Real-time attention monitoring using physiological signals is critical for next-generation wearable healthcare and human–computer interaction systems. However, most existing approaches rely on computationally intensive models limiting their applicability in resource-constrained edge environments. This paper presents a hardware-aware multimodal deep learning framework for real-time attention-state classification using photoplethysmography (PPG) and galvanic skin response (GSR) signals. The proposed approach integrates a lightweight convolutional neural network (CNN) with Bayesian hyperparameter optimization using Optuna’s Tree-structured Parzen Estimator (TPE)-driven tuning to jointly optimize classification performance and embedded deployment efficiency. A novel end-to-end pipeline is developed that encompasses signal preprocessing, multimodal encoding, model optimization, and on-device inference on a low-power nRF52840 microcontroller-based platform. Offline evaluation on the PhyAAt dataset achieves a validation accuracy of 90.28%. In a preliminary single-subject on-device evaluation, the deployed INT8 model achieved 81.48% real-time accuracy (95% binomial exact confidence interval computed using the Clopper–Pearson method) with an average inference latency of 118 ms under strict memory constraints (48 kB Flash, 27 kB SRAM). The moderate performance degradation between offline and real-time settings demonstrates the robustness and practical feasibility of the proposed framework. Compared to existing multimodal and edge-AI approaches, the proposed system explicitly balances accuracy, latency, and memory footprint, enabling continuous on-device attention monitoring without reliance on cloud infrastructure. Thiswork establishes a scalable pathway toward low-power, privacy-preserving, and real-time cognitive state monitoring in wearable and IoT-based healthcare systems.

This literature review critically examines the design, validation, and application of non-invasive in-ear electroencephalography (ear-EEG) systems as emerging wearable platforms for long-term neurophysiological monitoring and intervention. Following PRISMA guidelines, studies published between 2010 and 2025 were systematically selected from four major databases and organized into four thematic domains: in-ear wearable system design and validation, multimodal sensing and stimulation, embedded intelligence, and brain-state monitoring and rehabilitation. The review focuses exclusively on wearable, ear-centered EEG technologies, explicitly excluding cochlear implants and other invasive or behind-the-ear systems. We analyze key engineering challenges unique to ear-EEG, including electrode placement constraints, mechanical–electrical coupling, motion robustness, power efficiency, and long-term wearability. The review highlights a growing transition toward compact, wireless ear-EEG systems with on-device signal processing and embedded machine learning, enabling real-time brain-state estimation under ambulatory conditions. Multimodal integration, combining ear-EEG with complementary sensors such as EOG, inertial units, and cardiovascular signals is shown to improve artifact awareness, contextual interpretation, and closed-loop capability. Beyond summarizing existing technologies, this review identifies critical gaps limiting clinical translation, including the lack of standardized validation protocols, limited embedded autonomy, and underexplored closed-loop neurofeedback and neuromodulation architectures. By synthesizing advances across hardware design, signal processing, and intelligent system integration, this work provides a systems-level roadmap for the future development of wearable, intelligent, and clinically robust ear-EEG platforms for mental health, neurorehabilitation, and continuous brain monitoring.

IntroductionDistributed electroencephalography (EEG) brain-computer interface (BCI) systems increasingly transmit neural data and control messages between a head-worn device, a nearby hub, and cloud-assisted services. These links require integrity, authenticity, replay protection, and low latency, while migration to post-quantum cryptography can increase session-establishment cost.MethodsWe present PQ-NeuroLink, a latency-aware post-quantum-ready communication framework that separates authenticated session establishment from the symmetric streaming fast path. The design supports classical X25519, post-quantum ML-KEM-768, and hybrid ML-KEM-768 plus X25519 key establishment, with pinned ML-DSA public keys for endpoint identity binding. The prototype was evaluated using EEG-derived traffic from public datasets over BLE-like and Wi-Fi-like link profiles, 1-hop and 2-hop topologies, symmetric key-update and public-key refresh events, active tampering scenarios, and MCU-informed resource and processing-energy projections.ResultsIdentity pinning and transcript binding prevented successful modeled session-splicing attempts. Post-quantum and hybrid modes primarily increased handshake bytes rather than steady-state streaming latency. In the most constrained BLE 1-hop summary condition, the hybrid M3 mode increased p95 latency by 0.45 ms relative to the unsecured M0 baseline while maintaining 99.0% packet delivery.DiscussionPost-quantum-ready authenticated session establishment can be integrated into distributed EEG communication prototypes when public-key operations are kept off the per-frame path. PQ-NeuroLink provides a reproducible communication-layer basis for secure next-generation BCI deployment studies requiring low-latency operation and long-horizon post-quantum protection.

