The rapid development of wearable devices and remote monitoring technologies has accelerated the demand for automated analysis of physiological signals. In this process, inevitable signal loss leads to irregular data, which poses challenges to accurate real-time monitoring. However, existing approaches to handling irregular medical data mainly center on electronic health records (EHR). These records are discrete, non-periodic and unsuitable for direct use in physiological signal analysis. Although some techniques have been developed to model irregular data, they often demand high computational resources and fail to capture inherent periodic patterns. To address these limitations, we propose Structural-to-Temporal Cross-Attention (STCA) to fuse structural and temporal information. We also introduce a Self-Adaptive Interpolation (SAI) strategy that explicitly capitalizes on the periodicity of such signals. Extensive experiments demonstrate that our STCA-Net combined with SAI achieves competitive performance comparing with state-of-the-art methods while remaining lightweight.

Knowledge tracing tasks with label imbalance face two challenges: cognitive uncertainty and data sparsity. To tackle them, we propose Uncertainty-aware Dual Contrastive Learning for Knowledge Tracing (UDCL4KT). Essentially, the framework unifies uncertainty awareness and contrastive learning. This design focuses on the reliability of knowledge state estimation. In practice, we model student knowledge states as a Gaussian distribution, where the mean characterizes mastery levels and the covariance reflects cognitive uncertainty. To capture state evolution over time, we design an attention component based on KL-divergence. This component evaluates probabilistic shifts between consecutive states. To better refine representations of student knowledge states from interaction sequences, we leverage prior knowledge to model pairwise difficulty correlations between concepts. A dual-path encoder is additionally introduced to learn both global dependencies and local patterns in interaction sequences. This architecture employs a dual contrastive loss to enhance feature discriminability and mitigate feature redundancy. To alleviate label imbalance, we incorporate a class-balanced cross-entropy loss into the overall objective function. Comprehensive experiments on five public datasets show that UDCL4KT achieves significantly higher prediction accuracy compared with nine state-of-the-art knowledge tracing models. Further analysis examines the relationship between uncertainty scores and student behaviors. Ablation results also confirm the complementary effects of the uncertainty-aware mechanism and the dual contrastive learning components.

Medical time series analysis, particularly for electrocardiogram (ECG) and electroencephalogram (EEG) signals, is essential in modern diagnostics, supporting early detection of conditions such as arrhythmias and epileptic seizures. However, existing approaches often struggle to capture multi-scale periodic patterns and longrange dependencies while meeting real-time processing demands. The envisioned 6G networks, with their terahertz communication and integrated sensing and communication (ISAC) capabilities, will generate vast volumes of high-fidelity physiological data at the network edge. This paradigm shift intensifies the conflict between the computational complexity of advanced AI models and the limited resources of edge devices, creating a critical bottleneck for deploying sophisticated analytics in real-world healthcare scenarios. To overcome these limitations, this paper introduces a 6G-enabled hierarchical contrastive learning framework, referred to as Hierarchical Contrastive Learning for Multi-Scale Medical time series analysis (HCL-MSM), which integrates three core components: a signal-adaptive encoder based on multi-period decomposition and 2D convolution, a patient-level contrastive module enhanced with decomposable multi-scale mixing, and a 6G-edge deployment module optimized via quantization and pruning. The framework effectively models nested physiological rhythms and cross-time dependencies in medical data, while maintaining low-latency operation under resource-constrained edge environments. We evaluated HCL-MSM on multiple clinical datasets under simulated 6G settings. Our framework achieves significant gains in arrhythmia detection, seizure prediction, and neurological monitoring.We evaluated HCL-MSM on multiple clinical datasets under simulated 6G settings. Our framework achieves significant gains in arrhythmia detection (F1-score: 86.39 percent), seizure prediction (Recall: 87.72 percent), and neurological monitoring (Recall: 87.8 percent), outperforming existing state-of- the-art methods.

