Non-invasive BCI

An offline EEG study of 15 participants found that on-screen flickering targets stepping up or down in frequency can be recognized as brain-control commands, but they did not classify more accurately than conventional fixed-frequency flicker. Mean normalized accuracies were 89.7% for fixed frequency, 85.6% for ascending and 89.5% for descending patterns, a statistically insignificant difference. Low-frequency stimulation at 6–10 Hz reached 90.8%, versus 80.6% at 41–45 Hz. Sequences of three consecutive commands averaged 91.5 ± 9.6%, with accuracy falling as sequences grew longer. The researchers see the main value in more flexible design of consecutive commands and say larger samples and online closed-loop testing are needed.

The First Affiliated Hospital of Bengbu Medical University, in eastern China's Anhui province, is recruiting for a clinical trial that uses brain-computer interface (BCI) neurofeedback to treat the cognitive-control deficits that keep people with internet gaming disorder from stopping play. The system reads frontal midline theta rhythms and feeds them back in real time so participants learn to regulate their own brain activity, with social rewards built in as reinforcers to improve efficacy and adherence. The trial plans to enroll 120 patients across three arms: dual-target neurofeedback plus social reward, neurofeedback alone, and sham feedback. Primary completion is expected on December 30, 2028.

Researchers have developed an adaptive version of Artifact Subspace Reconstruction (ASR) that removed 53.1% of blink artifacts from EEG recordings, compared with 24.0% for the original ASR and 26.7% for a limited-memory variant. EEG signals drift over time and are often contaminated by blinks and other artifacts, a problem BCIs must solve for long-term use in real-world settings. Standard ASR suppresses artifacts automatically based on variance but relies on a static covariance model derived from calibration data, which becomes less robust as signals drift. The new method tracks signal structure with an exponential moving average, combining short-term covariance estimates with slow recursive updates so the model adapts on a fast and a slow timescale at once. It was tested on 24-channel EEG from 10 subjects across different cognitive states. The trade-off was larger reconstruction and spectral deviation. The study is a preprint and has not been peer reviewed.

Researchers testing 3 EEG foundation models, general-purpose models pretrained on large amounts of EEG data, on 235 subjects excluded from training found that a lightweight adapter trained for each individual raised mean balanced accuracy by 1.5 to 5.4 percentage points over the shared population model. Every new brain-computer interface user needs calibration, and a personally tuned model may be more accurate either because it captures the individual's traits or simply because the underlying general model is stronger. Personal adapters also beat adapters borrowed from other people by 2.3 to 7.3 points. When the population model's training was scaled up fourfold, the personalization gain stayed positive but narrowed. Low-cost calibration using only a few labeled samples gave inconsistent results. The study is a preprint and has not been peer reviewed.

Researchers have built AutoBCI, a framework in which AI agents design EEG decoding models, and report that it slightly outperformed the strongest existing model across 14 datasets. Building a decoder that works well across many EEG tasks usually takes repeated trial and error. AutoBCI splits the job between 2 agents: one generates and refines architectures, while the other forecasts final performance from early learning curves to pick out candidates worth training further. Across 14 EEG datasets covering motor imagery, emotion recognition and sleep staging, the architecture found with Claude Opus 5.5 reached an average test balanced accuracy of 64.16%, slightly above the 63.87% of REVE, the strongest baseline. Looking at only 10 training epochs, the forecasting agent cut prediction error from 2.20 to 1.36 percentage points. The study is a preprint and has not been peer reviewed.

When a motor imagery BCI is used on a different day, EEG signals drift, yet a standard classifier keeps issuing commands regardless of confidence. Researchers at the University of Notre Dame, reporting in the Journal of Neural Engineering, added a conformal prediction gate that withholds commands when the model is unsure. Using DeepConvNet as the base classifier, the gate cut the share of wrong commands from 37.5% to 9–16% on the OpenBMI dataset (54 subjects) and from 35.5% to 6–8% on BCI Competition IV Dataset 2a (9 subjects), at the cost of issuing fewer commands. All results come from offline public datasets.

A collaboration involving the University of Geneva and the CIBM Center for Biomedical Imaging compared resting-state EEG under dry, semi-dry and gel conditions in 22 healthy participants in one session. Hair-wetting with tap water reduced bad channels from about 40% to 15% and improved mean electrode impedance by 75.7%. Eyes-closed spectral-power similarity with gel recordings increased by 7.7%, while functional-connectivity similarity improved by about 56–70% across frequency bands. Signal quality did not fully match gel systems, and the study did not test BCI task performance.

