Decoding Algorithms

A study proposes a hierarchical alignment framework to retrieve a read sentence using EEG signals and context. It focuses on sentence-level EEG-to-text classification, rather than directly generating sentences or identifying arbitrary thoughts from EEG alone. The authors note that much prior decoding research has focused on phonemes, letters, and words, while sentences carry richer meaning. They also describe sentence classification as less studied than sentence generation. The framework explores EEG-text relationships to inform the choice between retrieval and generation. Non-invasive assistive communication is its application context, not an established benefit for people with communication disabilities.

Researchers from Wuyi University and partner institutions proposed a method that combines EEG and functional near-infrared spectroscopy signals, reporting mean accuracy above the best baselines by 2.34, 0.59 and 1.47 percentage points on public-dataset tasks involving mental arithmetic, motor imagery and emotion recognition, respectively. The latter two differences were not statistically significant. The method separates information shared by the two signal types from information specific to each, then combines these representations for prediction. It addresses redundant multimodal information and difficulties in capturing temporal and spatial relationships. The authors interpret the findings as evidence of improved hybrid BCI decoding and strong generalization. These public-dataset results do not establish real-time performance or clinical effectiveness.

Researchers propose VARY to address a problem in EEG emotion recognition: an emotion label assigned to an entire recording may not fit every short segment. The framework combines two-stage selection of segments with reliable labels and EEG-specific data augmentation. Tests used two public datasets, three network architectures, and multiple label-noise settings. The authors report maximum accuracy gains of 11.54% on DREAMER and 5.39% on SEED compared with the best-performing baseline. These are maximum gains in dataset tests, not average improvements across all settings.

A proof-of-concept study by Aalborg University authors tested two-dimensional cursor control in 10 able-bodied participants with no prior brain-computer interface experience. Participants used either executed or imagined tongue movements, with a state machine providing up, right, down and left options. All participants achieved control with executed movements, with a completion rate of 96.0 ± 4.9%; seven achieved control with motor imagery, with a completion rate of 41.0 ± 36.2%. Average completion times for successful targets were 30.8 ± 17.6 s and 66.5 ± 26.8 s, respectively; the latter included only the seven participants with at least one successful trial. Potential muscle artifacts during executed movements were not controlled for. This was a feasibility test in able-bodied participants, not evidence of clinical efficacy in people with severe motor impairments.

Researchers from Kunming University of Science and Technology and collaborating institutions proposed a theoretical system for basic BCI research, formalizing BCIs as closed-loop probabilistic information-processing models through information theory and statistical decision-making. The work addresses system decodability, performance boundaries and dynamic constraints. It organizes scientific hypotheses, neuroscience principles, fundamental theorems and basic laws, derives four results that can guide experimental design, and uses motor imagery BCIs to illustrate application boundaries. The authors aim to provide a unified analytical benchmark and support a shift from experience-driven algorithm optimization toward theory-driven system analysis and design.

A review surveys tools for studying the collective activity of neural populations and discusses how low-dimensional structure can help explain sensory coding, working memory, decisions, and sensorimotor transformations. The authors emphasize that cognitive signals can occupy separable or near-orthogonal population dimensions, reducing interference without requiring distinct cell classes: individual neurons in a shared population can still respond to multiple kinds of information. The review also considers recurrent neural networks as testbeds for mechanistic hypotheses and intracortical brain-computer interfaces as an application in which population geometry becomes a control signal. This is a review, not a new clinical efficacy study.

Researchers introduced a method for selecting signal lags to identify when one signal's history is most useful for predicting another. Across three benchmark tasks using existing monkey-experiment data, the selected lags achieved the highest or equal-best coefficient of determination, R², with mean R² improving by 4.1% relative to the strongest competing baseline. Called ATEMPO, the method estimates information relationships across history depths and temporal positions with a single fit, reducing the need for repeated fitting. On datasets with known lags, its lag-detection accuracy was 0.98±0.04. The findings concern dataset prediction performance, not clinical efficacy.

Researchers from Hengyang Normal University and Chongqing University of Education in China jointly proposed NCEL, a lightweight EEG decoder evaluated on five benchmark datasets. It achieved 82.75% mean classification accuracy on BCIC IV-2b. Relative to EEGNet reference values, mean performance improved by 12.53 percentage points on P300 and 9.54 percentage points on BCIC IV-2b, with approximately 0.008M parameters. However, additional subject-specific evaluations with disjoint sessions found cross-session performance comparable to EEGNet on BCIC IV-2a and a mean improvement of 1.80 percentage points on BCIC IV-2b. The study demonstrates classification capabilities in a compact model, while broader cross-session generalization still needs validation.

