Rapid and effective enhancement of memory function through the integration of neuromodulation and cognitive training has been a longstanding objective in cognitive science and translational medicine. Research indicates that efficient training strategies are associated with rapid improvements in working memory (WM) performance, with successful WM performance linked to neural communication between frontomedial theta oscillations and posterior gamma oscillations. However, direct evidence showing how these strategies regulate neural communication and integrate neuromodulation to boost WM performance remains limited. This study addresses this gap by analyzing electroencephalography (EEG) data from participants undergoing strategy-based WM training combined with neuromodulation. We observed that WM enhancements were accompanied by altered coupling of theta-gamma oscillations. Then, targeted HD-transcranial Direct Current Stimulation (HD-tDCS) was applied to the left dorsolateral prefrontal cortex (LPFC), a region exhibiting representational similarity with behavior induced by the cognitive strategy implementation. This intervention resulted in enhanced WM performance associated with slower frontomedial theta oscillations. This study illuminates that cognitive-neural pattern plasticity can be rapidly and selectively improved through neuromodulation by leveraging spatial frequency coupling within specific cortical circuits involved in strategic memory.

Emotional neuromodulation refers to the direct manipulation of the nervous system using techniques such as electrical or magnetic stimulation to manage and adjust an individual's emotional experiences. Transcranial electrical stimulation (tES) targeting the right ventrolateral prefrontal cortex (rVLPFC) has been widely used to modulate emotions. However, the impact of emotions on brain network changes and modulation during tES remains unclear. In this study, we developed a subject-adaptive dynamic graph convolution network with fused features (FusSADGCNN) to decode the impact of tES on neuromodulation for emotion recognition and emotion elicitation. Specifically, we developed a fused feature, CPE, which integrates the average sub-frequency phase-locking value representing global functional connectivity with differential entropy characterizing local activation to explore network differences across emotional states, while incorporating an improved dynamic graph convolution to adaptively integrate multireceptive neighborhood information for precise decoding of individual tES effects. On the SEED dataset and our laboratory data, the FusSADGCNN model outperforms the state-of-the-art methods. Furthermore, we utilized these tools to assess the emotional modulation states induced by tES. Results indicated that in the experiment involving music-elicited emotional modulation, the tools effectively identified improvements in negative emotions under true stimulation, with predictive accuracy significantly related to the average connectivity strength of the brain network. In the active facial emotion recognition modulation experiment, jointed stimulation of rVLPFC and temporo-parietal junction achieved better modulation effects. These findings highlight that the FusSADGCNN effectively evaluate the neuromodulation states during tES-induced emotional regulation, providing a reliable foundation for integrating emotion recognition and neuromodulation.

The detection of Working Memory Load WML plays a crucial role in neu-rofeedback processes and the treatment of disorders such as ADHD. How-ever, the performance of existing detection methods remains unsatisfactory. Neuropsychology research indicates that high-level cognitive processes are driven by both inter-regional collaborations across different brain functional areas and cross-frequency couplings. To comprehensively capture brain ac-tivities spanning both frequency domains and intra inter-regional interac-tions, we propose a novel cognitively-inspired neural network – the Wavelet-based Cross-Frequency and Cross-Region Interaction Convolutional Neural Network CFCRNet – for WML decoding. Specifically, CFCRNet first em-ploys predefined wavelet kernels to perform 1D convolution for time-frequency feature extraction, followed by multi-branch learning to model cross-frequency feature coupling with varying scales, and finally integrates intra- and inter-regional information interactions through spatial attention mechanisms. This architecture systematically fuses neurophysiologically meaningful cross-frequency coupling mechanisms with functional integra-tion principles across brain regions, constructing a network model capable of simultaneously resolving dynamic characteristics of neural signals across dif-ferent frequency bands and complex interactive relationships within between functional areas. Experimental validation on our collected working memory dataset and public benchmarks demonstrates that incorporating neuroscien-tific priors into neural network design enhances classification performance. Collectively, our findings establish an advanced framework for accurate WML detection that can be extended to explore detection tasks associated with other cognitive behaviors and neurological disorders.

