Log anomaly detection (LAD) is of paramount importance to enhance the reliability and stability of software systems. Current state-of-the-art LAD suffers a significant performance degradation when dealing with consistently evolving log events caused by system updates. To build a reliable LAD model under the context of log data evolution, we propose an incremental learning-based method for LAD, namely ilLog, to avoid catastrophic forgetting of previously learned knowledge while continuously updating the model for better detection when processing the evolving log events. In particular, we design a novel entropy-driven sorting algorithm for real log sample replay, which enables the preservation of old knowledge via storing representative samples with discrete sequence features from previous tasks. Additionally, we introduce a Halton-based low discrepancy sequence to better approximate the sliced Cram´ er distance between the probability distributions of two models, thus enhancing the model learning capability. Based on a standard incremental learning protocol setting, we evaluate the newly proposed ilLog method on three publicly available datasets. Experimental results demonstrate that our approach achieves the best performance compared to SOTA LAD methods and models by applying existing IL-based methods in evolving software systems.

Large language models and multimodal foundation models are enabling medical artificial intelligence (AI) systems to move beyond isolated prediction and undertake multistep clinical tasks that require planning, tool use, memory, iterative correction, and coordination among specialized agents. However, the scope of agentic AI in medicine remains unsettled, and current evaluation practices are not yet aligned with the requirements of clinical use. We conducted a scoping review with systematic evidence mapping across five electronic sources, screened 1,649 exportable records, and provisionally included 557 unique studies that met predefined criteria for goal-directed task execution, tool use, interaction with external resources, feedback-based refinement, or multi-agent collaboration. The included studies describe single agents that use external tools, workflows supported by retrieval and external knowledge, multimodal agents, and multi-agent systems applied to medical question answering, image interpretation, electronic health record analysis, drug safety, and clinical trial prediction. The evidence base remains dominated by public benchmarks, simulated settings, retrospective datasets, and small-scale expert evaluation. Process reliability, evidence traceability, uncertainty, safety, workflow impact, and external validity are evaluated less consistently. Clinical translation will depend on clearer definitions, reproducible evaluation, auditable oversight, interoperable system design, and prospective validation in real-world clinical workflows.

Background: With the expanding clinical application of cardiac computed tomography (CT) for myocardial tissue characterization, the accuracy of extracellular volume fraction (ECV) quantification is critically dependent on image quality. This study aimed to compare the performance of deep learning reconstruction (DLR), hybrid iterative reconstruction (HIR), and model-based iterative reconstruction (MBIR) in terms of image quality and the accuracy of ECV quantification in cardiac CT. Methods: Fifty patients with suspected cardiac disease who underwent both cardiac CT and magnetic resonance imaging (MRI) were prospectively enrolled. Non-contrast and delayed-phase CT images were reconstructed using HIR, MBIR, and DLR algorithms, respectively. Visual image noise, beam-hardening artifacts, and display of myocardium and blood pool were evaluated using a four-point scale. CT value, image noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) of six segments of myocardium were measured and calculated. CT-ECV of six segments of myocardium was calculated in subtraction images. Pearson or Spearman correlation and Bland-Altman analysis were used to evaluate the correlation and consistency between CT-ECV and MRI-ECV in six segments of myocardium. Results: DLR significantly decreased image noise and beam-hardening artifacts, increasing SNR and CNR, providing superior visualization of the myocardium and blood pool compared to HIR and MBIR (P<0.001). In ECV quantification, DLR demonstrated a significantly lower mean absolute difference between CT-ECV and MRI-ECV across all six segments of myocardium (P<0.05). Compared with HIR and MBIR, DLR improved the correlation between CT-ECV and MRI-ECV (P<0.001). DLR showed the highest consistency in assessing ECV in all segments of myocardium (bias: -0.183) compared to HIR (bias: 0.390) and MBIR (bias: 0.540). Conclusions: Cardiac CT combined with DLR enhances the myocardial image quality and improves the accuracy of CT-ECV quantification.