A CUHK-Shenzhen collaboration introduced OneEEG for multitask EEG question answering. Standard 19-channel tasks use numerical signals and waveform images; non-standard tasks use images only. On the closed-set OneEEG-Bench, Overall accuracy reached 65.67% and Macro accuracy 68.24%, versus 50.39% Overall for baseline GPT-5.5. OneEEG received EEG-specific training; baselines used only images and text. These differences preclude interpreting the result as a general model or clinical diagnostic ranking.

Researchers from Beijing University of Technology and collaborating institutions proposed REFS to select EEG features for multi-dimensional emotion recognition when some emotional labels are missing. The method reconstructs incomplete labels while reducing redundancy among selected features, addressing modeling challenges posed by high-dimensional features and relatively few high-quality samples. Tests used DREAMER, DEAP and HDED, three datasets with different EEG channel configurations, at varying label-missing ratios. REFS consistently outperformed the 13 advanced feature-selection methods compared in robustness under these conditions. These dataset findings do not establish performance in deployed affective BCIs.

A collaboration involving UC Berkeley, USC and Microsoft Research introduced MtReNIRS, pretrained on 12 public fNIRS datasets covering 268 subjects and about 567 recording hours, then evaluated on three unseen datasets. It aligns probe layouts for cross-subject transfer and within-subject calibration with two to eight labels per class. Cross-subject results were not uniformly significantly better than fully supervised baselines, and few-shot gains varied by dataset. Device effects were not isolated because device, task and montage differences co-occurred.

Researchers introduced ORBITAL, a framework for compressing EEG decoders while evaluating resistance to adversarial attacks and actual hardware costs. Across nine subjects and three random seeds on BCI Competition IV-2a, S25-TRADES achieved the highest observed mean true-union robustness among the tested compressed EEGNet models: 23.06%, versus 21.27% for standard Full, with 26.27% fewer parameters and 25.10% fewer multiply-accumulate operations. Hardware tests also showed that lower numerical precision did not necessarily save energy or time: native INT8 incurred about twice the FP16 latency and energy. These results do not establish comprehensive reliability in real-world use.

A study from The University of Electro-Communications in Japan proposes enhancing brain signals without model training. The method repeatedly samples and pastes EEG signals to strengthen steady-state visual evoked potential components, which visual brain-computer interfaces use to identify target commands. In experiments with synthesized signals and open benchmark datasets, it outperformed conventional methods at every tested signal length, improving information transfer rate by up to 11.30 bit/min. The approach addresses the training-data requirements and upfront costs of learning-based methods; the reported gains come from signal and dataset experiments.

A University at Buffalo author evaluated an assistive communication interface combining a BCI with eye tracking. Gaze narrows the target area and brain signals confirm selection. The study reports faster selection than BCI alone, better control than eye tracking alone, and fewer selection errors and corrections.

A Nanyang Technological University researcher combined EEG and EMG timing information in a gesture-recognition network, reaching 87.6% and 89.2% accuracy on datasets with 12 and 20 participants. The results exceeded the strongest baselines by 9.4 and 9.1 percentage points; adaptation required no labelled target-user data. Participants were healthy young adults.

University of Essex researchers recorded EEG from 14 participants retrieving concepts from verbal clues. Mean accuracies in three pairwise comparisons of food, musical instruments and body parts were 73.4%, 70.7% and 68.2%. The model learned representations from other participants before calibration with each test participant’s own labelled data.

A Zhejiang University of Technology team developed MOPC to adapt motor-imagery EEG decoders as new signals arrive, without revisiting the original training data. The authors report gains of 2.22% and 3.67% over prior methods on BCI IV IIb and OpenBMI, with adaptation taking about 40–60 ms per trial. The arXiv record states that the paper was accepted by IEEE Transactions on Instrumentation and Measurement.

Oxford researchers compared six aggregation strategies; training-time logit averaging led in six of seven settings. Aggregating 100 same-phoneme trials yielded 64.1% balanced accuracy in a public-test-split re-evaluation, 17.4 points above the competition champion model. The preprint experiments use MEG recordings.