Predicting herb-target interaction (HTI) plays a pivotal role in uncovering the pharmacological mechanisms underlying Traditional Chinese Medicine (TCM) and accelerating the development of new drugs. Although deep learning methods have demonstrated significant potential in this field, existing approaches still face several challenges, including the difficulty of modeling complex relations among TCM entities, the issue of hidden biases, and poor generalizability to unseen data. These challenges hinder the models’ ability to effectively capture complex interactions among biomedical entities and lead to disappointing performance when predicting newly discovered elements, such as molecules or protein targets. To address these challenges, we first construct a TCM-Target Heterogeneous Information Network (TTHIN) to model the complex relationships and inherent characteristics of heterogeneous entities in TCM. Based on the TTHIN, we propose MRHormer, a multi-scale relation-based heterogeneous graph transformer for HTI prediction. We introduce topology-robust local attention to mitigate hidden biases caused by type-wise feature disparities, and topological perturbations when new nodes are added to the graph. Then, we present a multi-relation global attention to capture high-order features and long-range dependencies, particularly for cold-start nodes in inductive learning scenarios. To facilitate future research, moreover, we construct three datasets tailored for heterogeneous graph-based HTI prediction studies. Experiments on these datasets demonstrate that MRHormer achieves state-of-the-art performance compared to other models under inductive link prediction settings. The model also achieves competitive results on a public dataset, confirming its robustness and generalizability.

The rapid growth of medical data and increasing complexity of clinical decisions highlight the need for advanced technologies in medical question answering. Pre-trained language models perform well in natural language processing but struggle with specialized medical datasets due to insufficient domain-specific knowledge, resulting in hallucinations and factual errors. This paper proposes Pyramid Graph Neural Network Knowledge Distillation (PyGNN-KD), a framework that distills knowledge from a fine-tuned pre-trained language model to enhance a Pyramid Graph Neural Network (PyGNN) for medical question answering. PyGNN-KD builds a joint knowledge subgraph centered on context nodes for efficient multi-hop reasoning, uses task-specific fine-tuning for domain-aligned node features, and integrates shallow and deep layer features via a pyramid network with adaptive gating to address over-smoothing. An annealing distillation strategy optimizes graph neural network learning by assimilating probability distributions from large language models, improving medical semantic understanding. These innovations advance artificial intelligence through knowledge distillation and adaptive features, with applications in clinical decision support. Evaluations on five medical question answering datasets and five language models show a 3.01% overall improvement. Notably, PyGNN-KD achieves 6.27% average gain on the medical subset of the Massive Multitask Language Understanding (MMLU) dataset and 2.74% on the Medical Question Answering for the United States Medical Licensing Examination (MedQA-USMLE) dataset, aiding accurate disease diagnosis and treatment planning.

The convergence of 6G networks and edge computing is set to revolutionize healthcare through applications such as remote surgery, real-time monitoring, and personalized treatment. These systems generate vast amounts of Irregularly sampled Time Series with Missing values (IMTS) from physiological and behavioral sensors, posing significant challenges for real-time processing and forecasting. Existing methods struggle to balance structural modeling with dynamic evolution, relying on biased discrete approximations and complex self attention that limit deployment in real time healthcare. To address these gaps, we introduce Graph neural Ordinary Differential Equation for Forecasting IMTS (GODEFITi), a Graph neural Ordinary Differential Equation model designed to transform IMTS into sparse bipartite graphs. By unifying structural sparse modeling with continuous time Neural ODEs, GODEFITi bypasses imputation biases and captures correlations between channels across uneven intervals. Within an edge enhanced 6G framework, the model enables minimal latency, interpretable, and resilient forecasting. Experiments demonstrate superior accuracy in critical use cases like remote patient monitoring, bridging the gap between complex physiological sensing and deterministic 6G requirements.

Medical image segmentation plays a critical role in clinical diagnostics, treatment planning, disease monitoring, and neurological disorder identification. This article presents a comprehensive review of its systematic development, covering widely used public datasets, representative methods built on the U-Net, Transformer, and SAM architectures, and key evaluation metrics with their differences, followed by an analysis of major challenges from multiple perspectives. Unlike surveys that focus on a single model family or a specific clinical application, this review organizes U-Net-, Transformer-, and SAM-based methods within a unified analytical framework, with a particular focus on their effectiveness in improving segmentation accuracy and efficiency. This work aims to guide future research and support clinical translation of medical image segmentation, with all related resources publicly available in our GitHub repository: https://github.com/andrew-pengyu/Awsome_MedSeg/tree/main.

Multimodal Large Language Models (MLLMs) may memorize sensitive cross-modal information during pretraining. However, existing MLLM unlearning benchmarks rely on synthetic knowledge injection or complete subject-level deletion, which fail to capture realistic, personalized deletion requests that require fine-grained factual control. In this paper, we introduce PPU-Bench, a real-world and fine-tuning-free benchmark for personalized partial unlearning in MLLMs. PPU-Bench contains 24K multimodal and unimodal samples derived from pre-existing knowledge of 500 public figures under three progressively challenging settings: Complete, Selective, and Personalized unlearning. The benchmark evaluates whether methods can remove target knowledge while preserving non-target facts, model utility, and cross-modal consistency. Extensive experiments show that Complete Unlearning often suppresses visual identity rather than factual knowledge, while Selective and Personalized Unlearning expose significant forget–retain trade-offs and challenges in intra-subject factual boundaries. Robustness analysis under cross-image and prompt-based attacks reveals distinct vulnerabilities across different unlearning settings. Motivated by these findings, we propose Boundary-Aware Optimization (BAO), which explicitly models intra-subject forget-retain boundaries. Experimental results on two representative methods demonstrate that BAO can effectively enforce intra-subject factual boundaries.

