Papers

A perspective review published in Computers in biology and medicine examines EEG non-stationarity across sessions, people, and recording conditions. It asks whether mechanisms that help the brain maintain functional stability can improve the robustness of brain-computer interface models. The review covers synaptic plasticity, homeostatic regulation, neural oscillations, and spiking representations, comparing bio-inspired approaches with conventional machine learning and transfer learning. It also considers hybrid designs that combine biologically grounded mechanisms with artificial neural networks. The author proposes operational definitions for bio-inspired, bio-plausible, and bio-realistic modeling, together with a minimum specification for continual EEG benchmarks. Because direct EEG evidence remains limited for several proposed mechanisms, the review stresses the need to distinguish empirical findings from hypotheses and future research directions.

A research team has developed Brain2Qwerty, a deep learning model that decodes sentences people type from non-invasive brain recordings. Thirty-five healthy volunteers typed sentences they had just memorized while their magnetoencephalography (MEG, which measures the brain's magnetic fields) or EEG signals were recorded. The average character error rate was 29% with MEG and 65% with EEG; for the best-performing participants it fell to as low as 18%, and the model fully decoded a range of sentences outside its training set. The researchers say the results narrow the gap between non-invasive and implanted methods.

Researchers propose an injectable, in-situ-gelling multifunctional hydrogel to address the failure modes of micro-ECoG cortical recording — dural barrier disruption, cortical micromotion that weakens device-tissue coupling, and biofouling that triggers a foreign-body response. Combining dopamine-grafted sodium alginate with branched polyethyleneimine, the hydrogel forms a quasi-zwitterionic network that resists nonspecific protein adsorption and provides catechol-mediated wet adhesion, gelling rapidly under surgical-compatible conditions through dual macromolecular crosslinking without diffusible small-molecule monomers. Integrated with a 128-channel flexible micro-ECoG mesh array, the platform reduced glial activation and fibrotic encapsulation and preserved stable, high-fidelity cortical recording over the 3-week early chronic period. The authors say co-designing barrier repair, interface adhesion and antifouling in a single material can improve long-term function.

Researchers transferred dexamethasone-loaded titania nanotube arrays (TNA) onto a mechanically adaptive polymer nanocomposite (NC) substrate and, in a mouse model, compared four implants — silicon, NC, TNA-NC Empty and TNA-NC DEX (10 mice per group) — for neuroinflammation around intracortical microelectrodes at 2 and 4 weeks. At 2 weeks the gene-expression profiles were broadly similar, reflecting an early acute injury response; by 4 weeks the patterns diverged, with NC-based implants showing fewer differentially expressed neuroinflammatory genes than rigid silicon, led by TNA-NC Empty, while the dexamethasone group showed no additional benefit, suggesting drug delivery still needs optimization. The authors conclude that adding a TNA layer to flexible materials promotes resolution of the neuroinflammatory response at 4 weeks.

MVCLDG combines raw EEG and Hilbert-derived phase information with domain-alignment and contrastive-learning constraints to improve classification across unseen users. It outperformed baseline and representative domain-generalization methods on a public error-related-negativity dataset and a semantic-syntactic-violation dataset without target-domain adaptation; ablation and activation-map analyses supported the contribution and neurophysiological plausibility of its components.

A study published in Computers in biology and medicine introduces LEGEND, a neural-decoding architecture that jointly models cortical EEG, spinal ESG, and peripheral-muscle EMG for a neural bypass in spinal cord injury rehabilitation. The model encodes the three signal modalities in Lorentz hyperbolic space, connects 51 channel nodes through a signed tri-layer phase-locking-value graph, and refines the representation with graph attention. Under strict leave-one-subject-out evaluation on the Steele dataset, LEGEND achieved 56.51%±12.27% accuracy, 23.4 percentage points above EEGNet. The researchers argue that hyperbolic representations can capture complex relationships across the motor hierarchy, providing a computational foundation for rehabilitation decoding that links brain, spinal, and muscle activity.

