Papers

Spotting a deviation of a few degrees can matter in precision tasks, yet conventional training that only reports whether an answer was right does little to sharpen sensitivity to small errors. Researchers at The University of Texas at Austin and the National Institute of Neurological Disorders and Stroke instead read the brain's own error signal, the error-related potential, from EEG and fed its presence or absence back to participants in real time. Over a five-day longitudinal study, this BCI training accelerated perceptual learning for 3° errors and improved performance for 6° errors compared with behavioral training alone, alongside a larger error positivity (Pe) amplitude.

Researchers developed a reproducible multi-layer brain-mimicking phantom that replicates the dimpling and rupture forces of rodent pia mater and dura mater during neural interface implantation, built from a 0.5% agarose skin layer, a 1.01% agarose pia layer and a pre-stretched PVC dura layer assembled in a simple benchtop process. Tested with a cantilever force system on microwires of 12–100 µm diameter (tungsten and stainless steel, various tip geometries) and segmented silicon probes, the phantom produced results within the range of in vivo Sprague-Dawley rat data with significantly lower insertion variability than in vivo testing. The authors say its modular design — layer thickness and stiffness can be tuned for different species or devices — makes it a low-cost early screening platform that can accelerate neural implant development while reducing animal use.

The study presents DSNet, a deep denoising structure-preserving neural encoding network that classifies privacy-preserving encoded EEG without decryption. Common spatial pattern (CSP) features are converted to irreversible neural codes — an irreversible neural transformation designed for privacy rather than a formal cryptographic guarantee — and two deep-learning architectures (a feedforward network and a recurrent RNN) classify in the encoded feature space. On public datasets, DSNet-NN exceeded 87% accuracy for every subject, outperforming the RNN variant and baselines while staying resilient to simulated privacy attacks; the study also integrated GPT-4 to generate clinical-style summaries from model outputs.

French neurotech company BIOSerenity ran three months of large-scale EEG AI training experiments on Jean Zay, France's most powerful supercomputer for research, including a 300-million-parameter model trained on 110,000 hours of EEG data across 64 GPUs. It also fully trained two model families, Mercury and Neptune, in three sizes each, from about 13 million to 100 million parameters and on about 5,000 to 40,000 hours of data. Neptune improved as it scaled up while Mercury did the opposite, its smallest version performing best, a sign that brain-signal AI does not scale the way language models do; the results have not been peer-reviewed.

Researchers at Beihang University and Tsinghua University in Beijing built a multimodal contrastive learning framework that pairs EEG with synchronized video and motion data to detect cybersickness in virtual reality, representing the EEG as a connectivity graph and using an attention-based encoder to map the video and motion streams onto the same structure. Fusing all three signals separated cybersick from non-cybersick states more cleanly than any single modality or pair, and the model automatically pruned prefrontal connections unrelated to cybersickness. The setup used mBrainTrain's Smarting PRO 32-channel wireless EEG system and a Pico 4 Ultra headset, with 29 healthy adults navigating a VR scene under their own control.

A preprint from the Utrecht-BCI Lab at the UMC Utrecht Brain Center in the Netherlands, the invasive-BCI group led by Nick Ramsey, asks why some people with ALS control an implanted brain-computer interface (BCI) better than others. It found that preservation of grey matter in the motor cortex is associated with higher-quality brain signals, suggesting an MRI-based measure could help identify who is a suitable candidate for implantation. The preprint was posted to medRxiv on June 23, 2026 (DOI 10.64898/2026.06.23.26355654).

French medtech company BIOSerenity has built an EEG foundation model, BIOSerenity-E1, pre-trained self-supervised on more than 4,000 hours of recordings, and reports strong results on four clinical tasks: normal/abnormal classification, Alzheimer's disease detection, pediatric sleep staging and seizure detection. The normal/abnormal algorithm is already built into a CE-marked medical device, and the seizure-detection algorithm is in clinical evaluation. The work will be shown as a poster at the 8th Journées de Neurophysiologie Clinique in Grenoble, France.