Journal of neural engineering最新文献

Objective. Epilepsy is a chronic brain disorder characterized by recurrent seizures due to abnormal neuronal firing. Electroencephalogram (EEG)-based seizure classification has become an important auxiliary tool in clinical practice. This study aims to reduce reliance on expert experience in diagnosis and to improve the automated classification of epileptic seizures using EEG signals.Approach. We propose a novel filter-bank multi-view and attention-based mechanism neural network model for seizure classification. The model employs a learnable filter bank to decompose the raw EEG into multiple frequency sub-bands, forming multi-view representations. A multi-branch group convolution network is designed to capture multi-scale frequency-spatial features, while temporal dependencies are extracted through a bidirectional long short-term memory with an attention mechanism. A shared attention module adaptively emphasizes the most informative sub-bands and time windows for classification.Main results. The proposed model achieves an overallF1score of 0.7105, a weightedF1(WF1) score of 0.8314, and a Cohen's kappa coefficient of 0.6345 on the TUSZ v1.5.2 dataset. Compared with the baseline method FBCNet, the proposed model outperform by 3.22% in overallF1score (p < 0.05), 1.42% inWF1score (p < 0.05), and 2.87% in Cohen's kappa coefficient (p < 0.05). The best results are also obtained on the CHB-MIT dataset.Significance. These results demonstrate the effectiveness of combining multi-view feature extraction with attention-enhanced temporal modeling.

Objective.Speech brain-computer interfaces (BCIs) aim to provide an alternative means of communication for individuals who are not able to speak. Remarkable progress has been achieved to decode attempted speech in individuals with severe anarthria. In contrast, imagined speech remains challenging to decode. The underlying neural mechanisms and relations to other speech modes are still elusive.Approach.In this study, we collected low-density electrocorticography signals from ten participants during a word repetition task. Electrodes were implanted for presurgical epilepsy evaluation in participants with preserved speech abilities. Models were developed using linear discriminant analysis to classify five words in response to different speech modes. We compared models trained during speaking, listening, imagining speaking, mouthing and reading. The relations between speech modes were investigated by transferring and augmenting models across speech modes.Main results.As expected, performed speech achieved the highest word classification accuracy followed by listening, mouthing, imagining and reading. While the accuracies obtained were not high enough for practical application, model transfer and augmentation could be investigated across speech modes. Transferring or augmenting models from one speech mode to another mode could significantly improve model performance. In particular, patterns learned from performed and perceived speech could generalize to imagined speech, leading to significantly improved imagined speech performance in seven participants. For four participants, imagined speech could be decoded above chance exclusively when models were transferred or augmented with performed or perceived speech.Significance.Imagined speech is often preferred by speech BCI users over attempted speech, as it requires less effort and can be produced more quickly. Transferring models across speech modes has the potential to facilitate and boost the development of imagined speech decoders.

Objective. Penetrating polymer-based microelectrode arrays (pMEAs) offer the potential for long term high-quality electrophysiological recordings of dynamic neural activity. Compared to rigid metal wire and silicon MEAs, improved device-tissue interface stability has been reported. However, accurate surgical placement of long, thin shanks in deeper brain regions is challenging as flexibility is achieved at the expense of axial stiffness. This study systematically evaluates then compares two pMEA placement strategies-dissolvable dip coating and molded brace, both with bare, exposed pMEA tips-to address the need for consistent, reliable, and accurate surgical targeting. These methods were selected based on the criteria of ease of fabrication, surgical feasibility, and mechanical performance.Approach. Sham (mechanical model with no electrodes) and fully functional pMEAs with shanks up to 5.5 mm long were fabricated and then modified using biodegradable polyethylene glycol (PEG) to support implantation. PEG was applied to shanks by motorized dip coating or a mechanical mold. Dissolution time and insertion in agarose gel brain models and rat cortex were evaluated followed by targeting of dip coated pMEAs to the rat hippocampus.Main results. Dip coating at high withdrawal speeds achieved uniform coating on shanks. Both strategies yielded similar critical buckling forces and insertion forces for single shank and arrayed pMEAs. Dip coated pMEAs were successfully placed in hippocampal regions without severe tissue damage as confirmed by histology and recordings obtained.Significance. Dip coating is a simpler method to prepare pMEAs for surgical targeting of deep brain regions compared to the bracing technique, as it does not require both a specialized mold and application process. This work provides a guide for researchers using single or multi-shank pMEAs to an accessible insertion strategy for implanting into deep brain regions in rodents and other small animal models.

