong inter-subject variability. We propose a cross-subject EEG-to-image retrieval framework that eliminates the need for subject-specific calibration by combining feature-level mixup augmentation and subject-adversarial learning. This encourages subject-invariant yet semantically meaningful EEG representations. Experiments on the THINGS-EEG dataset show that our method substantially improves cross-subject retrieval accuracy over prior approaches, and ablation studies confirm the complementary benefits of mixup and subject-adversarial learning for robustness and generalization to unseen subjects.

… To ensure a robust and realistic assessment of model performance, this study employs a rigorous Leave-One-Subject-Out (LOSO) cross-validation strategy. The models were trained on …

Consumer-grade wearable sensors may enable continuous monitoring of pilot workload and stress during flight training, yet most prior studies rely on simulators, raw-score labelling, and within-subject validation, limiting generalisability. This study evaluates whether electrodermal activity (EDA), electrocardiogram (ECG)-derived features, and wrist skin temperature, recorded from an Empatica Embrace Plus and a Polar H10 during real Cessna 172 flight training, can classify pilots' task-relative workload and stress deviations. Thirty-five pilots completed four flight segments and rated workload and stress after each. Fold-safe two-way residual binary labels removed inter-pilot scale-use differences and task-level effects, and five classifiers were evaluated under leave-one-subject-out (LOSO) cross-validation with Benjamini-Hochberg FDR correction. Under LOSO, a Linear SVC on combined features classified stress (macro F1 = 0.607) and XGBoost on EDA classified workload (macro F1 = 0.598) significantly above chance (padj=0.033); both remained stable under nested cross-validation with an inner hyperparameter search (nested 0.606 and 0.561). A LightGBM model on EDA gave a numerically higher stress score (0.611) that did not survive nested validation. Subject-dependent within-subject validation produced higher apparent performance (macro F1 = 0.853 for stress and 0.791 for workload), but a stricter within-pilot analysis was unstable. These contrasts indicate that personalised classification may be feasible after calibration, whereas uncalibrated cross-pilot prediction in real flight remains modest, with post-flight debriefing the most plausible near-term application.

… Riemannian geometry is technical, our aim here is to show the appeal of the framework on an intuitive geometrical … In this article we explain the key features of a Riemannian decoder, …

A notable challenge encountered by motor imagery decoding algorithms utilizing electroencephalography (EEG) signals is the substantial redundancy and inadequate geometric representation of spatiotemporal features, which stem from the volume conduction effects inherent to the human head. This phenomenon can obscure essential information regarding motor intentions with noise or non-discriminative features. Although traditional decoding models have sought to alleviate redundancy through shallow attention mechanisms or feature selection in Euclidean space, they frequently neglect the intrinsic manifold geometric properties of EEG signals, such as the positive definiteness of covariance matrices. Additionally, static attention weights are often insufficient in dynamically capturing the cross-domain dependencies between spatiotemporal and spectral features. To address these limitations, we propose a novel spatiotemporal dynamic attention fusion network grounded in Riemannian manifolds (ST-MA-SENet) for EEG motor imagery decoding. ST-MA-SENet adeptly assesses the spatiotemporal correlations among EEG features in both Euclidean and Riemannian spaces from a comprehensive perspective, thereby facilitating the selection of a distinctive and effective EEG fusion feature for motor imagery recognition. To evaluate the efficacy of ST-MA-SENet, we conducted experiments utilizing three motor imagery datasets (BCI IV 2a, BCI IV 2b, HGD), and the results demonstrate that ST-MA-SENet represents a highly promising approach for EEG signal decoding.

BACKGROUND Brain computer interface (BCI) utilizes brain signals to help users interact with external devices directly. EEG is one of the most commonly used techniques for brain signal acquisition in BCI. However, it is notoriously difficult to build a generic EEG recognition model due to significant non-stationarity and subject-to-subject variations, and the requirement for long time training. Transfer learning (TL) is particularly useful because it can alleviate the calibration requirement in EEG-based BCI applications by transferring the calibration information from existing subjects to new subject. To take advantage of geometric properties in Riemann manifold and joint distribution adaptation, a manifold embedded transfer learning (METL) framework was proposed for motor imagery (MI) EEG decoding. NEW METHOD First, the covariance matrices of the EEG trials are first aligned on the SPD manifold. Then the features are extracted from both the symmetric positive definite (SPD) manifold and Grassmann manifold. Finally, the classification model is learned by combining the structural risk minimization (SRM) of source domain and joint distribution alignment of source and target domains. RESULT Experimental results on two MI EEG datasets verify the effectiveness of the proposed METL. In particular, when there are a small amount of labeled samples in the target domain, METL demonstrated a more accurate and stable classification performance than conventional methods. COMPARISON WITH EXISTING METHODS Compared with several state-of-the-art methods, METL has achieved better classification accuracy, 71.81% and 69.06% in single-to-single (STS), 83.14% and 76.00% in multi-to-single (MTS) transfer tasks, respectively. CONCLUSIONS METL can cope with single source domain or multi-source domains and compared with single-source transfer learning, multi-source transfer learning can improve the performance effectively due to the data expansion. It is effective enough to achieve superior performance for classification of EEG signals.

In the field of brain–computer interfaces (BCIs), the recognition of electroencephalogram (EEG) signals lies at the heart of decoding neural activity patterns and facilitating efficient human–machine interaction. Geometric learning (GL) methods have garnered increasing attention due to their enhanced robustness in EEG signal decoding. However, existing GL methods lack the ability to extract multi-scale features and spatial and channel attention from manifold data. To address these two issues, this paper proposes a model called Riemannian multi-scale residual network (RMS-RestNet), which is built upon the MAtt. RMS-RestNet designed a Riemannian multi-scale residual module (RMSRM), which extends depthwise and pointwise convolutions to the space of symmetric positive definite (SPD) manifolds, referred to as SPD depthwise (SPD DW) and SPD pointwise (SPD PW) convolutions, respectively. By employing SPD convolution kernels of varying scales, the model facilitates more comprehensive and discriminative extraction of geometric features. On this basis, residual connections are employed to fuse the geometric information of the data before and after processing, thereby enhancing the feature representation ability of the module. In addition, by combining a tangent space pooling strategy with SPD convolutions, we construct a manifold convolutional block attention module (MCBA) to capture both channel-wise and spatial attention across multi-channel SPD manifold data. Extensive experiments on both temporally synchronous and asynchronous EEG datasets demonstrate the superiority of our approach over state-of-the-art methods.