Using implantable brain-machine interface (iBMI) data from a non-human primate two-dimensional finger task, the study tests whether brain-muscle synergies improve decoding performance and generalization. Principal component analysis (PCA), demixed PCA (dPCA) and non-negative matrix factorization (NMF) all compressed brain-muscle data effectively with minimal decoding-accuracy loss, but none improved performance through denoising or enhanced cross-task generalization. The authors conclude that extracting synergies alone does not yield a better or cleaner control space for linear decoding, and call for larger samples and more muscle channels.

Researchers in the Department of Brain Sciences at Israel's Weizmann Institute of Science developed BAMBI, a brain-computer interface that uses a head-mounted miniature microscope to image calcium signals from active neurons in real time and deliver feedback. Because it realigns to the same field of view in each session, it tracks the same group of neurons over multiple days without retraining. Tested in head-fixed and freely moving mice across different behavioral paradigms, the animals learned to stably modulate specific hippocampal neurons for rewards over several days.

Researchers at Lanzhou University, Zhejiang University, the University of Hong Kong and Sun Yat-sen University have built MindShow, a unified generative framework that reconstructs images from fMRI at cohort level rather than one subject at a time. A hierarchically conditioned mixture-of-experts encoder separates population-shared latent representations from subject-specific neural traits, so a new subject can be adapted with limited calibration data; a gated Perceiver bottleneck maps fMRI features into fixed-size image and text tokens, an optimal transport loss aligns them with a pretrained vision-language model, and a frozen diffusion model renders the image. The authors report better high-level reconstruction metrics with competitive structural fidelity, in a study published in Medical Image Analysis on June 20, 2026.

A study published in Computers in biology and medicine proposes dynamic wavelet-basis selection to improve non-invasive EEG imagined-speech classification. For each EEG epoch, the method minimizes wavelet entropy to select an informative basis and then injects Gaussian noise into the corresponding coefficients. A convolutional neural network with channel-wise excitation classifies the augmented signals. The dataset contains 32 channels, 8 stimuli, and recordings from 10 participants. The words-vowels combination reached a highest classification accuracy of 98% with a Cohen's kappa of 0.95, although performance was lower for the full class set. The researchers report that the method outperformed conventional augmentation strategies and static wavelet approaches, offering an adaptive way to address noise and non-stationarity in EEG decoding.

Published in Cell Reports Methods on June 12, 2026, the study presents high-density micro-electrocorticography arrays that integrate silicone elastomers (optical transparency, repeated penetration with intracortical arrays) and polyimide films (fine photolithographic feature definition) for multi-scale studies of brain activity. The combination facilitates high-throughput functional mapping to identify targets and insertion of intracortical arrays for dense local sampling. The authors demonstrated functional mapping in rats, cats and marmosets, guiding multi-area laminar recordings, and demonstrated local and feedforward optogenetic stimulation to investigate cortico-cortical interactions.

A team led by the UMC Utrecht Brain Center in the Netherlands analyzed 8 participants with high-density electrocorticography (ECoG) grids of 64 to 128 electrodes and found that a well-positioned 32-electrode grid classified 12 Dutch words with 76 ± 16% accuracy, comparable to 75 ± 17% for the larger grids. The best placement was vertical and centered on the central sulcus. Speech brain-computer interfaces aim to restore communication to people paralyzed and unable to speak, but larger grids carry higher surgical risk. The participants were not paralyzed and testing was limited to a fixed word set.

Researchers have found that whether people can learn to control a device with their thoughts may depend on whether the control scheme follows the brain's own activity patterns. Participants lying in an MRI scanner controlled a video game avatar in real time by modulating activity in brain regions responsible for spatial navigation. When the researchers then changed the mapping between brain activity and avatar movement, participants could readjust and regain control if the new mapping fell along directions in which their brain activity already varied widely, known as the intrinsic manifold; when the mapping departed from that structure, they never learned it. The researchers say the principle could be used to design BCIs that are easier to learn.

