Brain-computer interface (BCI) technology enables direct communication between the human brain and external devices by decoding electroencephalogram (EEG) signals into actionable commands. As a noninvasive and portable modality, EEG-based BCIs hold promise for applications ranging from neurorehabilitation to assistive technologies. However, their performance depends critically on the accurate extraction of relevant neural features and the reliable recognition of underlying patterns. Deep learning has transformed this process. By automatically learning complex, task-relevant representations from raw or minimally processed EEG data, deep neural networks have surpassed many traditional handcrafted feature approaches in both accuracy and adaptability. Yet, the substantial computational and memory demands of many deep learning architectures limit their deployment in portable or real-time BCI systems. This challenge has motivated a growing interest in lightweight models-architectures optimized to reduce complexity while preserving or even enhancing performance. This paper provides a systematic review of such lightweight deep learning models for EEG signal classification, with EEGNet serving as a representative baseline. To organize this landscape, existing approaches are categorized into three main strategies: (1) information integration through multi-scale feature fusion, (2) optimization of hidden layer design, and (3) hybrid strategies combining multiple structural enhancements. The review synthesizes recent advances, identifies emerging trends, and outlines potential directions for future research. These insights aim to inform the design of efficient and robust EEG classification architectures capable of meeting the practical demands of real-world BCI applications.

. A novel phase-amplitude coupling (PAC) estimator is proposed to address the limitations of existing PAC estimators in terms of insufficient application scenarios.. The polar mutual information (PoMI) method is compared with the currently dominant PAC estimators, mean vector length, Kullback-Libler distance, general linear model, and phase-locking value, focusing on analyzing its characteristics in terms of coupling strength sensitivity, data length dependency, noise resistance, and coupling frequency band sensitivity. We recruited 54 healthy controls and 41 mild cognitive impairment (MCI) patients and assessed their cognitive level and whole-brain PAC connectivity by neurophysiological tests and resting electroencephalography, respectively.. The PoMI algorithm is sensitive to changes in coupling strength, exhibits low dependence on data length, is insensitive to noise variations, and produces stable computational outcomes. Therefore, the PoMI algorithm can quantify the PAC phenomenon in neural oscillations. Furthermore, reduced PAC connectivity in the frontal lobe of patients with MCI, while PAC activity is enhanced in the parietal and occipital lobes. The results indicate that alterations in prefrontal PAC connectivity in MCI patients may represent one manifestation of neuronal group degeneration in the prefrontal cortex of these individuals.. The PoMI algorithm can effectively evaluate the PAC phenomenon in neural oscillations and can be used as a PAC estimator. (Approved No. of ethic committee: 2024-P2-210-02).

Brain networks communicate through long range phase coupling of low frequency oscillations (LFOs, less than 35 Hz) between brain regions. At the same time, phase-amplitude cross-frequency coupling (CFC), in which the phase of the same LFO have been shown to modulate the power of high frequency activity has also been reported across brain regions as a critical regulator of neural activity and excitability. While cross-regional CFC has been reported as a potential mechanism of long-distance modulation of neural excitability, the mechanism underlying this phenomenon has yet to be understood and methods to dissociate the effect of local vs remote LFO have not been developed. Cross-regional CFC can be a result of either LFOs in one region directly modulating high frequency oscillations in another region or due to a chain effect, in which apparent cross-regional CFC results from coupling of LFO across sites.A novel method of partial modulation index (PMI) is proposed as a derivation of modulation index (MI) and based on Pearl's do-calculus to remove the mathematical bias of simultaneous phase coupling and CFC measurements. Here, we first test the PMI on a simulated dataset, showing it can differentiate between biased and unbiased CFC. We then evaluate the method on intracranially collected local field potentials recorded simultaneously from thalamus and cortex in a patient undergoing deep brain stimulator implantation for essential tremor, demonstrating that the observed thalamocortical CFC was partially biased.For both simulated and human datasets, the PMI was compared to the conventional MI. In simulated data, the PMI was able to disentangle cross-regional phase coupling and focal CFC which is not possible using conventional MI. While there is no ground truth for comparison in human data, the results from the simulated data demonstrate the value of the proposed method in removing mathematical bias.This novel method facilitates a mathematically rigorous characterization of residual CFC, enabling investigations of differential contributions and roles of brain-wide LFO to CFC, which can lead to a more complete understanding of the pathophysiology of neurological processes and disorders.