Electroencephalography (EEG) data are pivotal in brain–computer interfaces (BCIs), yet their utility is hindered by data scarcity arising from high acquisition costs, noise susceptibility, and privacy constraints. Traditional augmentation methods, such as noise injection and signal transformations, often fail to preserve task-relevant structure in multichannel EEG, while deep generative models may suffer from mode collapse or produce physiologically inconsistent samples. To address these limitations, we propose a Riemannian Conditional Generative Adversarial Network (RC-GAN) that enforces geometric consistency during signal generation. RC-GAN leverages the manifold of symmetric positive definite (SPD) covariance matrices to regularize synthetic EEG trials according to covariance-based representations widely used in BCI decoding. Evaluated on the BNCI 2014-001 motor imagery dataset, the proposed method outperforms state-of-the-art augmentation techniques, achieving a 12.0% improvement in classification accuracy. Qualitative and quantitative analyses demonstrate that RC-GAN generates diverse and realistic EEG samples while enhancing robustness at different augmentation levels. These results highlight the benefit of incorporating Riemannian structure into generative models for EEG augmentation and provide a principled framework for improving the reliability of BCI systems.

Background: The decoding of motor imagery electroencephalography (MI-EEG) is constrained by core issues including low signal-to-noise ratio (SNR) and cross-session as well as cross-subject domain shift, which seriously impedes the practical deployment of brain–computer interfaces (BCIs). Methods: To address these challenges, this paper proposes a novel end-to-end MI-EEG decoding method named BARN-DA. Two innovative modules, Band-Aware Channel Attention (BACA) and Multi-Scale Kernel Perception (MSKP), are designed: one enhances discriminative channel features by modeling channel information fused with frequency band feature representation, and the other captures complex data correlations via multi-scale parallel convolutions to improve the discriminability of the network’s feature extraction. Subsequently, the features are mapped onto the Riemannian manifold. For the source and target domain features residing on this manifold, a Riemannian Maximum Mean Discrepancy (R-MMD) loss is designed based on the log-Euclidean metric. This approach enables the effective embedding of Symmetric Positive Definite (SPD) matrices into the Reproducing Kernel Hilbert Space (RKHS), thereby reducing cross-domain discrepancies. Results: Experimental results on four public datasets demonstrate that the BARN-DA method achieves average cross-session classification accuracies of 84.65% ± 8.97% (BCIC IV 2a), 89.19% ± 7.69% (BCIC IV 2b), and 61.76% ± 12.68% (SHU), as well as average cross-subject classification accuracies of 65.49% ± 11.64% (BCIC IV 2a), 78.78% ± 8.44% (BCIC IV 2b), and 78.14% ± 14.41% (BCIC III 4a). Compared with state-of-the-art methods, BARN-DA obtains higher classification accuracy and stronger cross-session and cross-subject generalization ability. Conclusions: These results confirm that BARN-DA effectively alleviates low SNR and domain shift problems in MI-EEG decoding, providing an efficient technical solution for practical BCI systems.

This study investigates the application of Riemannian geometry‐based methods for brain decoding using invasive electrophysiological recordings. While Riemannian geometry has been successfully applied in noninvasive settings, its utility for invasive datasets, which are typically smaller and scarcer, remains less explored. Herein, a minimum distance to mean (MDM) classifier is proposed using a Riemannian geometry approach based on covariance matrices extracted from intracortical local field potential (LFP) recordings across various regions during different brain state dynamics. For benchmarking, the performance of the approach is evaluated against convolutional neural networks (CNNs) and Euclidean MDM classifiers. The results indicate that the Riemannian geometry‐based classification not only achieves a superior mean F1 macro‐averaged score across different channel configurations but also requires up to two orders of magnitude less computational training time. Additionally, the geometric framework reveals distinct spatial contributions of brain regions across varying brain states, suggesting a state‐dependent organization that traditional time series‐based methods often fail to capture. The findings align with previous studies supporting the efficacy of geometry‐based methods and extend their application to invasive brain recordings, highlighting their potential for broader clinical use, such as brain‐computer interface applications.

Reliable and explainable detection of mental workload from EEG can advance both scientific understanding and practical monitoring. This study presents a compact pipeline that 1) employs directed effective connectivity (dDTF) across standard frequency bands to capture information flow, 2) encodes each time window as a $14\times 14$ directed adjacency graph, and 3) classifies workload using a hybrid Graph Convolutional + Graph Attention network with residual connections and dual global pooling. Hyperparameters were optimized automatically via Optuna, and model performance was assessed with cross-validation. Targeted ablation experiments and post-hoc interpretation using GNNExplainer were conducted to probe model behavior. The approach achieves strong within-dataset performance on STEW (90.1% accuracy, 87.5% precision, 93.6% recall, F1 score 90.43%). Ablations show that the attention layer, residual connections, and max-pooling each contribute measurable gains, while interpretability maps consistently highlight fronto-parietal and fronto-temporal connections as key discriminative features. These results link improved classification with neuroscience-aligned explanations.

Major depressive disorder (MDD) presents a substantial global health challenge due to its high prevalence of morbidity, disability, and mortality, along with frequent underdiagnosis. The profound impact of MDD on patients’ mental health emphasizes the necessity for precise and objective diagnostic tools. However, existing studies have mostly overlooked the topological relationships among brain channels and the spatiotemporal correlations in brain activity. To discover the latent patterns within the complex neural networks associated with depression, we present a novel model, the spatiotemporal graph convolutional network-convolutional neural network-attention mechanism (ST-GCA), based on deep learning and complex network theories, aiming to enhance MDD detection. The proposed model utilized brain network connectivity matrices to quantify the connection lengths between channels, followed by a convolutional neural network (CNN) to extract deep spatial features. Moreover, a spatiotemporal attention mechanism was integrated to adaptively capture critical sequential information from segmented electroencephalogram (EEG) signals of patients with MDD. This approach facilitated the analysis of distinct neural mechanisms in MDD patients compared to healthy controls (HCs). The ST-GCA framework achieved an accuracy of 97.68% in subject-dependent (SD) experiments and 93.93% in subject-independent (SI) experiments on a publicly available dataset, demonstrating its effectiveness in analyzing MDD-related EEG signals. These findings provide a promising tool to assist clinicians in diagnosing MDD, thereby offering valuable clinical guidance.

… a neural-attenuation prior dynamic graph neural network (… multi-channel EEG-based seizure classification. The proposed … graph generator to model multi-scale functional connectivity, (2…

The growing burden of anxiety disorders highlights the urgent need for scalable and non-invasive systems for mental health monitoring. Electroencephalography (EEG)-based Brain-Computer Interfaces (BCIs) offer a promising solution by capturing neural oscillations linked to anxiety. However, conventional detection methods oversimplify the brain's graphstructured connectivity and are computationally intensive, limiting their feasibility for real-time, edge-based deployment. To address these limitations, we propose two edge-optimized Graph Convolutional Network frameworks (GatedGCN and GAT) for anxiety classification using EEG signals. By modeling multichannel EEG data as dynamic graphs, the system captures spatial-temporal brain dynamics critical for detecting anxietyrelated patterns. The architecture incorporates adaptive graph construction, hierarchical spatio-temporal convolutions, and quantization-aware training to enable a reduced-size inference model with minimal accuracy loss. Our approach achieves realtime, resource-efficient performance on low-power edge devices that enables continuous, private, and accessible anxiety monitoring to pave the way for practical mental health interventions in wearable and mobile healthcare settings.