A study published in Computers in biology and medicine applies cross-frequency bispectral analysis to nonlinear EEG activity during the planning and execution of natural reach-to-grasp movements. The researchers extracted magnitude- and phase-based features from complex bicoherence matrices and assessed them through classification, permutation-based feature selection, and within-subject statistics. Execution showed stronger nonlinear coupling than planning, led mainly by beta- and gamma-driven interactions. Decoding precision versus power grasps performed similarly across the two stages, suggesting that grasp-type representations emerge during planning and persist into execution. Compared with conventional analytical baselines, bispectral features provided consistent advantages for grasp-type discrimination and multiclass classification. The findings offer a new set of motor-decoding features for future brain-computer interface and neuroprosthetic research.

Researchers proposed ResDSNet, a residual depthwise-separable network with a matching preprocessing module that filters noise while keeping motor imagery features, to break what they call the "depth barrier": deepening lightweight decoders fails when residual structures are poorly designed or mismatched with preprocessing. On 3 public datasets covering 127 subjects it reached 79.36%, 84.95% and 64.13% accuracy, 3.16%, 1.59% and 8.40% higher than the best existing methods. It was tested only offline on public data, not in real-time BCI systems or patients.

The study introduces RLED, a reinforcement learning-enhanced EEG diffusion framework for adaptive data augmentation in endogenous EEG tasks such as motor imagery and emotion recognition. Reinforcement learning dynamically regulates the diffusion training process to balance temporal, spectral and category-related features. Across four datasets, the high-quality synthetic EEG signals it generated consistently improved classification performance.

Researchers have built a hybrid batch-stream framework for EEG-based emotion recognition, which until now has had to choose between accurate but slow offline batch processing and fast but less accurate real-time streaming. The system estimates whether the streaming module is currently reliable from its past performance, then decides probabilistically whether to use the batch or the streaming output. It achieved classification accuracies of 85%, 94% and 74% on the public DEAP, AMIGOS and SEED benchmarks, and was also tested under concept drift, when data distributions shift over time. Results come from offline evaluation on public datasets and have not been validated with real users.

Intracranial recordings from ten epilepsy patients show that neural activity during cued limb motor imagery changes with the phase of the task: low-frequency (8-30 Hz) activity was mostly suppressed during preparation and switched to activation once imagery began, while high-frequency (60-115 Hz) responses were stronger, more widely distributed, and at some contacts followed an activation-then-suppression sequence. The stereoelectroencephalography (sEEG) data indicate that responses are not uniform but shift with task stage and brain region, which the authors tie to stage-aware feature design for BCIs and to neurorehabilitation.

ChatBCI-Assist integrates a locally deployed LLM with adaptive stopping, achieving record P300 speller online ITR of 105.2 bits/min and 30.7 CPM in semantic spelling.

Researchers in Chongqing, a municipality in southwestern China, have proposed SPD-DANN, a method that aims to decode motor imagery, the brain signals produced when a person imagines moving a hand or foot, in new users without per-person calibration. Because EEG varies between people and drifts over time, BCIs typically need fresh data from each new user. SPD-DANN represents EEG features on a symmetric positive definite (SPD) matrix manifold, a mathematical space that better fits their nonlinear structure, and uses adversarial learning to extract features shared across individuals. Tested on 4 BCI datasets, the method outperformed several existing unsupervised domain adaptation techniques, the researchers said. The abstract gives no accuracy figures, and all tests were offline.

Current hearing aids amplify every sound in a noisy room rather than the person a listener wants to hear. Using intracranial EEG in neurosurgical patients, researchers built a closed-loop system that decodes which talker a listener is attending to and amplifies that voice in real time. Across multiple experiments it improved speech intelligibility, reduced listening effort and was consistently preferred by subjects, and it followed both instructed and self-initiated shifts of attention.