Researchers at Tianjin University of Technology and Education in northern China built a 64-target visual evoked potential brain-computer interface that encodes stimuli as relative changes among five colors (red, green, blue, yellow and gray) with grouped modulation, replacing the high-contrast black-and-white flicker that tires the eyes and limits the number of options. Their decoder uses task-discriminant component analysis (TDCA) to extract target-related components from EEG, then classifies them with a lightweight multilayer perceptron (MLP). In experiments with 10 participants conducted from April to June 2026, the TDCA-MLP algorithm averaged 93.72% accuracy, outperforming 4 commonly used algorithms. The authors say the system requires less calibration and offers a better subjective viewing experience.

A new review surveys the past decade of signal-processing algorithms and hardware for EEG-based motor imagery brain-computer interfaces, which let users drive external devices or retrain motor function simply by imagining a movement. In these systems, users imagine moving a hand or foot while scalp electrodes record signals from the sensorimotor cortex, and algorithms translate them into commands. The authors survey the more practical algorithms of the past ten years, focusing on how machine learning and deep learning improve performance and efficiency. They also review hardware and supporting technologies, including system-on-chip designs, application-specific chips, programmable chips, wearables, the Internet of Things and augmented and virtual reality, discuss how these pair with algorithms, and outline research directions for better real-time performance. The paper is a preprint and has not been peer reviewed.

A review by a team including Tianjin University sorts ways to slim down deep-learning decoders for motor imagery EEG, which are accurate but too large for wearable devices, into two routes: redesigning convolutional neural networks (depthwise separable, multi-branch and dilated convolutions, attention mechanisms) and compressing existing models (pruning, quantization, knowledge distillation, low-rank decomposition). It compares representative lightweight models on parameter count, computational cost and classification accuracy, and recommends shifting toward automated neural architecture search and hardware-software co-design.

A network meta-analysis of 22 randomized study cohorts with 727 participants compared BCI-coupled robotic rehabilitation, robot-assisted therapy alone, motor imagery and conventional rehabilitation for arm weakness after stroke. BCI-coupled robotic rehabilitation ranked highest on upper-limb motor function scores, averaging 6.80 points above conventional rehabilitation, a point estimate above the 5.25-point minimal clinically important difference, although the lower confidence limit fell short of it. Its advantage over robot-assisted therapy alone was only 2.04 points, which the authors judged unlikely to be clinically important.

Researchers have proposed SICMD, a method that combines two non-invasive brain signals, functional magnetic resonance imaging (fMRI) and magnetoencephalography (MEG), to identify speech a person is hearing from brain activity with a single model that works across different people. fMRI pinpoints where activity occurs and MEG captures when, together giving fuller spatiotemporal information. In cross-subject perceived-speech decoding, the method improved Top-1, Top-10 and Rankacc by more than 10.6%, 10.1% and 1.7% over baseline methods, and cut training costs by 88.8% and 60.5% compared with multi-subject and intra-subject decoding respectively. The work is a preprint and has not been peer reviewed.

Researchers have shown that language models do more to help decode imagined speech than heard speech from non-invasive brain recordings. Imagined speech is harder to read than speech a person hears: the brain's response is weaker and noisier, and less data is available. Using a magnetoencephalography (MEG) dataset in which participants both listened to and imagined the same material, the team brought a language model into both training and decoding. Imagined speech gained more from the model than heard speech, and its best results required leaning more heavily on the model. The authors conclude that the weaker the neural evidence, the more useful a language model's contextual prior becomes. The study is a preprint and has not been peer reviewed.

A brain-computer interface that times a nerve stimulus to the moment the brain prepares a wrist movement raised the excitability of the motor pathway from cortex to muscle in 17 able-bodied volunteers, whether they actually moved the wrist or only imagined it. Gains persisted 30 minutes after the session, and the two approaches did not differ significantly. The authors say this gives clinicians flexibility for patients who cannot move, though the study did not include stroke patients and individual responses varied widely.

