Decoding Algorithms

Intracranial EEG (iEEG), recorded by electrodes implanted inside the brain, is widely regarded as an ideal signal for decoding intent, but differing datasets and preprocessing pipelines make it hard to tell whether models are actually improving. A research team built iMINDBench, a benchmark that brings together naturalistic movie-watching data from three institutions, 15 decoding tasks, standardized preprocessing and fixed evaluation splits. Pretrained systems generally beat baselines within their own preprocessing pipeline, but classic spectral baselines remained competitive on other institutions' data. Scaling up supervised data from other subjects or institutions to 25 times the volume yielded only limited, task-dependent gains over training on data from the same session.

Wearable neurotechnology and BCI applications, from assistive interfaces to clinical monitoring, favor single-channel EEG because it needs few electrodes and is light to wear. But muscle, eye and motion artifacts are hard to remove from a single channel, especially without auxiliary channels, artifact labels or hand-picked clean baseline segments. SPAR-EEG needs none of these: it applies three artifact-specific attenuation passes to each EEG segment, one based on variational mode decomposition for high-frequency muscle bursts and two based on singular spectrum analysis for blink-like ocular transients and slow motion drift. It achieved the best artifact-region SNR improvement at all 26 SNR levels tested and raised final accuracy by 7.8 percentage points in a P300 speller task using only the FP1 and FP2 leads.

BCI hardware and decoding algorithms have advanced quickly, but basic questions such as how much intent can be read from brain signals and how low decoding error can go still lack a systematic theoretical answer. This paper treats a BCI as a closed-loop stochastic process: a user forms an intention, the brain encodes it as neural activity, the device observes the signal, an algorithm decodes it statistically, and feedback returns to the user. Using information theory and statistical decision theory, the authors derive theorems on when intent is decodable, the upper bound of observable information, the lower bound of optimal decoding error, and the convergence of closed-loop learning, plus laws describing constraints from signal-to-noise ratio, class separability and non-stationary co-adaptation. A four-class SSVEP-BCI, in which users gaze at targets flickering at different frequencies, serves as a theoretical case study. The work is theoretical and reports no new experimental data.

In three participants with tetraplegia imagining object manipulation, neural activity across several brain regions represented the configuration of the whole hand during imagined grasps. The posterior parietal cortex simultaneously encoded grasp and object properties, largely through distinct neuronal subpopulations. Grasp-related activity varied with motor planning versus imagined execution and with object presence versus absence; some regions also represented object shape during planning. The authors suggest that separating grasp and object information could support more stable BCI grasp control. However, the motor imagery task does not establish restored hand function or improved robotic grasp performance. The study is a preprint that has not been peer reviewed.

Researchers introduced a framework for analyzing neural traveling waves and identified oppositely directed propagation modes in marmoset electrocorticography and human EEG recordings. The waves were concentrated in specific spatial and temporal frequency ranges, involving alpha and low-beta activity, with propagation speeds the authors considered physiologically plausible. The framework analyzes temporal frequency, spatial frequency, and propagation direction, accommodates irregular sensor layouts, and helps distinguish traveling from standing waves. The study also introduces a single-trial detection method. These findings concern animal and human neural signal analysis, not demonstrated improvements in BCI control or clinical outcomes. The work is a preprint that has not been peer reviewed.

During removal of a right parafalcine meningioma, surgeons placed four Layer 7 μECoG arrays, 4,096 electrodes in total, on either side of the central sulcus and recorded somatosensory evoked potentials under 5 contralateral stimulation conditions. Using a 2 MΩ impedance cutoff, 3,739 channels were usable, a 91.3% yield; phase-reversal latencies were 19, 21, 25, 26 and 25 ms, consistent with the standard intraoperative mapping performed in the same operation. The researchers caution that the setup provides dense spatial sampling rather than submillimeter physiological resolution: measured responses were correlated across roughly 3–4 mm of cortex.

After days of practicing a brain-computer interface with adaptive decoders, rhesus macaques needed significantly fewer motor cortex neurons to reach 80% task classification accuracy than in early sessions (p = 9.8 x 10^-4). Fixed decoders produced no such compaction. A neural network model suggests the adaptive decoders directly contribute to these more compact representations; the authors propose this may be because the decoders alter the error information used for credit assignment. The finding implies that assistive algorithms do not merely boost immediate BCI performance but actively reshape how the brain learns over time -- a consideration for real-world decoder design. The study used 128-channel microwire arrays in motor and premotor cortices of two monkeys and has not been tested in humans.

Researchers introduced NeuroGate, a cross-regional signal translation framework that predicts an unrecorded brain region's waveform from a recorded region. Across 56 pathways spanning rodent local field potentials, human stereo-EEG and electrocorticography, and scalp EEG, its waveform predictions were more accurate than those of 13 deep-learning and linear baselines. To test directionality, the researchers silenced piriform-cortex output. Piriform-to-bulb predictions collapsed in all 8 folds, while reverse predictions were preserved. The authors interpret this as consistent with monosynaptic anatomy and demonstrate applications in assessing recording redundancy and testing directional translation. This study is a preprint and has not been peer reviewed.

