EEG signal generation is hindered by challenges such as complex time-frequency structures, the lack of explicit spectral modeling, limited data availability, and limited generalization across subjects and tasks. To address these issues, we propose EEGDTF, a diffusion-based generative framework for synthesizing high-fidelity EEG signals with improved time-frequency modeling. EEGDTF first employs a multi-scale residual encoder to enhance temporal representation learning and training stability. It further introduces a dual-branch encoder-decoder architecture for time-frequency disentanglement: the frequency branch models both periodic and aperiodic components via power spectral parameterization, while the temporal branch captures waveform continuity and long-range dependencies. A frequency-guided cross-attention mechanism integrates both branches effectively. The model is optimized through a joint waveform and spectral loss, enabling stable clean-signal estimation during reverse sampling. Experiments on four benchmark datasets demonstrate that EEGDTF achieves state-of-the-art performance in both time and frequency domains, particularly under cross-subject conditions. These results underscore the model's robustness and generalizability, positioning EEGDTF as a reliable tool for EEG data augmentation and BCI-related applications.

Lung cancer patients frequently experience breakthrough pain episodes, with up to 91% requiring timely intervention. To enable proactive pain management, we propose a hybrid machine learning and large language model pipeline that predicts pain episodes within 48 and 72 hours of hospitalization using both structured and unstructured electronic health record data. A retrospective cohort of 266 inpatients was analyzed, with features including demographics, tumor stage, vital signs, and WHO-tiered analgesic use. The machine learning module captured temporal medication trends, while the large language model interpreted ambiguous dosing records and free-text clinical notes. Integrating these modalities improved sensitivity and interpretability. Our framework achieved an accuracy of 0.876 (48h) and 0.917 (72h), with improvements in sensitivity of 10.6% and 10.7%, respectively, attributable to large language model augmentation. This hybrid approach offers a clinically interpretable and scalable tool for early pain episode forecasting, with potential to enhance treatment precision and optimize resource allocation in oncology care.

Survival prediction using Whole-Slide Images (WSIs) is a cornerstone of precision oncology, yet it remains a formidable challenge due to the massive scale of gigapixel images and the intricate heterogeneity of the tumor microenvironment. While Multiple Instance Learning (MIL) has emerged as a predominant paradigm, existing methods often struggle to balance the capture of fine-grained morphological gradients with the efficient modeling of long-range prognostic dependencies in ultra-long sequences. This study presents MILA-MIL, a novel dual-branch framework designed to bridge the gap between local micro-anatomical nuances and global survival outcomes. To address the inherent directional sensitivity of pathological features, we propose a Pinwheel Convolution (P-Conv) module that introduces a pathology-specific inductive bias for capturing directional morphological gradients. Complementing this, a Mamba-inspired Linear Attention (MILA) branch is developed to facilitate efficient global context modeling with strictly linear complexity, ensuring robust scalability to large-scale clinical datasets. A gated fusion mechanism is further utilized to dynamically integrate these decoupled representations into a discriminative prognostic manifold. Extensive evaluations across six diverse cancer cohorts demonstrate that MILA-MIL consistently achieves state-of-the-art performance, offering superior stability and predictive power over existing MIL aggregators. By synergizing directional morphology with scalable global modeling, this work provides a robust and interpretable solution for computational pathology, with significant potential to enhance clinical decision-making in personalized cancer management.

Functional brain network analysis is crucial for understanding brain operating mechanism, aging, sexual distinction and brain disorders. As a powerful neuroimaging technique, resting-state functional Magnetic ResonanceImaging(rs-fMRI) measuresbloodoxygenlevel dependent (BOLD) signals which reflect the interactions between brain regions. Existing methods representsrs fMRI based functional brain network as a fully connected graph with brain regions as nodes and functional connectivity as edges but neglects the heterogeneity of brain activities. Furthermore, current studies of functional brain network mainly focus on single tasks like age regression, gender classification and disease diagnosis, while these tasks share latent related features. To this end, we pro pose a Multi-Task Heterogeneous Path graph Network (MT HPN) for functional brain network analysis. Considering that the information of brain network mostly contains in edges, we utilize a novel Path-Based Heterogeneous Graph Convolution (PB-HGC) to fuse compact edge features of heterogeneous graph paths and generate global brain representation. Then, Path-Based Cross-Attention Block (PB CAB)is applied to exchange useful information between related tasks and emphasize the exclusive features each task respectively. PB-CAB consists of a unique cross-attention transformer designed for graph algorithm to fuse edge features and highlight crucial paths. We estimate the MT-HPN on the ADHD200 dataset and ADNI dataset for gender and disease classification. Our method shows strong capability in multi-task functional brain network analysis and gender related disease diagnosis.

