Decoding Algorithms

Researchers developed an EEG transform that requires no trained privacy model to reduce features linking recordings from the same person across sessions. Tests on two public datasets used five attackers spanning four covariance-space representations. The strongest attacker's accuracy fell from 90.8% to 15.6% on BCI Competition IV-2a and from 90.8% to 9.9% on OpenBMI. Relative to the 8-30 Hz input, motor imagery decoding accuracy decreased by 7.0 and 2.1 percentage points, respectively. The transform was fitted on one session and applied unchanged to a held-out later session. This study is a preprint and has not been peer reviewed.

A validation study compared the visual BCI experiment platform Nimbus Studio with independent, hand-written reference implementations. It tested six decoding pipelines on three public datasets spanning motor imagery and P300 tasks, using within-subject session holdout and cross-subject evaluation. Deterministic pipelines matched exactly across all 54 session-holdout subject-runs and all CSP/MDM grouped cross-validation folds, with a maximum difference of 0.00 points. The remaining grouped-CV cells differed by at most 0.06 points. Stochastic deep models agreed within seed-to-seed training variability. The study assessed computational agreement with reference implementations, not clinical efficacy. This study is a preprint and has not been peer reviewed.

Researchers introduced Neurons++, a framework for reconstructing static images and dynamic videos from functional magnetic resonance imaging recordings while examining how the outputs relate to brain signals. It uses image-fMRI pretraining followed by subject-specific fine-tuning for image and video reconstruction separately. The authors report state-of-the-art video reconstruction and competitive image reconstruction on standard benchmarks. Neural perturbation tests, cortical-feature sensitivity analyses, and brain-region-guided semantic modulation probe how outputs respond to neural signals. These model-level analyses do not establish that generated content faithfully reproduces a person's visual experience. This study is a preprint and has not been peer reviewed.

Researchers used natural-language class descriptions to help models learn the correspondence between brain signals and target classes, evaluating the approach in visual, motor, and speech decoding. The study tested seven encoder architectures with independently trained models using scalp and intracranial recordings. Speech-decoding accuracy improved by up to 17.15 percentage points over matched class-label-supervision baselines. The learned representations also supported candidate-set reuse and label-supervised subject and session adaptation. In steady-state visual evoked potential tasks, the method enabled unseen-frequency decoding under specific conditions. These findings do not establish a single model that performs every task or adapts to new subjects without labels. This study is a preprint and has not been peer reviewed.

The study proposes inferring cognitive states using brain-intrinsic prior constraints rather than simply mapping neural signals to predefined external labels. Across five neural recording modalities and three cognitive domains—motor, perception and internal mentation—the method reorganized neural representations. The resulting high-level semantic representations showed consistent geometric relationships across tasks and enabled recovery of stable cognitive states from variable observations. The authors interpret these findings as suggesting that relatively stable cognition may be maintained through inference without fixed neural activity patterns; this interpretation is distinct from the experimental observations.

A contextual prior can raise overall accuracy in MEG speech decoding while making some correctable errors more confident, a preprint reports. Tracking candidate scores across two magnetoencephalography datasets, the authors describe a reversal: when the correct candidate starts near the top of the local ranking, a larger fused margin makes a repair more likely, but when it starts lower, a larger margin makes the error more likely to persist. On MEG-MASC, pooled correctness AUROC is 0.87, while the AUROC separating repairs from residual errors falls from 0.70 to 0.39; errors starting beyond rank 20 account for 46.6% of post-fusion errors. The authors argue confidence should retain local and contextual evidence rather than fused scores alone; reading the two scores separately lifts decoder coverage to 74.5% of windows from 56.7%. The study is a preprint and has not been peer reviewed.

EEG-based brain-computer interfaces lose accuracy when the user or session changes, because the signal distribution shifts, and users usually have to sit through labeled calibration. SPDAlign tackles this without labels. The researchers show theoretically that shifts caused by the forward process, linked to factors such as head geometry, can be recovered solely through linear transformations on the Symmetric Positive Definite manifold. The framework aligns domain-specific means, then corrects global rotations with an optimal transport technique called Wasserstein Procrustes. The authors report competitive performance on public EEG datasets and say the method can identify relevant frequency ranges and spatial patterns. The study is a preprint and has not been peer reviewed.

