Papers

A preprint presents STEAM, a hierarchical transfer framework for EEG decoding that pairs a dual-branch spatio-temporal encoder with a shared soft mixture-of-experts (SSMoE) module to reconcile general representation learning with paradigm-specific specialization in EEG foundation models, letting complementary representations exchange information through a compact set of soft slots. The researchers report that STEAM attained the best average rank at competitive inference cost across 7 downstream datasets and 14 evaluation settings, and that its hierarchical pretraining further lifted decoding accuracy without retraining from scratch. The study is a preprint and has not been peer-reviewed.

CORTIVA, a candidate-score fusion framework for EEG- and MEG-to-image retrieval, reached 73.5% Top-1 and 95.3% Top-5 accuracy across ten participants on the 200-way THINGS-EEG2 benchmark, beating the strongest reported baseline by 10.3 and 5.4 percentage points. Instead of compressing heterogeneous visual supervision into a single embedding before ranking, the authors let separate decoding routes align to different visual targets and score the same candidate pool independently, merging only the temperature-scaled score vectors, which they say preserves complementary evidence; with a modality-specific encoder the same approach reached 42.4% Top-1 on THINGS-MEG. The work is a preprint and has not been peer reviewed.

The study presents a spatial harmonization framework built on a two-layer graph convolutional network (GCN) that maps heterogeneous EEG recordings onto a unified physical electrode layout while preserving motor imagery information, addressing electrode-configuration mismatches across MI-BCI datasets. Each trial is modeled as a graph so the GCN captures spatio-temporal relations, and harmonized EEG showed lower error than spherical spline interpolation while retaining the key temporal-spectral-spatial features. Combining real and harmonized EEG lifted EEGNet accuracy from 56.57% to 66.20% and FBCNet from 61.96% to 72.54% in within-session classification on Dataset A, and supported source-only cross-dataset transfer and target-domain fine-tuning.

A preprint uses radio-frequency backscatter and near-field wireless charging to tackle the wireless-link and power constraints of implantable brain-machine interfaces, noting that while high-resolution microelectrode arrays enable precise brain readout and stimulation, the 32–128 Mbps links they need are too power-hungry for a long-lived implanted battery. The approach strips the transceiver electronics out of the implant, moving complexity to the external reader to cut implant power, and powers the neural recording and stimulation chips via magnetic coupling. Preliminary tests validate the feasibility of the design; the study has not been peer reviewed.

This preprint introduces EasyBCI, which the authors say automates BCI preprocessing across six signal types with a two-phase large language model agent. The study is a preprint and has not been peer reviewed. On EEG with a fixed linear classifier, the authors report that all five EasyBCI backbones preserve more task-relevant separability than the manual pipeline, and that the system extends to five additional modalities spanning nearly three orders of magnitude in sampling rate.

Researchers released a multi-paradigm longitudinal EEG dataset from 24 stroke patients, covering a novel 'sixth-finger' motor imagery paradigm and affected-hand motor imagery. The dataset spans the full pre-training, post-training and follow-up stages and includes raw EEG, preprocessed data and patient clinical information. Preliminary analysis with classical classifiers (CSP+SVM, CSP+LDA) kept average cross-paradigm classification accuracy at roughly 85%–86%.

Researchers evaluated the technical feasibility of a multimodal imaging workflow for deploying Synchron's Stentrode endovascular brain-computer interface in a human head phantom. The workflow — thin-slice CT, transfer of DICOM data to an external core laboratory for target identification and marking, re-import of the marked CT, 3D rotational angiography, and registration with the marked reconstruction and intraoperative fluoroscopy — produced a dataset suitable for anatomical analysis, with the core laboratory marking the intended deployment region and the markings persisting through transfer and re-import, and the annotated CT fusing with 3D angiography without significant misregistration or artifacts. The authors say the workflow can generate intraprocedural targeting guidance for Stentrode deployment, as a preclinical technical validation.

Researchers achieved closed-loop speed control of freely swimming eels through lateral-line nerve stimulation, with a proportional-derivative feedback controller dynamically adjusting stimulation frequency. The biohybrid system reduced velocity deviations by 63.7% (forward) and 50.6% (backward) compared to fixed-frequency stimulation.

Researchers at the University of Michigan, Ann Arbor have built SpikeCleaner, an algorithm that grades neural units after automated spike sorting, reaching 97% accuracy and a 92% F1 score on single units in benchmarking. It combines spike rate, spike-timing metrics and waveform features to classify units as good, noise or multi-unit activity, a job that otherwise falls to manual curation.

Researchers in Pakistan and South Korea have built EmoWrite, a brain-computer interface that uses sentiment analysis to turn thought into text, typing at 6.6 words per minute with 90.36% accuracy in tests with 72 volunteers. The system adds a dynamic keyboard and visual feedback to make the interface easier to use.