… , a Graph Attention Convolutional Neural Network (GAT-CNN) … EEG signals are classified into two states namely alert and fatigued by fully connected layers and a softmax classification …

The healthcare field, human-computer interaction, and affective computing development require accurate human emotional recognition through electroencephalography (EEG) signals. EEG signals present intricate spatial temporal and spectral dependencies that both traditional machine learning and deep learning methods fail to capture resulting in restricted cross-subject generalizability and limited understanding. We introduce MER-TFC, a new EEG-based emotion recognition system which uses GraphNeuralNetworks (GNNs) with attention mechanisms to create inter-channel connectivity and temporal models from EEG data while connecting theta, alpha, beta, and gamma frequency bands. The framework is evaluated using leave-one-subject-out cross-validation (LOSOCV) on DEAP and AMIGOS.benchmark datasets. The experimental results show that MER-TFC achieves 88.2 % average accuracy with 87.7 % F1-score and 0.93 ROC-AUC and 87.9 % balanced accuracy which outperforms Transformer-CNN, Multilayer Graph Transformer, Frequency-Adaptive Dynamic Graph Transformer (FreqDGT) and Adaptive Progressive Attention GNN (APAGNN). Attention analysis reveals brain regions that play critical roles and shows how emotional states correspond to functional brain network connections. The proposed framework also maintains high performance when EEG signals experience background noise, which demonstrates its ability to function well on different subjects.

… (GNNs) to model functional connectivity. However, most existing approaches operate at the … of multiple nodes to generate a feature for the entire graph, followed by graph classification. …

… While the current study focuses on fixed spectral connectivity … leverages the strengths of graph neural networks (GNNs) and … pipeline from raw EEG signals to emotion classification. The …

Imagined swallowing (MI-SW) is a promising paradigm for brain–computer interface (BCI)-assisted dysphagia rehabilitation; however, reliable EEG-based decoding remains challenging because of the low signal-to-noise ratio and high inter-subject variability. In addition, prior EEG-based MI-SW studies have not explicitly modeled interregional brain interactions and have rarely employed subject-independent validation frameworks, limiting their generalizability. To address these limitations, this study proposes a regional graph-based EEG decoding framework that integrates (i) region-wise graph modeling of EEG channels, (ii) leave-one-subject-out (LOSO) subject-independent validation, and (iii) occlusion-based interpretability analysis. EEG data were collected from 30 healthy participants under two paradigms: volitional imagined swallowing and sensory-induced imagined swallowing, which was achieved by holding water in the mouth before imagery. Spectral, event-related desynchronization, entropy-based, and nonlinear features were extracted and classified using Graph Convolutional Networks (GCN). The classification task was formulated as a binary problem distinguishing Rest versus Imagined Swallowing conditions. All experiments were evaluated using leave-one-subject-out (LOSO) cross-validation to assess subject-independent decoding performance. The proposed framework achieved robust subject-independent performance across regions, with the sensory-induced paradigm consistently outperforming volitional conditions. Peak classification accuracy of >87% was obtained in the premotor region under sensory-induced conditions, whereas the volitional paradigm showed optimal performance using whole-head electrode integration. No direct comparison with conventional CNN or CSP-based baselines was performed in this study; therefore, performance gains are reported within the proposed framework under a consistent evaluation setting. Occlusion-based interpretability analysis further confirmed physiologically meaningful patterns, with low-frequency ( $\delta $ – $\theta $ ) desynchronization and entropy/complexity features emerging as the dominant discriminative markers of imagined swallowing. These findings demonstrate that integrating spatial graph representations with the peripheral somatosensory context enhances the EEG-based decoding of swallowing motor imagery and supports the development of reliable and interpretable BCI systems for dysphagia rehabilitation.

Accurate seizure detection and type classification from Electroencephalography (EEG) are crucial for epilepsy diagnosis and clinical treatment. The intricate nature of seizure dynamics presents a significant challenge in effectively extracting distinguishing features from multivariate EEG signals, mainly due to long-range temporal dynamics and complex spatial dependencies between electrodes. To tackle these challenges, we introduce TS-S4GNN, a two-stream graph neural network (GNN) with structured state space modeling. Specifically, we combine channel-independent 1D-convolutional neural network with structured state space models to capture local and longrange temporal dependencies. The complex spatial dependencies between electrodes are learned with two GNN layers in parallel, one with static graph structure constructed according to the distance between electrodes, one with dynamically evolving graph structures learned from data. We validated TS-S4GNN on TUSZv1.5.2, the largest public EEG corpus with seizure type annotations. Experiments demonstrate that TS-S4GNN achieves an AUROC score of 0.913 in seizure detection and a weighted F1-score of 0.781 in seizure type classification, surpassing the current state-of-the-art models. Ablation studies validate the effectiveness of multi-scale temporal modeling and the twostream GNN approach. This study provides new insights into spatiotemporal modeling of multivariate signals and can be easily extended to other relevant learning tasks. The code is available at https://github.com/XploreAI-Lab/TS-S4GNN.

… pattern contradicts the inherently sparse connectivity among EEG channels. Overall, these … CNN-based method that has demonstrated effectiveness in EEG classification. FSTNet …

Electroencephalography signal classification is essential for the diagnosis and monitoring of neurological disorders, with significant implications for patient treatment. Despite the progress made, existing methods face challenges such as capturing the complex dynamics of Electroencephalogram (EEG) signals and generalizing across diverse patient populations. In this study, the graph attention network and the transformer model are integrated for EEG signal classification, leveraging the enhanced capability to dynamically compute attention weights and adapt to the variable relevance of brain regions. The proposed approach is capable of modeling the intricate relationships within EEG activities by learning context-dependent attention scores. We conducted a comprehensive evaluation of the proposed approach comparing with the state-of-the-art algorithms. Experimental outcomes show that it surpasses the competing models. The superior performance is attributed to the proposed approach's dynamic attention mechanism, which better captures the nuanced patterns in EEG signals across different subjects and seizure types. In the experiments, the CHB-MIT dataset was exploited, which served as a benchmark for evaluating the performance of the proposed framework in distinguishing interictal, ictal, and normal EEG patterns. The results prove the usefulness of our work in advancing EEG signal classification. The findings suggest that the combination of graph attention and self-attention mechanisms is a promising approach for improving the accuracy and reliability of EEG-based diagnostics, potentially improving the management of neurological disorders.