Knowledge distillation (KD) has achieved remarkable success in model compression, yet its extension to cross-architecture scenarios remains a formidable challenge. This difficulty primarily arises from the severe representation gap and distinct inductive biases inherent in heterogeneous architectures. Existing approaches often struggle, as direct intermediate feature alignment suffers from architecture-specific noise, while logit-based distillation inevitably discards crucial structural information. To elegantly bridge this gap without relying on complex auxiliary modules, we propose Reciprocal Correlation Distillation (RCD). RCD innovatively anchors multi-stage student features to the teacher's pre-classifier manifold, a unified semantic space that effectively filters shallow heterogeneous noise. Upon this shared manifold, RCD tackles feature misalignment through a meticulously designed dual-constraint objective. First, it employs a reciprocal alignment mechanism to explicitly decouple feature channels by minimizing off-diagonal cross-correlations, thereby filtering out architecture-specific redundancy. Simultaneously, it enforces topology consistency by aligning inter-channel Gram matrices, compelling the student to faithfully reconstruct the teacher's semantic co-activation patterns and reasoning logic. Extensive experiments demonstrate that RCD establishes new state-of-the-art benchmarks in both homogeneous and heterogeneous settings, achieving remarkable average accuracy gains of 11.62% on CIFAR-100 and 2.34% on ImageNet-1K over student baselines.

As Internet of Things (IoT) ecosystems expand into critical infrastructure, the accurate identification of attack patterns, known as vulnerability-to-tactic and technique (VTT) mapping, becomes a critical requirement for cybersecurity. However, applying VTT mapping to the IoT domain faces problems: the data sparsity, where fragmented vulnerability reports lack standardized descriptions, and the explainability deficit, where "blackbox" predictions pose risks of physical operational disruption. To address these challenges, we propose VDM-IoT, a framework that synergizes knowledge graphs (KGs) with large language models (LLMs). The framework constructs an IoT-enhanced vulnerability KG (VKG) from MITRE database/National Vulnerability Database (NVD). For each target vulnerability, it 1) employs a context-aware similarity mechanism to compensate for information sparsity by retrieving topologically related neighbors; 2) filters explicit reasoning paths to potential targets; and 3) guides the LLM to generate verifiable mappings with natural language justifications. Evaluations on the BRON-based benchmark and our constructed expert-labeled VTT dataset show that VDM-IoT achieves 35.37% Hit@1 in technique mapping and 98.11% accuracy in tactic mapping.

Threat Intelligence Platforms (TIPs) play an essential role in proactive cybersecurity operations by systematically gathering, analyzing, and disseminating threat data, thereby offering critical insights into the evolving threat landscape. However, the significant variation in TIP quality and the inherent limitations of any single platform underscore the pressing need for a standardized methodology to effectively assess and compare TIP capabilities. In this paper, we propose a novel comprehensive evaluation framework that assesses TIPs from dual perspectives of platform performance and intelligence quality. The framework integrates eight core metrics to systematically measure detection capability, alert uniqueness, consistency, and information depth, while allowing customizable weight adjustment based on specific operational requirements. Furthermore, we introduce TIPRank, a manipulation-resistant ranking mechanism inspired by PageRank. By synthesizing composite quality scores with citation relationships via a weighted directed graph, TIPRank generates stable and well-balanced rankings. We evaluate seven commercial and open-source TIPs using two real-world malicious-IP datasets (768 and 17,120 IPs) derived from attack and honeypot logs to validate their detection capabilities and quantify performance disparities. Robustness analysis, including metric ablation, data sampling, and weight sensitivity experiments, confirms that the proposed framework produces consistent results across varied configurations. The experimental results provide actionable insights for scientifically grounded TIP selection and contribute to strengthening proactive cyber defense.

