Sleep is essential for health, yet polysomnography (PSG), the clinical gold standard for comprehensive sleep assessment, is intrusive and impractical for routine use. Sleep sounds offer a non-intrusive alternative, but acoustic signals alone may provide insufficient information for reliable assessment. To address this limitation, we propose a multimodal knowledge distillation (MKD) framework for estimating subjective sleep quality, defined in this study as participants' self-reported satisfaction with each night's sleep. The framework uses richer multimodal information only during training and deploys a sound-based student model at inference time. The teacher model combines sleep-stage features, PSG features, subjective factors, sound features, and demographic factors using hierarchical gated variable selection networks (GVSNs). The student model is restricted to sound features and demographic factors, and response-based distillation transfers the teacher's multimodal decision behavior to this practical input setting. Using 198 nights from 101 adults, we evaluate the framework under both night-wise and subject-wise protocols. Distillation improves the student over the non-distilled model in the repeated splits, with the larger observed gain occurring when test participants are unseen during training. Further analyses show that the preferred distillation weight, the teacher's learned modality weights, and the effect of removing each teacher modality differ between the evaluation protocols. In particular, the ablation results show that a modality's effect on standalone teacher performance does not necessarily match its effect on the distilled student's performance. These findings provide preliminary, dataset-specific evidence of multimodal distillation but do not establish clinical utility or the transfer of any specific modality.
Nonlinear spiking neural P (NSNP) systems offer a biologically inspired framework for modeling nonlinear temporal dynamics via spike consumption and generation. Existing NSNP-based recurrent architectures, such as the long short-term memory model inspired from spiking neural P systems (LSTM-SNP), primarily rely on implicit gating mechanisms and lack explicit state correction during recurrent propagation, which can lead to unstable state evolution and excessive information decay in long-term temporal modeling. To address this limitation, a compensation-balanced LSTM-SNP (CBLSTM-SNP) architecture is proposed. The model introduces a compensation-driven regulation mechanism, comprising a regulation branch and an adaptive compensation branch, to dynamically adjust hidden state evolution. From a nonlinear dynamical systems perspective, CBLSTM-SNP establishes a balanced recurrent state transition integrating memory preservation, nonlinear spike processing, and adaptive state restoration. Theoretical analysis confirms the boundedness and Lipschitz continuity of the proposed recurrent dynamics. Empirical evaluations on five benchmark time series datasets demonstrate that CBLSTM-SNP produces smoother hidden state trajectories and achieves comparable forecasting performance compared with several representative recurrent and hybrid models.
Accurate segmentation of brain MRI images is essential for early diagnosis and prevention of brain tumors and strokes. However, it is challenging due to data scarcity, domain shift, complexity, and blurred edges. Recent methods use mean-teacher model learning for the semantic details of an individual image based on the efficiency of teacher architecture. However, to enhance the segmentation in brain MRI, the paper introduces SemS4, a semi-supervised method on the basis of cross-image architecture. SemS4 uses a single architecture integrating an Edge Prototype Attention (EPA) and a foreground prototype attention (FPA). More specifically, EPA uses edge prototype integrated with adaptive edge container to improve the edge features for which global edge features are stored to produce edge prototype. Moreover, FPA stabilizes the segmentation by transferring complementary foreground information. Furthermore, the proposed method uses Pixel Affinity Loss (PAL) for missing contextual correlation during supervision to enhance segmentation performance on edges. The performance of the proposed approach in experimental results for two brain MRI benchmarks, LGG and BRISC, validates the superiority over existing methods under different partitioning settings.
Asynchronous steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs) show great potential for real-world control, yet their performance is often impaired by phase shifts between ongoing electroencephalogram (EEG) signals and fixed-phase templates. In this study, we observed and validated a phase-sliding oscillation (PSO) phenomenon: when a fixed-phase template is used, phase-sliding trajectories derived from sliding-window sequences with different initial phases exhibit stable periodic oscillations locked to the stimulus frequency. Based on this finding, we propose a phase-sliding oscillation-based asynchronous classification (PSO-AC) method. Specifically, EEG signals are reconstructed into sequences with varying initial phases and transformed into stable time-frequency representations via wavelet synchrosqueezing transform to extract phase-decoupled features. A probabilistic temporal model is then used to characterize temporal dynamics and distinguish control from non-control states. Offline experiments (n=22) showed that PSO-AC achieved a mean control/non-control accuracy of 94.2% with a 2-s data length, outperforming a state-of-the-art baseline. It also maintained superior performance in nine-class classification under different time delays. In online robotic-arm experiments , PSO-AC enabled more stable command triggering and higher control efficiency, reducing the command cost per successful trial from 4.30 to 1.14.
