Electroencephalography (EEG)-based motor imagery (MI) brain–computer interfaces (BCIs) hold significant potential for applications in neuro-rehabilitation and assistive technologies. Yet, their development remains constrained by challenges such as low spatial resolution, vulnerability to noise and artifacts, and pronounced inter-subject variability. Conventional approaches, including common spatial patterns (CSP) and convolutional neural networks (CNNs), often exhibit limited robustness, weak generalization, and reduced interpretability. To overcome these limitations, we introduce EEG-GCIRNet, a Gaussian connectivity-driven EEG imaging representation network coupled with a regularized LeNet architecture for MI classification. Our method integrates raw EEG signals with topographic maps derived from functional connectivity into a unified variational autoencoder framework. The network is trained with a multi-objective loss that jointly optimizes reconstruction fidelity, classification accuracy, and latent space regularization. The model’s interpretability is enhanced through its variational autoencoder design, allowing for qualitative validation of its learned representations. Experimental evaluations demonstrate that EEG-GCIRNet consistently outperforms both baseline models and state-of-the-art MI classification methods. Beyond classification accuracy, analysis of the reconstructed connectivity maps and latent space visualizations confirms that the model captures physiologically relevant patterns and learns a well-disentangled feature space. These findings suggest that EEG-GCIRNet provides a robust and interpretable end-to-end framework for EEG-based BCIs, advancing the development of reliable neurotechnology for rehabilitation and assistive applications.

Electroencephalography (EEG)-based motor imagery (MI) brain–computer interfaces (BCIs) hold considerable potential for applications in neuro-rehabilitation and assistive technologies. Yet, their development remains constrained by challenges such as low spatial resolution, vulnerability to noise and artifacts, and pronounced inter-subject variability. Conventional approaches, including common spatial patterns (CSP) and convolutional neural networks (CNNs), often exhibit limited robustness, weak generalization, and reduced interpretability. To overcome these limitations, we introduce EEG-GCIRNet, a Gaussian connectivity-driven EEG imaging representation network coupled with a regularized LeNet architecture for MI classification. Our method integrates raw EEG signals with topographic maps derived from functional connectivity into a unified variational autoencoder framework. The network is trained with a multi-objective loss that jointly optimizes reconstruction fidelity, classification accuracy, and latent space regularization. The model’s interpretability is enhanced through its variational autoencoder design, allowing for qualitative validation of its learned representations. Experimental evaluations demonstrate that EEG-GCIRNet outperforms state-of-the-art methods, achieving the highest average accuracy (81.82%) and lowest variability (±10.15) in binary classification. Most notably, it effectively mitigates BCI illiteracy by completely eliminating the “Bad” performance group (<60% accuracy), yielding substantial gains of ∼22% for these challenging users. Furthermore, the framework demonstrates good scalability in complex 5-class scenarios, performing competitive classification accuracy (75.20% ± 4.63) with notable statistical superiority (p = 0.002) against advanced baselines. Extensive interpretability analyses, including analysis of the reconstructed connectivity maps, latent space visualizations, Grad-CAM++ and functional connectivity patterns, confirm that the model captures genuine neurophysiological mechanisms, correctly identifying integrated fronto-centro-parietal networks in high performers and compensatory midline circuits in mid-performers. These findings suggest that EEG-GCIRNet provides a robust and interpretable end-to-end framework for EEG-based BCIs, advancing the development of reliable neurotechnology for rehabilitation and assistive applications.

As a cognitive process, motor imagery (MI) includes mentally simulating motor actions in the absence of physical movement. It has a variety of uses, including assistive technologies, medical diagnostics, and rehabilitation. MI paradigms are utilized in conjunction with brain-computer interfaces (BCI), which use electroencephalographic recordings (EEG) because of their high temporal resolution, cheap cost, portability, and non-invasiveness. BCIs apply MI paradigms by directly connecting the human brain to a computer. However, because scalp readings are non-stationary and non-linear, real-time processing of EEG signals is challenging. Furthermore, in order to minimize the impact of outside noise and artifacts, clinical MI methods must be implemented under carefully monitored laboratory conditions. A deep learning model-based approach is shown for analyzing EEG data and giving real-time feedback to a brain-computer interface. Generally, the system's design is portable and low-cost, allowing the MI paradigm to perform under poorly regulated sampling conditions.

