International journal of neural systems最新文献

Estimating the speed of nearby vehicles is essential for driver assistance. A real-time, camera-only pipeline is presented that uses an onboard monocular camera: vehicles are detected and tracked with an off-the-shelf one-stage CNN (YOLOv8); distance is approximated from bounding-box angular width using class-dependent priors (2.0[Formula: see text]m cars; 2.5[Formula: see text]m larger vehicles) and camera intrinsics; a Nadaraya-Watson kernel smooths the distance sequence, and its analytic derivative yields relative speed. The approach supports multiple targets without dedicated ranging hardware. Evaluation on CARLA synthetic video with ground truth analyzes estimated versus ground-truth distance, estimate/ground-truth ratio versus image-center displacement, and kernel-based speed versus a polynomial trend. Results show a positional bias away from the image center and a stability-lag trade-off due to smoothing. The contribution is a detector-agnostic distance-speed head that couples angular geometry with analytic Nadaraya-Watson smoothing and differentiation for real-time operation, positioned as a low-cost alternative or complement to active sensors, with limitations and paths to real-world validation outlined.

EEG recordings obtained before medication are regarded as valuable biological indicators for depression detection. Currently, depression diagnosis based on EEG using convolutional neural networks (CNNs) has achieved relatively high detection performance, but some issues remain unresolved. CNNs are constrained by their limited receptive fields, which restrict them to capturing local rather than global dependencies. In addition, the complex features learned by CNNs are often hard to interpret and typically require a substantial number of trainable parameters. To tackle these issues, an interpretable hybrid neural network named SINCFORMER-SHAP is proposed. SINCFORMER-SHAP comprises two main components, namely the spatial-frequency and temporal feature extraction modules. The spatial-frequency feature extraction module leverages a hybrid design, where temporal filtering through a sinc-based convolution is coupled with spatial convolution, enabling the model to learn fine-grained spatial-spectral patterns. The sinc-convolutional layer helps constrain the parameter count, enhancing model efficiency. Subsequently, the temporal domain feature extraction module utilizes Transformer to capture global time-domain dependencies. Kernel visualization is used to provide direct insights into the spectral features learned by the spatial-frequency feature extraction module. To further enhance interpretability on the spatial domain, a post-hoc analysis is conducted using SHAP method. Based on the results of interpretability analysis, potential biomarkers have been observed within alpha and gamma rhythms across the frontal, parietal, temporal, and occipital areas. Comprehensive experiments conducted on public MODMA, EDRA and Mumtaz datasets were used to assess the performance of the proposed approach. The experimental outcomes provide compelling evidence that the proposed method not only surpasses multiple state-of-the-art approaches in performance, but also contributes a significant advancement toward the development of interpretable diagnostic technique for depression, thereby bridging the gap between computational methodologies and practical psychiatric applications.

With the increase in work-related stress, the issue of psychological pressure in occupational environments has gained increasing attention. This paper proposes an enhanced Informer stress recognition and classification method based on deep learning, which guarantees performance by integrating tailored spatial and channel attention mechanisms (SAM/CAM) with the Informer backbone. Unlike existing attention-augmented models, the proposed SAM is designed to prioritize time-sensitive physiological signal segments, while CAM dynamically weights complementary stress-related features, enabling precise capture of subtle stress-related patterns. With this dual attention mechanism, the proposed model can capture subtle changes associated with stress states accurately. To evaluate the performance of the proposed method, the experiments on one publicly available dataset were conducted. Experimental results demonstrate that the proposed method has outperformed existing approaches in terms of accuracy, recall, and F1-score for stress recognition. Additionally, we performed ablation studies to verify the contributions of spatial attention module and channel attention module to the proposed model. In conclusion, this study not only provides an effective technical means for the automatic detection of psychological stress, but also lays a foundation for the application of deep learning model in a broader range of health monitoring applications.

