Alzheimer's Disease (AD) is a progressive neurodegenerative disorder primarily affecting the elderly, characterized by gradually worsening memory loss, cognitive impairment, and behavioral changes, ultimately leading to the loss of ability to perform daily activities. As global populations age, AD has emerged as a significant challenge in public health. While there is currently no cure, early diagnosis and intervention can substantially improve patients' quality of life. This study focuses on the temporal cumulative effects during the progression of AD from Cognitively Normal (CN) to Mild Cognitive Impairment (MCI), and then to AD. We propose a predictive model for the progression of Alzheimer's disease that takes into account the temporal cumulative effects, which can effectively capture the role of time accumulation in the development process of the disease. Our model was trained using longitudinal data from Alzheimer's patients to predict disease progression at three future time points. Extensive experiments have shown that the proposed algorithm achieves optimal performance in terms of precision and recall. Over three time periods, the average prediction precision for CN, MCI, and AD is 0.805, 0.812, and 0.812, respectively, while the average recall rate is 0.862, 0.801, and 0.816, respectively.

Dengue fever prevention and control is a critical task in subtropical and tropical countries. Public participation is widespread, with measures such as clearing water from vases and wearing long sleeves and long pants when outdoors to avoid bites. Public health agencies are implementing measures such as widespread insecticide spraying, releasing sterilized male mosquitoes, and distributing testing kits. However, the annual dengue fever infection rate remains unchecked. Dengue fever peaks between May and October and causes flu-like symptoms such as high fever, severe headache, and body aches. In extreme cases, it can lead to bleeding, difficulty breathing, and even death. This case study focuses on leveraging artificial intelligence technology to develop a public-engaged dengue fever prevention and control system, enabling citizens to actively participate in dengue fever prevention and control, both during outdoor activities and at home. The approach involves deploying multiple models trained using the ResNeXt-101 residual network framework on a public dengue fever prevention and AI control platform. The system is designed to instantly upload mosquito photos, identify the mosquito species, and provide alerts indicating the risk of infection and the number of cases in the area, thereby increasing public engagement and improving prevention and control effectiveness. The government, through the platform's integration with the city's extensive dengue monitoring infrastructure, can conduct targeted area disinfection and distribute dengue testing kits after a case occurs, enabling early detection and treatment. Implementing this system minimizes environmental impact, enabling targeted dengue prevention and control.

The integration of electronic health records (EHRs) into clinical research represents a pivotal advance toward improving data quality, operational efficiency, and regulatory compliance. This review systematically explores recent technological progress in automated data extraction and integration from EHRs, with a focus on three core domains: the development of electronic case report forms (eCRFs), dynamic data exchange mechanisms between EHR and electronic data capture (EDC) systems, and automated standardization strategies to achieve interoperability. eCRFs have evolved from static, manually populated templates to dynamic, technology-enhanced tools that streamline data collection and ensure alignment with clinical protocols and regulatory requirements. Emerging systems increasingly leverage artificial intelligence techniques—including large language models and natural language processing—to automate the generation and population of eCRFs from unstructured EHR data. In parallel, dynamic data pipelines enable real-time, accurate data transfer from EHRs to EDCs, improving data fidelity and reducing clinician workload. Furthermore, the adoption of common data models and standardization frameworks supports multi-center interoperability for large-scale translational research. Despite these advances, widespread implementation is impeded by persistent challenges such as data heterogeneity, governance constraints, and limited scalability. Ongoing efforts to address existing technical challenges are essential optimize the reuse of EHR data in clinical research workflows.

This study investigates the evolving dynamics of public attention toward multiple influenza subtypes in China from 2019 to early 2025, using Baidu Search Index (BSI) as a proxy for collective health awareness. Drawing upon official surveillance data from the Chinese Center for Disease Control and Prevention (CDC), we examine the temporal and directional relationships between public search behavior and influenza epidemic indicators across multiple time scales and platforms. Using correlation analysis, lagged cross-correlation, and Granger causality modeling, we identify both synchronous co-movements and predictive lead-lag structures between BSI and surveillance-based indicators. Influenza A, Swine, and Avian Influenza consistently emerge as early signals of rising public concern, with Swine and Avian strains often showing directional influence on broader influenza-related search trends. In contrast, Influenza B remains peripheral, with weaker and less interconnected public attention. Our findings highlight a post-pandemic amplification of influenza awareness during the 2023-2024 season, marked by historically high search volumes and the re-establishment of predictive lag structures, particularly at Lag = 2-3 weeks. Regional disparities are evident: southern provinces exhibited more resilient BSI-CDC coherence during the pandemic, while northern regions showed delayed recovery. These results underscore the utility of search engine data for real-time infodemiological surveillance and early warning. From a policy perspective, integrating BSI into national influenza monitoring frameworks can enhance the timeliness and responsiveness of public health interventions. This study provides actionable insights for designing adaptive, multi-scale digital surveillance systems in the post-pandemic era.