Background and Objective: Accurate classification of intestinal polyps is crucial for preventing colorectal cancer but is hindered by visual similarity among subtypes and endoscopic variability. While deep learning aids in diagnosis, single-modal models face efficiency-accuracy trade-offs and ignore pathological semantics. We propose a multimodal framework that integrates endoscopic images with structured pathological descriptions to bridge this gap. Methods: We propose LPA-Tuning CLIP, which incorporates three key innovations: replacing CLIP's instance-level contrastive loss with cross-modal projection matching (CMPM) with ID loss to explicitly optimize intraclass compactness and interclass separation through label-aware image-text similarity matrices; introducing structured clinical semantic templates that encode WHO diagnostic criteria into hierarchical text prompts for consistent pathology annotations; and developing medical-aware augmentation that preserves lesion features while reducing domain shifts. Results: The experimental results demonstrate that our proposed method achieves an accuracy of 85.8% and an F1 score of 0.862 on the internal test set, establishing a new state-of-the-art performance for intestinal polyp classification. Conclusions: This study proposes a multimodal polyp classification paradigm that achieves 85.8% accuracy on three-subtype classification via endoscopic image-pathology text joint representation learning, outperforming unimodal baselines by 8.7% and a multimodal baseline by 4.3%.

Coronary artery disease (CAD) is a leading cause of death worldwide, and electrocardiography (ECG) serves as a crucial diagnostic tool. However, current deep learning methods for automated CAD diagnosis often focus on intra-lead features, overlooking inter-lead relationships. Additionally, their black-box nature limits clinical applicability. To address these issues, we propose a graph neural network (GNN)-based interpretable model for CAD diagnosis. We propose a feature extractor that integrates attention mechanisms with one-dimensional convolutional neural networks (1D CNNs) to effectively capture temporal features within leads over multiple time steps. To model inter-lead dependencies, we construct graph structures based on lead similarity and leverage a graph convolutional network (GCN) to aggregate spatial information across the leads. Furthermore, an interpretability analysis utilizing attention weights highlights critical ECG segments that influence model decisions, offering insights consistent with established medical knowledge. Experimental results on two public datasets demonstrate state-of-the-art performance, achieving an accuracy of 0.995 and an AUC of 1.0 on the INCART dataset, and an accuracy of 0.833 and an AUC of 0.901 on the STC dataset, significantly outperforming existing approaches on both benchmarks. This method provides an effective and interpretable diagnostic tool to support cardiologists in CAD detection and prevention.

Time series data is pervasive across various digital twin applications, where precise forecasting is essential for tasks such as traffic management, energy optimization, weather prediction, and healthcare. However, the non-stationary nature of time series data poses significant challenges for accurate forecasting, particularly in long-term scenarios. Recent models have attempted to incorporate frequency domain information to address these challenges. Nevertheless, they often introduce excessive computational burdens or risk losing vital time domain information. To tackle these issues, we propose a novel model, Temporal Wavelet-Refine Network (TWRNet), designed to mitigate the adverse effects of non-stationary components on forecasting performance. A pivotal component of TWRNet is the Adaptive Multi-Res Wavelet (AMRW) filter. It employs learnable wavelet filters to selectively transmit or attenuate certain components of the time series signal, thereby extracting key temporal patterns. Our method achieves state-of-the-art performance while maintaining linear time complexity, making it highly suitable for digital twin applications in industrial scenarios where real-time efficiency and robustness to non-stationarity are critical. Finally, we conduct extensive experiments on eight time series forecasting benchmarks, and the results demonstrate that our method outperforms state-of-the-art methods in terms of effectiveness and flexibility.

Log anomaly detection (LAD) is essential to ensure the safe and stable operation of Cyber-physical systems. Although current LAD methods exhibit significant potential in addressing challenges posed by unstable log events and temporal sequence patterns, their limitations in detection efficiency and generalization ability present a formidable challenge when dealing with evolving systems. To construct a real-time and reliable online LAD model, we propose OMLog, a semi-supervised online meta-learning method, to effectively tackle the distribution shift issue caused by changes in log event types and frequencies. Specifically, we introduce a maximum mean discrepancy-based distribution shift detection method to identify distribution changes in unseen log sequences. Depending on the identified distribution gap, the method can automatically trigger online fine-grained detection or offline fast inference. Furthermore, we design an online learning mechanism based on meta-learning, which can effectively learn the highly repetitive patterns of log sequences in the feature space, thereby enhancing the generalization ability of the model to evolving data. Extensive experiments conducted on two publicly available log datasets, HDFS and BGL, validate the effectiveness of the OMLog approach. When trained using only normal log sequences, the proposed approach achieves the F1-Score of 93.7% and 64.9%, respectively, surpassing the performance of the state-of-the-art (SOTA) LAD methods and demonstrating superior detection efficiency.

