Cognitive Neurodynamics最新文献

Motor imagery (MI) decoding remains a critical challenge in brain-computer interface (BCI) systems due to the low signal-to-noise ratio, non-stationarity, and complex spatiotemporal dynamics of electroencephalography (EEG) signals. Although deep learning architectures have advanced MI-EEG decoding, existing approaches-including convolutional neural networks (CNNs), Transformers, and recurrent neural networks (RNNs)-still face limitations in capturing global temporal dependencies, maintaining positional coherence, and ensuring computational efficiency. To address these challenges, we propose MIFNet, a MamBa-based Interactive Frequency Convolutional Neural Network that systematically integrates spectral, spatial, and temporal feature extraction. Specifically, MIFNet incorporates: non-overlapping frequency decomposition, which selectively extracts motor imagery-related mu (8-12 Hz) and beta (12-32 Hz) rhythms; a ConvEncoder module, which autonomously learns to fuse spectral-spatial features from both frequency bands; and a MamBa-based temporal module, leveraging selective state-space models (SSMs) to efficiently capture long-range dependencies with linear complexity. Extensive experiments on three public MI-EEG datasets (BCIC-IV-2A, OpenBMI, and High Gamma) demonstrate that MIFNet outperforms existing models, achieving an average classification accuracy improvement of 12.3%, 8.3%, 4.7%, and 5.5% over EEGNet, FBCNet, IFNet, and Conformer, respectively. Ablation studies further validate the necessity of each component, with the MamBa module contributing a 5.5% improvement in accuracy on the BCIC-IV-2A dataset. Moreover, MIFNet exhibits strong generalization performance in cross-validation settings, establishing a robust foundation for real-time BCI applications. Our findings highlight the potential of hybridizing CNNs with state-space models (SSMs) for improving EEG decoding performance, effectively bridging the gap between localized feature extraction and global temporal modeling.

Charge-trap transistors are widely used for the simulation of biological synaptic functions. However, the unique structure of silicon-oxide-nitride-oxide-silicon (SONOS) makes it difficult to simulate short-term memory (STM). Based on simulation modeling, this work proposes a cylindrical surrounding double-gate (CSDG) nanowire synaptic transistor with a Si $_3^{}$ N $_4^{}$ charge trap layer in direct contact with the channel. The synaptic functions of the enhanced weights are mimicked by modulating electrical impulses to achieve the short-term potentiation (STP) to long-term potentiation (LTP) transition. In addition, the post-synaptic response changes with light intensity and wavelength under light illumination, which is phenomenologically similar to light-assisted dopamine-promoted synaptic activity. Furthermore, the high blue light responsiveness successfully exhibits the physiological characteristic that blue light promotes more dopamine secretion in the retina of the human eye. This model introduces additional light stimulation to achieve dopamine dynamics driven learning acceleration, providing a foundation for improving the rapid recognition and learning ability of neural computing systems in the next step.

Overweight individuals often experience impairments in executive function, particularly working memory. Physical exercise has been shown to mitigate such cognitive decline and modulate brain activities. This study aimed to investigate whether a six-week high-intensity interval (HIIT) Tabata exercise could improve working memory performance in overweight individuals and explore the associated neural mechanisms. To achieve this aim, two experiments were conducted. In Experiment 1, 20 overweight (Body Mass Index, BMI ≥ 24) and 20 health-weight university students completed the n-back task (n = 0 ~ 2) to assess working memory. Results confirmed that overweight participants exhibited lower accuracy (ACC) in the 2-back task compared with health-weight participants. Accordingly, in Experiment 2, another 40 overweight university students were randomly assigned into the training group (six-week HIIT Tabata) or control group (no physical exercise). All the participants performed the 2-back task with EEG recording at two points: before and after the six-week intervention (pre-test vs. post-test). Results showed that compared to pre-test, the training group showed higher accuracy at the post-test, whereas no such change was observed in the control group. Moreover, ERP results revealed a reduction in post-test P2 amplitude in the training group. Overall, this study demonstrates that being overweight negatively impacts working memory, while a six-week HIIT Tabata intervention may help alleviate these deficits, possibly through more efficient neural resource utilization.