Researchers in Germany fitted participants with lightweight EEG electrodes around the ear and a portable hearing aid research platform, then tested auditory attention while they sat and walked. Speech reconstructed from the EEG followed earlier lab findings: most accurate for a single talker, then the attended talker of two, and least for the ignored one. Walking introduced a speech-synchronized artifact, possibly from hearing aid cables touching the electrodes, which affected accuracy; the researchers conclude feasibility is shown only for stationary settings.

EEG signals drift, so a brain-computer interface model calibrated in the lab often loses accuracy in everyday use. This preprint puts Common Spatial Patterns (CSP) and Riemannian covariance-based pipelines through the same time-ordered evaluation: predict first, then update after the true label arrives. Across 4 public datasets, label-revealed updates lifted accuracy in 13 of 14 model/dataset pairs on the 2 largest streams, up to about 18% over a frozen model. A Shapley-based valuation gave the most recent time block the highest mean value in all 3 datasets analyzed, while older blocks retained positive value.

Researchers recorded EEG during overt and imagined vocalisation to distinguish five emotion categories: anger, happiness, neutral, sadness, and pleasure. The overt experiment involved 14 participants and the imagined experiment involved 21, yielding single-trial decoding accuracies of 78.1% and 36.4%, respectively, against 20% chance. Facial surface EMG recorded in five participants during imagined vocalisation indicated that involuntary muscle activity contributed to EEG spectral and spatial patterns used for classification. One lateral-channel pattern was stronger during imagined happy vocalisations. The results mean that decoding accuracy cannot be attributed entirely to brain-derived emotion information.

An ISAE-SUPAERO study tested a brain-computer interface combining semi-dry EEG acquisition with visual stimulation. Thirty participants completed an 11-class keypad typing task on a standard screen. With 33 s of calibration data, the system achieved 99.8% accuracy and a mean decoding time of 4.3 s. No spurious commands occurred during a 2-min video-viewing idle condition between two online typing sessions. A virtual reality proof of concept reached 100% online accuracy with a mean decoding time of 5.6 s. Participants reported high visual comfort and little peripheral distraction in both conditions, but greater fatigue in VR. The short idle test does not establish long-term performance.

Researchers found that a decoder trained on peripheral physiological signals (heart rate, heart rate variability, respiration, electrodermal activity and pupil diameter) read arousal more accurately than EEG, reaching a within-subject area under the curve (AUC) of 93.0%, versus 85.2% for an EEG-only decoder and 79.8% for the original filter-bank common spatial pattern decoder. The offline analysis used a public simulated-flight dataset. Only the peripheral decoder reproduced the inverted-U relationship between arousal and performance, and EEG neurofeedback did not change overall arousal trajectories, suggesting its effect may come from regulation at task-critical moments. The study is a preprint and has not been peer reviewed.

When a user looks at targets flickering at different frequencies, the visual cortex produces EEG responses at matching frequencies, known as steady-state visual evoked potentials (SSVEP), which a brain-computer interface can use to infer the intended target. Such systems usually need calibration data, and more data means more user fatigue. Researchers propose SCR-FTSDNN, a deep convolutional network that moves EEG into the frequency domain, then extracts temporal and spatial features with added spatial and channel reconstruction modules. Tested on the public Benchmark and BETA datasets, the authors report higher recognition accuracy and stability than several existing methods, especially with limited calibration data. The abstract gives no specific figures, and the results come from offline tests.

A study combines EEG data with multilingual text to classify emotions, since EEG alone is prone to noise while language models read sentiment in words but miss implicit changes in the brain. The author reports 89.7% accuracy for the fused model, versus 75% for EEG alone and 79% for text alone, with an F1-score of 0.88 and an average inference latency of 180 ms, enough for real-time recognition. The abstract does not state the number of participants or the datasets used.

Researchers have proposed CortexBridge, a lightweight adapter that lets EEG foundation models, usually pretrained on a single fixed electrode layout, handle any montage. It combines EEG features with electrode and brain-atlas coordinates to map signals into a shared cortical latent space before passing them to the pretrained model, which is left unchanged. On 5 BCI datasets from the public MOABB benchmark, tested with 3 frozen foundation models, EEGPT, LaBraM and CBraMod, it improved performance in 13 of 15 evaluations, raising balanced accuracy by an average of 0.80%, 0.70% and 3.26% respectively, with a maximum gain of 13.02% on 12-class steady-state visually evoked potential classification. The study is a preprint and has not been peer reviewed.