Once we learn a reliable gait, we no longer have to consciously contract individual muscles to walk, or think about the fine-grained low-level control of our joints. Instead, we mainly make decisions on where we want to end up, at what pace and through which path. Estimating this high-level intent may provide the necessary input to wearable robotic devices to adapt to their user's needs. We introduce a continuous representation of locomotion goals and investigate how it may be estimated from muscle signals and body posture. Approach: This study investigated methods to estimate a representation of high-level locomotion intent, the horizontal walking path. We collected full-body motion capture and bipolar surface electromyography data from 6 subjects during non-steady-state gait. We trained temporal convolutional networks to causally predict the walking path directly or parametrically with a critically damped trajectory model, using a mixture of muscle and body posture signals. Main results: We achieved a mean trajectory estimation accuracy for a 1-second walking path corresponding to $r^2=0.89$ using a multimodal model. We simultaneously provided estimates for current and desired walking velocities as constrained by the walking path model, aiding interpretability of the estimator's output. Significance: Our approach could provide user interfacing in a subject-independent format for wearable robotic devices. Moreover, this high-level intent representation is flexible and able to be synthesized in virtual environments, where it can serve as a surrogate for biosignals of simulated intent-driven robotics.

Rhythmic and periodic patterns (RPP) are harmful brain activity observed on electroencephalography (EEG) recordings of critically ill patients. This work describes automatic methods for detection of the frequency and spatial extent of specific RPPs: lateralized and generalized rhythmic delta activity (LRDA, GRDA) and lateralized and generalized periodic discharges (LPD, GPD).

Chronic brain sensing devices, such as the Medtronic Percept™ or Neuropace RNS system, record local field potentials (LFPs) that may be vulnerable to interference and noise due to hardware limitations, environmental factors, movement, stimulation, cardiac signals, and analytical procedures. Although onboard hardware filters can attenuate some unwanted signals, additional processing is often required. Here we demonstrate that cardiac artifacts significantly alter the power spectral density (PSD) of neural activity within the theta (4-8 Hz), alpha (8-12 Hz), and beta (12-30 Hz) bands. We introduce a time-domain template subtraction method specifically designed to remove QRS complex cardiac artifacts. Separately, we describe techniques for transforming time domain data to the frequency domain and mitigating transient artifacts by estimating background neural activity-either through window rejection based on PSD characteristics or via principal component analysis. Finally, we present an approach to isolate oscillatory neural activity by subtracting the aperiodic 1/f component from the power spectrum by fitting the FOOOF logarithmic function. While filter selection must be tailored to the specific device and participant environment to avoid over-filtering, these interference and noise mitigation strategies are crucial for ensuring the integrity of LFP recordings.

Objective Automated EEG-based epilepsy diagnosis has reported near-perfect accuracies for almost two decades on a benchmark dataset, yet virtually no system is used in routine care. We critically re-examined this translation gap by reproducing five widely cited AI models spanning statistical feature extraction, classical machine-learning and deep-learning paradigms, and assessed their ability to generalise from the benchmark dataset to a newly curated, clinically verified scalp-EEG cohort. Approach All models were implemented as originally described and trained with 10-fold cross-validation on the benchmark dataset. External validation was performed on our independently curated dataset made publicly available, comprising 30 subjects (15 epilepsy, 15 healthy) recorded with 19-channel scalp EEG under standard clinical protocols. We further examined the influence of subject-level data leakage by contrasting performance when training/testing samples overlapped with those when completely independent patient partitions were enforced. Accuracy, sensitivity, specificity, and AUC of ROC (with 95% Confidence Intervals) were the primary metrics. Main results When transferred unchanged to the external cohort, overall accuracy fell from >= 94% on the benchmark dataset to 42-53%, with sensitivities as low as 0.97% for the deep CNN and specificities dropping to 4% for time-frequency methods. Permitting subject overlap artificially elevated accuracy to 59-96%, whereas strict patient separation reduced it to 41-53% in 95% Confidence Interval. Deep-learning models exhibited the steepest decline, confirming over-fitting to subject-specific artefacts. Statistical feature-based approaches, though less affected, still under-performed clinically acceptable thresholds. Significance Our results expose key translational barriers in AI for EEG-based epilepsy diagnosis--data leakage, acquisition bias, and overfitting to patient idiosyncrasies--leading to severe performance erosion on clinical data. Rigorous patient-independent validation, transparent reporting (aligned with CARE principles), and well-curated multi-channel scalp-EEG datasets are essential to ensure clinically dependable AI tools for epilepsy diagnosis.