… Graph Neural Network (GNN)-based approach tailored for single-channel EEG classification … time-frequency features with functional connectivity measures. Their method achieved a …

… , epileptic EEG signal and seizure classification can also be … tional connectivity from EEG signals [21–23]. The functional … EEG with deep convolutional neural network. In ICEIC 2018: …

To address the inherent complexity and nonlinearity of electroencephalogram (EEG) signals, this study proposes a refined classification framework, NeuroNetFusion, which strategically integrates and selects multi-context network-based features for improved performance. This framework enhances the performance of bio-signal classification models by integrating multi-directionally expressed cross-dependence information. Unlike prior EEG classification studies that mainly relied on single-domain or uni-directional features, our framework introduces a systematic multi-context integration strategy, which constitutes a primary contribution of this work. The process begins with preprocessing the EEG signals using a Savitzky–Golay (SG) filter to reduce noise. Next, the signals are decomposed into multiple frequency bands using the Discrete Wavelet Transform (DWT). The resulting data are then reconstructed from five bands into corresponding adjacency matrices. Following this, the signals are represented as two distinct types of networks: a causality network based on the Directed Transfer Function (DTF) and a correlation network using the Pearson correlation coefficient. To combine features from these two networks, we utilize the TF-IDF method to vectorize the non-zero element index sequences from the adjacency matrices. This procedure transforms sparse adjacency matrices into a quantitative representation, allowing us to assess the importance of each connection within the network. Additionally, a genetic algorithm is employed to select important TF–IDF features, optimizing them for classification tasks. Performance is evaluated by comparing several conventional machine learning models using the EEG source dataset (MTOUH), employing standard evaluation metrics. The proposed model achieved a final accuracy of 88.05%, representing an 8.73% absolute improvement over the baseline TCN model. This demonstrates the effectiveness of our method in identifying abnormalities in EEG signals. The key contribution lies in bridging causality- and correlation-based representations through TF-IDF-driven feature encoding, offering a novel pathway for interpretable and scalable EEG analysis. Our approach holds promise for future applications in cross-dependent bio-signal classification problems, paving the way for further research and development in this area.

Electroencephalogram (EEG) signals are increasingly used for emotion recognition because of their non-invasive nature and high temporal resolution. However, their nonlinear and complex dynamics remain challenging to analyze. This study applies Graph Neural Networks (GNNs) to classify emotions from EEG by representing neural activity as graphs. Using the SEED-V dataset, the preprocessing included filtering, downsampling, segmentation, labelling, and feature extraction (DE, PSD, ApEn, and statistical metrics). These graph-based datasets were then evaluated with four models: GCN, GAT, ST-GCN, and DCNN. Experimental results demonstrated that GAT consistently outperformed the others, achieving validation accuracy above 97% at two to three layers while maintaining stable convergence. ST-GCN provided stronger modelling of temporal dependencies but required longer training, whereas GCN showed steady yet modest improvement, and DCNN yielded the weakest performance. These findings highlight the importance of model depth and attention mechanisms in EEG-based emotion recognition and provide practical insights for designing intelligent systems capable of interpreting human emotions more reliably. Limitations include the absence of independent test sets and efficiency analysis, which open avenues for future research on generalisation, multimodal integration, and dynamic graph learning.

Schizophrenia is associated with disrupted neural connectivity and abnormal brain dynamics that can be captured through electroencephalogram (EEG) analysis. However, traditional EEG analysis methods often fail to represent the complex spatial and temporal dependencies in brain activity. This paper addresses this challenge by modeling EEG signals as graph signals and comparing three graph construction strategies for schizophrenia classification: a Gaussian kernelbased similarity graph, a functional-causal fusion graph, and a Semilocal graph. EEG recordings from healthy controls and schizophrenia patients were preprocessed, band-pass filtered into canonical frequency bands (delta, theta, alpha, beta, gamma), and represented as graphs according to each construction method. Graph-based features were extracted and used for classification with a support vector machine under fivefold stratified cross-validation. Experimental results demonstrate that the functional-causal fusion graph consistently achieved the highest accuracy across all frequency bands, reaching perfect classification in several cases, while the Gaussian kernel and Semilocal graphs produced slightly lower but competitive results. These findings indicate that integrating functional and causal connectivity information provides a more discriminative graph representation for schizophrenia EEG classification and emphasize the importance of graph construction strategy in EEG analysis using graph signal processing.

Robust cross-subject emotion recognition from multimodal physiological signals remains a challenging problem, primarily due to modality heterogeneity and inter-subject distribution shift. To tackle these challenges, we propose a novel adaptive learning framework named Hierarchical Attention and Dynamic Uniform Alignment (HADUA). Our approach unifies the learning of multimodal representations with domain adaptation. First, we design a hierarchical attention module that explicitly models intra-modal temporal dynamics and inter-modal semantic interactions (e.g., between electroencephalogram(EEG) and eye movement(EM)), yielding discriminative and semantically coherent fused features. Second, to overcome the noise inherent in pseudo-labels during adaptation, we introduce a confidence-aware Gaussian weighting scheme that smooths the supervision from target-domain samples by down-weighting uncertain instances. Third, a uniform alignment loss is employed to regularize the distribution of pseudo-labels across classes, thereby mitigating imbalance and stabilizing conditional distribution matching. Extensive experiments on multiple cross-subject emotion recognition benchmarks show that HADUA consistently surpasses existing state-of-the-art methods in both accuracy and robustness, validating its effectiveness in handling modality gaps, noisy pseudo-labels, and class imbalance. Taken together, these contributions offer a practical and generalizable solution for building robust cross-subject affective computing systems.

Neonates are highly susceptible to seizures, often leading to short or long-term neurological impairments. However, clinical manifestations of neonatal seizures are subtle and often lead to misdiagnoses. This increases the risk of prolonged, untreated seizure activity and subsequent brain injury. Continuous video electroencephalogram (cEEG) monitoring is the gold standard for seizure detection. However, this is an expensive evaluation that requires expertise and time. In this study, we propose a convolutional neural network-based model for early prediction of neonatal seizures by distinguishing between interictal and preictal states of the EEG. Our model is patient-independent, enabling generalization across multiple subjects, and utilizes mel-frequency cepstral coefficient matrices extracted from multichannel EEG and electrocardiogram (ECG) signals as input features. Trained and validated on the Helsinki neonatal EEG dataset with 10-fold cross-validation, the proposed model achieved an average accuracy of 97.52%, sensitivity of 98.31%, specificity of 96.39%, and F1-score of 97.95%, enabling accurate seizure prediction up to 30 minutes before onset. The inclusion of ECG alongside EEG improved the F1-score by 1.42%, while the incorporation of an attention mechanism yielded an additional 0.5% improvement. To enhance transparency, we incorporated SHapley Additive exPlanations (SHAP) as an explainable artificial intelligence method to interpret the model and provided localization of seizure focus using scalp plots. The overall results demonstrate the model's potential for minimally supervised deployment in neonatal intensive care units, enabling timely and reliable prediction of neonatal seizures, while demonstrating strong generalization capability across unseen subjects through transfer learning.