Social conversational robots are emerging as significant instruments in healthcare, capable of improving patient experiences and operational efficiency. This research examines the integration of the Pepper robot with ChatGPT and hospitalspecific data to enhance interactions in hospital waiting areas. The enhanced Pepper robot utilizes advanced natural language processing and real-time contextual data to deliver timely, empathetic, and context-aware responses, effectively addressing patient anxiety and disengagement issues. The results indicate that incorporating conversational Artificial Intelligence (AI) enhances the robot's capabilities, allowing for multilingual, personalized, and adaptive interactions. This research emphasizes the insufficient academic investigation of the Pepper robot, noting that only two studies have examined its application and highlighting the necessity for focused initiatives to enhance its utilization in healthcare. A framework is proposed for a systematic approach to integrate ChatGPT prompts with hospital databases, facilitating seamless interaction workflows. Feedback loops are integrated to facilitate ongoing enhancement of the robot's functionality while ensuring alignment with the hospital's specific requirements. This study establishes a basis for developing social conversational robots, facilitating innovative approaches to transforming patient care and healthcare interactions.

We investigate the neural mechanisms through which music influences mental fatigue using electroencephalography (EEG) and deep learning. We collected EEG data from 49 participants during periods with and without musical auditory stimulation, categorized by mental fatigue status based on standardized questionnaires. Following preprocessing and feature extraction, we implemented three deep learning architectures (Convolutional Neural Network [CNN], Gated Recurrent Unit [GRU], Long Short-Term Memory [LSTM]) to decode and classify four distinct conditions: with/without music and with/without mental fatigue. The GRU model achieved superior classification performance (96.92% accuracy), followed closely by LSTM (96.03%), while CNN showed significantly lower performance (58.97%). The high decoding accuracy demonstrates that music exposure produces distinctly identifiable neural patterns that can be differentiated from those associated with mental fatigue. The superior performance of recurrent architectures highlights the importance of modeling temporal dependencies in EEG data associated with these cognitive states. Our findings provide neural evidence supporting music’s potential as an intervention for mental fatigue and establish a foundation for developing personalized music-based strategies to enhance cognitive well-being.

BackgroundThis paper addresses a critical challenge in developing practical EEG-based brain-computer interfaces (BCIs): enhancing cross-subject generalization by mitigating individual differences in brain signals. How can we effectively leverage data from existing subjects to improve performance for a new user with minimal subject-specific calibration?MethodsWe systematically compare and optimize three prominent data alignment techniques, Riemannian Procrustes Analysis (RPA), Euclidean Alignment (EA), and Correlation Alignment (CORAL), designed to transform EEG data from multiple source subjects and a target subject into a common representation space, mitigating variability.EvaluationWe employed leave-one-subject-out cross-validation (LOSO-CV) framework on EEG-based attention decoding data to empirically evaluate the effectiveness of each alignment method compared to a baseline condition with no alignment. Key parameters, specifically the regularization parameter α for EA, were optimized to maximize cross-subject transfer performance.ResultsThe study demonstrates that alignment methods improve classification accuracy compared to the baseline. Notably, EA evaluated at α = 100 the scaling value at which the largest fraction of subjects attained their best accuracy in our parameter sweep yielded the largest mean improvement, increasing classification accuracy by 3.44% over the no alignment baseline (paired t(17)≈2.48, p≈0.024; Cohen's dz≈0.59; 95% confidence interval for the mean improvement [0.52%, 6.36%]). Because this α value was identified from the same sweep that produced the per-subject accuracies, this estimate together with the per-subject “best-parameter” results should be interpreted as an oracle sensitivity-analysis upper bound on subject-specific tuning rather than as a leakage-free LOSO estimate. While optimized EA showed the best mean performance, the analysis also demonstrated subject-specific differences in the most ideal alignment strategy.ConclusionThis comparison framework quantifies the benefits of different alignment approaches and highlights the valuable contribution of parameter optimization, particularly for EA.SignificanceThese results indicate the potential of optimized alignment techniques, EA in particular, to significantly enhance cross-subject transfer learning in EEG-based BCIs. This has practical ramifications for methodology selection and tuning, and maps a path toward more robust and generalizable BCI systems requiring less subject-specific calibration for real-world applications.