Prosthesis control can be seen as a new skill to be learned. To enhance learning, both internal and augmented feedback are exploited. The latter represents external feedback sources that can be designed to enhance learning, e.g. biofeedback. Previous research has shown that augmented feedback protocols can be designed to induce retention by adhering to the guidance hypothesis, but it is not clear yet if that also results in transfer of those skills to prosthesis control. In this study, we test if a training paradigm optimised for retention allows for the transfer of myoelectric skill to prosthesis control. Approach. Twelve limb-intact participants learned a novel myoelectric skill during five one hour training sessions. To induce retention of the novel myoelectric skill, we used a delayed feedback paradigm. Prosthesis transfer was tested through pre-and post-tests with a prosthesis. Prosthesis control tests included a grasp matching task, the modified box and blocks test, and an object manipulation task, requiring five grasps in total ('power', 'tripod', 'pointer', 'lateral grip', and 'hand open'). Main results. We found that prosthesis control improved significantly following five days of training. Importantly, the prosthesis control metrics were significantly related to the retention metric during training, but not to the prosthesis performance during the pre-test. Significance. This study shows that transfer of novel, abstract myoelectric control from a computer interface to prosthetic control is possible if the training paradigm is designed to induce retention. These results highlight the importance of approaching myoelectric and prosthetic skills from a skill acquisition standpoint, and open up new avenues for the design of prosthetic training protocols.

Foundation models have demonstrated transformative potential in medical AI but remain underexplored in functional neuroimaging, particularly magnetoencephalography (MEG). This study aims to develop a domain-specific, self-supervised MEG clinical foundation model tailored for headache disorders to address the challenges of high-dimensional data and limited labeled datasets in clinical research.

: Myoelectric control systems translate electromyographic (EMG) signals into control commands, enabling immersive human-robot interactions in the real world and the Metaverse. The variability of EMG due to various confounding factors leads to significant performance degradation. Such variability can be mitigated by training a highly generalisable but massively parameterized deep neural network, which can be effectively scaled using a vast dataset. We aim to find an alternative simple, explainable, efficient and parallelisable model, which can flexibly scale up with a larger dataset and scale down to reduce model size, and thereby will significantly facilitate the practical implementation of myoelectric control.: In this work, we discuss the scalability of a random forest (RF) for myoelectric control. We show how to scale an RF up and down during the process of pre-training, fine-tuning, and automatic self-calibration. The effects of diverse factors such as bootstrapping, decision tree editing (pre-training, pruning, grafting, appending), and the size of training data are systematically studied using EMG data from 106 participants including both low- and high-density electrodes.: We examined several factors that affect the size and accuracy of the model. The best solution could reduce the size of RF models by ≈500×, with the accuracy reduced by only 1.5%. Importantly, for the first time we report the merit of RF that with more EMG electrodes (higher input dimension), the RF model size would be reduced.: All of these findings contribute to the real time deployment RF models in real world myoelectric control applications.

To enhance frequency recognition in Steady-State Visual Evoked Potential (SSVEP)-based Brain-Computer Interfaces (BCIs), particularly under short data acquisition and complex environmental conditions.