Researchers showed that seizure-prediction methods lose much of their accuracy in continuous use, with mean AUROC dropping from 90% offline to 57.9% online across 664.9 hours of intracranial EEG from 16 patients with drug-resistant epilepsy, because brain signals drift over time. An online evolutionary framework that keeps selecting and updating models on the data stream predicted every seizure in 75% of patients and missed only one in 87.5%. Reading only contacts in the epileptogenic network cut the required contacts by 79.12%. The team has released the dataset. The study is a preprint and has not been peer reviewed.

Training a decoder to read brain signals usually means collecting a large labeled dataset from every new subject, which is slow and a burden on patients. The proposed method, MAPA, first runs self-supervised pretraining on unlabeled intracranial EEG recordings pooled across subjects, then transfers to new ones. The difficulty is that electrode contact placement and neuroanatomy vary from person to person, so MAPA adds two spatial encodings, an anatomical region embedding and a relative positional encoding, to a standard masked autoencoder. In cross-subject tests, about 164 labeled trials were enough to reach the accuracy that otherwise takes 3,500. The team reports that MAPA set new best results on the Neuroprobe benchmark in all three settings, within-session, cross-session and cross-subject, without fine-tuning, suggesting that calibration for implanted BCIs could become much shorter.

Because the fingers' representations sit close together in the motor cortex, telling individual finger movements apart from scalp EEG has long been difficult. Researchers first learned a set of spectral projections from simultaneously recorded EEG and functional MRI (fMRI), then used the class geometry derived from fMRI to correct EEG predictions. In tests on 12 healthy participants, group average accuracy for two-class movement execution rose from 66.93% to 74.53%, and for three-class execution from 44.83% to 56.58%; for two-class motor imagery it rose from 80.78% to 85.63%. At inference the system uses EEG alone, with no paired fMRI data required.

Researchers recorded passengers' EEG as they watched driving scenes in a highly automated vehicle, then trained models to judge whether a risk lay ahead and where the hazard appeared. A 3D-CRNN reached 95.3% ± 2.7% balanced accuracy in risk prediction and raised hazard identification from 80.9% to 85.0%. In cross-subject tests on passengers the model had not seen, balanced accuracy fell to 64.9% ± 8.5%, showing the approach is still some way from deployment.

P300 BCI spellers are often built by switching on every component that works on its own, on the assumption that more is better. A preprint study tested that assumption with a four-component full-factorial experiment and found that a component's value is conditional, not additive. Subject calibration was the strongest single contributor; Euclidean Alignment made up for the lack of calibration in zero-calibration settings; and stacking components that are each useful on their own could lower performance, an effect the authors call component anti-synergy. Language model support was not universally beneficial either: its effect depended strongly on how strong the underlying EEG pipeline was.

Ten native Mandarin speakers read aloud six single-vowel rhymes while optically pumped magnetometer magnetoencephalography (OPM-MEG) and EEG recorded the same task. In the 150-500 ms window after stimulus onset, OPM-MEG decoded significantly more accurately than EEG with all five classifiers tested. The best-performing combination, common spatial pattern features with a linear support vector machine, reached a mean accuracy of 60.3% for OPM-MEG versus 56.7% for EEG. That suggests speech BCIs that need no implanted electrodes and do not restrict head movement may have a clearer signal path than EEG.

Participants steering a virtual avatar with a brain-computer interface reported a stronger sense that the avatar's body was their own when they imagined jogging than when they imagined opening their right hand, even though jogging imagery produced weaker EEG signals. Forward movement was driven by motor imagery-related sensorimotor rhythm event-related desynchronization from scalp EEG and direction by eye gaze, as participants guided a jogging avatar along a curved course before rating their embodiment in questionnaires. The researchers say neural signal strength and embodied experience can diverge, so the congruence between imagery and action should be weighed alongside standard decoding metrics.

Researchers have mapped the brain-computer interface (BCI) attack surface along five dimensions (forged neural signals, desynchronization-based evasion, replay hijacking, "Vein Tapping" eavesdropping and embedded backdoors), which they group as "NERVE Attacks" and describe as orthogonal and together spanning the full BCI stack. Their EEGle framework, which the team is releasing to the community for building and verifying device security, surfaced 17 new neuro-specific attack instances and a stealth-versus-effectiveness trade-off in backdoor design. The authors warn that generative AI is lowering the barrier for non-expert attackers, with risks to cognitive autonomy, mental privacy and physical safety, from neural data exfiltration to malicious control of BCI-connected devices. The preprint, by Zahra Tarkhani, Georgios Akkogiounoglou, Lorena Qendro, Isabel Tscherniak and Anil Madhavapeddy, has not been peer reviewed.

Even when the eyes stay fixed on one point, the brain tracks moving objects in a video, and that tracking grows stronger with attention. But researchers at KU Leuven found that EEG tracking of an object's motion weakens the farther the object sits from the fixation point, meaning decoding methods that read attention from tracking strength may mistake where an object is for where attention is.

