Papers

A research team in Poland used just two occipital electrodes (O1 and O2) on an EEG headband to distinguish four classes of eye events (blinks, left saccades, right saccades and a neutral state) with 72.37% test accuracy. Instead of training a single stronger model, they split four one-dimensional convolutional neural networks into binary experts and used stacked generalization, letting a meta-classifier resolve conflicts between them. Notably, the experts that performed worst on their own produced the best results when combined. The work was published by Neurotechnology, an EEG hardware maker whose BrainAccess brand makes consumer EEG headbands; HALO is its four-channel portable headband, and the accompanying software is free. Eye tracking usually requires dedicated equipment, and this study shows that two occipital channels on a consumer EEG headband can also pick up blinks and saccades. But the data come from only 4 healthy participants and 1,284 samples, so the approach is still a long way from replacing an eye tracker.

Translating scalp EEG directly into free-form text has long been seen as one of the most ambitious goals for non-invasive brain-computer interfaces. But when researchers fed random noise instead of real EEG into several published decoders, some models still produced fluent sentences and scored about as well as they did on real brain data, suggesting the language model was doing most of the work. The field has since made noise-baseline tests and decoding without teacher forcing standard validation practice, and has used magnetoencephalography (MEG) as a comparison to quantify how far EEG trails cleaner signals.

Public datasets for EEG-based brain-computer interfaces (BCIs) keep multiplying, but electrode layouts, task definitions, preprocessing and participant records differ from study to study, making the data hard to reuse across datasets and the results hard to replicate. This review screened 16,920 records down to 129 publications, mapped the fragmentation across structural, semantic, procedural, human/contextual and computational layers, and scored reporting transparency against 10 criteria, with a median score of 9. Data availability was stated in 61.2% of the papers; code or pipeline availability in just 20.9%. Existing standards, ontologies, software platforms and transfer-learning methods each solve part of the problem, the authors conclude, but none yet delivers full semantic interoperability.

Researchers paired an 8-channel OpenBCI Cyton board and an EEG cap laid out on the international 10-20 system with a Meta Quest 2 headset to capture motor imagery and attention signals, reaching up to 90% subject-specific online classification accuracy with the EEGNet model when the signal was stable. The work targets cognitive control training for attention-deficit/hyperactivity disorder (ADHD), building EEG sensing into a virtual reality (VR) serious game so users can regulate their mental state in real time; earlier approaches mostly relied on hardware-heavy multichannel systems and did not combine motor imagery with attention levels. The team also documented artifacts from high impedance, channel saturation and mechanical tension from the headset strap, showing that a low-cost setup is usable only once skin-electrode impedance and mechanical interference are under control. This is a single-subject feasibility study with no clinical evaluation.

Combining EEG with functional near-infrared spectroscopy (fNIRS) did not make brain-controlled electrical stimulation more accurate in a blinded randomized trial of 16 healthy volunteers. The hybrid and EEG-only groups showed no statistically significant differences in real-time three-class recall, sense of agency, attention or physical comfort; median recall was 53.5% in the hybrid group and 57.3% with EEG alone. The researchers also released the full EEG-fNIRS dataset.

Can a machine tell from EEG alone whether a person is actually moving or only imagining the movement? A preprint study tested a Random Forest classifier on consumer-grade and research-grade EEG systems and found it could separate the two kinds of activity and identify which body part was involved. EEG is a basic tool for studying the brain's electrical activity during movement, and BCIs use that activity to build assistive technologies, particularly for people with physical disabilities; because extracting features and patterns from the signal remains complex, the task is often handed to machine learning. The study also found that differences in EEG between individuals drag down classification accuracy, meaning a model that works for one person may perform worse on another. The work is a preprint and has not been peer reviewed.

Which new technology does most for upper-limb recovery after stroke: virtual reality, robotics or a brain-computer interface (BCI)? A systematic review and network meta-analysis placed all three in a single evidence network, pooling 25 randomized controlled trials and 1,145 stroke survivors. The authors searched PubMed, Web of Science, the Cochrane Library and Embase from inception to October 2025, used conventional physical therapy as the common comparator in a star-shaped network, and applied a Bayesian random-effects model to estimate relative efficacy and calculate SUCRA rankings. Robot-assisted training produced the most robust findings, with two studies supporting robotics plus conventional physical therapy and three supporting robotics alone; only one BCI study yielded extractable data, too little to judge efficacy. The authors caution that the top-ranked intervention, robotics combined with rehabilitative functional electrical stimulation, rests on a single trial and should not be read as definitive evidence of superiority.