Researchers have used a generative adversarial network, an AI model that can produce synthetic data, to generate artificial primary motor cortex signals from other functionally related cortical regions and add them to training data, addressing the shortage of task-relevant signals in any single subject's recordings. On MEG data recorded while subjects performed reaching tasks, motor decoding performance rose by about 10% (p < 0.05) after augmentation, and the gain held even without real primary motor cortex signals. The method was also applied to a motor imagery BCI competition dataset to improve classification accuracy.

Researchers combined high-resolution surface electrophysiology recordings from the first 70 patients implanted with an anatomically conformal, thin-film neural interface to build a functional atlas of the human central nervous system. The interface was positioned at more than 100 locations, with >100,000 electrode deployments spanning motor, sensory, language and other functional regions, as well as the brainstem and spinal cord. Aggregated recordings showed spatially ordered digit representations, movement-related frequency-band changes and task-modulated neural population activity. The study also demonstrated real-time neuroprosthetic control and decoding of cortical activity associated with speech, facial movements and emotional expressions. This approach combines recordings across patients rather than simultaneously reading the entire central nervous system in one person. This study is a preprint and has not been peer reviewed.

Color vision tests usually rely on what the subject reports, while a brain-computer interface can read the brain's response to color differences directly: the more two flickering light sources differ in perceived color, the stronger the steady-state visual evoked potential in the EEG. Existing systems sample color space evenly and repeat measurements at each point, which is accurate but slow. MetaIDNet, a closed-loop system, instead uses Gaussian process regression and a neural network to decide where to measure next. Offline, it made assessment 70% faster at the same accuracy; online, it improved accuracy by 40% at the same speed. The offline and online experiments involved only 10 and 8 participants.

Researchers have proposed NEUROTOKEN, a single network that uses conditional flow matching to score candidate voices and identify, from EEG, which speaker a listener is attending to in a noisy room, the so-called cocktail-party problem. On the public KU Leuven, DTU and NJU datasets with 5-second windows, it raised per-segment accuracy by 9% to 16% over the strongest non-generative baseline and narrowed across-subject variance by about 3x. The authors also report that the 95% to 97% accuracies often cited for decoding attention direction fall by 17% to 45% once training and test trials are strictly separated. The work is a preprint and has not been peer reviewed.

A meta-analysis of 26 randomized controlled trials covering 1100 stroke patients found that brain-computer interface (BCI) rehabilitation improved upper-extremity motor function compared with control interventions, but showed no significant benefit for lower-limb function or walking tests. A dose-response analysis suggested an inverted U-shaped relationship, with the peak effect at approximately 48 min per session.

Speech brain-computer interfaces aim to restore communication by turning neural activity tied to speech, language or communicative intent into text, synthesized voice or avatar movements. This review argues they are not simple neural-to-text decoders but clinical systems that change over time, in which neural representations, recording hardware, decoding models, language models, feedback and the user's own learning interact. The authors say the next generation should be judged not only on offline accuracy but on stability across sessions, calibration burden, latency, the ability to signal uncertainty, usability and safeguards against decoding words a user did not mean to say. The paper is a preprint and has not been peer reviewed.

Researchers have released BrainNet Studio, a single toolkit that takes brain network data from construction and feature extraction through predictive modeling and visualization. Brain networks map structural and functional links between brain regions, but most existing tools handle only static networks and miss how those links change over time. The toolkit bundles 27 algorithms spanning deep learning, graph neural networks and spatiotemporal sequence models, and uses a large language model to generate connectivity summaries that researchers can check. The developers say it can support cognitive neuroscience, exploratory brain disorder studies and brain-computer interface research. The work is a preprint and has not been peer reviewed.

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.

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 proposed BrainVLA, a framework that borrows a vision-language-action model pretrained on large amounts of robot data to decode what a paralyzed user intends to do. A key obstacle for BCIs that control external devices is the scarcity of paired neural and movement data; the usual approach of learning a task-specific map from neural activity to kinematic parameters is held back by that shortage. The team first built a paired dataset of neural activity, action signals, language instructions and rendered visual observations, then adapted the OpenVLA-OFT policy to the target action space through LoRA fine-tuning. Neural activity is aligned with language representations to extract an intention signal, which the model combines with visual input to generate executable actions. On 2 neural motor datasets with different action dimensions, the framework outperformed the evaluated baselines in cross-session decoding R² and task success rate, and was data-efficient to train. 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.