Electrical status epilepticus during sleep (ESES) is a severe but often underdiagnosed epileptic encephalopathy that can significantly impair cognitive development in children. Current diagnostic practices rely heavily on expert visual inspection of spike-and-slow-wave complexes in electroencephalogram (EEG) recordings. This process is time-consuming, labor-intensive, and subjective, often resulting in poor inter-rater agreement. To address these limitations, this paper introduces a novel ESES automatic detection framework based on a large pre-trained EEG model. The framework employs a pre-trained NeuroEncoder to extract temporal and frequency features from spike-and-slow-wave complexes, and uniquely incorporates a temporal and spatial embedding module to enhance the model's ability to recognize spatiotemporal patterns. By leveraging the multi-head attention mechanism of stacked Transformers, the framework further captures complex relationships and deep features. A symmetric label smoothing cross-entropy is introduced, which applies both forward and backward constraints to reduce overfitting to noisy labels and improve robustness to label noise and individual variability. The proposed framework can automatically quantify spike-and-slow-wave complexes and calculate the spike-wave index (SWI) for ESES diagnosis. Evaluation was conducted on EEG recordings from 35 patients using a leave-one-subject-out cross-subject strategy. The results demonstrate excellent performance, with an accuracy of 91.91%, sensitivity of 91.52%, and specificity of 92.37%. The high agreement between the model and expert annotation is confirmed by a Kappa coefficient of 0.8368 and Gwet's AC1 of 0.8394. The proposed method demonstrates superior generalization across subjects and outperforms existing approaches. The proposed framework provides an efficient and reliable tool for automatic ESES detection, supporting early intervention and improved clinical management for patients with ESES.
Understanding how the human brain encodes complex natural scenes remains a central problem in computational neuroscience and artificial intelligence. Existing visual encoding models often rely on a single dominant feature representation and may insufficiently characterize how saliency-guided spatial information and high-level semantic context jointly contribute to cortical response prediction. To address this issue, this study proposes a saliency-guided multimodal visual encoding model, termed SMG-MVEM, to predict voxel-wise cortical responses to natural scene stimuli. The model integrates image features, saliency cues, and text-derived semantic representations through a hierarchical fusion architecture, followed by a Transformer-based brain mapper. Experiments on the Natural Scenes Dataset (NSD) show that SMG-MVEM improves prediction performance over representative neural encoding baselines and internal control variants, with the average PCC increasing from [Formula: see text] for the best-performing baseline to [Formula: see text]. Regional analyses further show that saliency contributed more strongly to early visual areas, whereas semantic features provided greater benefits in higher-order regions. Representational analyses also suggest that the model-predicted responses preserved aspects of hierarchical and category-related organization across the visual cortex. These findings indicate that structured integration of saliency and semantic context can improve cortical response prediction and provide interpretable representational patterns for natural vision.
This study addresses the entity identification problem in information propagation networks and proposes a novel dual-track interactive-fusion directed graph neural network (GNN). The method characterizes node features from two complementary perspectives: static topology and dynamic path dependency. On the one hand, it considers the relative position of nodes in directed network path sequences and enhances the model's ability to represent directed structural relations through static-track encoding. On the other hand, dynamic-track encoding is used to learn the sequential associations among nodes in network path sequences. Based on this, a bidirectional cross-track interactive fusion mechanism is introduced to facilitate the interaction, fusion, and collaborative enhancement of latent information between the static and dynamic tracks, thereby improving node representations in directed graphs. Experiments on the Cora, Reddit, PubMed, and Ogbn-arxiv datasets show that, compared with the best baseline, the proposed dual-track interactive-fusion method improves the F1-score by 1.83%, 3.32%, 3.19%, and 2.61%, respectively. The corresponding reduction rates in the remaining performance gap are 5.93%, 11.98%, 13.98%, and 8.55%, respectively. This study provides a new theoretical perspective and technical approach for handling complex dynamic directed graphs and has important application value in social media analysis, public opinion monitoring, and network security.