Data-driven modeling is a core enabling technology of UAV digital twin systems, playing a crucial role in flight state prediction, trajectory optimization, and real-time monitoring tasks. By simulating UAV flight trajectories, states, and key control parameters in the digital twin system, the digital twin provides auxiliary decision-making support, optimizes scheduling, and reduces wear and tear on physical equipment. However, in the early stages of system deployment, insufficient accumulation of historical data, along with sparse and incomplete sensor data, limits the integrity of system modeling and the accuracy of physical-world predictions. To address this issue, this paper proposes a diffusion model-based time-series data synthesis method to generate additional samples that align with real data distributions, enhancing the early-stage modeling capability of UAV digital twin systems. This approach enables model training under data-scarce conditions, improving the stability and generalization of flight predictions. Experimental results demonstrate that the proposed method offers significant advantages in real-time prediction and optimized scheduling of UAV digital twin systems, not only accelerating the construction of the digital twin system but also effectively enhancing prediction accuracy and overall system reliability.

Log-based anomaly detection (LAD) is one of the dominant approaches to improving the reliability and security of software systems. Presently, despite the efficacy demonstrated by state-of-the-art LAD approaches in processing static log events, their performance significantly degrades when confronting changes of log event types from system updates. To construct a reliable LAD model that could adapt well to the evolution of log data, we propose a method grounded in semi-supervised domain adaptation on the rationale of incremental log anomaly detection dubbed as SSDALog, which dynamically updates the model utilizing limited labeled samples to reconcile distributional shifts between evolving and historical data. Specifically, the proposed approach addresses the issue through two primary mechanisms: (i) creation of a cross-domain mixup algorithm, which computes the feature salience of log discrete sequences through occlusion strategy, thus enhancing the adaptability of the model to unknown patterns by mixing evolving features; and (ii) design of an incremental semi-supervised domain adaptation training framework based on noisy label learning to obtain a robust feature extractor, thus improving the generalization ability of the detection model. We empirically assess the efficacy of the SSDALog approach across two publicly available datasets. The experimental results show that our method outperforms the SOTA LAD approach, particularly for evolving systems.

Named entity recognition (NER) is a critical task in natural language processing. It extracts entity information such as person, location, and organization by predicting various categories of label types and entity spans in text. Nowadays, NER has achieved good recognition results in English text by machine learning. However, satisfactory recognition results cannot be achieved when processing text in Japanese, due to the diversity of the text composition and the particularity of the language itself. Compared with English text, which different words are marked by spaces, there is no clear separation mark between two words in Japanese. Simultaneously, Japanese text includes three types of representation methods, which is different from English text which only consists of English alphabet. In order to solve the above problems, a feature integration network with BERT called FINB is introduced in this paper based on multi-feature integration, which can integrate pronunciation features and glyph features of Japanese into the model to obtain more semantic information. The experiments for verification are conducted on the Kyoto University Web Document Leads Corpus called KWDLC and the Japanese Wikipedia dataset, which both prove that the proposed method can improve the recognition of named entities in Japanese effectively.

Hypertension is a core risk factor for cardiovascular and cerebrovascular events, and its early and accurate identification is clinically important for curbing target organ damage. Traditional cuff-based blood pressure detection methods have limitations such as large bias in occasional measurements and inability to reflect cardiac electrophysiological changes caused by long-term blood pressure loading, while existing electrocardiogram (ECG) diagnostic models are limited by insufficient feature extraction capability and poor noise robustness, which make it difficult to satisfy the demand for accurate clinical screening. In this paper, we propose a hypertension classification framework HDNet based on deep residual network and hybrid attention mechanism, which breaks through the bottleneck of existing technology through multi-level feature fusion and dynamic enhancement mechanism. The model is innovative in three aspects: (1) constructing a multi-scale feature extraction network, based on the improved residual architecture, synchronously capturing the local details and global rhythm features of ECG images through multilevel convolution and skip connections, and dynamically enhancing the pathologically sensitive regions by combining with the channel-spatial attention mechanism, so as to improve the detection ability of small electrophysiological changes in early hypertension; (2) Introducing a cascading dual-path attention module, which strengthens the features of pathologically sensitive areas such as left ventricular high-voltage region through the spatial path, and suppresses the lead displacement noise by the channel path to achieve dynamic optimization of feature expression; (3) Designing a depth-optimized classifier: adopting a global feature compression and a multi-level regularization strategy to achieve high-precision classification under limited medical data, and effectively avoiding overfitting. Experiments show that the model achieves an AUC value of 0.80 on the independent test set with a specificity of 81.0%. This result provides a highly robust tool for early noninvasive screening of hypertension.