Adaptive neuromodulation systems and implantable brain-computer interfaces have made notable strides in recent years, translating experimental prototypes into clinical applications and garnering substantial attention from the public. This surge in interest is accompanied by increased scrutiny related to the safety, efficacy, and ethical implications of these systems, all of which must be directly addressed as we introduce new neurotechnologies. In response, we have synthesized the insights resulting from discussions between groups of experts in the field and summarized them into five key domains essential to therapeutic device development: (1) analyzing current landscape of neuromodulation devices and translational platforms (2) identifying clinical need, (3) understanding neural mechanisms, (4) designing viable technologies, and (5) addressing ethical concerns. The role of translational research platforms that allow rapid, iterative testing of hypotheses in both preclinical and clinical settings is emphasized. These platforms must balance experimental flexibility with patient safety and clear clinical benefit. Furthermore, requirements for interoperability, modularity, and wireless communication protocols are explored to support long-term usability and scalability. The current regulatory processes and funding models are examined alongside the ethical responsibilities of researchers and device manufacturers. Special attention is given to the role of patients as active contributors to research and to the long-term obligations we have to them as the primary burden-bearers of the implanted neurotechnologies. This article represents a synthesis of scientific, engineering, and clinical viewpoints to inform key stakeholders in the neuromodulation and brain-computer interface spaces.

As the prevalence of dementia continues to rise, the need for accurate and early diagnostic tools becomes increasingly critical. Despite diverse underlying causes, dementia types share common cognitive symptoms, making accurate diagnosis essential for effective treatment.: This study investigates electroencephalographic (EEG)-based spectral brain connectivity in individuals with Alzheimer's disease (AD,N=36), frontotemporal dementia (FTD,N=23), and healthy controls (HCs,N=29), with the dual aim of identifying condition-specific connectivity patterns and evaluating three coherence-based connectivity measures: coherence, imaginary coherence, and partial coherence. Resting-state, eyes-closed EEG data (19 channels) were analyzed, and connectivity was estimated across frequencies to assess both global and local network alterations.The results indicate that dementias (both AD and FTD) are characterized by decreased connectivity in higher frequency bands and increased connectivity in lower frequencies, reflecting respectively impaired neural communication and neurodegeneration. Moreover, the severity of cognitive impairment correlates with the spatial extent and magnitude of connectivity disruptions. Notably, partial coherence-unlike coherence and imaginary coherence-effectively distinguishes between the AD and FTD groups, suggesting that direct connectivity measures may provide more discriminative information for differential diagnosis.These findings highlight the potential of EEG-based spectral connectivity analysis, particularly partial coherence, as a non-invasive tool to aid in the diagnosis and differential diagnosis of dementia subtypes, supporting early clinical decision-making.

With the advancement of deep learning technologies, more and more researchers have begun developing end-to-end automatic sleep stage classification frameworks. However, these frameworks typically require access to large electroencephalogram (EEG) datasets for training, which imposes a significant computational burden. Furthermore, EEG data contains patient privacy information, and using such data for training raises concerns about privacy infringement. To address these issues, we propose a hybrid data distillation method. We aim to enable single-channel EEG sleep stage classification with less training cost and privacy risk by distilling large real datasets into a tiny, privacy-preserving synthetic set for training from scratch.We first apply the gradient matching method to optimize the randomly initialized synthetic dataset. The gradient changes in the early stages of model training can quickly reduce the performance gap between the synthetic dataset and the source dataset. Subsequently, to avoid oscillations near the optimal solution during gradient matching, we switch to distribution matching to further optimize the synthetic dataset. This method aligns the data distribution at a global level, enhancing overall consistency. In addition, we adopt a novel mini-batch iteration method to assist the synthetic dataset in learning temporal dependencies.We validated our framework on three public datasets and achieved robust results.This study proposes an efficient and robust hybrid data distillation algorithm, providing a feasible approach for implementing sleep stage staging based on privacy protection.