How can people with severe motor disorders bypass damaged nerve pathways and steer a robotic arm with their thoughts? This review traces the full signal chain: recording neural activity from the motor cortex, preprocessing it, extracting features, decoding it and generating continuous, discrete or multi-degree-of-freedom arm commands. The author compares signal acquisition in invasive, non-invasive and intravascular BCIs and argues that jointly optimizing signal quality, decoding algorithms and shared control is key to more natural and reliable arm operation, while noting that long-term stability, control accuracy and clinical safety remain limited.

A team at the Institute of Biomedical Engineering, Chinese Academy of Medical Sciences and Peking Union Medical College, has built a hybrid brain-computer interface (BCI) that combines surface electromyography (sEMG) with steady-state visual evoked potentials (SSVEP), encoding 480 targets with 120 flicker frequencies and four hand gestures. In online experiments it achieved a mean classification accuracy of 84.55 ± 7.23% and a mean information transfer rate (ITR), a key measure of practical BCI performance, of 260.07 ± 30.41 bits/min. Its command set matches the largest target counts in existing systems, and the authors place its performance among the top, offering a technical reference for large-command-set BCIs. The work appeared in Cognitive Neurodynamics on September 7, 2026.

Researchers have proposed a compact hybrid deep learning model for classifying four-class motor imagery EEG signals. Combining spatial convolutional filtering, bidirectional temporal modeling and an attention mechanism, the end-to-end model reached a mean accuracy of 83.89% on the public BCI Competition IV Dataset 2a, outperforming conventional baselines. The authors say bidirectional temporal modeling and attention weighting make motor imagery classification more robust, which could help BCIs assist patients with motor disabilities, and that the model offers an efficient option for decoding in resource-constrained settings.

A transformer neural network trained on single-neuron spikes, recorded with intracranial microelectrodes while participants watched an audiovisual episode once, predicted which specific people or places they would name during later free recall. The brain regions carrying that information shifted across sleep: medial temporal lobe signals decoded memories before sleep, frontal cortex signals after. The authors say such decoding could point toward brain-computer interfaces for episodic memory disorders. The study is a preprint and has not been peer reviewed.

Speech imagery (SI) brain-computer interface results may be hard to reproduce, a University of Essex team reports. When the researchers re-ran published decoding pipelines, classification accuracy averaged 11.25 percentage points below the original reports, with gaps ranging from 2% to 39%; in a replication analysis, only 36% of SI participants cleared the threshold for statistical significance, against 91% in motor imagery (MI) datasets. Published in the Journal of Neural Engineering on September 4, 2026, the study is the first to assess both the reproducibility and the replicability of SI decoding. Every SI study evaluated was missing methodological details, and the authors conclude that SI's feasibility as a practical BCI paradigm may have been overestimated, a warning for clinical applications and follow-up research that rely on it.

Researchers have added a dedicated readout neuron that outputs decision confidence to an echo state network, a neuro-inspired computational model suited to time-series data, so that confidence is updated continuously as a decision forms. In a simulated decision task, confidence activity rose as stimuli became stronger and easier to distinguish. In EEG-based motor imagery classification, confidence activity increased with decision accuracy and distinguished correct from incorrect decisions, especially in better-performing participants. The study is a preprint and has not been peer reviewed.

Can people with poor eyesight still steer a brain-computer interface (BCI) just by shifting their attention? Researchers had healthy volunteers covertly attend to colored targets in their peripheral vision under simulated blurred and monocular vision. The impairments changed the brain's electrical responses to the stimuli and shrank some of them, but the BCI decoded the users' choices just as accurately. The results suggest that as long as a target's color can be recognized, a gaze-independent, non-invasive BCI may stay usable for people with visual impairments. The study is a preprint and has not been peer reviewed.

A preprint posted to arXiv on September 3, 2026, sets out the 2026 PNPL competition, built on an extended LibriBrain100 dataset that adds 32 subjects at about 40 minutes each and roughly 80 more hours of within-subject data. The competition runs two tracks: a Deep track for within-subject word classification at scale, and a Broad track for cross-subject generalization that steps subject-specific fine-tuning data down from about 40 minutes to 20 and then 10, a clinically feasible range. Winning 2025 submissions reached F1-macro scores of 95.6% on speech detection and 73.6% on phoneme classification.

A wireless implantable neural interface, the Brain Interchange (BIC), paired with artifact-removal algorithms captured about 82% of the high-frequency oscillations (HFOs) that a clinical-grade amplifier detected across 24-hour intracranial EEG recordings in 10 patients with drug-resistant epilepsy, according to a preprint. Using those HFOs, the system localized the seizure onset zone with 84% sensitivity and 90% specificity, comparable to clinical equipment. The authors present the work as a translational framework for chronic tracking of epileptogenic networks and for future biomarker-guided adaptive neuromodulation.

Most speech brain-computer interfaces for people who have lost the ability to speak first decode neural signals into text and then read that text aloud, a two-step chain that adds delay and passes errors from one stage to the next. Brain2Speech-Net skips the text step, mapping long neural recordings directly into the internal representation of a speech synthesis model. The authors report it was the only system in their comparison to produce intelligible speech while generating faster than real time. The work is a preprint and has not been peer reviewed.