Accurate decoding in electroencephalography (EEG) technology, particularly for rapid visual stimuli, remains challenging due to the low signal-to-noise ratio (SNR). Additionally, existing neural networks struggle with issues related to generalization and interpretability. This article proposes a cross-modal aligned network, E2IVAE, which leverages shared information from multiple modalities for self-supervised alignment of EEG to images for extracting visual perceptual information and features a novel EEG encoder, ISTANet, based on algorithm unrolling. This network framework significantly enhances the accuracy and stability of EEG decoding for object recognition in novel classes while reducing the extensive neural data typically required for training neural decoders. The proposed ISTANet employs algorithm unrolling to transform the multilayer sparse coding algorithm into an end-to-end format, extracting features from noisy EEG signals while incorporating the interpretability of traditional machine learning. The experimental results demonstrate that our method achieves SOTA top-1 accuracy of 62.39% and top-5 accuracy of 88.98% on a comprehensive rapid serial visual presentation (RSVP) dataset for public comparison in a 200-class zero-shot neural decoding task. Additionally, ISTANet enables visualization and analysis of multiscale atom features and overall reconstruction features, exploring biological plausibility across temporal, spatial, and spectral dimensions. On another more challenging RSVP large-scale dataset, the proposed framework also achieves significantly above chance-level performance, proving its robustness and generalization. This research provides critical insights into neural decoding and brain-computer interfaces (BCIs) within the fields of cognitive science and artificial intelligence.

Emotion recognition based on electroencephalography (EEG) has garnered increasing attention for its ability to reflect human emotional states objectively and in real time. However, EEG signals exhibit significant variations across different subjects and experimental sessions, posing challenges to the generalization of emotion recognition algorithms to unseen scenarios. To relieve this issue, we designed a novel hybrid EEG emotion recognition model named DGLFS, which integrates domain-invariant feature selection, label propagation, and adaptive dual-graph regularization into a unified optimization framework. Domain-invariant feature selection projects EEG data from different domains into a shared subspace, capturing emotion-related features that are domain-independent, thereby effectively mitigating data differences across subjects and sessions. Adaptive dual-graph learning simultaneously constructs a local similarity graph and a global structural graph to comprehensively capture both local similarities and global dependencies among EEG samples. Additionally, a graph-based semi-supervised label propagation method is employed, leveraging both global and local structural information embedded in the dual graphs to propagate emotional labels from a small subset of labeled data to unlabeled samples, thereby enabling more accurate emotion estimation in the target domain. We conducted extensive experiments on SEED-IV and SEED-V datasets involving cross-subject and cross-session tasks. The experimental results validate that DGLFS are superior to competitive algorithms in classification accuracy. Moreover, the intrinsic activation patterns revealed by DGLFS are consistent with emotional cognition. The code of DGLFS will be available at https://github.com/czxyhll/DGLFS.

Working memory, a fundamental cognitive function of the brain, necessitates the evaluation of cognitive load intensity due to limited cognitive resources. Optimizing cognitive load can enhance task performance efficiency by preventing resource waste and overload. Therefore, identifying working memory load is an essential area of research. Deep learning models have demonstrated remarkable potential in identifying the intensity of working memory load. However, existing neural networks based on electroencephalogram (EEG) decoding primarily focus on temporal and spatial characteristics while neglecting frequency characteristics. Drawing inspiration from the role of cross-frequency coupling in the hippocampal region, which plays a crucial role in advanced cognitive processes such as working memory, this study proposes a Multi-Band Multi-Scale Hybrid Sinc Convolutional Neural Network (MBSincNex). This model integrates multi-frequency and multi-scale Sinc convolution to facilitate time-frequency conversion and extract time-frequency information from multiple rhythms and regions of the EEG data with the aim of effectively model the cross-frequency coupling across different cognitive domains. Due to its unique structural design, the proposed model proficiently extracts features in temporal, frequency, and spatial domains while its feature extraction capability is validated through post-hoc interpretability techniques. On our collected three-class working memory dataset, the proposed model achieved higher classification accuracy compared to other state-of-the-art methods. Furthermore, by analyzing the model’s classification performance during different stages of working memory processes, this study emphasizes the significance of the encoding phase and confirms that behavioral response does not accurately reflect cognitive load.

The complexity of Alzheimer's disease (AD) is influenced by population heterogeneity, prompting the study of subtypes through imaging phenotypes. However, the role of neural morphological heterogeneity and the correlated gene expression (CGE) in AD subtypes remains unclear. This study links cortical thickness deviations of AD subtypes with CGE connectivity patterns. Transcriptional activity was measured based on the Allen Human Brain Atlas, and a density-based clustering algorithm identified AD subtypes. Subtype 1 mainly exhibits cortical thinning, while subtype 2 shows cortical thickening. Using whole-brain gene expression data, we found that deviations in transcriptionally connected neighboring regions predicted regional deviations in both subtypes. Gene enrichment analysis revealed that epicenter regions were associated with biological processes like synaptic dysfunction and phosphorylation regulation. Our study establishes associations between CGE connectivity and neural morphological alterations, identifies distinct epicenter regions in AD subtypes, and provides novel insights into how molecular-level gene expression shapes subtype-specific pathology.