PURPOSE:In cancer therapy, tumor cell heterogeneity and dynamics influence gene sequencing and immunohistochemical staining. Importantly, patients treated with epidermal growth factor receptor (EGFR)-tyrosine kinase inhibitors (TKIs) have not demonstrated a favorable long-term prognosis. Therefore, this study proposes an integrated framework for artificial intelligence (IFAI) to explore new molecular detection methods. MATERIALS AND METHODS:Our study integrated data from 506 non-small cell lung cancer (NSCLC) patients across three institutions in China and the USA. To fuse radiomics scores and deep network features from both tumors and surrounding tissues, we developed the IFAI with an attention-based DenseNet 121 as the backbone network. We also explored the synergy between IFAI and clinical factors (IFAI-C). Additionally, we gained further insights into the biological mechanisms of IFAI by analyzing patient RNA sequencing data. RESULTS:In independent test data, the IFAI-C demonstrated notable predictive performance, boasting an area under the curve of 0.912 for EGFR, 0.911 for exon 19 deletion (19Del), 0.905 for exon 21 mutation (L858R), 0.911 for T790M, and 0.904 for programmed cell death protein 1 (PD-1) or its ligand 1 (PD-L1). This capability is a crucial complement to traditional methods like gene sequencing and immunohistochemistry. Our analysis revealed that radiomics scores and deep network features in IFAI were significantly associated with EGFR genotypes, drug resistance mutations, and immune molecule expression. Furthermore, these features displayed robust connections with multiple genotypes associated with drug resistance and cancer progression mechanisms. CONCLUSION:IFAI-C introduces a novel method with performance advantages, accompanied by biological analyses demonstrating the extraction of genotypic and immunomolecular information from both tumors and surrounding tissues. This discovery holds potential value in guiding therapeutic decisions for lung cancer.

With advances in multimodal pre-training, more efforts focus on integrating it into recommendation models. Current methods mainly focus on utilizing multimodal pre-training models to obtain multimodal representations of items and designing specific model architectures for downstream tasks. However, these methods often neglect the suitability of multimodal representations for recommendation systems since the pre-training is not conducted on recommendation datasets, making the directly obtained representations potentially suboptimal due to semantic biases from domain discrepancy and noise interference. Furthermore, collaborative information, a key element in recommendation systems, significantly impacts the effectiveness of recommendation models, but existing advanced multimodal pre-training models (e.g., CLIP) are unable to capture the collaborative information of items. To bridge the gap between multimodal pre-training models and recommendation systems, we propose a novel multimodal pre-training framework CLIP-based Pre-training MultiModal (CPMM) item representations model for recommendation. First, the representations of images, text, and IDs are mapped to a new low-dimensional contrastive representation space for alignment and semantic enhancement, ensuring the consistency and robustness of the multimodal contrastive representation (MCR). A contrastive learning approach is designed to regulate the inter-modal distances, mitigating the impact of noise on recommendation performance. Finally, modeling of the first-order similarities of the items is conducted, thereby integrating the collaborative information of the items into the multimodal contrastive representations. Extensive experiments on Amazon benchmark datasets (Beauty, Toys, Tools) validate CPMM's effectiveness across three core recommendation tasks: sequential recommendation, collaborative filtering, and click-through rate prediction.

As a lightweight message transmission protocol, MQTT protocol has been widely used in various application scenarios of the IoT due to its high efficiency and flexibility. With the rapid increase in the number of IoT devices and the diversification of application scenarios, security issues have become particularly important. However, the MQTT protocol’s own security mechanism has security loopholes, such as data tampering during data transmission. This paper focuses on the security of the MQTT protocol, designs and proposes a protocol security mechanism scheme called MQTT-SA, which completes the identity authentication of the communicating parties and ensures that the key generation of both parties is equal and fair, and encrypts the transmitted data. In order to verify whether the protocol is secure, we use formal modeling methods to verify the security of the protocol and verify the protocol model. The results show that the MQTT-SA protocol is secure.

Privacy-Preserving Data Publishing (PPDP) faces significant challenges in dynamic scenarios where evolving datasets require continuous anonymization while maintaining data utility and resisting temporal attacks. This paper introduces an Adaptive Knowledge Inheritance Evolutionary Framework (AKIEF), a novel approach employing an Adaptive Differential Evolution (ADE) strategy tailored for dynamic PPDP under l -diversity constraints. AKIEF integrates several key innovations to enhance optimization efficiency and effectiveness: (1) a Depth-First Search (DFS) combined initialization strategy to improve the quality of the starting population at the initial time step ( t = 1); (2) a knowledge inheritance mechanism that leverages elite solutions from previous time windows to accelerate adaptation and convergence in subsequent steps ( t > 1); and (3) a novel two-dimension adaptive mutation strategy that dynamically and independently adjusts exploration-exploitation balance for generalization and suppression components within the ADE process. Comprehensive experiments are conducted using 16 test configurations derived from a large, real-world healthcare dataset, comparing AKIEF against baseline methods and static state-of-the-art algorithms. The results demonstrate the AKIEF frequently achieves superior data utility, exhibits robustness by consistently finding high-quality solutions where others struggle, and significantly accelerates convergence in dynamic environments. These findings demonstrate the advantages of the AKIEF as an effective and robust solution for the demanding challenges of dynamic PPDP.