Transcutaneous spinal cord stimulation, a non-invasive spinal cord neuromodulation method holds tremendous promise and hope to restore functions in individuals with paralysis resulting from spinal cord injury (SCI), cerebral palsy, stroke and other neurological conditions. Yet, there are relatively few options for such stimulation devices compared to conventional stimulators commonly used for neuromuscular electrical stimulation, transcutaneous electrical nerve stimulation, and functional electrical stimulation, particularly for people with neurological conditions in the developing countries.In this report, we present OpenXstim, an open-source, two-channel programmable electrical stimulator developed to advance research in non-invasive muscle, nerve, or spinal cord stimulation treatments.. OpenXstim can deliver current pulses up to 110 mA with a compliance voltage of 96 V per channel. In benchtop testing, we found that the stimulator successfully generates high frequency (9 kHz) burst stimulation, a mode commonly used for spinal cord neuromodulation. The stimulator was further tested in two individuals with SCI and showed preliminary indications of functional improvement. However, large controlled trials are needed to establish efficacy. Although special care was taken in the design of the stimulator to ensure user safety, users are strongly warned to handle the device with utmost caution, as it can generate high voltage and current that may cause adverse health effects if not used properly.This programmable, open-source stimulator offers tangible hope for improving the accessibility of non-invasive neuromodulation treatments for people with paralysis worldwide. The design and complete source-code of the stimulator are freely available online in a public repository:https://github.com/OpenMedTech-Lab/OpenXstim.

Domain adaptation (DA) has achieved remarkable performance in cross-subject electroencephalogram (EEG) decoding by mitigating the inter-subject data distribution discrepancies. However, when exploring the feature alignment subspace and performing self-supervised pseudo-labeling in an iterative way, two difficulties are often encountered: one is that unreliable target labeling results inevitably mislead the domain-free feature learning process in the early stage and the other is that the contribution of source and target samples should be balanced in the later stage.To address both issues, this paper proposes prototype-based progressive confident target sample labeling (PCSL) method to use subspace class prototypes to assist in labeling target samples under the unified framework of domain-invariant EEG feature learning and the self-supervised target sample labeling, and progressively incorporate confident target samples into DA model fitting. The underlying rationality is that early-stage pseudo-labels from unconverged models are prone to error propagation, requiring auxiliary mechanisms to ensure their reliability and stabilize training. With the gradual alignment of cross-subject features, the estimated pseudo-label information of target domain will be more reliable, meaning that more target samples should be involved in model training.Experiments on emotion recognition and inner speech decoding demonstrate the competitive performance of PCSL in cross-subject EEG classification in comparison with SOTA methods.Our study indicates the effectiveness of jointly considering the reliability of target samples and their contribution to model training in the context of DA. In addition, some fine-grained results including the sample confidence allocation strategy, the DA effects, and the dynamic model optimization process are provided to further illustrate the model execution details.

Despite decades of electroencephalography (EEG) research, the relationship between EEG and underlying spiking dynamics remains unclear. This limits our ability to infer neural dynamics reflected in intracranial signals from EEG, a critical step to bridge electrophysiological findings across species and to develop non-invasive brain-machine interfaces (BMIs). In this study, we aimed to estimate spiking activity in the visual cortex using non-invasive scalp EEG.

As an emerging paradigm of brain-computer interfaces (BCIs), speech BCI has the potential to directly reflect auditory perception and thoughts, offering a promising communication alternative for patients with aphasia. Chinese is one of the most widely spoken languages in the world, whereas there is very limited research on speech BCIs for Chinese language.This paper reports a text-magnetoencephalography (MEG) dataset for non-invasive Chinese speech BCIs. It also proposes a multi-modality assisted speech decoding (MASD) algorithm to capture both text and acoustic information embedded in brain signals during speech activities.Experiment results demonstrated the effectiveness of both our text-MEG dataset and our proposed MASD algorithm.To our knowledge, this is the first study on multi-modality assisted decoding for non-invasive Chinese speech BCIs.