EEG-based emotion recognition is hampered by profound dataset heterogeneity (channel/subject variability), hindering generalizable models. Existing approaches struggle to transfer knowledge effectively. We propose'One Model for All', a universal pre-training framework for EEG analysis across disparate datasets. Our paradigm decouples learning into two stages: (1) Univariate pre-training via self-supervised contrastive learning on individual channels, enabled by a Unified Channel Schema (UCS) that leverages the channel union (e.g., SEED-62ch, DEAP-32ch); (2) Multivariate fine-tuning with a novel'ART'(Adaptive Resampling Transformer) and'GAT'(Graph Attention Network) architecture to capture complex spatio-temporal dependencies. Experiments show universal pre-training is an essential stabilizer, preventing collapse on SEED (vs. scratch) and yielding substantial gains on DEAP (+7.65%) and DREAMER (+3.55%). Our framework achieves new SOTA performance on all within-subject benchmarks: SEED (99.27%), DEAP (93.69%), and DREAMER (93.93%). We also show SOTA cross-dataset transfer, achieving 94.08% (intersection) and 93.05% (UCS) on the unseen DREAMER dataset, with the former surpassing the within-domain pre-training benchmark. Ablation studies validate our architecture: the GAT module is critical, yielding a +22.19% gain over GCN on the high-noise DEAP dataset, and its removal causes a catastrophic -16.44% performance drop. This work paves the way for more universal, scalable, and effective pre-trained models for diverse EEG analysis tasks.

Personalised music-based interventions offer a powerful means of supporting motor rehabilitation by dynamically tailoring auditory stimuli to provide external timekeeping cues, modulate affective states, and stabilise gait patterns. Generalisable Brain-Computer Interfaces (BCIs) thus hold promise for adapting these interventions across individuals. However, inter-subject variability in EEG signals, further compounded by movement-induced artefacts and motor planning differences, hinders the generalisability of BCIs and results in lengthy calibration processes. We propose Individual Tangent Space Alignment (ITSA), a novel pre-alignment strategy incorporating subject-specific recentering, distribution matching, and supervised rotational alignment to enhance cross-subject generalisation. Our hybrid architecture fuses Regularised Common Spatial Patterns (RCSP) with Riemannian geometry in parallel and sequential configurations, improving class separability while maintaining the geometric structure of covariance matrices for robust statistical computation. Using leave-one-subject-out cross-validation, `ITSA'demonstrates significant performance improvements across subjects and conditions. The parallel fusion approach shows the greatest enhancement over its sequential counterpart, with robust performance maintained across varying data conditions and electrode configurations. The code will be made publicly available at the time of publication.

In this paper, we focus on the challenge of individual variability in affective brain-computer interfaces (aBCI), which employs electroencephalogram (EEG) signals to monitor and recognize human emotional states, thereby facilitating the advancement of emotion-aware technologies. The variability in EEG data across individuals poses a significant barrier to the development of effective and widely applicable aBCI models. To tackle this issue, we propose a novel transfer learning framework called Semi-supervised Domain Adaptation with Dynamic Distribution Alignment (SDA-DDA). This approach aligns the marginal and conditional probability distribution of source and target domains using maximum mean discrepancy (MMD) and conditional maximum mean discrepancy (CMMD). We introduce a dynamic distribution alignment mechanism to adjust differences throughout training and enhance adaptation. Additionally, a pseudo-label confidence filtering module is integrated into the semi-supervised process to refine pseudo-label generation and improve the estimation of conditional distributions. Extensive experiments on EEG benchmark databases (SEED, SEED-IV and DEAP) validate the robustness and effectiveness of SDA-DDA. The results demonstrate its superiority over existing methods in emotion recognition across various scenarios, including cross-subject and cross-session conditions. This advancement enhances the generalization and accuracy of emotion recognition, potentially fostering the development of personalized aBCI applications. The source code is accessible at https://github.com/XuanSuTrum/SDA-DDA.

Decoding natural language from brain activity using non-invasive electroencephalography (EEG) remains a significant challenge in neuroscience and machine learning, particularly for open-vocabulary scenarios where traditional methods struggle with noise and variability. Previous studies have achieved high accuracy on small-closed vocabularies, but it still struggles on open vocabularies. In this study, we propose ETS, a framework that integrates EEG with synchronized eye-tracking data to address two critical tasks: (1) open-vocabulary text generation and (2) sentiment classification of perceived language. Our model achieves a superior performance on BLEU and Rouge score for EEG-To-Text decoding and up to 10% F1 score on EEG-based ternary sentiment classification, which significantly outperforms supervised baselines. Furthermore, we show that our proposed model can handle data from various subjects and sources, showing great potential for high performance open vocabulary eeg-to-text system.

Electroencephalography provides a non-invasive window into brain activity, offering valuable insights for neurological research, brain-computer interfaces, and clinical diagnostics. However, the development of robust machine learning models for EEG analysis is hindered by the scarcity of large-scale, well-annotated datasets and the inherent variability of EEG signals across subjects and recording conditions. Inspired by the success of foundation models in natural language processing and computer vision, we propose the Large Cognition Model-a transformer-based foundation model designed to generalize across diverse EEG datasets and downstream tasks. Unlike traditional approaches, our proposed transformer-based architecture demonstrates strong generalization capabilities across datasets and tasks, even without pretraining, surpassing some existing EEG universal models on specific downstream applications. LCM leverages large-scale self-supervised learning techniques to capture universal EEG representations, enabling efficient fine-tuning for applications such as cognitive state decoding, disease classification, and neurofeedback systems. We introduce a novel architecture that integrates temporal and spectral attention mechanisms, optimizing the model's ability to extract meaningful features from raw EEG signals. Extensive evaluations demonstrate that LCM outperforms state-of-the-art approaches across multiple EEG benchmarks, exhibiting strong cross-subject and cross-task generalization. Our findings highlight the potential of pretrained EEG foundation models to accelerate advancements in neuroscience, personalized medicine, and BCI technology.

Carrying conversations in multi-sound environments is one of the more challenging tasks, since the sounds overlap across time and frequency making it difficult to understand a single sound source. One proposed approach to help isolate an attended speech source is through decoding the electroencephalogram (EEG) and identifying the attended audio source using statistical or machine learning techniques. However, the limited amount of data in comparison to other machine learning problems and the distributional shift between different EEG recordings emphasizes the need for a self supervised approach that works with limited data to achieve a more robust solution. In this paper, we propose a method based on self supervised learning to minimize the difference between the latent representations of an attended speech signal and the corresponding EEG signal. This network is further finetuned for the auditory attention classification task. We compare our results with previously published methods and achieve state-of-the-art performance on the validation set.

Electroencephalografic (EEG) data are complex multi-dimensional time-series that are very useful in many applications, from diagnostics to driving brain-computer interface systems. Their classification is still a challenging task, due to the inherent within- and between-subject variability and their low signal-to-noise ratio. On the other hand, the reconstruction of raw EEG data is even more difficult because of the high temporal resolution of these signals. Recent literature has proposed numerous machine and deep learning models that could classify, e.g., different types of movements, with an accuracy in the range 70% to 80% (with 4 classes). On the other hand, a limited number of works targeted the reconstruction problem, with very limited results. In this work, we propose vEEGNet, a DL architecture with two modules, i.e., an unsupervised module based on variational autoencoders to extract a latent representation of the data, and a supervised module based on a feed-forward neural network to classify different movements. To build the encoder and the decoder of VAE we exploited the well-known EEGNet network. We implemented two slightly different architectures of vEEGNet, thus showing state-of-the-art classification performance, and the ability to reconstruct both low-frequency and middle-range components of the raw EEG. Although preliminary, this work is promising as we found out that the low-frequency reconstructed signals are consistent with the so-called motor-related cortical potentials, well-known motor-related EEG patterns and we could improve over previous literature by reconstructing faster EEG components, too. Further investigations are needed to explore the potentialities of vEEGNet in reconstructing the full EEG data, generating new samples, and studying the relationship between classification and reconstruction performance.