In recent years, the incidence of Chronic Kidney Disease (CKD) has been increasing, establishing CKD as one of the leading causes of mortality worldwide. Early detection and treatment of CKD can mitigate kidney damage. However, early-stage CKD often presents no overt symptoms, posing challenges for its prediction. Deep learning have shown promising results in the prediction of many chronic diseases. Currently, many scholars have used deep learning methods to predict CKD and achieved good results. However, scholars often only use the patient’s biochemical test data and ignore the chief complaint data when using deep learning to process patient data. We believe that the chief complaint data can provide additional information for model prediction, which is beneficial for the model’s prediction. Therefore, We developed a hybrid model-TG.Net. The attention mechanism module is designed in this model, which can effectively extract information from chief complaint data and improve the model’s ability to predict the five stages of CKD by combining it with biochemical test data. In this study, the model was trained on a clinical dataset and its predictive performance was evaluated using accuracy, precision, recall, and F1 Score. Compared to baselines, TG.Net exhibits better predictive performance, achieving an accuracy of 94

Context-aware social media recommendation has been important in many applications such as e-commerce and entertainment. However, existing systems consider pre-specified contexts and cannot well handle user preferences, which negatively affects the recommendation quality and efficiency, and causes them not extendable to various applications. In this demo, we design RECARS, the first responsible and extendable context-aware recommender system. RECARS is designed with novel techniques, including efficient data organization over MongoDB and Apache Flink, and effective responsible recommendation generation that supports the system interactions with users. It allows users to perform iteratively refining and explaining the results by active learning and large language model (LLM). We demonstrate the usage of RECARS via YouTube.

Each year, a significant number of people succumb to cardiovascular diseases (CVD), making early and accurate prediction of CVD highly clinically significant. Methods based on electronic health records (EHR) and deep learning are increasingly being adopted for this purpose. However, current CVD prediction methods often rely solely on biochemical test data, lacking comprehensive integration of multimodal information, and face challenges in capturing long-range dependencies within complex EHR data. In this paper, we propose a deep learning approach based on the Chinese BERT model and multi-head self-attention mechanism to enhance the accuracy and reliability of CVD prediction. Specifically, we transfer a pre-trained Chinese BERT model to our task, integrating patient chief complaints, demographic data, vital signs such as blood pressure, pulse, and temperature, as well as textual descriptions of preliminary diagnoses by physicians. Features are extracted using the RoBERTa-wwm-ext-large model, and key features and long-range dependencies within the data are captured using a multi-head self-attention mechanism. We evaluated our model on three real-world clinical datasets comprising data from 12,180 patients. Experimental results demonstrate that our proposed method performs excellently across different datasets and text lengths, achieving an AUC of 0.98 and an accuracy of 92.93

Recent advancements in fully convolutional networks (FCNs) have significantly improved medical image segmentation. Ensemble methods are often used to further enhance performance, with diversity among learners being a critical factor. However, many current approaches focus on diversifying training samples or predictions while overlooking the diversity of internal multi-scale features. This oversight can lead to high correlations among features across different learners, limiting overall effectiveness. Additionally, traditional quantization methods aim to minimize accuracy loss by maintaining a rigid quantization process. This rigidity can eliminate the randomness introduced by quantization, further reducing ensemble diversity and effectiveness. In this paper, we propose a novel approach called Quantization-based Deep Diversified Ensemble (QDD-Ens) for medical image segmentation. Our method enhances the diversity of internal features among ensemble learners through two mechanisms: deep diversified loss, which focuses on feature diversity rather than segmentation accuracy, and deep diversified quantization, which preserves beneficial randomness in quantization process. Furthermore, QDD-Ens facilitates a deeper form of ensemble learning by employing a meta-learner to integrate diversified features at multiple resolution levels from various base learners, which are diversified by two above diversify enhancement mechanisms. Extensive experiments on five public medical image segmentation datasets show that our method significantly improves segmentation accuracy and outperforms existing ensemble techniques. The source code is publicly available to support future research. (https://github.com/JerRuy/QDD-Ens)