Accurate biomedical Knowledge Graph (KG) alignment is critical for integrating heterogeneous biomedical information and enabling reliable downstream applications, such as clinical decision support, biomedical search, and personalized healthcare. However, variations in terminologies, KG structures, and semantic granularity across biomedical sources introduce substantial heterogeneity, making the construction of high-quality Similarity Features (SFs) a major challenge. Although the Generative Pre-trained Transformer (GPT) has shown strong potential in capturing nuanced biomedical semantics, each GPT-augmented SF typically reflects only one semantic perspective, and combining multiple SFs in a meaningful manner remains non-trivial because of the expanded symbolic search space and potential feature conflicts. To address these challenges, we propose a novel Tree-based Particle Swarm Optimization with Adaptive Fitness Optimization (T-PSO-AFO) framework for GPT-augmented SF construction. First, a GPT-based semantic feature construction method is introduced to extract expressive and context-aware SFs that better capture biomedical entity equivalence. Then, a tree-based symbolic representation within a PSO framework is developed to explore diverse and complex SF combinations more effectively. Finally, an adaptive fitness landscape optimization mechanism is proposed to dynamically reshape the symbolic search space during evolution, improving the convergence stability and alignment performance. Experiments on OAEI's LargeBio and Disease and Phenotype datasets demonstrate that T-PSO-AFO significantly outperforms state-of-the-art biomedical entity matching approaches, validating its robustness, effectiveness, and scalability in aligning heterogeneous biomedical KGs.

Psychiatric inpatient monitoring generates multimodal data, but privacy constraints and cross-hospital heterogeneity limit centralized learning. We propose FedTFT, a federated Temporal Fusion Transformer for multi-horizon psychiatric risk prediction with horizon-decoupled prediction heads for one-hour, one-day, and one-week forecasting and an area under the receiver operating characteristic curve (AUROC)-weighted server aggregation strategy for non-independent and identically distributed hospital data. Each horizon uses its own linear output head to reduce cross-horizon gradient interference, while local training uses proximal updates. We trained and evaluated the model on 246 patients from three South Korean hospitals without sharing raw records. On the global holdout set, FedTFT achieved 93.9% accuracy, AUROC 0.9054, event-F1 0.8242, and Brier score 0.0680. Under matched federated settings, FedTFT improved event-F1 by 19.58 percentage points(PP) over the best competing federated baseline in event-F1 and improved AUROC by 2.74 pp over the highest-AUROC competing federated baseline, while maintaining the lowest Brier score. Ablations confirmed contributions from both the horizon-decoupled design and the AUROC-weighted aggregation strategy. Gradient SHapley Additive exPlanations (SHAP) analysis identified significant predictors such as treatment time, circadian heart-rate fluctuations, and mobility changes. These findings support accurate, calibrated, and interpretable privacy-preserving psychiatric risk forecasting for proactive intervention.

Driver fatigue detection has become increasingly critical for healthcare 4.0 systems, as it enables real-time monitoring of internal cognitive states to ensure road safety. However, existing methods often suffer from two critical limitations, i.e., insufficient modeling of time-varying dynamics, and ineffective fusion of multi-modal signals due to neglect of intra- and inter-modality dependencies. To address these challenges, we propose an explainable AI (XAI) framework, named LTC-DFD, for multi-modal driving fatigue detection. It is composed of five parallel branches to process distinct physiological modalities, each equipped with a Liquid Time-Constant block to model temporal dynamics using trainable differential equations. A dual-level attention mechanism is introduced, combining channel attention to emphasize salient intra-modal features and token-level attention to capture cross-modal dependencies. The fused representation is then passed through a fully connected regression head to estimate the driver's fatigue level. We evaluate LTC-DFD on the SEED-VIG dataset under a cross-subject protocol. It achieves an accuracy of 96.5%, RMSE of 0.22, and parameter count of only 0.42 M, demonstrating superior performance over existing state-of-the-art (SOTA) methods. In addition, the learned temporal dynamics and attention patterns are consistent with known neurophysiological markers of drowsiness, supporting trustworthy deployment of LTC-DFD in healthcare 4.0 driver-monitoring services.