Objective.During speech perception, properties of the acoustic stimulus can be reconstructed from the listener's brain using methods such as electroencephalography (EEG). Most studies employ the amplitude envelope as a target for decoding; however, speech acoustics can be characterised on multiple dimensions, including as spectral descriptors. The current study assesses how robustly an extended acoustic feature set can be decoded from EEG under varying levels of intelligibility and acoustic clarity.Approach.Analysis was conducted using EEG from 38 young adults who heard intelligible and non-intelligible speech that was either unprocessed or spectrally degraded using vocoding. We extracted a set of acoustic features which, alongside the envelope, characterised instantaneous properties of the speech spectrum (e.g. spectral slope) or spectral change over time (e.g. spectral flux). We establish the robustness of feature decoding by employing multiple model architectures and, in the case of linear decoders, by standardising decoding accuracy (Pearson'sr) using randomly permuted surrogate data.Main results. Linear models yielded the highestrrelative to non-linear models. However, the separate decoder architectures produced a similar pattern of results across features and experimental conditions. After convertingrvalues toZ-scores scaled by random data, we observed substantive differences in the noise floor between features. Decoding accuracy significantly varies by spectral degradation and speech intelligibility for some features, but such differences are reduced in the most robustly decoded features. This suggests acoustic feature reconstruction is primarily driven by generalised auditory processing.Significance. Our results demonstrate that linear decoders perform comparably to non-linear decoders in capturing the EEG response to speech acoustic properties beyond the amplitude envelope, with the reconstructive accuracy of some features also associated with understanding and spectral clarity. This sheds light on how sound properties are differentially represented by the brain and shows potential for clinical applications moving forward.

Objective.Temporal interference stimulation (TIS) has emerged as an innovative and promising approach for non-invasive stimulation. While previous studies have demonstrated the efficacy and performance of TIS using benchtop instruments, a dedicated system-on-chip for TIS applications has not yet been reported. This work addresses this gap by presenting a design for a TIS chip that enhances portability, thereby facilitating wearable applications of TIS.Approach.A miniaturized dual-channel temporal interference stimulator for non-invasive neuro-modulation is proposed and fabricated in a 0.18µm CMOS BCD process. The TIS chip occupies the silicon area of only 2.66 mm². It generates output signals with a maximum amplitude of ±5 V and reliable frequency, with programmable input parameters to accommodate diverse biomedical applications. The carrier frequencies of the generated signals include 1 kHz, 2 kHz, and 3 kHz, combined with beat frequencies of 5 Hz, 10 Hz, and 20 Hz. This results in a total of nine available operation modes, enabling effective TIS.Main results.The proposed chip has effectively generated temporally interfering signals with reliable frequency and amplitude. To validate the efficacy of the TIS chip,in-vivoanimal experiments have been conducted, demonstrating its ability to produce effective electrical stimulation signals that successfully elicit neural responses in the deep brain of a pig.Significance.This work has replaced the bulky external stimulator with a fully integrated silicon chip, significantly enhancing portability and supporting future wearable clinical applications.

Objective.Myoelectric control systems translate electromyographic (EMG) signals into control commands, enabling immersive human-robot interactions in the real world and the Metaverse. The variability of EMG due to various confounding factors leads to significant performance degradation. Such variability can be mitigated by training a highly generalizable but massively parameterized deep neural network, which can be effectively scaled using a vast dataset. We aim to find an alternative simple, explainable, efficient and parallelizable model, which can flexibly scale up with a larger dataset and scale down to reduce model size, and thereby will significantly facilitate the practical implementation of myoelectric control.Approach.In this work, we discuss the scalability of a random forest (RF) for myoelectric control. We show how to scale an RF up and down during the process of pre-training, fine-tuning, and automatic self-calibration. The effects of diverse factors such as bootstrapping, decision tree editing (pre-training, pruning, grafting, appending), and the size of training data are systematically studied using EMG data from 106 participants including both low- and high-density electrodes.Main results.We examined several factors that affect the size and accuracy of the model. The best solution could reduce the size of RF models by≈500×, with the accuracy reduced by only 1.5%. Importantly, for the first time we report the merit of RF that with more EMG electrodes (higher input dimension), the RF model size would be reduced.Significance.All of these findings contribute to the real time deployment RF models in real world myoelectric control applications.

Objective.Deep learning (DL) exhibits considerable potential for steady-state visual evoked potential (SSVEP) classification in electroencephalography-based brain-computer interfaces (BCIs). SSVEP signals contain both frequency and phase characteristics that correspond to the visual stimuli. However, existing DL training strategies typically focus on either frequency or phase information alone, thus failing to fully exploit these dual inherent properties and substantially limiting classification accuracy.Approach.To tackle this limitation, this study proposes a joint frequency-phase training strategy (JFPTS), which comprises two complementary stages with distinct time-window sampling schemes. The first stage adopts a frequency prior-driven sampling scheme to improve frequency component utilization, whereas the second stage employs a phase-locked sampling scheme to enhance intra-category phase consistency. This design enables JFPTS to effectively leverage both frequency and phase properties of SSVEP signals.Main results.Comprehensive experiments on two well-established public datasets validate the effectiveness of JFPTS. The results demonstrate that the JFPTS-enhanced model achieves a marked superiority over the current state-of-the-art classification approaches, notably surpassing the long-standing performance benchmark set by task-discriminative component analysis (TDCA).Significance.Overall, JFPTS establishes a new training paradigm that advances DL approaches for SSVEP classification and promotes the broader adoption of SSVEP-BCIs.