Gesture recognition based on electromyography (sEMG) has garnered significant interest due to its immense potential in motor rehabilitation and auxiliary prosthesis movement. However, conventional deep learning models often struggle to fully capture the complex spatial and temporal dependencies in muscle activity, which can limit their performance in gesture recognition tasks. To address these challenges, we propose a novel CNN-AttST (Convolutional Neural Network with Spatiotemporal Attention) model that integrates a feature extraction module and a classification module. The feature extraction module employs a convolutional neural network to automatically learn relevant features from the raw sEMG data. In parallel, a spatiotemporal attention mechanism is implemented to emphasize critical temporal and spatial aspects of the signals, enhancing the model's ability to capture complex muscle activation patterns. Experimental results show that the CNN-AttST model was validated on both the Ninapro DB2 public dataset and our own collected dataset. On the Ninapro DB2 dataset, the model achieved recognition accuracies of 84.79% and 87.29% for time window lengths of 200ms and 300ms, respectively, outperforming similar works in the field.

Processing high-level cognitive functions like emotion involves dynamic interaction among multiple brain regions. Interactions involving within- and cross-frequency couplings across these regions are paramount in supporting brain functions. Existing emotion recognition models predominantly focus on within-frequency couplings. However, they lack the incorporation of cross-frequency couplings and within-frequency interactions, essential for providing a comprehensive representation of emotional states. To address this limitation, we propose a novel semi-supervised model for emotion recognition that incorporates a multi-layer network and coupling redundancy minimization (JMNCRM) into a unified framework. First, we construct a generalized multilayer network that embeds rich coupling information about within- and cross-frequency couplings through cosine similarity of features. Then, without increasing the feature dimensionality, the multi-layer network is incorporated into a discriminative linear regression model as a redundant minimum regularization term. During the optimization process, our model selects the most discriminative and non-redundant feature subsets for emotion recognition while retaining the rich structural, discriminative, and coupling information of electroencephalogram (EEG) data in the learned projection subspace. Extensive experimental results on two public datasets and our music-evoked emotion dataset demonstrate that the JMNCRM model outperforms other state-of-the-art algorithms regarding classification performance. Additionally, the intrinsic activation patterns revealed by JMNCRM are consistent with emotional cognition. The code for JMNCRM will be available at https://github. com/czxyhll/JMNCRM.

Sleep staging plays a crucial role in assessing sleep quality and diagnosing sleep-related disorders. Although previous studies have attempted to recognize sleep stages using automatic classification methods, achieving relatively high classification accuracy, most current automatic staging algorithms mainly rely on unimodal data and are limited to time-domain or frequency-domain features. To overcome these limitations, this study proposes a deep neural network model based on a multimodal fusion state-space sequence mechanism (AdaptiveSleepNet), designed to utilize multimodal signals for automatic sleep stage classification. AdaptiveSleepNet integrates a masking module, Adaptive Feature Recalibration (AFR) module, and State Space Model(SSM). The masking module is capable of masking missing modality signals. Then, features are extracted from multiple frequency bands of the signal through the Adaptive Feature Recalibration module, and channel feature weights are obtained using the residual Squeeze-and-Excitation (SE) attention mechanism. Next, the state-space sequence coupling module is used to learn cross-modal sequence relationships between signals. Finally, the fully connected layer outputs the sleep stage classification results. Evaluation of the Sleep-EDF-20 and Sleep-EDF-78 datasets shows that AdaptiveSleepNet achieved a classification accuracy of 84.74% and 82.34% through five-fold cross-subject classification with the EMG signal masked. Experimental results demonstrate that the AdaptiveSleepNet model proposed in this study effectively optimizes the performance of sleep staging and outperforms current state-of-the-art methods in staging performance.

Significant progress has been made in the study of Alzheimer’s disease (AD) and Mild Cognitive Impairment(MCI) progression using multimodal approaches. Recent studies have identified peripheral blood gene expression data as valuable biomarkers for distinguishing AD and MCI progression subtypes. However, these studies either rely entirely on prior data for gene selection or are purely data-driven. These strategies are not conducive to discovering potentially important genes or may lead to results unrelated to brain neural function pathways. This study adopts a data-driven approach based on prior knowledge, using genes mapped onto morphologically different brain regions as features for selection. These features are then input into our Generative Adversarial Network framework to obtain attention masks for cortical morphological indicators, which are weighted and used in training the structural MRI feature extraction main network. Our method ensures that the selected genes are correlated with brain regions and group differences, and through post-hoc interpretability analysis, we identify potential biomarkers in both genes and brain regions across two modalities.

Although current research has made significant progress in decoding working memory (WM) representation information, this area still faces several limitations that cannot be ignored. These limitations mainly include a lack of understanding of the dynamic neural mechanisms of WM and the limited feature extraction capability that traditional machine learning encounters when dealing with nonlinear temporal data. Inspired by the deep learning model - attention-based temporal convolutional networks (ATCNet), the decoding of representation information for WM has seen a new light. The ATCNet, through its unique attention mechanism and temporal convolutional networks (TCN) structure, demonstrates potential advantages in processing nonlinear temporal data and capturing dynamic patterns. In this deep learning framework, we successfully decoded the orientation information during a WM task based on single-trial electroencephalogram (EEG) with the average accuracy of 85.81%. The findings demonstrate the potential applications of ATCNet in decoding WM and suggest that EEG signals carrying task-related information can be effectively utilized for robustly decoding memory representations.