Coronary artery disease (CAD), also known as coronary atherosclerotic heart disease, is a common cardiac pathology. Although the degree of coronary stenosis can be identified through invasive coronary angiography, it is often difficult to correctly identify many angiograms by the naked eye. Therefore, we propose a novel classification model based on graph neural network for coronary artery stenosis. The network defines multi-head attention mechanisms, residual blocks, and graph convolutional networks, which are combined with many layers. Through a series of processing input signals, the model ultimately yields classification results. We conducted experiments using a dataset provided by Zhongshan Hospital affiliated with Dalian University, achieving a recall rate of 0.806 and an F1 score of 0.754. These performances offer certain reference value for doctors in diagnosing coronary diseases, enabling better resolution of issues that are imperceptible in angiograms even to the naked eye, which significantly improves the accuracy of coronary disease recognition.

Clinical diagnosis of Parkinson’s disease(PD) is highly dependent on subjective assessment by experts, which suffers from resource scarcity and outcome heterogeneity. Existing unimodal based machine learning methods are difficult to comprehensively capture the complex motor symptoms of PD due to modal homogeneity and spatio-temporal alignment challenges of heterogeneous data. In this paper, we propose Parkinson’s Heterogeneous Stream Fusion Framework (PHSFF), which for the first time unites event camera gait data and RGB hand video data to achieve high-precision assisted diagnosis through dual-stream network architecture and dynamic decision-level fusion. PHSFF innovatively designs: Firstly, for event stream gait data, a feature extraction module is constructed based on the Spiking Neural Network (SNN) with Temporal-Channel Joint Attention (TCJA). This module utilizes the bio-temporal coding properties of spiking neurons to enhance the gait phase modeling capability. Secondly, for RGB hand videos, a cascade architecture of Graph Attention Network variant 2 (GATv2) with BiLSTM is proposed, which is derived from the spatial relationships of hand joint points and temporal motion to extract tremor and bradykinesia features; Thirdly, a dynamic weighted fusion strategy is designed to integrate complementary information from heterogeneous modalities at the decision-level to avoid the temporal alignment problem. A series of experiments on the self- constructed multimodal dataset PHG-MD have shown that PHSFF is able to achieve better performance.

Parkinson’s disease (PD), a prevalent neurodegenerative disorder, significantly impairs motor functions, necessitating effective rehabilitation management. This paper presents the design of a Parkinson’s rehabilitation management system leveraging multimodal learning and digital twin technology. By integrating data from diverse sources—such as speech signals, hand motion data, and gait videos—this system comprehensively captures PD symptoms, facilitating early detection and accurate diagnosis. Employing deep learning techniques, the system analyzes and fuses multimodal data, enhancing diagnostic precision and providing personalized rehabilitation plans. The digital twin component creates a virtual model of the patient, continuously monitoring motor functions, medication responses, and symptom changes, thus optimizing treatment plans and improving quality of life. Case studies demonstrate the system’s efficacy in speech recognition, tremor identification, and tremor severity assessment, highlighting its role in enhancing patient self-management and rehabilitation outcomes. The proposed system not only advances PD rehabilitation management, but also sets a precedent for intelligent healthcare solutions by integrating multimodal data analysis with digital twin technology.

Colorectal cancer is a malignant tumor that poses a serious threat to human health. Screening for colorectal tumors from abdominal CT images is a common task for radiologists. However, utilizing non-enhanced CT scans without bowel preparation or contrast agents for colorectal cancer screening poses a substantial challenge for clinicians with limited experience. Addressing the issues of poor performance on complex training images, low recall rates, and instability in current methods, we propose the CSAM-YOLO algorithm. Specifically, we construct a dilated convolution sequence to form the D-ELAN convolution layer, thereby expanding the receptive field. Subsequently, the CSAM module infers attention maps along the channel and spatial dimensions separately, and multiplies these attention maps with the input features for feature optimization. This module is then added to the end of the backbone network to better extract target information. Additionally, we introduce Slide-Loss and DyHead to balance samples and improve model performance. Experimental results demonstrate that compared to YOLOv7, the proposed CSAM-YOLO network improves the recall rate by 12.7%, mAP0.5 by 7.4%, and mAP0.5:0.95 by 15.2%, validating the effectiveness of the proposed algorithm.