In motor imagery (MI)-based brain-computer interfaces (BCIs), convolutional neural networks (CNNs) are widely employed to decode electroencephalogram (EEG) signals. However, due to their fixed kernel sizes and uniform attention to features, CNNs struggle to fully capture the time-frequency features of EEG signals. To address this limitation, this paper proposes the Multi-Domain Dynamic Weighted Network (MD-DWNet), which integrates multimodal complementary feature information across time, frequency, and spatial domains through a branch structure to enhance decoding performance. Specifically, MD-DWNet combines multi-band filtering, spatial convolution, and temporal variance calculation to extract spatial-spectral features, while a dual-scale CNN captures local spatiotemporal features at different time scales. A dynamic global filter is designed to optimize fused features, improving the adaptive modeling capability for dynamic changes in frequency band energy. A lightweight mixed attention mechanism selectively enhances salient channel and spatial features. The dual-branch joint loss function adaptively balances contributions through a task uncertainty mechanism, thereby enhancing optimization efficiency and generalization capability. Experimental results on the BCI Competition IV 2a, IV 2b, OpenBMI, and a self-collected laboratory dataset demonstrate that MD-DWNet achieves classification accuracies of 83.86%, 88.67%, 75.25% and 84.85%, respectively, outperforming several advanced methods and validating its superior performance in MI signal decoding.

Clinically, epilepsy manifests as a chronic condition marked by unprovoked, recurrent seizures, plaguing over 70 million individuals with debilitating seizures and life-threatening complications. Approximately 30% of patients with epilepsy do not respond to conventional antiepileptic drugs, indicating the limited efficacy of these medications in controlling seizures universally. Therefore, seizure prediction has become a key factor in enabling timely intervention for epilepsy patients, which can provide crucial time for clinical treatment and preventive measures. This study aimed to propose a lightweight seizure prediction model integrating a residual network (ResNet) with a kernel-enhanced global temporal attention Block (GTA Block). The ResNet extracts electroencephalogram (EEG) features while maintaining gradient stability, and the GTA mechanism constructs full-sequence temporal association matrices to capture the dynamic evolution of EEG patterns. Then a kernel function is embedded into GTA Block for mapping EEG samples into a high-dimensional space in which the distinction between preictal and interictal states is enhanced. The model significantly outperforms existing methods while maintaining a lightweight architecture suitable for embedded systems. With only 1.94 million parameters and an inference time of 0.00207[Formula: see text]s, this lightweight design facilitates real-time deployment on wearable devices, enhancing feasibility for continuous clinical monitoring in resource-constrained settings.

This study aims to develop a multimodal driver emotion recognition system that accurately identifies a driver's emotional state during the driving process by integrating facial expressions, ElectroCardioGram (ECG) and ElectroEncephaloGram (EEG) signals. Specifically, this study proposes a model that employs a Conformer for analyzing facial images to extract visual cues related to the driver's emotions. Additionally, two Autoformers are utilized to process ECG and EEG signals. The embeddings from these three modalities are then fused using a cross-attention mechanism. The integrated features from the cross-attention mechanism are passed through a fully connected layer and classified to determine the driver's emotional state. The experimental results demonstrate that the fusion of visual, physiological and neurological modalities significantly improves the reliability and accuracy of emotion detection. The proposed approach not only offers insights into the emotional processes critical for driver assistance systems and vehicle safety but also lays the foundation for further advancements in emotion recognition area.

Accurate localization of the theoretical epileptogenic zone in cingulate epilepsy is particularly challenging due to the region's deep anatomical location and complex connectivity. While invasive stereoelectroencephalography (sEEG) methodology offers excellent spatiotemporal sampling of deep intracerebral structures, interpretation of these high-dimensional recordings remains largely qualitative and subject to interpretation by clinician experts. To address this limitation, we propose a quantitative, biomarker-based framework using phase-amplitude coupling (PAC) to investigate 25 seizures recorded from four patients with complex cingulate epilepsy who underwent sEEG followed by surgical treatment (either laser ablation or open resection), achieving ≥ 1 year of sustained seizure freedom. PAC values were computed from sEEG electrode contacts across multiple seizures during pre-ictal and ictal phases, employing wide-frequency and band-specific frequency coupling approaches. Among frequency pairs, theta-beta ([Formula: see text]-[Formula: see text]) coupling consistently demonstrated the most robust differentiation between surgically-treated and untreated contact sites. Our findings highlight frequency-specific PAC-based metrics as a potential tool for mapping dynamic epileptiform activity in brain networks, offering quantitative insight that may refine surgical planning and decision-making in challenging cases of cingulate epilepsy.