Implanted electrodes for nerve stimulation and myoelectric recording facilitate bidirectional sensory feedback and control for neuromuscular conditions such as limb loss. While increasing implanted electrode channel count offers potential benefits, it also presents engineering and implementation challenges. This case study examines how increasing implanted electrode channel count affects sensory perception and myoelectric controller performance, thereby supporting the value of these advancements. Approach: One participant with upper extremity transradial limb loss received a percutaneous implanted system with two 8-channel extraneural cuff electrodes on the median and ulnar nerves, totaling 16 stimulating channels. The individual later received a wirelessly connected implanted system featuring four 16-channel extraneural cuff electrodes on the median, ulnar, and radial nerves, totaling 64 stimulation channels, and four tetra-intramuscular recording electrodes in residual muscles, totaling 32 sensing channels configured into 16 bipolar pairs. To compare sensory perception between the 16- and 64-channel stimulation systems, we assessed cumulative percept location coverage and the number of unique percept locations, estimated through hierarchical clustering. We compared performance across three myoelectric control algorithms that mapped 8, 10, or 14 intramuscular signal inputs through an artificial neural network to control a virtual hand in 4 degrees-of-freedom (DOFs), with simultaneous, independent, and proportional control. Main results: Increasing stimulation channel count expanded cumulative percept location coverage and increased the number of unique percept locations on the hand. Adding intramuscular recording channel inputs improved 4-DOF myoelectric control of a virtual hand, increasing target posture match percentage and path efficiency. Significance: This case study demonstrates that increasing the number of implanted electrodes can advance sensory restoration and myoelectric control for bidirectional upper limb prostheses. Continued development of more complex systems with higher channel counts may further improve outcomes for individuals with limb loss and enhance the function of sensorimotor restoration systems. Trial registration: ClinicalTrials.gov ID: NCT04430218, 2020-06-30.

Group interactions capture cooperative dynamics among neural populations quantitatively, while also enabling precise detection of ensemble-level synchrony patterns and transcending the limitations of node-level relationships. To evaluate higher-order group interactions, we propose the PLASSO-homophily framework using multichannel stereo-electroencephalography (SEEG) recorded from patients with epilepsy.Specifically, we use phase locking value to improve least absolute shrinkage and selection operator method for constructing hypergraphs. Afterwards, we calculate affinity ratios between brain zones. Finally, we investigate higher-order interactions among different groups from a homophily perspective. The extremal result of strict homophily serves as a crucial theoretical framework for understanding homophily concepts, reflecting the constraints that different groups follow in higher-order interactions.It is observed that group interactions between seizure onset zones (SOZ), propagation zones (PZ) and non-involved zones (NIZ) present significant distinction across different seizure phases. In particular, the homophily of SOZ reaches a peak point during the seizure and sharply decreases in the post-seizure, with the most statistically significant differences onθandγbands. Furthermore, during the seizure, SOZ-PZ exhibits enhanced coupling while SOZ-NIZ exhibits impaired functional integration. Finally, among three groups, only SOZ exhibits strict monotonic and majority homophily.By analyzing changes in in-class and out-class connectivity, we quantitatively assess the activity levels and combinatorial constraints of the SOZ, PZ, and NIZ, thereby providing a novel perspective for exploring seizure mechanisms and developing epilepsy treatments.

Damage to the peripheral nerves frequently leads to significant impairments in their functional capacity, highlighting the need for effective treatments that can facilitate nerve repair. This study explores the potential of grape skin extract (Ex), alone and in combination with zinc oxide nanoparticles (ZnO NPs), to enhance regeneration following sciatic nerve injury in rat. Approach: ZnO NPs were synthesized using both a conventional chemical route and a green synthesis method in which Ex served as a natural reducing and capping agent. The synthesized nanoparticles were characterized by Fourier-transform infrared spectroscopy (FTIR), Scanning electron microscopy (SEM), X-ray diffraction (XRD), Thermogravimetric analysis (TGA), Energy-dispersive X-ray spectroscopy (EDX), zeta potential, and Gas chromatography-mass spectrometry (GC-MS) analyses to confirm the role of Ex in shaping nanoparticle morphology and surface properties. Functional recovery and histological outcomes were then assessed in a murine sciatic nerve injury model. Main results: Treatment with Ex and ZnO/Ex significantly reduced collagen accumulation, fibrosis, and tissue vacuolization compared to untreated controls. Both interventions also improved myelination and enhanced the sciatic function index (SFI), indicating improved neural repair. Significance: These findings demonstrate that Ex and ZnO/Ex promote nerve regeneration and highlight their potential as promising candidates for the development of biogenic nanotherapeutics targeting peripheral nerve injuries.