Fuzzy logic provides a robust framework for enhancing explainability, particularly in domains requiring the interpretation of complex and ambiguous signals, such as brain-computer interface (BCI) systems. Despite significant advances in deep learning, interpreting human emotions remains a formidable challenge. In this work, we present iFuzzyAffectDuo, a novel computational model that integrates a dual-filter fuzzy neural network architecture for improved detection and interpretation of emotional states from neuroimaging data. The model introduces a new membership function (MF) based on the Laplace distribution, achieving superior accuracy and interpretability compared to traditional approaches. By refining the extraction of neural signals associated with specific emotions, iFuzzyAffectDuo offers a human-understandable framework that unravels the underlying decision-making processes. We validate our approach across three neuroimaging datasets using functional Near-Infrared Spectroscopy (fNIRS) and Electroencephalography (EEG), demonstrating its potential to advance affective computing. These findings open new pathways for understanding the neural basis of emotions and their application in enhancing human-computer interaction.

The method of Common Spatial Patterns (CSP) is widely used for feature extraction of electroencephalography (EEG) data, such as in motor imagery brain-computer interface (BCI) systems. It is a data-driven method estimating a set of spatial filters so that the power of the filtered EEG signal is maximized for one motor imagery class and minimized for the other. This method, however, is prone to overfitting and is known to suffer from poor generalization especially with limited calibration data. Additionally, due to the high heterogeneity in brain data and the non-stationarity of brain activity, CSP is usually trained for each user separately resulting in long calibration sessions or frequent re-calibrations that are tiring for the user. In this work, we propose a novel algorithm called Spectrally Adaptive Common Spatial Patterns (SACSP) that improves CSP by learning a temporal/spectral filter for each spatial filter so that the spatial filters are concentrated on the most relevant temporal frequencies for each user. We show the efficacy of SACSP in providing better generalizability and higher classification accuracy from calibration to online control compared to existing methods. Furthermore, we show that SACSP provides neurophysiologically relevant information about the temporal frequencies of the filtered signals. Our results highlight the differences in the motor imagery signal among BCI users as well as spectral differences in the signals generated for each class, and show the importance of learning robust user-specific features in a data-driven manner.

Diagnosing epilepsy is challenging when routine EEGs lack interictal epileptiform discharges (IEDs). Intermittent photic stimulation (IPS) and hyperventilation (HV) can increase diagnostic yield, but their interpretation is subjective. We propose a reproducible pipeline that classifies EEG recordings acquired during stimulation procedures, using machine-learning features spanning temporal, spectral, wavelet, and connectivity domains, and a stacked ensemble to combine complementary feature sets. Performance is evaluated with leave-one-subject-out (LOSO) cross-validation on the TUH Epilepsy Corpus and a clinical Erasmus MC (EMC) cohort, including IED-free analyses on TUH. On TUH, ensembles achieve up to 97.8\% AUC / 93.1\% BAC on IED-free resting-state EEG and 94.1\% AUC / 86.8\% BAC on IED-free IPS. On EMC, IPS provides the strongest discrimination (79.4\% AUC / 73.9\% BAC), while HV performance benefits from stratifying subjects by responsiveness. These results indicate that stimulation-evoked activity, particularly IPS, contains meaningful discriminative information for IED-free epilepsy classification and that multi-domain ensembling improves robustness.

Electroencephalogram (EEG) classification plays a key role in medical diagnosis and brain-computer interfaces, but remains challenging due to low signal-to-noise ratios and high inter-subject variability. As a result, many existing approaches rely on subject-specific models, which fail to exploit shared structure in neural signals and do not generalize to unseen subjects. To address these limitations, we propose LAtte, a framework that combines Lorentz attention with a hyperbolic InceptionTime-based encoder to improve cross-subject generalization in EEG classification. The model explicitly decomposes EEG signals into a learned baseline component and task-relevant deviations, enabling more structured representation learning. To further improve robustness and adaptability, we incorporate subject-specific low-rank adaptation (LoRA) modules at both encoder and decoder levels, augmented with a Lorentz boost-based LoRA mechanism and hyperbolic projection layers to reduce overfitting in geometric representations. We evaluate LAtte with and without finetuning in three settings: subject-specific, subject-conditional, and leave-one-subject-out (LOSO) on five established EEG datasets, achieving a consistent improvement in performance over current state-of-the-art methods for smaller datasets and maintaining performance for larger datasets.

Electroencephalography(EEG)-basedemotionrecognitionre- mains challenging in cross-subject settings due to severe inter-subject variability. Existing methods mainly learn subject-invariant features, but often under-exploit stimulus-locked group regularities shared across sub- jects. To address this issue, we propose the Group Resonance Network (GRN), which integrates individual EEG dynamics with offline group resonance modeling. GRN contains three components: an individual en- coder for band-wise EEG features, a set of learnable group prototypes for prototype-induced resonance, and a multi-subject resonance branch that encodes PLV/coherence-based synchrony with a small reference set. A resonance-aware fusion module combines individual and group-level rep- resentations for final classification. Experiments on SEED and DEAP under both subject-dependent and leave-one-subject-out protocols show that GRN consistently outperforms competitive baselines, while abla- tion studies confirm the complementary benefits of prototype learning and multi-subject resonance modeling.

A major issue in Motor Imagery Brain-Computer Interfaces (MI-BCIs) is their poor classification accuracy and the large amount of data that is required for subject-specific calibration. This makes BCIs less accessible to general users in out-of-the-lab applications. This study employed deep transfer learning for development of calibration-free subject-independent MI-BCI classifiers. Unlike earlier works that applied signal preprocessing and feature engineering steps in transfer learning, this study adopted an end-to-end deep learning approach on raw EEG signals. Three deep learning models (MIN2Net, EEGNet and DeepConvNet) were trained and compared using an openly available dataset. The dataset contained EEG signals from 55 subjects who conducted a left- vs. right-hand motor imagery task. To evaluate the performance of each model, a leave-one-subject-out cross validation was used. The results of the models differed significantly. MIN2Net was not able to differentiate right- vs. left-hand motor imagery of new users, with a median accuracy of 51.7%. The other two models performed better, with median accuracies of 62.5% for EEGNet and 59.2% for DeepConvNet. These accuracies do not reach the required threshold of 70% needed for significant control, however, they are similar to the accuracies of these models when tested on other datasets without transfer learning.