Automatic seizure detection holds significant importance for epilepsy diagnosis and treatment. Convolutional neural networks (CNNs) have shown immense potential in seizure detection. Though traditional CNN-based seizure detection models have achieved significant advancements, they often suffer from excessive parameters and limited interpretability, thus hindering their reliability and practical deployment on edge computing devices. Therefore, this study introduces an innovative Morlet convolutional neural network (Morlet-CNN) framework with its effectiveness demonstrated in seizure detection tasks. Unlike traditional CNNs, the convolutional kernels in the Morlet-CNN contain only two learnable parameters, allowing for a lightweight architecture. Additionally, we propose a frequency-domain-response-based kernel pruning algorithm for Morlet-CNN and implement an INT8 quantization algorithm by incorporating Kullback-Leibler (KL) divergence calibration with a Morlet lookup table (LUT). With the pruning and quantization algorithms, the model's parameter scale achieves over 90% reduction while maintaining minimal accuracy loss. Furthermore, the model exhibits enhanced interpretability from a signal processing perspective, distinguishing it from many previous CNN models. Extensive experimental validation on the Bonn and CHB-MIT datasets confirms the Morlet-CNN model's efficacy with a compact Kilobyte (KB)-level model size, making it highly suitable for real-world applications.

Knowledge graphs (KGs) which represent entities and their relations in a structured manner, have become a fundamental resource for various natural language processing tasks. However, the incompleteness of KGs significantly hinders their effectiveness, thereby reducing their practical utility. The challenge of predicting missing relations between entities and performing these predictions efficiently has become a focal point of research. To address the challenge of incomplete KGs, we propose GEGS, a novel KG embedding framework that enhances scalability and expressiveness for relation prediction. GEGS introduces GAT-SNP, a graph attention network that, for the first time, integrates nonlinear spiking neural P (SNP) mechanisms into graph attention models and applies them to the KG domain, effectively capturing complex relational structures. The GAT-SNP network assigns distinct attention weights to each node, enabling the model to focus on the most relevant nodes in the graph. To mitigate information loss in long-range and sequential path features, we incorporate a BiLSTM-SNP component, which alleviates long-term dependency issues while preserving global path information. By leveraging GAT-SNP and BiLSTM-SNP, GEGS achieves superior performance in link prediction tasks, paving the way for applications in large-scale knowledge base completion. Kinship, FB15k-237, and WN18RR are used to evaluate the proposed GEGS model. The experimental results indicate that the proposed GEGS model has achieved state-of-the-art results in multiple evaluation metrics(e.g. Hits@10 and MRR).

Electroencephalogram (EEG) plays a vital role in seizure detection, yet existing methods often fail to adequately capture the spatiotemporal characteristics of EEG signals, leading to limited performance. Moreover, most current models depend on supervised learning and thus require large amounts of labeled data. To address these issues, this paper introduces the Long Short-Term Memory-Transformer (LTformer) encoder, designed to model long-term temporal dependencies in EEG signals while retaining spatial information across electrode channels. We further propose a dual-stream self-supervised learning (SSL) strategy to pretrain the model, enabling the LTformer encoder to learn discriminative representations from extensive unlabeled EEG data. After pretext training, the encoder is transferred and fine-tuned for downstream seizure detection. The proposed method, termed Self-Supervised Attention LTformer (SALT), is evaluated on two public EEG datasets using both segment-based and event-based experimental protocols. In the segment-based evaluation, SALT achieves 98.87% sensitivity, 99.15% accuracy, and 99.41% specificity on CHB-MIT, and 98.04% sensitivity, 97.72% accuracy, and 97.62% specificity on Siena. In the event-based evaluation, SALT attains 98.57% sensitivity with a false discovery rate (FDR) of 0.26 on CHB-MIT, and 98.65% sensitivity with an FDR of 0.25 on Siena. The code is publicly available at https://github.com/peutim114/SALT.