Researchers at Xi’an Jiaotong University and collaborating institutions developed E-STACNet to model spatial relationships between EEG channels alongside changes over time. Mean accuracies were 88.00% on BCI Competition IV 2a and 80.62% on Giga DB; DEAP valence and arousal classification reached 96.49% and 96.46%.

Researchers at Huazhong University of Science and Technology proposed KDHNet, a hybrid network combining analytical signal processing with neural networks to decode steady-state visual evoked potentials, stable brain responses to visual stimulation, with limited individual calibration data. Across three public datasets, it achieved the highest mean accuracy and information transfer rate among the compared methods in 14 of 15 analysis-window settings and all ten calibration-ratio settings in within-subject evaluation. In cross-subject evaluation, performance ranked first or second, with generally larger gains when less calibration data were available. These dataset results do not establish real-world plug-and-play operation.

Authors affiliated with Tongji University, Ant Group, Shanghai Jiao Tong University and collaborating institutions introduced the NeuroAct EEG research agent. In an experience-reuse experiment, BCIC-IV-2a balanced accuracy rose from 34.66% to 50.81% while LLM token use fell 64.02%. The preprint also evaluates general analysis, cross-task optimization, single-task research and collaborative prediction.

Researchers developed a motor imagery EEG classification framework, tested it on two competition datasets, and deployed it on an NVIDIA Jetson Nano. Subject-dependent accuracy reached 93.67% on BCIC III IVa and 79.14% on BCIC IV 2a. Under leave-one-subject-out evaluation, which tests generalization to a person excluded from training, average accuracy was 76.49% and 55.50%, respectively. Deployment tests also measured complete pipeline latency and throughput. These findings describe dataset classification performance and device processing speed, not demonstrated improvements in motor impairments.

Researchers proposed TriRouter to classify motor imagery when EEG channels are sparse or corrupted, allocating contributions from different processing paths according to channel availability. On BNCI2014-001, accuracy exceeded the study's strongest corrupted-EEG baseline by 3.68, 4.02, and 6.46 percentage points at nominal missingness settings of 25%, 50%, and 75%, respectively. The authors also report effectiveness under structured regional missingness and in an additional low-channel dataset evaluation. These dataset results do not directly establish BCI performance in practical use.

Researchers proposed BEST, a two-stage method that converts neural activity into text, to examine whether speech brain-computer interfaces need large models. On the Brain-to-Text '24 and '25 benchmarks, BEST achieved word error rates of 6.19% and 1.62%, respectively, using only 3.85% of the parameters required by comparable methods. Word error rate measures errors relative to reference text; lower values indicate fewer errors. The authors argue that larger decoders do not consistently perform better and that alignment between neural signals, speech, and text matters more than model size. These benchmark results do not directly establish everyday communication performance.

Moving BCI computation from the cloud onto devices still requires trade-offs among energy use, response latency, and accuracy. A systematic review surveys edge-AI accelerators for brain-computer interfaces and emphasizes joint algorithm and chip design to reduce both computational demands and data-movement overhead. It discusses algorithmic compression, compute-in-memory, and neuromorphic architectures, identifying on-device adaptive learning as important for addressing neural signal variability across sessions. The author also proposes privacy-preserving hardware and multimodal sensor fusion as development directions, not as established evidence of clinical effectiveness or security.

Researchers proposed NeuroDETR, a framework that identifies event classes and estimates their onsets and offsets directly from continuous EEG, without predefined event boundaries at inference time. It combines information across temporal scales and spatial relationships among EEG channels, using predicted event presence to guide feature processing. The authors report state-of-the-art performance on eight EEG datasets spanning diverse paradigms. These dataset experiments address event recognition for continuous interaction and monitoring; they do not directly establish performance in deployed brain-computer interface systems.

Researchers propose an adaptive EEG window-selection framework for motor imagery classification. Instead of independently retaining or discarding each segment, it considers event phases and relationships between successive windows. Each window receives a Discard, Onset, Sustain, or Offset role, with constraints blocking selections that violate phase order. The authors report consistent decoding improvements across multiple benchmark datasets in both cross-session and cross-subject settings. Visualizations further suggest that the learned policy aligns with neurophysiological principles. These dataset experiments do not establish equivalent gains during online BCI use. This study is a preprint and has not been peer reviewed.