Facial expressions enable individuals to assess and understand emotions conveyed by others. Two crucial sources of expressive cues on the human face-the eyes and the mouth-capture attention and serve as reliable shortcuts for expression recognition. However, how the brain effectively extracts emotional information from these diagnostic features remains unknown. We investigated this issue using an electroencephalogram combined with a rapid serial visual presentation task in which participants were asked to recognize facial expressions (fear, happiness, and neutrality) from three formats (whole face, eye region, and mouth region). We found that participants recognized happy expressions from the mouth region more accurately than the other expressions, affirming the role of diagnostic features in facilitating bottom-up attentional capture. The isolated eye region with higher visual saliency induced the largest P1 component. Diagnostic features, such as a happy mouth and fearful eyes, elicited a larger N170 component compared to non-diagnostic features, such as a fearful mouth and happy eyes. Source analysis of N170 showed that the fusiform gyrus exhibited similar patterns in response to these emotional features. The P3 was effective in discriminating between different emotional content. When whole faces were visible, fearful and happy expressions were not distinguishable in the N170, while the P3 amplitude was larger when induced by fearful faces than by happy faces. Our study contributes to understanding how facial features play distinct roles in emotional perception, attention, and facial processing.

Improving the accuracy of non-invasive brain-computer interface (BCI) and promoting their daily use can be achieved by developing an individualized model training framework, where individual training means that the model is based on small-sample learning from individual data. In the process of data augmentation through synthetic data, the criteria for data generation needs to be further specified according to the requirements. Therefore, in this study, the proposed BCI model utilizes dynamic networks to describe electroencephalogram (EEG) activity during the motor imagery (MI) task, innovatively generates individualized dynamic networks from individual data, and ultimately achieves EEG-controlled grasping through model training. Specifically, this study involves the EEG signals of the right-hand grasping movements of eight subjects and proposes using morphological pattern spectrum (MPS) to encode EEG potentials during MI processes. The MI condition representation was achieved by combining the dynamic networks with MPS encoding, and more dynamic network EEG encoding samples were synthesized through generative adversarial network (GAN) or variational autoencoder (VAE). The AUCs based on the long short-term memory (LSTM) architecture for generating and classifying can be improved by 0.003-0.07. The optimal BCI model based on the Wasserstein GAN and Granger causality (GC) dynamic network encoded by MPS achieved a mean true/false positive rate (TPR/FPR) of 90.0%/0.0%, far better than the 52.9%/4.4% achieved without individualized modeling. Moreover, the BCI establishment of handling multi-task and complex command outputs further demonstrates the reliability of MPS encoding of the GC dynamic network in BCI modeling. The advantage of this "generative-individual" approach is that it not only reduces the sample size requirement while ensuring accuracy but also avoids building models that are applicable to all individuals, which leads to difficult convergence.

This study explores graph theory methods for analyzing the functional connectivity of multiple spike trains. We study simultaneously recorded multiple spike trains recorded from the visual cortex of a cat under different visual stimuli. To find the functional connectivity for a given visual stimulus we use the Cox method (Masud and Borisyuk, J Neurosci Methods 196:201-219, 2011). The application of graph theory methods for analysing the connectivity circuit, revealed that the functional connectivity of multiple spike trains is characterized by low density, long communication distances, and weak interconnectivity. Nevertheless, some spike trains also exhibit high degrees of centrality, including betweenness centrality, expansiveness coefficient, and attractiveness coefficient. Additionally, the analysis also identified significant motifs within the functional connections. Thus, our approach allows to describe the correspondence between the stimulus and functional connectivity diagram and compare functional connections under different stimuli.

Motor imagery (MI) is a fundamental paradigm in brain-computer interfaces (BCIs), extensively employed to assist individuals with disabilities to operate external devices. Accurate decoding of MI signals is essential for effective interaction. However, robust decoding remains a challenge due to the inherent complexity and variability of MI-EEG signals. To address this issue, we propose an innovative Dual-Branch Multi-Attention Temporal Convolutional Network (DBMATCN) to improve the performance of MI-EEG signal classification. First, the dual-branch structure extracts rich spatial-temporal features. Then, the channel attention enhances local channel feature extraction and calibrate feature mapping. Next, by combining a sliding window technique and multi-head locality self-attention improves the feature representation of MI-EEG signals by emphasizing the most relevant features. Finally, the temporal convolution fusion network decoding module is used to extensively capture comprehensive temporal features from MI data and carry out the classification task. DBMATCN achieves average accuracies of 88.08%, 96.83%, and 89.71% in inter-session validation on the BCI-IV-2a, HGD, and BCI-IV-2b datasets, respectively. In cross-validation, the model reaches an accuracy of 85.14%, and in the subject-independent scenario, it attains 71.78%. DBMATCN outperforms all baseline models in these cases. These results suggest that our model is effective in decoding MI signals.