Researchers assessed the feasibility of a hybrid brain-computer interface that integrates motor imagery (MI) and steady-state visual evoked potentials (SSVEP) with robotic glove feedback for upper limb motor rehabilitation in 32 stroke patients, split into a conventional-treatment control group and an experimental group receiving 10- or 20-day BCI interventions. The experimental group showed considerable improvement in Fugl-Meyer scores over the control group, and the BCI achieved EEG classification accuracy up to 98.08% with stable operation; after longer training, accuracy rose, the laterality coefficient moved toward normal, and task-related brain connectivity strengthened. The authors say the hybrid system may overcome the limits of conventional therapy and single-modality BCIs.

A Nature Communications study tested ultra-Flexible Implantable Neural Electrode arrays during awake surgery in 16 patients. Valid single-unit recordings were obtained in 11 patients, yielding 719 isolated neurons and as many as 135 simultaneously, while five early cases failed because of operating-room noise or damaged insertion needles. The work was intraoperative and does not establish long-term implant performance.

Researchers at Xidian University built DisCo-Former, a Transformer framework for single-trial rapid serial visual presentation (RSVP) EEG decoding with three components guided by neurophysiological priors, then found its attention consistently collapsed: attention maps went nearly uniform and value-projection weights shrank toward 0. Stripping out the Transformer encoder left DisCo-MLP, a pure multilayer perceptron that matched or beat the Transformer version across two datasets and three evaluation regimes, with within-subject mean AUCs of about 0.94 to 0.98. For RSVP-EEG, the authors argue, modeling the signal's structure matters more than architectural complexity. The study was published in the International Journal of Neural Systems on April 10, 2026.

To reduce negative transfer in motor imagery BCIs, the study presents an adaptive dynamic transfer learning framework that decomposes EEG time-frequency features via wavelet convolution, matches source and target samples through a dynamic transfer attention module, and uses a joint loss to shrink marginal and class-conditional differences. On BNCI2014001, BNCI2014002 and BNCI2015001 it reached 78.78%, 82.11% and 78.19% accuracy, averaging 0.13% to 27.7% above baseline algorithms.

Researchers propose DPMS-Net, a deep network that uses dynamic convolution to mine discriminative cues across temporal, spatial and frequency dimensions, combines channel and temporal attention, and adds a spectral-domain analysis component to surface subtle oscillatory features hidden in the EEG spectrum. On the BCI Competition IV 2a and 2b datasets it reached subject-dependent accuracies of 83.93% and 88.38%, and 67.67% on a self-collected stroke-patient dataset. The authors say its efficient decoding and robustness suit neurorehabilitation BCI systems.

The study presents MAPS-CS, an online SSVEP brain-computer interface that selects channels dynamically during the experiment. A multi-dimensional feature framework covering signal energy, stability and inter-channel correlation quantifies anomalies and generates scores that a hierarchical decision step combines into a channel-quality score to identify and remove bad channels — with no training required. Against the channel ensemble (CE) method, MAPS-CS lifted standard FBCCA accuracy by 3.5%, 4.1%, 4.4% and 6.5% at stimulus durations of 2 s, 1.5 s, 1 s and 0.5 s, the best among the CE, binary harmony search and TOP-K local optimization methods compared.

Two people with tetraplegia typed on a virtual two-handed QWERTY keyboard simply by attempting finger movements, which electrodes implanted in the brain decoded into keystrokes. Typing speed was set by the users and reached 110 characters per minute, or 22 words per minute, with a word error rate of 1.6%, which the authors say resembles able-bodied typing accuracy. Accurate decoding required as few as 30 calibration sentences, and a 5-gram language model improved sentence output. One participant has amyotrophic lateral sclerosis (ALS) and the other a spinal cord injury.

Researchers built STAND-Net, a spiking neural network that cleans blink, muscle and other artifacts from EEG and computes only when the signal changes, combining a leaky integrate-and-fire encoder-decoder, a dilated-convolution residual backbone and an attention mechanism that locates artifacts from neuron firing patterns. Against state-of-the-art methods it raised signal-to-distortion ratio by more than 3.7 dB across artifact types and used 97.98% less power than comparable deep neural networks; BCI classification on the cleaned signals was 6.64% more accurate. The results come from the paper's abstract, with no data from wearers.