Researchers writing in npj Biomedical Innovations have proposed DSTF-Net, a framework that decodes steady-state visual evoked potentials (SSVEP) from frontal EEG alone, removing the need for electrodes over the occipital cortex. In cross-subject transfer to 20 new users, including eight brain-injured patients lying supine, it improved decoding accuracy by as much as 33.47% over baseline methods.

Researchers at the University of Washington have built mechanically compliant, precision-porous brain implants and tested them in rat brains. At 4 weeks, the porous scaffolds drew less astrocyte encapsulation than solid hydrogel rods, softer hydrogels reduced pro-inflammatory macrophage polarization, and new blood vessels, neuronal markers and neurogenesis appeared inside the pores. The authors present the design as a route to limiting glial scarring and improving regeneration in implant-based central nervous system therapies.

A preprint proposes three regression models — a partial least squares regressor, a multilayer perceptron and an attention-based regressor — to decode multiple kinematic and kinetic parameters of grasp-and-lift tasks simultaneously from EEG signals. Evaluated on the WAY EEG GAL dataset, the attention-based regressor performed best with an R² of 0.8 and 29.2 ms latency, markedly improving simultaneous multi-parameter decoding, though per-parameter decoding declined; the multilayer perceptron was more consistent across the two settings but less accurate (R² = 0.49). The study has not been peer reviewed.

A preprint proposes EEGForceFusion, a hybrid EEG decoding framework that jointly models continuous and tokenized representations for grasp-force decoding, where continuous decoding is limited by complex temporal dynamics, high inter-subject variability and poor generalization. The framework combines convolutional-recurrent representation learning, quantized tokenization and Transformer-based temporal modeling in a unified fusion regression architecture to capture both fine-grained neural structure and long-range temporal dependencies. Under strict leave-one-subject-out cross-validation on the WAY-EEG-GAL dataset, it reached an offline R² of 0.817 and a simulated real-time R² of 0.793, with latency suited to real-time deployment. The study has not been peer reviewed.

A preprint proposes UnSPC (uncertainty-guided self-paced cyclic learning), a framework that integrates domain adaptation (DA) and domain generalization (DG) in an iterative cycle to address neural drift in long-term invasive brain-machine interfaces, where drift erodes decoding performance and forces frequent recalibration and where existing methods rely on DA or DG alone. Under an uncertainty-guided, self-paced pseudo-labeling scheme with a noise-robust ranking strategy, UnSPC iteratively mines reliable pseudo-labeled samples and, through cyclic adaptation and generalization, gradually mitigates both global and sub-domain drift. Experiments on multiple neural decoding datasets validated its effectiveness and robustness; the authors say it is the first method to integrate DA and DG through a pseudo-label loop. The study has not been peer reviewed.

The study uses spatial proteomics to assess tissue around non-functional intracortical microelectrodes implanted for four weeks in mice, measuring neuronal integrity, immune-cell activation and local cytokine expression around probes coated with drug-loaded, controlled-release titanium dioxide nanotube array (TNA) coatings. The authors note that blood-brain barrier disruption can translocate gut-derived bacteria to the implant site and sustain chronic inflammation, and that the TNA coating's therapeutic loading and controlled release further damp residual neuroinflammation. They conclude that TNA offers a multifunctional, tunable interface for locally regulating the neuroimmune microenvironment, a step toward long-term reliable intracortical recordings.

Researchers evaluated ruthenium-oxide-coated amorphous silicon-carbide microelectrode arrays during 25 weeks of intracortical microstimulation in rodents. Perception thresholds stabilized at about 0.4 nC per phase per electrode by week nine, behavioral performance remained around 91%, and reliable sensation persisted through week 25.

Researchers paired interquartile-range clipping for outlier removal with a random forest classifier on the UCI machine learning repository's EEG-Eye-State dataset, classifying eyes-open versus eyes-closed at 92.49% accuracy with a ROC-AUC of 0.9791. Cross-validation put mean accuracy at 92.86%. The authors present the pipeline as a stable, interpretable option for BCI uses such as drowsiness monitoring and assistive technology.

A preprint proposes SSCDL, a self-supervised consistency-enhanced disentangled learning framework that decomposes motor signals into velocity, direction, and speed to capture representations invariant to neural drift and significantly enhance cross-day decoding generalization for invasive BMIs. The researchers report state-of-the-art decoding performance with high robustness and cross-day stability across extensive experiments; the work has not yet been peer reviewed.

UC Irvine researchers tested an EEG category-decoding task in 30 participants viewing pictures and words from five semantic groups. All picture-category pairs were statistically separable, but only one word-category pair was; parietal and left-temporal electrodes contributed more to picture decoding than frontal and right-temporal sites. The bioRxiv preprint has not been peer reviewed.