Enhanced beta oscillations (12-25 Hz) within the cortico-basal ganglia-thalamic network are significantly associated with motor deficits and are a prominent characteristic of the neural dynamic pathology in Parkinson's disease. Although the striatum has been proposed as a promising origin for enhanced beta oscillations, the precise mechanism through which distinct striatal neurons collaborate to orchestrate beta oscillations remains elusive. This study constructs a biophysical neural network model of the striatum based on experimental constraints. The model faithfully reproduces various experimental observations, including dopamine-dependent beta oscillations and phase-locked firing patterns. Through both theoretical and numerical analysis, our analysis reveals that striatal beta oscillations emerge from interactions within the cellular architecture, particularly the somatostatin-expressing interneurons (SOM) driven choline acetyltransferase-expressing interneurons (ChAT)-indirect pathway striatal projection neurons (iSPN) loop. Our results underscore the critical role of ChATs in enhancing beta oscillations. ChATs, instead of passively providing excitatory drive, actively amplify beta oscillations by enhancing their excitation efficacy through a phase-locked mode. Additionally, the inhibitory interactions among iSPNs, with robust and slow inhibitory recovery dynamics within iSPNs, potentially result in beta oscillations. The slow inhibitory recovery is likely attributed to the slow dynamics of the KCNQ current. SOMs further modulate the beta oscillations by affecting their downstream ChAT-iSPN loop. These results provide novel insights into the mechanism underlying striatal beta oscillations, shedding light on the processes involved in beta oscillations generation during pathological states.

Accurate decoding and strong feature interpretability of Motor Imagery (MI) are expected to drive MI applications in stroke rehabilitation. However, the inherent nonstationarity and high intra-class variability of MI-EEG pose significant challenges in extracting reliable spatio-temporal features. We proposed the Dynamic Spatio-Temporal Feature Augmentation Network (DSTA-Net), which combines DSTA and the Spatio-Temporal Convolution (STC) modules. In DSTA module, multi-scale temporal convolutional kernels tailored to the α and β frequency bands of MI neurophysiological characteristics, while raw EEG serve as a baseline feature layer to retain original information. Next, Grouped Spatial Convolutions extract multi-level spatial features, combined with weight constraints to prevent overfitting. Spatial convolution kernels map EEG channel information into a new spatial domain, enabling further feature extraction through dimensional transformation. And STC module further extracts features and conducts classification. We evaluated DSTA-Net on three public datasets and applied it to a self-collected stroke dataset. In tenfold cross-validation, DSTA-Net achieved average accuracy improvements of 6.29% (p < 0.01), 3.05% (p < 0.01), 5.26% (p < 0.01), and 2.25% over the ShallowConvNet on the BCI-IV-2a, OpenBMI, CASIA, and stroke dataset, respectively. In hold-out validation, DSTA-Net achieved average accuracy improvements of 3.99% (p < 0.01) and 4.2% (p < 0.01) over the ShallowConvNet on the OpenBMI and CASIA datasets, respectively. Finally, we applied DeepLIFT, Common Spatial Pattern, and t-SNE to analyze the contributions of individual EEG channels, extract spatial patterns, and visualize features. The superiority of DSTA-Net offers new insights for further research and application in MI. The code is available in https://github.com/CL-Cloud-BCI/DSTANet-code.

Adolescent brain development is characterized by significant anatomical and physiological alterations, but little is known whether and how these alterations impact the neural network. Here we investigated the development of functional networks by measuring synaptic plasticity and neural synchrony of local filed potentials (LFPs), and further explored the underlying mechanisms. LFPs in the hippocampus were recorded in young (21 ~ 25 days), adolescent (1.5 months) and adult (3 months) rats. Long term potentiation (LTP) and neural synchrony were analyzed. The results showed that the LTP was the lowest in adolescent rats. During development, the theta coupling strength was increased progressively but there was no significant change of gamma coupling between young rats and adolescent rats. The density of dendrite spines was decreased progressively during development. The lowest levels of NR2A, NR2B and PSD95 were detected in adolescent rats. Importantly, it was found that the expression levels of autophagy markers were the highest during adolescent compared to that in other developmental stages. Moreover, there were more co-localization of autophagosome and PSD95 in adolescent rats. It suggests that autophagy is possibly involved in synaptic elimination during adolescence, and further impacts synaptic plasticity and neural synchrony.