Objective.Emotional states and mood disorders are closely interconnected, and their joint recognition serves as a critical pathway to uncovering their intrinsic relationship. Currently, deep learning (DL) models based on electroencephalogram (EEG) have achieved significant progress in single tasks such as emotion recognition or mood disorder (MD) recognition. However, most existing models are limited to handling only one of these tasks independently and fail to effectively leverage the shared features in EEG data related to both emotions and mood disorders. This limitation hinders the in-depth exploration of the complex interplay between emotions and mood disorders. Therefore, this study aims to develop an EEG-based DL framework for the joint recognition of emotions and mood disorders, thereby providing a foundation for further investigation into their interaction.Approach.We design a multi-gate mixture-of-experts graph convolutional network model(MMoGCN) for joint emotion and MD recognition. MMoGCN comprises three key modules: (1) a feature extraction module based on differential entropy to robustly represent EEG signals; (2) a Multi-gated shared experts module, which integrates two experts, and combines them through a gating mechanism to extract shared representations across tasks; and (3) adaptive task-specific towers, which consist of individual classification towers for each task and incorporate an adaptive weighting loss function to dynamically adjust task contributions. MMoGCN is evaluated on a self-collected dataset and further validated on the public DEAP dataset.Main results.MMoGCN achieves superior performance compared with state-of-the-art single-task and multi-task baselines in both emotion and MD recognition. Validation experiments on DEAP further demonstrate the scalability and generalization of MMoGCN.Significance.An effective multi-task learning model is proposed for joint emotion and MD recognition based on EEG. Additionally, the cognitive differences are also analyzed in emotional responses between healthy controls and subjects with mood disorders, providing methodological insights and potential assistance for cognitive rehabilitation from both cognitive and emotional perspectives.

Objective.Parkinson's disease (PD) is a common neurodegenerative disease best known for its defining motor symptoms. However, it is also associated with significant cognitive impairment at all stages of the disease, with many patients eventually progressing to dementia. Therefore, there exists a significant need to identify objective functional biomarkers that better predict and monitor cognitive decline. While methods that analyse either spontaneous or evoked electroencephalogram (EEG), due to increasing practical usability and ostensible objectivity, have been investigated, current approaches are limited in that the associated measures are, in the absence of a theoretical basis, purely correlative.Approach.To address this shortcoming, we propose calculating changes in evoked EEG amplitude variability, quantified using information theoretic differential entropy (DE), during a three-level passive auditory oddball task, as it is argued this will directly index functional changes in cognition. We therefore estimate changes in stimulus-evoked DE in cognitively normal PD participants (N= 25), both on and off their medication, and in healthy age-matched controls (N= 25), and find substantial stimulus (standard, target, novel) and group differences.Main results.Notably, we find the time-course of the return of post-stimulus reductions in DE (i.e. information processing) to pre-stimulus levels delayed in PD compared to healthy controls, thus mirroring the assumed bradyphrenia. The observed changes in DE, together with the corollary increases in resting alpha (8-13 Hz) band activity seen in PD, are explained in the context of a well-known macroscopic theory of mammalian electrocortical activity, in terms of reduced tonic thalamo-cortical drive.Significance.This method of task-evoked DE EEG amplitude variability is expected to generalise to any situation where the objective determination of cognitive function is sought.

Objective.Motor imagery brain-computer interfaces hold significant promise for neurorehabilitation, yet their performance is often compromised by electroencephalography (EEG) non-stationarity, low signal-to-noise ratios, and severe cross-session variability. Current decoding methods typically suffer from fragmented optimization, treating temporal, spectral, and spatial features in isolation.Approach.We propose common temporal-spectral-spatial patterns (CTSSP), a unified framework that jointly optimizes filters across all three domains. The algorithm integrates: (1) multi-scale temporal segmentation to capture dynamic neural evolution, (2) channel-adaptive finite impulse response filters to enhance task-relevant rhythms, and (3) low-rank regularization to improve generalization.Main results.Evaluated across five public datasets, CTSSP achieves state-of-the-art performance. It yielded mean accuracies of 76.9% (within-subject), 68.8% (cross-session), and 69.8% (cross-subject). In within-subject and cross-session scenarios, CTSSP significantly outperformed competing baselines by margins of 2.6%-14.6% (p< 0.001) and 2.3%-13.8% (p< 0.05), respectively. In cross-subject tasks, it achieved the highest average accuracy, proving competitive against deep learning models. Neurophysiological visualization confirms that the learned filters align closely with motor cortex activation mechanisms.Significance.CTSSP effectively overcomes the limitations of decoupled feature extraction by extracting robust, interpretable, and coupled temporal-spectral-spatial patterns. It offers a powerful, data-efficient solution for decoding MI EEG in noisy, non-stationary environments. The code is available athttps://github.com/PLC-TJU/CTSSP.