Joint learning on diverse ultrasonic source scenarios presents a challenge in preserving stable generalization due to the combination of heterogeneity of different sources and the inconsistency of joint learning features. Previous joint learning studies, which are not source-resilient frameworks, may not preserve stable generalization when trained on diverse source scenarios. Furthermore, the limited variations in single-source data and the interference from ultrasound imaging, which are common in ultrasonic source scenarios, further decrease generalization. To address these problems, we proposed a source-resilient joint learning framework consisting of three stages: 1) Source transforming, where our 1-to-N transformation unifies diverse source scenarios for source-resiliency. 2) Our feature enhancement modules model the source-resilient joint learning network, including a manifold-constraint normalization module (MCNM) for addressing heterogeneity by minimizing manifold-based loss, a task-consistent attention module (TCAM) shares the multi-scale features with self-attention to address inconsistency, and an adaptive feature-shifting module (AFSM) for feature-level augmentation to overcome single-source data. 3) Our ultrasound-hybrid linear mapping (USmapping) cascades speckle randomization and mask-guiding Monge-Kantorovitch linear mapping to achieve ultrasonic style randomization for addressing the interference of ultrasonic data. Our framework was evaluated on eight ultrasound datasets from various scanners at multiple centers and surpassed previous comparable studies in both segmentation (DSC$WAvg$ of 75.7%) and classification (AUROC$WAvg$ of 68.8%) tasks. Our framework has the potential to serve as a general framework for enhancing the performance of joint learning under diverse ultrasonic source scenarios.

Prosthetics, exoskeletons, and rehabilitation devices that seamlessly respond to the user's motion intent are essential for improving the quality of life for individuals with physical disabilities. However, existing motion intent recognition systems based on Electromyography (EMG) often experience significant performance degradation over time and limb positions. To address this, this paper proposed a Hierarchical EMG-FMG Fusion (HEFF) framework that incorporates Force Myography (FMG) as a complementary modality to enhance robustness in long-term and multi-position motion intent recognition. The HEFF framework introduced three new strategies: (1) Complementary Pyramid Fusion, a neural network architecture that effectively integrates the complementary characteristics of EMG and FMG; (2) Position Compensation and Pairing, to increase the generalizability of the model by producing augmented training data mimicking the temporal and positional variability; and (3) Feedback-based Baseline Correction, to dynamically counteract FMG baseline drift. The robustness of the system was evaluated through 10 distinct motions performed in 3 different limb positions, retested after an interval of approximately 7 days. Notably, the training data were only collected in a single limb position on the first day. The real-time experiments with 15 subjects showed that the HEFF framework maintained 91.89% classification accuracy in the primary limb position and 85.39% across multiple unseen positions. According to offline analysis, HEFF improved classification accuracy by up to 34.99% across various limb positions compared to the conventional fusion method. These results highlight HEFF's robustness across various conditions, being well-suited for assistive technology applications.

Accurately predicting protein functions is critical for understanding disease mechanisms and discovering potential drug targets. Gene Ontology (GO), with its hierarchical and semantic information, provides valuable context that can be integrated to improve prediction accuracy. Recently, several existing methods have attempted to integrate GO knowledge with protein sequence features for function prediction. However, these methods ignore the fact that GO embeddings should be tailored to proteins to reflect protein-specific functional relevance. To address this limitation, we proposed DPGOK, a deep learning-based method that fused protein-aware GO representations with protein features for function prediction. DPGOK first learns GO semantic representations with a knowledge graph loss and further generates protein-aware GO embeddings under the guidance of protein features. Results show that DPGOK outperforms state-of-the-art methods across all GO domains. Additional experiments demonstrated that DPGOK is capable of discovering hierarchically deeper and more informative functions for target proteins. Ablation studies revealed that the knowledge graph loss we introduced contributes to more stable and semantically coherent GO representations across different domains. Finally, we find that the predictive performance can be further improved when DPGOK is combined with homology-based approaches.