Electroencephalogram (EEG) brain networks describe the driving and synchronous relationships among multiple brain regions and can be used to identify different emotional states. However, methods for extracting interpretable structural features from brain networks are still lacking. In the current study, a novel deep learning structure comprising both an attention mechanism and a domain adversarial strategy is proposed to extract discriminant and interpretable features from brain networks. Specifically, the attention mechanism enhances the contribution of crucial rhythms and subnetworks for emotion recognition, whereas the domain-adversarial module improves the generalization performance of our proposed model for cross-subject tasks. We validated the effectiveness of the proposed method for subject-independent emotion recognition tasks with the SJTU Emotion EEG Dataset (SEED) and the EEGs recorded in our laboratory. The experimental results showed that the proposed method can effectively improve the classification accuracy of different emotions compared with commonly used methods such as domain adversarial neural networks. On the basis of the extracted network features, we also revealed crucial rhythms and subnetwork structures for emotion processing, which are consistent with those found in previous studies. Our proposed method not only improves the classification performance of brain networks but also provides a novel tool for revealing emotion processing mechanisms.

Deciphering brain activity evoked by visual stimuli has consistently been a popular pursuit in cognitive neuroscience. Due to the elusive foundations of visual formation, research on reconstructing visual stimuli encounters challenges. With the advancement of deep learning, several studies have successfully reconstructed scenes resembling visual stimuli from functional magnetic resonance imaging (fMRI). However, substantial dissimilarities persist in terms of contour representation. Furthermore, the majority of existing research primarily focuses on within-subject decoding. In this study, we propose a novel approach - MindLDM that permits cross-subject vision reconstruction. It first employs a Masked Autoencoder (MAE) to obtain the latent features of fMRI and align them into the Contrastive Language-Image Pre-Training (CLIP) text feature space. Then, the Very Deep Variational Auto-Encoders (VDVAE) is utilized to get the contour information of the visual input. Finally, a latent diffusion model combined with ControlNet is proposed to reconstruct the visual stimuli. The MindLDM successfully achieves image reconstruction on the publicly available Natural Scenes Dataset, generating images that exhibit a high degree of semantic correlation with the visual stimuli and demonstrate improved restoration of scene details. Quantitative and qualitative results demonstrate the effectiveness of the proposed method. An exhaustive ablation study was also conducted to analyze our framework.

In the current research on brain-computer interface (BCI), the electroencephalography (EEG) signal is usually only represented by a two-dimensional matrix, and the installation position of the EEG electrodes and the correlation between them are not considered. Actually, the cerebral cortex is a continuous potential surface and the information collected directly from each electrode is influenced by the other electrodes, so direct use of the raw data results in information redundancy. This paper converts the EEG signal into a graph, and then creates an information separator (IS) based on the Laplace matrix of the graph to obtain the independent source information of electrode nodes, and propose an IS-based area convolutional network (IS-ACN). Integrating the proposed IS with some advanced methods, the experimental results show that the incorporation of IS can enhance the performance of these methods. By observing and tracking samples with abnormal noise in the BCI competition IV dataset 2a, it is demonstrated that the proposed method can greatly reduce the influence of noise and effectively obtain the source features of EEG signals with low signal-to-noise ratio, and the average accuracy and kappa coefficient of the proposed IS-ACN on this dataset are 80.59% and 74.1%, respectively.

Emotion refers to the subjective emotional experience that humans generate in specific situations, typically accompanied by physiological and psychological changes. In the field of emotions, multi-channel EEG emotional features can better reflect the collaborative mechanisms across multiple brain regions. Therefore, we propose a novel feature called asPLV (averaged sub-frequency phase locking value) based on the Morlet transform method to construct functional network edges. The proposed feature encompasses a comprehensive analysis of phase synchronization across sub-frequency bands spanning a wide range of frequencies and has the potential to reduce fluctuations arising from reliance on a single frequency band. We designed a music-evoked emotion experiment aimed at inducing corresponding emotions in participants while simultaneously recording their electroencephalogram (EEG) signals and extracted our proposed asPLV feature for classification. The results show that the proposed feature displays superior classification performance and generalization compared to other state-of-the-art methods. The proposed method is not only effective in successfully distinguishing between different emotions but also introduces a novel brain network metric to elucidate the collaboration and information exchange among emotion-related brain regions. Moreover, the asPLV feature could provide new insights for the development of an emotional brain-computer interface (BCI).