The cross-subject application of EEG-based brain-computer interface (BCI) has always been limited by large individual difference and complex characteristics that are difficult to perceive. Therefore, it takes a long time to collect the training data of each user for calibration. Even transfer learning method pre-training with amounts of subject-independent data cannot decode different EEG signal categories without enough subject-specific data. Hence, we proposed a cross-subject EEG classification framework with a generative adversarial networks (GANs) based method named common spatial GAN (CS-GAN), which used adversarial training between a generator and a discriminator to obtain high-quality data for augmentation. A particular module in the discriminator was employed to maintain the spatial features of the EEG signals and increase the difference between different categories, with two losses for further enhancement. Through adaptive training with sufficient augmentation data, our cross-subject classification accuracy yielded a significant improvement of 15.85% than leave-one subject-out (LOO) test and 8.57% than just adapting 100 original samples on the dataset 2a of BCI competition IV. Moreover, We designed a convolutional neural networks (CNNs) based classification method as a benchmark with a similar spatial enhancement idea, which achieved remarkable results to classify motor imagery EEG data. In summary, our framework provides a promising way to deal with the cross-subject problem and promote the practical application of BCI.

Use of the electroencephalogram (EEG) and machine learning approaches to recognize emotions can facilitate affective human computer interactions. However, the type of EEG data constitutes an obstacle for cross-individual EEG feature modelling and classification. To address this issue, we propose a deep-learning framework denoted as a dynamic entropy-based pattern learning (DEPL) to abstract informative indicators pertaining to the neurophysiological features among multiple individuals. DEPL enhanced the capability of representations generated by a deep convolutional neural network by modelling the interdependencies between the cortical locations of dynamical entropy based features. The effectiveness of the DEPL has been validated with two public databases, commonly referred to as the DEAP and MAHNOB-HCI multimodal tagging databases. Specifically, the leave one subject out training and testing paradigm has been applied. Numerous experiments on EEG emotion recognition demonstrate that the proposed DEPL is superior to those traditional machine learning (ML) methods, and could learn between electrode dependencies w.r.t. different emotions, which is meaningful for developing the effective human-computer interaction systems by adapting to human emotions in the real world applications.

Training for telerobotic systems often makes heavy use of simulated platforms, which ensure safe operation during the learning process. Outer space is one domain in which such a simulated training platform would be useful, as On-Orbit Operations (O3) can be costly, inefficient, or even dangerous if not performed properly. In this paper, we present a new telerobotic training simulator for the Canadarm2 on the International Space Station (ISS), which is able to modulate workload through the addition of confounding factors such as latency, obstacles, and time pressure. In addition, multimodal physiological data is collected from subjects as they perform a task from the simulator under these different conditions. As most current workload measures are subjective, we analyse objective measures from the simulator and EEG data that can provide a reliable measure. ANOVA of task data revealed which simulator-based performance measures could predict the presence of latency and time pressure. Furthermore, EEG classification using a Riemannian classifier and Leave-One-Subject-Out cross-validation showed promising classification performance and allowed for comparison of different channel configurations and preprocessing methods. Additionally, Riemannian distance and beta power of EEG data were investigated as potential cross-trial and continuous workload measures.

Emotion recognition based on EEG has become an active research area. As one of the machine learning models, CNN has been utilized to solve diverse problems including issues in this domain. In this work, a study of CNN and its spatiotemporal feature extraction has been conducted in order to explore capabilities of the model in varied window sizes and electrode orders. Our investigation was conducted in subject-independent fashion. Results have shown that temporal information in distinct window sizes significantly affects recognition performance in both 10-fold and leave-one-subject-out cross validation. Spatial information from varying electrode order has modicum effect on classification. SVM classifier depending on spatiotemporal knowledge on the same dataset was previously employed and compared to these empirical results. Even though CNN and SVM have a homologous trend in window size effect, CNN outperformed SVM using leave-one-subject-out cross validation. This could be caused by different extracted features in the elicitation process.

Robust decoding and classification of brain patterns measured with electroencephalography (EEG) remains a major challenge for real-world (i.e. outside scientific lab and medical facilities) brain-computer interface (BCI) applications due to well documented inter- and intra-participant variability. Here, we present a large-scale benchmark evaluating over 340,000+ unique combinations of spatial and nonlinear EEG classification. Our methodological pipeline consists in combinations of Common Spatial Patterns (CSP), Riemannian geometry, functional connectivity, and fractal- or entropy-based features across three open-access EEG datasets. Unlike prior studies, our analysis operates at the per-participant level and across multiple frequency bands (8-15 Hz and 8-30 Hz), enabling direct assessment of both group-level performance and individual variability. Covariance tangent space projection (cov-tgsp) and CSP consistently achieved the highest average classification accuracies. However, their effectiveness was strongly dataset-dependent, and marked participant-level differences persisted, particularly in the most heterogeneous of the datasets. Importantly, nonlinear methods outperformed spatial approaches for specific individuals, underscoring the need for personalized pipeline selection. Our findings highlight that no universal'one-size-fits-all'method can optimally decode EEG motor imagery patterns across all users or datasets. Future work will require adaptive, multimodal, and possibly novel approaches to fully address neurophysiological variability in practical BCI applications where the system can automatically adapt to what makes each user unique.

Cross-subject motor imagery decoding remains a fundamental challenge in EEG-based brain-computer interfaces due to substantial inter-subject variability. Recent approaches have leveraged Riemannian geometry by representing EEG signals as covariance matrices on the symmetric positive definite (SPD) manifold. However, existing methods primarily focus on manifold-based representations while largely overlooking subject-specific variations in covariance dispersion and orientation. In this work, we address these challenges through geometry-aware congruence transformations and propose three complementary models: (i) Discriminative Congruence Transform (DCT), (ii) Deep Linear DCT (DLDCT), and (iii) Deep DCT-UNet (DDCT-UNet). The proposed models are evaluated both as manifold alignment modules for downstream classifiers and as end-to-end discriminative architectures optimized via cross-entropy with a custom logistic regression head. Experiments on challenging cross-subject motor imagery benchmarks demonstrate consistent improvements in transductive decoding performance, achieving 2-3% higher accuracy than strong baselines. These results highlight the effectiveness of geometry-aware congruence learning for mitigating inter-subject variability in EEG decoding.

The application of Riemannian geometry in the decoding of brain-computer interfaces (BCIs) has swiftly garnered attention because of its straightforwardness, precision, and resilience, along with its aptitude for transfer learning, which has been demonstrated through significant achievements in global BCI competitions. This paper presents a comprehensive review of recent advancements in the integration of deep learning with Riemannian geometry to enhance EEG signal decoding in BCIs. Our review updates the findings since the last major review in 2017, comparing modern approaches that utilize deep learning to improve the handling of non-Euclidean data structures inherent in EEG signals. We discuss how these approaches not only tackle the traditional challenges of noise sensitivity, non-stationarity, and lengthy calibration times but also introduce novel classification frameworks and signal processing techniques to reduce these limitations significantly. Furthermore, we identify current shortcomings and propose future research directions in manifold learning and riemannian-based classification, focusing on practical implementations and theoretical expansions, such as feature tracking on manifolds, multitask learning, feature extraction, and transfer learning. This review aims to bridge the gap between theoretical research and practical, real-world applications, making sophisticated mathematical approaches accessible and actionable for BCI enhancements.