Voice-based analysis is attracting growing interest as a noninvasive means of identifying early markers of autism spectrum disorder (ASD). While pretrained audio models such as YAMNet and VGGish provide complementary views of children's speech, most existing studies rely on a single representation and do not explore how these embeddings may be combined in a structured manner. This work introduces the multibranch attention and fusion network (MBAFNet), an architecture designed to make fuller use of heterogeneous embeddings by processing each stream through its own convolutional encoder, extracting temporal cues at multiple scales, and modeling cross-representation interactions through a self-attention layer. A gating mechanism then regulates the relative contribution of each embedding before classification. Experiments conducted on the CASD-SC corpus under a subject-disjoint five-fold evaluation protocol show that MBAFNet achieves 94.17% accuracy, outperforming all evaluated baselines and previously reported state-of-the-art approaches. These findings indicate that carefully designed selective fusion can reveal complementary information contained in pretrained embeddings and supports the development of more robust speech-based ASD assessment frameworks, while broader validation remains necessary before screening-oriented use.
Characterized by recurrent fluctuations in mood states, bipolar disorder (BD) is widely conceptualized as a disconnection syndrome associated with dysregulated brain dynamics. Nevertheless, the molecular mechanisms underlying this aberrant connectivity dynamics in BD remain elusive. Using resting-state electroencephalography (EEG) data from BD patients and healthy controls, this study first delineated the characteristic alterations in temporal variability of functional connectivity in BD and further elucidated their underlying molecular mechanisms and clinical relevance. Current findings revealed significantly reduced temporal variability within large-scale brain subnetworks, most notably in the dorsal attention, somatomotor, and visual networks. Importantly, these neurodynamic signatures effectively predicted the symptom severity in individuals with BD. Moreover, the spatial patterns of these dynamic alterations are associated with the expression of BD risk genes enriched in synaptic function and metabolic pathways, as well as with the spatial organizations of various neurotransmitter receptors, including CB1, mGluR5, H3, and MOR. Collectively, these results provide evidence for a multiscale pathophysiological framework that links genetic susceptibility and chemoarchitectural alterations to dynamic brain network instability, ultimately underpinning the core clinical manifestations in BD.
This study proposes an innovative WalkBCI real-time motor imagery brain-computer interface system to solve the problems of time-consuming model calibration and low EEG decoding accuracy in stroke patients because of individual differences. WalkBCI integrates generative adversarial networks and transfer-learning techniques to generate motor imagery feature data via RM-GAN using the resting data of target subjects, which shortens the time of calibration data acquisition and reduces patient fatigue. Meanwhile, the system enhances classification performance by combining source domain data during transfer learning optimization. The system is evaluated in both offline and online experiments to assess its practicality and stability. The study results show that WalkBCI outperformed traditional methods on the stroke patient dataset. In the offline experiments, the model's accuracy, precision, recall, and F 1-score of the model increased by 3.3%, 2.6%, 3.8%, and 3.4%, respectively. In online experiments, the system maintained stable classification performance with a maximum accuracy of 72.5%, while reducing calibration time by 58%. WalkBCI effectively lowered subjects' fatigue, and the comparison experiments revealed average reductions of 7.5 points in FS-14 scores and 74.05 in frontal theta/beta values. This study offers an efficient and stable solution for real-time EEG decoding, particularly well-suited for motor-imagery tasks in stroke patients.