Brain computer interface (BCI) enables the brain to directly control an external device by converting neural signals into actionable outputs. However, effective real-time translation of brain activity strongly depends on the quality of neural communication between the brain and the external device. 6G is the next generation of wireless communication, expected to provide unprecedented levels of data rates, data security, and automation capabilities. In this context, integrating 6G into BCI systems would not only enhance the performance of brain-device communication, but would also create new opportunities for innovative applications. This work provides a comprehensive study on how BCI technology can be built effectively on top of 6G wireless networks by introducing several technical aspects and use cases. We first provide an overview of BCI and 6G, following their progression from early development to convergence through cognitive communication and advanced neural interfaces. We then highlight the need for the upcoming 6G systems toward BCI technology in every aspect, including 6G technologies such as intelligent edge and zero-touch networks, and 6G use cases such as digital twin, immersive communication, and internet of minds. Furthermore, we identify key technical challenges, open issues, and future research directions related to the 6G-enabled BCI paradigm.

Electroencephalography (EEG) is a scientific technique used to analyze and decode brain activity to identify different neurodegenerative diseases and psychiatric disorders. The insufficiency of specialists and neurologists is attributed to the prevalence of intricate disorders. Prompt and accurate diagnosis is crucial in managing any healthcare issue. EEG signal has shown its value in the prompt identification and prediction of diseases, hence providing valuable assistance to specialists and neurologists. This systematic review examines and analyzes the use of EEG signals for the diagnosis of different disorders and the identification of their biomarkers, drawing on existing research. The research aims to examine the use of EEG signals in the diagnostic process and determine whether a disorder may be predicted using EEG signals in advance. This raises novel research inquiries, which are investigated via 100 academic publications. Findings: EEG signals have been shown to be effective not only for early-stage diagnosis but also for predicting diseases and disorders before they arise, such as Bipolar detection, cancer detection, and the prediction of harmful brain activities. The results indicate that EEG signals can provide novel prospects for more exploration, such as the prediction and prognosis of diseases.

Federated Learning (FL) allows healthcare organizations to train models using diverse datasets while maintaining patient confidentiality collaboratively. While promising, FL faces challenges in optimizing model accuracy and communication efficiency. To address these, we propose an algorithm that combines feature selection with synthetic data generation, specifically targeting medical datasets. Our method eliminates irrelevant local features, identifies globally relevant ones, and uses synthetic data to initialize model parameters, improving convergence. It also employs a zero-trust model, ensuring that data remain on local devices and only learned weights are shared with the central server, enhancing security. The algorithm improves accuracy and computational efficiency, achieving communication efficiency gains of 4 to 14 through backward elimination and threshold variation techniques. Tested on a federated diabetic dataset, the approach demonstrates significant improvements in the performance and trustworthiness of FL systems for medical applications.

This work presents a novel hybrid deep learning architecture combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) for automated mental health assessment using electroencephalogram (EEG) signals. The proposed framework addresses the limitations of traditional deep learning approaches by integrating hierarchical feature extraction with transformer-based sequence modeling. Using a dataset of 48 subjects and a 16-channel dry electrode EEG system, our model simultaneously classifies anxiety, depression, and stress conditions. To address data imbalance, we implement a Conditional Generative Adversarial Network (cGAN) for synthetic data generation. The model achieves remarkable classification performance on the held-out test set, with accuracies of 95.77%, 90.21%, and 90.48% for anxiety, depression, and stress, respectively. SHapley Additive exPlanations (SHAP) analysis reveals distinct EEG feature patterns associated with each condition, providing interpretable insights into the model's decision-making process. Our results demonstrate the potential of hybrid CNN-ViT architectures in advancing automated mental health diagnostics through EEG analysis.