The study tested seven detection methods to assess whether motor imagery BCIs can identify movement classes absent from training rather than force them into known categories. For many users, model uncertainty was higher for familiar classes than for unfamiliar ones, undermining methods that work well in other machine learning domains. Detection performance was correlated with performance on the trained task. Deep Ensemble and Monte Carlo-Dropout models achieved an on-task AUC of 0.9 and out-of-distribution detection ability up to 0.7. The authors conclude that rejecting unfamiliar cognitive states is feasible for models and participants with high task performance.

A team including the University of Science and Technology Beijing reviewed the open-source software ecosystem for non-invasive, EEG-based brain-computer interfaces, which read brain signals through scalp electrodes without surgery. The review covers four dimensions, technical foundations, tool taxonomy, scenario adaptation and future evolution, shows how open-source tools span the full workflow from signal acquisition to real-time interaction, classifies them by application such as healthcare, education and entertainment, and identifies the ecosystem's current technical challenges, offering developers a method for choosing tools.

Researchers re-examined a multi-class imagined speech EEG dataset, words spoken silently in the mind, under a strictly subject-independent protocol in which a model is never trained on data from the person it is tested on. Spectral bandpower features reached a mean trial-wise accuracy of 49.03% ± 4.18% in coarse-level classification, clearly ahead of the 37.97% ± 3.79% from time-domain statistical features, and a few frequency bands carried most of the discriminative information. The study is a preprint and has not been peer reviewed.

Researchers tested whether shrinking brain-computer interface (BCI) decoders through pruning or low-precision quantization, common steps for running them on low-power devices, changes how they hold up against adversarial attacks, in which imperceptible perturbations to EEG signals trick a model into misclassifying. Compressing a model while preserving accuracy did not make it more robust: at a perturbation strength of ε=0.005, EEGNet accuracy fell to 22% to 24% in its original, 50% pruned and two INT8 quantized versions alike. Attacks transferred markedly less well between pruned and original models, while quantization left transfer almost unchanged. The study is a preprint and has not been peer reviewed.

Researchers propose DenoFlow, an EEG denoising method designed to remove interference from muscle and eye electrical activity while preserving information that a brain-computer interface can decode. The authors report better signal fidelity and decoding accuracy than seven baseline denoising models in tests on two public steady-state visual evoked potential (SSVEP) datasets with five decoders. They created paired contaminated and original signals by injecting recorded muscle and eye artifacts at controlled signal-to-noise targets. The findings come from these controlled dataset tests, and the study has not been peer reviewed.

Spanish researchers asked 15 healthy participants to imagine raising an arm while watching a humanoid robot do it, viewed in immersive virtual reality, on 2D video or live in person, and recorded their brain activity with EEG. Immersive VR produced the strongest activation of the sensorimotor cortex, measured as desynchronization, a weakening of EEG rhythms that indicates a more active region. The approach of engaging motor areas by watching and imagining, without moving, underlies rehabilitation for people with limited or lost mobility, and the researchers say immersive VR could help optimize such programs and brain-computer interface applications. The study included only healthy people and has not been tested in patients.

Researchers at Lanzhou Jiaotong University proposed a hierarchical graph learning framework to decode motor imagery EEG recorded while people watch a movement in virtual reality and imagine performing it. The method splits scalp EEG into rhythmic components, selects key features, represents each trial's brain-region connectivity as a graph and classifies it with a hierarchical graph convolutional network. On the team's self-collected VR motor imagery dataset, it outperformed common baselines including EEGNet in three-class decoding. The team also found that observing plus imagining in VR strengthened connectivity among sensorimotor regions, engaging a wider brain network. The abstract does not give specific accuracy figures or the number of participants.

A systematic review of 32 studies asks whether stroke survivors can retrain paralyzed limbs by imagining movement. Most approaches had patients imagine an action, read the signal from scalp EEG, and used it to drive virtual reality, functional electrical stimulation or robotic exoskeletons, closing the feedback loop. Training ranged from small pilots to as many as 80 sessions, averaging about 30. The authors conclude such methods can significantly improve clinical outcomes in stroke survivors, though the benefit depends on the specific modality and stroke characteristics, and there is still no agreed treatment dose.