University of British Columbia researchers released BEND-BCI, an open benchmark comparing 23 neural decoders across 16 real or synthetic recordings in motor, visual, speech and spatial tasks. Rankings often changed when robustness, computational cost and cross-recording representation consistency were considered alongside held-out accuracy, and simpler baselines sometimes matched or beat much larger deep networks. The bioRxiv preprint has not been peer reviewed.

A preprint proposes an adaptive EEG generation framework based on conditional flow matching to ease data scarcity in brain-computer interfaces and support large-scale neural modeling. Noting that existing flow methods assume one global time course across all channels and time segments, the framework adds position-adaptive time scheduling that tracks per-position reconstruction error to modulate each position's time course, plus decomposed spatiotemporal attention and a frequency-aligned multi-resolution spectral consistency loss to model cross-channel dependencies and compensate for EEG's power-law spectral bias. Across three EEG datasets with different acquisition protocols and task semantics, it consistently beat the strongest baselines, cutting TS-FID by up to 62.2% and lifting downstream classification accuracy by up to 6.77 percentage points. The study has not been peer reviewed.

Researchers report a preprint, posted to arXiv on July 25, 2026 and not yet peer reviewed, that contrasts Bayesian complete-pooling models against frequentist baselines for cross-subject, left-hand versus right-hand motor imagery EEG classification across 20 datasets. Six frequentist pipelines were each paired with an analogous Bayesian pipeline sharing identical feature engineering and fit via Markov chain Monte Carlo posterior sampling. Bayesian complete-pooling produced statistically but not practically significant improvements in reliability and increased predictive uncertainty, with no significant differences in Brier score, resolution, or discrimination. Bayesian pipelines consumed roughly 13 times more energy than their frequentist counterparts, and the authors conclude that complete pooling alone offers limited practical benefit, pointing to partial pooling as a more promising direction.

A registry-based cross-sectional analysis found only 11 pediatric brain-computer interface (pBCI) clinical trials worldwide, spanning 7 countries. Eight evaluated non-implanted devices and 3 evaluated implanted systems. Non-implanted trials had a median enrollment of 29 participants and median duration of 56.0 days; implanted trials had a median enrollment of 8 and median duration of 365.3 days. Only 4 studies enrolled exclusively pediatric participants; the rest recruited both children and adults. The authors conclude that current pBCI research is limited in scope and that children may be inadequately prioritized.

The researchers developed a deep learning architecture that analyzes multichannel interictal intracranial EEG to localize epileptogenic zones without requiring seizure-period recordings or manual annotation. The model combines a Morlet-wavelet temporal Transformer with a spatial attention encoder. It was evaluated with 50.5 hours of recordings from 161 patients and 17,012 channels at 7 independent centers. Leave-one-center-out validation produced a pooled AUROC of 0.778 with a 95% confidence interval of 0.748 to 0.808, and discrimination remained above chance at every held-out center. The results indicate performance comparable to established electrophysiological baselines across centers and implantation modalities, but prospective clinical validation is still needed. This medRxiv preprint has not been peer reviewed.

Researchers ran 11 healthy participants through a personalized rapid serial auditory presentation paradigm and decoded the resulting weak auditory evoked responses trial by trial with a new model, the Multi-Scale Spatial-Temporal Dual Attention Network (MS-STAMNet), reaching an unweighted average recall of 69.67±6.12% and an AUC of 0.7618±0.07, ahead of the EEGNet and PLNet baselines. Conventional measurement of the auditory frequency difference limen (FDL) depends on participants actively reporting what they hear, which leaves it open to subjective bias. Regression analysis found that neural decoding and behavioral performance came apart, suggesting the model picks up frequency deviations too small to reach conscious report.

A preprint proposes interpretable metrics that separately quantify temporal, spatial, and frequency variability in BCI-related brain activity, tested across two motor-imagery BCI datasets totaling 133 participants and validated through within-subject and cross-subject classification experiments. The researchers report negative correlations of -0.2 to -0.4 under most conditions, indicating lower variability tracks higher BCI performance, and note the metrics show deep-learning and Riemannian classifiers differ in robustness to variability, with weaker correlations for the former. The work has not been peer reviewed.

Researchers have paired steady-state motion visual evoked potentials (SSMVEP) with electromyography (EMG) in a hybrid brain-computer interface, using a parallel architecture to expand the command space from 15 targets to 60. The multimodal setup reached an information transfer rate of 62.33 bits/min, against 42.49 bits/min for the best single-modality condition. Deep-learning decoding of the two signal streams held classification accuracy steady while lowering the effort required of users with severe motor impairment.

A post from SpikeGadgets surveys how far multichannel electrophysiology has scaled: its Modular Stacking Headstage supports chronic recording of 1,024 channels across 10 brain regions; the University of Pittsburgh's MePhys platform uses 992 electrode contacts to cover an entire macaque hemisphere; and a Rice University preprint describes a custom ASIC that supports 5,376 simultaneous recording channels.