Brain computer interface (BCI) technologies have moved from lab tests to becoming key players in medical, commercial, and consumer markets. This shift raises important questions about ethics, privacy, security, and real-world use. A review of 29 research studies from 2023 to 2025 highlights four main areas in BCI development: ethical frameworks, potential dual-use issues, privacy protection, and security risks, along with market needs for BCI. Even with rapid advancements in neural data processing and classification, significant ethical issues still exist across different cultures. Recent studies indicate effective ways to protect privacy using federated learning methods and perturbative techniques that maintain utility while ensuring privacy. The survey points out major and ongoing market failures in standardized security practices, ethical standards relevant to various cultures, and user-friendly BCI systems that cater to different user needs. Additionally, security researchers have found vulnerabilities at every stage of the BCI process, from signal acquisition to signal classification. The article wraps up by suggesting future BCI developments that balance progress and privacy, advocating for thoughtful governance, robust privacy measures, and teamwork across fields to address the complex issues that lie at the intersection of neuroscience, computer science, and ethics.

The use of electroencephalography (EEG)-based analysis in educational research offers valuable insight into the dynamics of virtual and robot-based tutoring systems. This study aims to classify neural patterns associated with two distinct tutoring modalities (virtual tutor versus robot tutor) to understand how different educational interfaces affect brain activity and engagement biomarkers. This classification enables adaptive learning systems that can optimize educational environments based on real-time neural responses, potentially enhancing learning outcomes by matching tutoring modalities to individual neural preferences. We examine the performance of multiple EEG features such as power spectral density, Fast Fourier Transform (FFT) magnitude coefficients, amplitude and variance, coherence, Hjorth parameters, and wavelet transform coefficients in distinguishing between these two tutoring modalities. A transformer-based binary classifier was employed to evaluate the effectiveness of these characteristics in classifying EEG data collected during interactions with both types of tutors. The wavelet transform coefficients demonstrated the highest classification accuracy of 98.75%, precision of 99.17%, recall of 98.18%, F1 score of 98.61%, and Area Under the Curve (AUC) of 1.00, indicating exceptional performance in all metrics. This high accuracy represents a significant advancement in EEG-based educational classification systems. The Hjorth parameters (90.13% accuracy, 89.91% precision, 95.45% recall) and preprocessed EEG (91.73% accuracy, 91.57% precision, 90.91% recall) also showed strong performance. Simpler features, such as amplitude and variance, exhibited limited discriminatory power with only 72.67% accuracy. These findings underscore the importance of feature engineering and robust preprocessing in EEG-based educational studies and demonstrate the feasibility of creating personalized learning environments that can adapt to individual students’ neural preferences. Limitations such as a small sample size and topic-specific focus are noted, paving the way for future research to generalize findings and optimize methodologies.

Magnetotellurics (MT) investigates the subsurface resistivity from naturally occurring electromagnetic field fluctuations. The linearity and noise in MT data cause traditional inversion methods to fail, resulting in poor resistivity reconstruction. While convolutional and recurrent neural networks have been applied to invert MT, they cannot be used to include complex, non-local relationships over depth and frequency. In this research, we propose a hybrid model that combines graph attention networks (GATs) and multihead attention mechanisms. The GAT module serves as both denoiser and encoder, structuring the input frequency–phase data into a compressed representation, and the transformer-based attention selectively emphasizes the most informative spectral elements. This allows the model to predict resistivity distributions more accurately with depth. Experiments on both synthetic and field datasets demonstrate that the proposed model outperforms the Convolutional Neural Network (CNN) Mean Squared Error (MSE) (0.0030), with correlation coefficient R = 0.78, and Vanilla Neural Network (VNN) MSE (0.0173), with correlation coefficient R = 0.71 baselines with a preeminence of MSE of 0.0018 and a correlation coefficient of 0.85. The integration of graph attention with frequency-aware multihead attention allows the current model to capture the underlying physics of MT data in a far more effective way and provides resistivity profiles that are accurate and less noisy. This graphical abstract summarizes a study on the one-dimensional (1D) inversion of magnetotelluric (MT) data using a novel deep learning model that combines Graph Attention Networks (GATs) with multi-head attention mechanisms. The GATs denoise and encode the MT data by structuring it as a graph, which captures dependencies between frequencies. The multihead attention component then dynamically focuses on the most informative frequency elements to predict subsurface resistivity distributions more accurately with depth. Evaluated on both synthetic and real-world datasets from the United Arab Emirates, the model demonstrated superior robustness to noise and better generalization than current deep learning models. This approach offers a more scalable, interpretable, and physically informed solution for MT inversion. A novel deep learning model using Graph Attention Networks (GATs) and multi-head attention is proposed for 1D MT inversion. The novel ML method enhanced robustness to noise and better generalization compared to existing deep learning models in MT inversion. The proposed GAT-attention model offers a more interpretable and scalable solution for inverting MT data.