Current dementia diagnostic methods can be costly, invasive, or limited in their ability to distinguish between disorders with overlapping clinical symptoms. Dysregulated microRNAs (miRNAs) have emerged as promising noninvasive biomarkers for neurodegenerative disease, but individual miRNA changes alone may not capture the complex molecular patterns needed for accurate disease classification. Machine learning provides a way to integrate multiple layers of miRNA-derived information and identify disease-specific biomarker signatures. In this study, we developed machine learning models to classify dysregulated miRNAs associated with Alzheimer’s disease dementia (AD), Lewy body dementia (LBD), and frontotemporal dementia (FTD). Each miRNA was represented using sequence-derived descriptors, predicted gene targets, and KEGG pathway features. The highest-performing models trained on AD, LBD, and FTD achieved 10-fold cross-validation accuracies of 90.6%, 92.9%, and 100%, respectively. When evaluated on independent datasets, the AD, LBD, and FTD models achieved accuracies of 88.9%, 77.8%, and 90.9%, respectively. Cross-disease testing showed reduced performance when models were applied across dementia types, suggesting partially disease-specific miRNA patterns while also indicating overlap among the molecular signatures of AD, LBD, and FTD. These results suggest that machine learning-based integration of miRNA sequence, target-gene, and pathway information can improve the identification of dementia-associated biomarker signatures and may support the future development of noninvasive diagnostic tools for dementia.
Modeling deep neural networks as continuous-time dynamical systems has emerged as a powerful approach for analyzing their evolution and robustness. By leveraging the superior numerical stability of implicit methods, several studies have proposed the implicit Euler-based architecture, which significantly enhances both robustness and generalization under adversarial attack. However, existing methods require solving implicit equations iteratively, which substantially increases the computational cost. To overcome these limitations, this paper proposes an explicit reformulation of the implicit Euler method, termed Implicit-to-EXplicit ResNet (IEX-ResNet). IEX-ResNet preserves the inherent numerical stability of implicit methods while eliminating the need for iterative solvers, thereby enabling the efficient training of robust neural models. Theoretical analysis demonstrates that IEX-ResNet, under local linearization and spectral radius constraint, achieves Lipschitz-bounded perturbation propagation and maintains a stable approximation of the underlying implicit dynamics. Experimental results show that, compared to conventional implicit Euler architectures, IEX-ResNet achieves improved robustness across diverse challenging settings and maintains competitive robustness while reducing training time by nearly 50% under standard evaluation. Moreover, IEX-ResNet provides practical benefits in a real-world document image restoration task. These results suggest that implicit Euler explicitization is a practical route toward efficient robust neural architecture design.
Scientific conceptual change poses a cognitive challenge in science learning, requiring cognitive conflict resolution and restructuring of entrenched semantic knowledge. Although neuroscience implicates coordinated prefrontal cognitive control and hippocampal-temporal systems to conflict resolution and semantic memory updating, direct neural evidence of their interaction during conceptual change remains limited. Using functional MRI and Granger causality (GC) analyses, we examined neural mechanisms distinguishing successful from unsuccessful conceptual change across task complexities in 48 undergraduate and graduate STEM majors who completed 39 conceptual change tasks (19 weak restructuring task, 20 radical restructuring task). Learners exhibited greater conceptual change success during weak than radical restructuring, with enhanced BOLD activation in the superior, middle, and inferior frontal gyri (SFG, MFG, IFG), hippocampus (HIP), and superior temporal gyrus (STG) during successful versus unsuccessful conceptual change. GC analysis identified a minimal network (SFG, MFG→HIP, STG→MTG, MTG→HIP) common to both conditions, whereas successful conceptual change engaged a denser prefrontal-hippocampal-temporal network, marked by additional unidirectional (MFG→SFG, MFG→IFG, IFG→HIP) and bidirectional (MFG↔STG, STG↔HIP, MTG↔HIP) connectivity. Unsuccessful conceptual change was characterized by shallow, fragmented connectivity and limited cross-regional integration. Collectively, these findings suggest that effective conceptual change arises from coordinated neural networks integrating executive control, semantic processing, and memory updating.
Artifacts are noisy signals that commonly contaminate electroencephalographic (EEG) recordings, mixing with underlying brain activity and degrading the quality of neurophysiological data. Previous research on epileptic Anomaly Detection has shown that this approach is also sensitive to unlabelled artifacts, often leading to an increased False Positive rate. While most methods focus on detecting or removing a single type of artifact, this work proposes a unified multi-class framework to classify several artifact types alongside normal and pathological brain activity within a recording using a single, simple Machine Learning classifier. A set of spectral, temporal and statistical features commonly associated with different artifact types is extracted. KNN and XGBoost classifiers are trained and evaluated under a cross-validation scheme. The results demonstrate strong performance for both models, achieving approximately 90% sensitivity across all classes while maintaining a 100% specificity. These findings highlight the effectiveness of a unified and interpretable approach for multi-class EEG artifact classification.