In cross-subject transfer learning for motor imagery BCIs, picking the right 'golden subject', the person whose EEG best serves as a training template for others, directly affects decoding accuracy for other users. An earlier study using the fuzzy decision method FDOSM named S5, but that method did not separate weighting from ranking. Researchers at Imam Ja'afar Al-Sadiq University in Iraq built a unified 9 x 18 golden-subject decision matrix covering nine subjects on 18 training- and evaluation-phase metrics, including accuracy, recall, F1 score, log loss and specificity. Using FWZIC for weighting and VIKOR for compromise ranking, they found the best candidate shifted to S1 (Q = 0.0000), which ranked first in three of four weighting scenarios in a sensitivity analysis. The work is an early methodological exploration involving only nine subjects and has not been validated on larger datasets.

Muscle activity can swamp scalp EEG across a broad frequency range and bury the neural information that analysis and brain-computer interfaces rely on. Researchers built P2P Trans, a conditional adversarial denoiser that combines a U-shaped encoder-decoder, multi-head self-attention at the bottleneck, cross-attention-guided skip fusion and a PatchGAN discriminator. On synthetic data built from EEGdenoiseNet across 10 signal-to-noise levels from -7 to 2 dB, it beat 5 convolutional encoder-decoder baselines with the best mean correlation coefficient (0.8645) and lowest mean spectral error (0.3596), with the margin widening as contamination grew. It did not lead on pointwise waveform reconstruction, pointing to a trade-off between suppressing spectral noise and preserving the waveform. The work is a preprint and has not been peer reviewed.

Decoding individual words from scalp EEG and MEG is hard because word-level neural evidence is weak, spread out in time, and entangled with acoustic, lexical and semantic structure. Deep learning can already decode language from intracranial recordings, but extending that to non-invasive recordings remains unsolved. Hierarchical Dynamic Neural Decoding (HDND), proposed by Yueyang Li, Shuran Chen, Wai Ting Siok and Nizhuan Wang, treats word decoding as step-by-step refinement rather than one-shot label retrieval. Across seven EEG and MEG datasets in English, Dutch, Mandarin and Cantonese, covering nine listening, reading and reading-aloud conditions, HDND beat the matched contextual baseline on participant-averaged balanced Top-10 accuracy in all nine and had the highest mean of all compared methods in eight. The work is a preprint and has not been peer reviewed.

Researchers in Portugal had seven people with chronic stroke imagine arm movements while an EEG-based brain-computer interface gave some of them real-time feedback and others sham feedback; both groups also received four weeks of standard physiotherapy. Patients receiving real feedback improved beyond the minimal clinically important difference on at least one functional measure, most clearly grip strength, and their brain activity became more symmetrical between hemispheres. As a small case series, the study shows the method is feasible, and its results are preliminary.

Portable EEG headsets have few channels and capture limited cortical information, making emotion decoding difficult. A team at Xiamen University of Technology in the eastern Chinese province of Fujian developed the Spatiotemporal Spectral Asymmetric Fusion Network (STSANet), which models nonlinear hemispheric lateralization between homologous frontotemporal electrodes. It reached 86.80% accuracy on the public SEED dataset and 87.43% on a self-collected dataset. The team also found consistent spectral energy distributions between the portable Xmuse headset and the professional-grade Enobio, suggesting the method could run on consumer EEG hardware. The work is a preprint and has not been peer reviewed.

A team at Chongqing University's School of Electrical Engineering in southwestern China built MTANet, an end-to-end neural network for decoding motor imagery EEG. An EEG inception attention module extracts and weights coupled time-frequency-spatial features, and a parallel temporal attention module runs several temporal convolutional networks over weighted segments of different feature sequences. In offline tests on three public datasets, MTANet outperformed seven existing models, with average accuracies of 89.47% on BCI Competition IV dataset IIb, 83.39% on dataset IIa and 77.06% on BCI Competition III dataset IIIa. It has not been tested in an online BCI setting.

Pretrained EEG foundation models are increasingly seen as general-purpose encoders for brain-computer interfaces, but recent benchmarks disagree on how well they transfer. Evaluated under a validation-locked protocol on the four-class BCI Competition IV-2a motor imagery dataset, two leading models, LaBraM and CBraMod, underperformed every supervised comparator in every configuration, including validation-selected fine-tuning. A key confound emerged: foundation models and task-specific decoders are usually evaluated with different input pipelines. When three supervised architectures were retrained on the broadband input the foundation models use, accuracy moved in opposite directions (ATCNet gained 0.078, EEG Conformer lost 0.088), and none of the changes was significant after multiple-comparison correction at n = 9. Measuring the foundation-model gap against a single comparator may therefore be unreliable. The work is a preprint and has not been peer reviewed.