Objective: Motor Imagery (MI) serves as a crucial experimental paradigm within the realm of Brain Computer Interfaces (BCIs), aiming to decoding motor intentions from electroencephalogram (EEG) signals. Method: Drawing inspiration from Riemannian geometry and Cross-Frequency Coupling (CFC), this paper introduces a novel approach termed Riemann Tangent Space Mapping using Dichotomous Filter Bank with Convolutional Neural Network (DFBRTS) to enhance the representation quality and decoding capability pertaining to MI features. DFBRTS first initiates the process by meticulously filtering EEG signals through a Dichotomous Filter Bank, structured in the fashion of a complete binary tree. Subsequently, it employs Riemann Tangent Space Mapping to extract salient EEG signal features within each sub-band. Finally, a lightweight convolutional neural network is employed for further feature extraction and classification, operating under the joint supervision of cross-entropy and center loss. To validate the efficacy, extensive experiments were conducted using DFBRTS on two well-established benchmark datasets: the BCI competition IV 2a (BCIC-IV-2a) dataset and the OpenBMI dataset. The performance of DFBRTS was benchmarked against several state-of-the-art MI decoding methods, alongside other Riemannian geometry-based MI decoding approaches. Results: DFBRTS significantly outperforms other MI decoding algorithms on both datasets, achieving a remarkable classification accuracy of 78.16% for four-class and 71.58% for two-class hold-out classification, as compared to the existing benchmarks.

Modelling dynamically evolving spatio-temporal signals is a prominent challenge in the Graph Neural Network (GNN) literature. Notably, GNNs assume an existing underlying graph structure. While this underlying structure may not always exist or is derived independently from the signal, a temporally evolving functional network can always be constructed from multi-channel data. Graph Variate Signal Analysis (GVSA) defines a unified framework consisting of a network tensor of instantaneous connectivity profiles against a stable support usually constructed from the signal itself. Building on GVSA and tools from graph signal processing, we introduce Graph-Variate Neural Networks (GVNNs): layers that convolve spatio-temporal signals with a signal-dependent connectivity tensor combining a stable long-term support with instantaneous, data-driven interactions. This design captures dynamic statistical interdependencies at each time step without ad hoc sliding windows and admits an efficient implementation with linear complexity in sequence length. Across forecasting benchmarks, GVNNs consistently outperform strong graph-based baselines and are competitive with widely used sequence models such as LSTMs and Transformers. On EEG motor-imagery classification, GVNNs achieve strong accuracy highlighting their potential for brain-computer interface applications.

Motor imagery (MI) based brain-computer interfaces (BCIs) hold significant potential for assistive technologies and neurorehabilitation. However, the precise and efficient decoding of MI remains challenging due to their non-stationary nature and low signal-to-noise ratio. This paper introduces a novel end-to-end deep learning framework of Discriminative Residual Dense Convolutional Autoencoder with Spatio-Temporal Graph Neural Network (DRDCAE-STGNN) to enhance the MI feature learning and classification. Specifically, the DRDCAE module leverages residual-dense connections to learn discriminative latent representations through joint reconstruction and classifica-tion, while the STGNN module captures dynamic spatial dependencies via a learnable graph adjacency matrix and models temporal dynamics using bidirectional long short-term memory (LSTM). Extensive evaluations on BCI Competition IV 2a, 2b, and PhysioNet datasets demonstrate state-of-the-art performance, with average accuracies of 95.42%, 97.51%, and 90.15%, respectively. Ablation studies confirm the contribution of each component, and interpreta-bility analysis reveals neurophysiologically meaningful connectivity patterns. Moreover, despite its complexity, the model maintains a feasible parameter count and an inference time of 0.32 ms per sample. These results indicate that our method offers a robust, accurate, and interpretable solution for MI-EEG decoding, with strong generalizability across subjects and tasks and meeting the requirements for potential real-time BCI applications.

Alzheimer's Disease is a progressive neurological disorder that is one of the most common forms of dementia. It leads to a decline in memory, reasoning ability, and behavior, especially in older people. The cause of Alzheimer's Disease is still under exploration and there is no all-inclusive theory that can explain the pathologies in each individual patient. Nevertheless, early intervention has been found to be effective in managing symptoms and slowing down the disease's progression. Recent research has utilized electroencephalography (EEG) data to identify biomarkers that distinguish Alzheimer's Disease patients from healthy individuals. Prior studies have used various machine learning methods, including deep learning and graph neural networks, to examine electroencephalography-based signals for identifying Alzheimer's Disease patients. In our research, we proposed a Flexible and Explainable Gated Graph Convolutional Network (GGCN) with Multi-Objective Tree-Structured Parzen Estimator (MOTPE) hyperparameter tuning. This provides a flexible solution that efficiently identifies the optimal number of GGCN blocks to achieve the optimized precision, specificity, and recall outcomes, as well as the optimized area under the Receiver Operating Characteristic (AUC). Our findings demonstrated a high efficacy with an over 0.9 Receiver Operating Characteristic score, alongside precision, specificity, and recall scores in distinguishing health control with Alzheimer's Disease patients in Moderate to Severe Dementia using the power spectrum density (PSD) of electroencephalography signals across various frequency bands. Moreover, our research enhanced the interpretability of the embedded adjacency matrices, revealing connectivity differences in frontal and parietal brain regions between Alzheimer's patients and healthy individuals.

Objective The electrical characteristics of the EEG signals can be used for seizure detection. Statistical independence between different brain regions is measured by functional brain connectivity (FBC). Specific directional effects can't consider by FBC and thus effective brain connectivity (EBC) is used to measure causal intervention between one neuronal region and the rest of the neuronal regions. Our main purpose is to provide a reliable automatic seizure detection approach. Methods In this study, three new methods are provided. Deep modular neural network (DMNN) is developed based on a combination of various EBC classification results in the different frequencies. Another method is named "modular effective neural networks (MENN)". This method combines the classification results of the three different EBC in the specific frequency. "Modular frequency neural networks (MFNN)" is another method that combines the classification results of the specific EBC in the seven different frequencies. Results The mean accuracy of the MFNN are 97.14%, 98.53%, and 97.91% using directed transfer function, directed coherence, and generalized partial directed coherence, respectively. Using the MENN, the highest mean accuracy is 98.34%. Finally, DMNN has the highest mean accuracy which is equal to 99.43. To our best knowledge, the proposed method is a new method that provides the high accuracy in comparison to other studies which used MIT-CHB database. Conclusion and significance The knowledge of structure-function relationships between different areas of the brain is necessary for characterizing the underlying dynamics. Hence, features based on EBC can provide a reliable automatic seizure detection approach.