Electrical status epilepticus in sleep (ESES) is an epilepsy syndrome requiring early diagnosis. Its diagnostic subjectivity and therapeutic inadequacy necessitate the urgent discovery of objective markers. This study developed a microstate-centered multidimensional EEG framework to distinguish epilepsy patients with ESES from those without ESES and to identify interpretable neurophysiological alterations. Broadband spatiotemporal parameters, frequency-specific microstate features, nonlinear complexity, symbolic-sequence dynamics, Spike-Wave Index (SWI) correlations, and exploratory classification performance were evaluated. ESES-related alterations were concentrated in microstate B and C, involving prolonged and more variable temporal persistence, altered occurrence frequency, increased global field strength, frequency-sensitive duration changes, and reduced frequency chaos game representation (FCGR) entropy. Markov order features were positively associated with SWI-related electrophysiological burden, suggesting potential severity-related sequence dynamics. A five-feature L2-regularized logistic regression model achieved promising exploratory discrimination, with [Formula: see text] and [Formula: see text]. This study indicated that ESES is associated with more constrained and less flexible microstate dynamics, while a compact, interpretable microstate-derived model achieved promising exploratory discrimination. These findings may provide novel analytical perspectives for deciphering ESES pathophysiology and identifying candidate neurophysiological markers for future assessment.
Polysomnography (PSG) is the current gold standard for sleep monitoring; however, it is complex, burdensome, and costly. In contrast, commercial health monitoring devices, such as the Apple Watch, offer sleep scoring at a low cost over a few days. This paper proposes a dual-attention mechanism-based sleep staging method using single-lead electrocardiogram (ECG) signals, named SleepECGNet. The method extracts multi-scale features by employing convolutional kernels of different sizes, thereby capturing subtle details in the ECG signals. The model then integrates an SE module and a Luong attention-based encoder-decoder architecture to filter redundant information and dynamically focus on different sleep stages. Experiments were conducted using the HMC, MGH, and SHHS1 datasets for sleep staging tasks and the Apnea-ECG corpus was employed to execute transfer learning strategies directed at sleep apnea recognition. Comprehensive ablation experiments were conducted to verify the effectiveness of each component. Moreover, an analysis of sequence duration demonstrated that expanding the input timeframe directly improves classification outcomes.
Deep learning (DL) has shown considerable promise for EEG-based dementia assessment; however, rigorous cross-family comparisons under leakage-free and clinically meaningful evaluation protocols remain limited. To address this gap, we benchmarked twelve DL architectures from four major families - recurrent neural networks (RNNs), temporal convolutional networks (TCNs), Transformers, and self-supervised learning (SSL) - for the multi-class differentiation of Alzheimer's disease (AD), frontotemporal dementia (FTD), and healthy controls (HC). The analysis was conducted on a clinically acquired EEG dataset comprising 36 AD patients, 23 FTD patients, and 29[Formula: see text]HC participants. To prevent subject-level data leakage, we employed strict subject-independent five-fold cross-validation, and to improve clinical relevance, final predictions were aggregated at the subject level using majority voting across EEG segments. Model performance was evaluated using classification metrics, learning stability, statistical ranking, computational efficiency, and interpretability analysis. Among all benchmarked architectures, the self-supervised SimCLR-Time framework achieved the strongest subject-level performance, reaching a mean accuracy of [Formula: see text] and a weighted F1-score of [Formula: see text]. The time-series transformer (TST) was the best-performing fully supervised model, with an accuracy of [Formula: see text], followed closely by the hybrid LSTM-FCN model at [Formula: see text]. In contrast, the EEG-Seq2Seq model showed the weakest performance ([Formula: see text]). Computational profiling further showed that TCN-Attention offered a favorable trade-off between accuracy and edge-oriented efficiency, while SimCLR-Time combined strong diagnostic performance with the fastest inference latency. In addition, Integrated Gradients-based saliency analysis suggested that the top-performing model relied on structured spatiotemporal EEG patterns rather than arbitrary temporal artifacts. Overall, these findings provide a controlled within-dataset benchmark for multi-class dementia classification from raw EEG and suggest that self-supervised and attention-based architectures are promising directions for clinically oriented EEG decision-support systems. However, external validation on larger and more heterogeneous cohorts remains necessary before broader generalization can be claimed.