In this paper, we examine the effectiveness of neurofeedback in improving attention in remote learning environments using a passive brain computer interface (BCI). Participants had three EEG-based neurofeedback training sessions, receiving immediate feedback on their cognitive states to improve attention and diminish inattentiveness by self-regulating brain activity. The results indicated substantial enhancements in attention during the sessions, as shown by elevated beta and gamma activity, which are essential for attention regulation and cognitive processing. Power spectral density (PSD) measurement demonstrated sustained increased brainwave power over time, corroborating individuals' capacity to modify brain activity successfully. An analysis of variance (ANOVA) with a single factor revealed a significant increase in attention levels throughout sessions (F-(2,F-,F- 27) = 95.06 , p > 0.0001), with post-hoc Tukey's tests demonstrating considerable improvements between Sessions 1 and 3 (mean difference = 5.261, p = 0.001). The beta and gamma activity study corroborated the participants' cognitive engagement enhancement, evidenced by less inattentiveness and improved higher-order processing. The negligible latency in processing real-time feedback guaranteed that the feedback stayed pertinent to the task. This study underscores the efficacy of neurofeedback as a valuable instrument for improving attention and cognitive engagement in online learning settings. Future studies should investigate long-term impacts, individualized feedback systems, and the incorporation of additional cognitive training methodologies to optimize the advantages of neurofeedback in educational settings.

The use of electroencephalography (EEG)-based analysis in educational research offers valuable insight into the dynamics of virtual- and robot-based tutoring systems. This study aims to classify neural patterns associated with two distinct tutoring modalities (virtual tutor vs. robot tutor) to understand how different educational interfaces affect brain activity and engagement. Such classification enables adaptive learning systems that can optimize educational environments based on real-time neural responses, potentially enhancing learning outcomes by matching tutoring modalities to individual neural preferences. We examine the performance of multiple EEG features such as power spectral density, Fast Fourier Transform (FFT) magnitude coefficients, amplitude and variance, coherence, Hjorth parameters, and wavelet transform coefficients in distinguishing between these two tutoring modalities. A transformer-based binary classifier was employed to evaluate the effectiveness of these features in classifying EEG data collected during interactions with both types of tutors. The wavelet transform coefficients demonstrated the highest classification accuracy of 98.75%, precision of 99.17%, recall of 98.18%, F1 score of 98.61%, and Area Under the Curve (AUC) of 1.00, indicating exceptional performance across all metrics. This high accuracy represents a significant advancement in EEG-based educational classification systems. Hjorth parameters (90.13% accuracy, 89.91% precision, 95.45% recall) and preprocessed EEG (91.73% accuracy, 91.57% precision, 90.91% recall) also showed strong performance. Simpler features, such as amplitude and variance, exhibited limited discriminatory power with only 72.67% accuracy. These findings underscore the importance of feature engineering and robust preprocessing in EEG-based educational studies and demonstrate the feasibility of creating personalized learning environments that can adapt to individual students’ neural preferences. Limitations such as a small sample size and topic-specific focus are noted, paving the way for future research to generalize findings and optimize methodologies.