Reliable decoding of motor imagery (MI) from electroencephalographic signals remains a challenging problem due to their nonlinear, noisy, and non-stationary nature. To address this issue, this work proposes an end-to-end deep learning model, termed TEKTE-Net, that integrates time embeddings with a kernelized Transfer Entropy estimator to infer directed functional connectivity in MI-based brain–computer interface (BCI) systems. The proposed model incorporates a customized convolutional module that performs Takens' embedding, enabling the decoding of the underlying EEG activity without requiring explicit preprocessing. Further, the architecture estimates nonlinear and time-delayed interactions between cortical regions using Rational Quadratic kernels within a differentiable framework. Evaluation of TEKTE-Net on semi-synthetic causal benchmarks and the BCI Competition IV 2a dataset demonstrates robustness to low signal-to-noise conditions and interpretability through temporal, spatial, and spectral analyses of learned connectivity patterns. In particular, the model automatically highlights contralateral activations during MI and promotes spectral selectivity for the beta and gamma bands. Overall, TEKTE-Net offers a fully trainable estimator of functional brain connectivity for decoding EEG activity, supporting MI-BCI applications, and promoting interpretability of deep learning models.

Reliable decoding of motor imagery (MI) from electroencephalographic signals remains a challenging problem due to their nonlinear, noisy, and non-stationary nature. To address this issue, this work proposes an end-to-end deep learning model, termed TEKTE-Net, that integrates time embeddings with a kernelized Transfer Entropy estimator to infer directed functional connectivity in MI-based brain–computer interface (BCI) systems. The proposed model incorporates a customized convolutional module that performs Takens’ embedding, enabling the decoding of the underlying EEG activity without requiring explicit preprocessing. Further, the architecture estimates nonlinear and time-delayed interactions between cortical regions using Rational Quadratic kernels within a differentiable framework. Evaluation of TEKTE-Net on semi-synthetic causal benchmarks and the BCI Competition IV 2a dataset demonstrates robustness to low signal-to-noise conditions and interpretability through temporal, spatial, and spectral analyses of learned connectivity patterns. In particular, the model automatically highlights contralateral activations during MI and promotes spectral selectivity for the beta and gamma bands. Overall, TEKTE-Net offers a fully trainable estimator of functional brain connectivity for decoding EEG activity, supporting MI-BCI applications, and promoting interpretability of deep learning models.

Attention Deficit Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental condition commonly assessed through clinical interviews, behavioral observation, and rating scales. Although electroencephalography (EEG) has emerged as a promising complementary tool for ADHD assessment, robust subject-independent classification remains challenging due to inter-subject variability, limited datasets, and the need for interpretable computational models. This work introduces EEG-TACT, a compact end-to-end deep learning architecture for identifying ADHD subjects from EEG epochs. The proposed model integrates an EEGNet-inspired convolutional embedding, a Transformer encoder operator, and an attention-based pooling mechanism to jointly capture local spatiotemporal EEG patterns, contextual temporal dependencies, and task-relevant latent representations. EEG-TACT was evaluated on a publicly available EEG dataset using strict, subject-independent stratified group partitions, ensuring that data from the same subject were never shared across the training, validation, and test subsets. Model interpretability is examined through learned temporal filter responses, class-conditioned self-attention maps, and latent-space projections, while an ablation study quantifies the contribution of each architectural component. Performance was assessed at the fold, subject, and epoch levels, together with statistical significance comparisons against representative state-of-the-art architectures. EEG-TACT outperformed the contrasted models, achieving subject-level accuracy of 87.5%, recall of 96.0%, and precision of 82.8%, while requiring only a few thousand trainable parameters. By exhaustively repeating the initialization, the proposed model demonstrated improved labeling reliability and achieved the best average ranking among the evaluated architectures. The reported results therefore prove that EEG-TACT provides a compact, robust, and interpretable model for EEG-based ADHD identification under subject-independent evaluation settings.