Large-scale datasets impose substantial training costs on machine learning models. Dataset distillation addresses this issue by synthesizing compact datasets that can achieve performance comparable to that of the original data. However, text dataset distillation remains challenging: the discrete nature of text renders traditional gradient-matching methods ineffective, while embedding optimization approaches are often inefficient and exhibit limited generalization. To address these challenges, this paper proposes an LLM-native distillation framework based on dual-agent collaboration. Our framework decomposes the distillation process into two stages: selection and improvement. The Selector identifies high-quality samples through multi-dimensional scoring, while the Improver enhances data density and clarity under semantic consistency constraints and a candidate generation mechanism. The entire pipeline is automated through a self-iterative cycle of generation and selection, and leverages scoring signals to drive agent self-reward iteration, thereby avoiding differentiable optimization and trajectory matching. Experiments show that starting from a random subsampling, using only 2.5% of original data and only three iterations, the distilled dataset enables Llama-2-7B, Mistral-7B, and Qwen2.5-7B to match full-dataset performance on MMLU and Winogrande. It also improves training stability, convergence efficiency, and cross-model generalization, especially among architecturally similar models, while maintaining a competitive distillation cost of 143 GPU h. Overall, this study provides a practical solution for efficient, automated, and general-purpose text dataset distillation.
This study proposes a novel enhancement to the Conformer architecture for epileptic seizure identification by replacing the standard Vision Transformer (ViT) with the Swin Transformer. The proposed Swin-Conformer model leverages the hierarchical patch merging and shifted-window self-attention mechanisms of the Swin Transformer to better capture both local electrophysiological patterns and long-range temporal dependencies in electroencephalogram (EEG) signals. To handle the significant class imbalance inherent in epileptic EEG data, a weighted focal loss function is employed during training. The model is evaluated on the CHB-MIT dataset through stratified 10-fold cross-validation and validated on the independent Bonn EEG dataset. Experimental results demonstrate that the proposed model achieves accuracy of 99.24%, specificity of 99.55%, and sensitivity of 98.47% on the segment-based evaluation, as well as an event-based sensitivity of 99.50%, outperforming the original Conformer and state-of-the-art baseline methods. Ablation studies confirm that the performance gains originate from the Swin Transformer's hierarchical multi-scale feature representation and efficient local-global attention mechanism. Statistical analysis confirms that the improvements are statistically significant ([Formula: see text]).
This study introduces a deep learning framework for the inferential exploration of latent representations in 3D brain MRI, leveraging a simple convolutional autoencoder with a hierarchical encoder and a compact latent space. Trained on segmented gray matter images from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, the model learns latent representations that preserve neuroanatomical structure and reflect clinical variability across cognitive status. Dimensionality reduction techniques (PCA, t-SNE, PLS, UMAP) were applied to visualize and interpret the latent space, correlating it with anatomical regions defined by the AAL atlas. As a novel contribution, the Latent-Regional Correlation Profiling (LRCP) framework, which combines statistical association and supervised discriminability to identify brain regions that encode clinically relevant latent information is proposed. Our results show that even minimal architectures capture meaningful patterns associated with progression to Alzheimer's disease. Interpretability is assessed by applying SHAP-based regression to a post-hoc model that predicts reconstruction error from atlas-based regional gray matter intensities, thereby identifying anatomically meaningful regions involved in class-specific reconstruction strategies. These findings are further validated using statistical agnostic methods, highlighting the importance of rigorous evaluation in neuroimaging. This work demonstrates the potential of autoencoders as exploratory tools for biomarker discovery and hypothesis generation in clinical neuroscience.
