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.

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.

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.

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.

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]).

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.

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.

Generative AI systems increasingly support users in coding, data analysis, and creative tasks through natural-language interaction. However, user prompts are often underspecified or ambiguous, and current LLM-based assistants typically proceed with a plausible interpretation, leaving misalignments to be discovered only after the output is inspected or executed. This behavior can trigger costly trial-and-error prompt revisions and may be amplified by sycophantic tendencies that reinforce incorrect assumptions. We present a progressive disambiguation framework that resolves prompt uncertainty before producing a final answer. The framework separates intent clarification from solution synthesis by guiding users through a structured pre-generation dialogue that elicits missing constraints, compares alternative interpretations, and illustrates consequences via targeted input-output examples and representative edge cases. In addition, it performs incremental constraint consistency checks that flag implausible or conflicting user-provided assumptions and request explicit confirmation before generation. After incompatible interpretations are pruned and constraints are validated, the system generates a single, intent-aligned solution. Experiments on a diverse benchmark covering coding, data analysis, and creative writing show that our approach improves output accuracy, reduces corrective iterations and overall user effort, and achieves competitive end-to-end resolution time and higher user satisfaction compared to one-shot and reactive clarification baselines.

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.

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 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.

The automatic detection of anomalies in medical images is a significant challenge in the assisted diagnosis of neurodegenerative diseases such as Alzheimer's. This paper presents an anomaly detection model based on Transformers for the analysis of brain magnetic resonance images. The proposed architecture combines a Vision Transformer as an encoder with a memory bank module that allows modeling the distribution of healthy brains and detecting deviations through reconstruction error. The model is trained using a one-class learning approach, using only images considered normal, with the aim of learning the representation of normality and automatically flagging atypical structural patterns. To adapt volumetric studies to the architecture, a preprocessing procedure is designed that transforms three-dimensional information into a two-dimensional representation compatible with the model. The results obtained demonstrate a solid ability to characterize normality and generate reliable predictions, confirming the viability of Transformer-based architectures for unsupervised anomaly detection in neuroimaging. This approach lays the foundation for future extensions in clinical settings and other medical imaging applications.

Accurate and adaptive time-frequency representation is essential for analyzing nonstationary signals in critical applications, such as epileptic seizure prediction utilizing electroencephalogram (EEG) data. However, existing deep learning approaches often suffer from two major limitations: the reliance on fixed, nonadaptive feature extraction methods and the lack of model interpretability. To bridge these gaps, we propose Adaptive Discrete Wavelet Convolution (ADWConv), a general-purpose trainable multi-level wavelet decomposition front-end, with validation in an end-to-end seizure prediction framework. ADWConv parameterizes the low-frequency and high-frequency filters within each decomposition level and decouples their kernel shapes from level-specific scaling factors, enabling task-related adaptive weighting of frequency bands. This end-to-end learning is guided by a joint optimization strategy, which incorporates a regularization loss with wavelet priors to ensure the learned filters are both discriminative and physically interpretable. Extensive experiments were conducted on the public CHB-MIT database and the private SH-SDU database. On the CHB-MIT database, the ADWConv-based model attained an event-based sensitivity of 98.65%, while maintaining a false prediction rate (FPR) of only 0.014/h. Comparable results were also demonstrated on the SH-SDU database, attaining 96.53% event-based sensitivity with 0.028/h FPR. Ablation and visualization analyses confirm that learning the wavelet kernels is the primary driver of performance gains and supports task-specific adaptation of discriminative frequency bands. Overall, ADWConv improves the feature-level interpretability of a deep learning framework for seizure prediction while maintaining competitive accuracy and robustness.

Epilepsy manifests as a chronic neurological condition marked by recurrent seizures. Recent advances in computational analysis of Electroencephalography (EEG) signals have enabled new possibilities for identifying ictal events in extended recordings. This work develops a novel deep learning architecture that simultaneously resolves two fundamental challenges in automated seizure detection: comprehensive feature representation and class distribution imbalance. First, a multibranch neural network structure is proposed to process EEG signals across varying spectral and temporal resolutions. Then, an attention-based feature refinement mechanism is utilized to automatically emphasize clinically relevant signal characteristics. Finally, a modified loss function is leveraged to incorporate class-specific margin adjustments to handle data imbalance scenarios. Analysis of scalp EEG recordings yields detection accuracy with 96.06% sensitivity and 98.50% specificity, and false detection rate (FDR) is maintained at a low level of 0.34 events per hour. When applied to intracranial EEG data, the algorithm demonstrates similar efficacy (95.90% sensitivity, 98.65% specificity) with further reduced false detections (0.18/h). The consistent efficacy validated on diverse EEG modalities (scalp and intracranial) supports its clinical utility as a practical diagnostic tool.