Federated learning (FL)–a distributed machine learning that offers collaborative training of global models across multiple clients. FL has been considered for the design and development of many FL systems in various domains. Hence, we present a comprehensive survey and analysis of existing FL systems, drawing insights from more than 250 articles published in 2019-2024. Our review elucidates the functioning of FL systems, particularly in comparison with alternative distributed learning approaches. Considering the healthcare domain as an example, we define the building blocks of a typical FL healthcare system, including system architecture, federation scale, data partitioning, open-source frameworks, ML models, and aggregation algorithms. Furthermore, we identify and discuss key challenges associated with the design and implementation of FL systems within the healthcare sector while outlining the directions of future research. In general, through systematic categorization and analysis of existing FL systems, we offer insights to design efficient, accurate, and privacy-preserving healthcare applications using cutting-edge FL techniques.

This work proposes a deep-learning framework for the real-time auditory attention classification using multimodal physiological signals, integrating electroencephalogram (EEG) and photoplethysmogram (PPG) signals. The architecture combines a multi-scale one-dimensional (1D) convolutional backbone to capture features across varying temporal resolutions, followed by a channel-wise attention mechanism using a Squeeze-and-Excite (SE) block to enhance feature selectivity. A Bi-directional Gated Recurrent Unit (Bi-GRU) is incorporated to model sequential dependencies within the multimodal time-series data. The trained model is quantised, optimised, and deployed on the STM32L496G-DISCO microcontroller using the X-CUBE-AI toolchain. Deployment on board demonstrated a low memory footprint (RAM: 57.91 KiB, Flash: 31.47 KiB) and real-time inference capability with an average latency of 476.9 ms and accuracy of 79.75%. The proposed model achieved an overall classification accuracy of 93.39%, Sensitivity of 88.72%, Precision of 91.69%, and F1-score of 92.07%, and AUC values of 0.96 across all classes.

The electrocardiogram (ECG) is an important tool for exploring the structure and function of the heart due to its low cost, ease of use, efficiency, and non-invasive nature. With the rapid development of artificial intelligence (AI) in the medical field, ECG beat classification has emerged as a key area of research for performing accurate, automated, and interpretable cardiac analysis. According to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses criteria, we examined a total of 106 relevant articles published between 2014 and 2024. This study investigates ECG signal analysis to identify and categorize various beats with better accuracy and efficiency, by emphasizing and applying vital pre-processing techniques for denoising the raw data. Particular attention is given to the evolution from traditional feature-engineering methods toward advanced architectures with automated feature extraction and classification, such as convolutional neural networks, recurrent neural networks, and hybrid frameworks with attention mechanisms. In addition, this review article investigates the common challenges observed in the existing studies, including data imbalance, inter-patient variability, and the absence of unified evaluation metrics, which restrict fair comparison and clinical translation. To address these gaps, future research directions are proposed, focusing on the development of standardized multi-center datasets, cross-modal fusion of physiological signals, and interpretable AI models to facilitate real-world deployment in healthcare systems. This systematic review provides a structured overview of the current state and emerging trends in ECG beat classification, offering clear insights for researchers and clinicians to guide future advancements in intelligent cardiac diagnostics.

Image inpainting is currently a hot topic within the field of computer vision. It offers a viable solution for various applications, including photographic restoration, video editing, and medical imaging. Deep learning advancements, notably convolutional neural networks (CNNs) and generative adversarial networks (GANs), have significantly enhanced the inpainting task with an improved capability to fill missing or damaged regions in an image or a video through the incorporation of contextually appropriate details. These advancements have improved other aspects, including efficiency, information preservation, and achieving both realistic textures and structures. Recently, Vision Transformers (ViTs) have been exploited and offer some improvements to image or video inpainting. The advent of transformer-based architectures, which were initially designed for natural language processing, has also been integrated into computer vision tasks. These methods utilize self-attention mechanisms that excel in capturing long-range dependencies within data; therefore, they are particularly effective for tasks requiring a comprehensive understanding of the global context of an image or video. In this paper, we provide a comprehensive review of the current image/video inpainting approaches, with a specific focus on Vision Transformer (ViT) techniques, with the goal to highlight the significant improvements and provide a guideline for new researchers in the field of image/video inpainting using vision transformers. We categorized the transformer-based techniques by their architectural configurations, types of damage, and performance metrics. Furthermore, we present an organized synthesis of the current challenges, and suggest directions for future research in the field of image or video inpainting.