The electrocardiogram (ECG) is a ubiquitous medical diagnostic tool employed to identify arrhythmias that are characterized by anomalous waveform morphology and erratic intervals. Current ECG analysis methods primarily rely on the feature extraction of single leads or scales, thereby overlooking the critical complementary data obtainable from multiple channels and scales. This paper introduces the Multi-Scale Grid Transformer (MSGformer) network, which extracts spatial features from limb and chest leads and employs a multi-scale grid attention mechanism to capture temporal features. The self-attention mechanism-based multi-lead feature fusion approach leverages diverse leads' perspectives to reflect each lead's heart's comprehensive state and extract unique essential features. Furthermore, MSGformer incorporates a multi-scale grid attention feature extraction strategy that employs multi-head and multi-scale attention mechanisms to extract multi-scale temporal features from two dimensions. The MSGformer network combines these feature extraction strategies, resulting in simultaneous capturing of morphological characteristics across different leads and temporal characteristics within the same lead in ECG. This integration facilitates the effective detection of morphological abnormalities and erratic intervals in cardiac electrical activity. Utilizing the publicly available 2018 China Physiological Signal Challenge (CPSC 2018) and MIT-BIH electrocardiogram datasets, the performance of MSGformer was evaluated and compared to existing ECG classification models. Experimental results demonstrate that MSGformer achieved an F1 score of 0.86, while on the MIT-BIH dataset, it attained accuracy, sensitivity, and positive predictive value of 99.28%, 97.13%, and 97.87%, respectively, outperforming other current models.

Identifying essential proteins is vital for deciphering the intricacies of disease mechanisms and devising efficacious therapeutic strategies. Over the past several decades, a plethora of algorithms have been proposed, aimed at synthesizing topological and biological information to address the complex challenge of essential protein identification. Nevertheless, a critical examination of the current methodologies reveals certain limitations: (1) the aggregation of diverse features in various methods often requires parameter tuning to maintain balance, potentially introducing instability and increasing complexity in practical scenarios; (2) traditional methods commonly combine various features without in-depth mathematical or physical interpretation, possibly falling short of fully revealing the principles behind the observed phenomena. Hence, we propose a new algorithm for essential protein detection, which is named ITBSE. The basic idea behind this method is to reconstruct the PPI network by removing false positive edges and the subsequent allocation of a protein score. This scoring process integrates both the topological attributes of the reconstructed PPI network and biological data, utilizing the computational framework of Shannon entropy for a comprehensive assessment. To evaluate the effectiveness of our method, we conduct the experiments on real PPI networks and compare with 10 popular methods including DC, BC, CC, LID, PR, DMNC, LAC, NC, PeC and esPOS. The comparison results demonstrate that ITBSE is able to achieve better performance than those competing algorithms.

The Transformer has emerged as the predominant model in Natural Language Processing due to its exceptional performance in various sequence modeling tasks, particularly in handling long-term dependencies. However, the traditional absolute and relative position encoding methods, which do not learn from data, tend to ignore the inherent structure of natural language sequences due to the position embedding layer at the input end of the Transformer model. This paper introduces a novel learnable neural Ordinary Differential Equation Position Encoding (ODEPE) method that can implicitly capture the natural position relationships within a sequence without requiring additional position embeddings. ODEPE can model continuous sequences and leverage differential equations to simulate the evolution of position information along the sequence, enabling position information to flow seamlessly between sequences. Additionally, a highly effective recurrent attention framework is proposed, which hybridizes attention with the ODEPE method to improve model performance. Compared to the Transformer-based sequence modeling network, our framework demonstrates a performance improvement of 24.0 points on the WikiText-103 dataset, while also achieving a performance improvement of 1.06 points on the Enwik8 dataset. This corresponds to an improvement of 4.9% and 0.17%, respectively.

Current SNN models typically use only one type of spiking neuron, such as the Leaky Integrate-and-Fire (LIF) neuron. However, we argue that biological systems are more complex and include a diverse range of neuron types. To tackle this, we propose the Multi-Spiking Neural Network (MSNN) model, which incorporates various neuron types, including the widely-used parametric LIF and Integrate-and-Fire neurons. Our experiments show that using multiple neuron types can improve accuracy compared to using only one type. We also introduce a new training algorithm called Average Firing Rate Temporal Efficient Training (AFR-TET) algorithm, which uses the average firing rates of neurons instead of spike times to update network parameters, resulting in more efficient training. We evaluate MSNN on two image classification benchmarks, achieving state-of-the-art performance on both. Our findings demonstrate the potential of MSNN and AFR-TET for improving SNN accuracy and efficiency.

针对患者的多个电子医疗病历中单关键字检索的安全性不足和检测效率低的问题,提出一种区块链下的电子病历(EMRs)聚合密钥可搜索加密方案.采用星际文件系统(IPFS)和云服务器结合存储EMRs和密文,由患者在区块链上设置智能合约和存储交易单,对多文档多关键字的EMRs提出聚合密钥可搜索加密方案,数据请求方通过两次陷门匹配保证数据有针对性且安全的共享.安全性分析和数值模拟实验结果表明,该方案在保证聚合密钥安全和密文安全的前提下,具有较高的计算效率.