Graves' orbitopathy (GO) is an major ocular complication of thyroid-associated autoimmune diseases. Accurate evaluation of disease activity is crucial for treatment planning and prognosis. However, current scoring systems remain subjective and labor-intensive. This study proposes an automated deep learning framework for objective GO activity classification based on multimodal single-photon emission computed tomography/computed tomography (SPECT/CT) images and extraocular muscle masks (EOM), integrated with uncertainty quantification (UQ). A retrospective dataset of 478 GO patients was collected, and pretrained segmentation models were used to generate EOM masks. SPECT images, CT images, and EOM masks were concatenated as multi-channel inputs to a three-dimensional convolutional neural network (3D-CNN) for disease activity classification. The model outputs both the predicted activity status and a case-wise uncertainty score derived from a Dirichlet distribution. Five-fold cross-validation was performed to evaluate model performance, and the relationship between uncertainty scores and predictive accuracy was analyzed. The proposed model achieved an accuracy of 85.3% and an area under the receiver operating characteristic curve (AUC) of 0.91 on the test set, demonstrating strong generalizability. Prediction accuracy decreased with lower confidence levels, validating the clinical utility of the uncertainty quantification mechanism. Overall, the model enables automated, accurate GO activity staging and case-level uncertainty assessment, supporting safer and more reliable clinical decision-making.

Computational pathology has become an indispensable component of cancer survival analysis, yet state-of-the-art multimodal pipelines still depend heavily on paired whole-slide images (WSIs) and in situ genomic or immunohistochemistry (IHC) profiles. In practice, however, IHC and high-throughput sequencing are costly, tissue-consuming, and frequently unavailable, creating a critical performance gap for cases lacking molecular data. To bridge this gap, we introduce HMSP (Hierarchical Multimodal Survival Prediction), a framework that leverages richly structured pathology reports as a flexible linguistic surrogate for missing genomic information. Concretely, HMSP formulates WSIs, free-text pathology narratives, and (when present) genomic features as a three-level hierarchy of visual, lexical, and molecular semantics. We enforce hierarchical consistency with a Tri-Modal Ranking Contrastive Loss and a Cone-Based Hyperbolic (H-Cone) regulariser, enabling (i) fine-grained alignment between image regions and report phrases and (ii) high-level alignment between report embeddings and genomic biomarkers. Once trained, the report encoder alone can act as a plug-in genomic proxy, allowing robust survival prediction even when molecular assays are absent. Extensive experiments on five public cohorts demonstrate that HMSP not only surpasses previous multimodal baselines under full-data settings, but also maintains state-of-the-art performance when genomic channels are wholly ablated—highlighting the practicality of text-driven surrogates in real-world resource-constrained scenarios. Code and pretrained weights will be released upon publication.

The prediction of thyroid cancer recurrence is a critical task in clinical decision-making, yet traditional machine learning models face significant challenges, particularly around data privacy, model generalization with huge datasets, and interpretability. In healthcare, patient data is sensitive and sharing it across institutions for model training raises privacy concerns. This research addresses these issues by utilizing federated learning (FL), a decentralized machine learning approach that allows institutions to collaboratively train a model while ensuring patient data remains private. FL enables local model training at each institution, with only model updates shared across participants, safeguarding sensitive data. Alongside federated learning, the study incorporates Explainable AI (XAI) techniques to enhance the transparency of predictions, enabling clinicians to interpret and trust the model's decision-making process. By combining multiple machine learning models in an ensemble approach, the research improves the prediction accuracy and robustness of thyroid cancer recurrence, even with limited data. The method is evaluated using a cohort dataset of thyroid cancer patients, with synthetic data augmentation addressing data scarcity. The results demonstrate that the approach outperforms traditional models while addressing critical challenges of data privacy and model interpretability. This proposed model outperforms previous methods, significantly higher than the best result on same dataset from prior work. Additionally, our model shows superior performance even when trained on larger datasets, confirming its generalizability and robustness.

This paper presents the development of sentence-level medical text classifiers by fine-tuning eight pre-trained transformer-based models on the PubMed 20k RCT dataset. The models span both general-purpose and biomedical-specific architectures. To enhance performance and address class imbalance, a composite loss function combining cross-entropy, focal loss, and dice loss was applied during training. The fine-tuned models were trained on PubMed 20k RCT and then applied, without further adaptation, to the MTSamples dataset using balanced and imbalanced test subsets. ClinicalBERT achieved the highest results, reaching 97.15% accuracy and 96.93% F1-score on PubMed 20k RCT, 95.20% accuracy and 95.10% F1-score on the balanced MTSamples subset, and 91.80% accuracy and 90.60% F1-score on the imbalanced subset, indicating strong transferability across structured and unstructured medical texts. These outcomes highlight the effectiveness of domain-specific fine-tuning combined with optimized training strategies in building accurate and adaptable medical sentence classifiers.

Depression is a common mental health disorder that significantly affects the physical and psychological well-being of hundreds of millions of people worldwide. Recent advances in deep learning have driven notable progress in automatic depression detection (ADD) systems. These systems are capable of efficiently analyzing multimodal data, including electroencephalography (EEG), audio, video, and text, to identify signs of depression with high accuracy. Existing reviews have introduced various methods. However, they often lack comprehensive coverage of data sources across all these modalities. This academic review aims to address that gap by examining publicly available datasets used in multimodal machine learning for depression detection. It systematically analyzes the characteristics, strengths, and limitations of datasets involving different modalities, including textual, auditory, visual, and EEG signals. Integrating multiple data modalities holds promise for improving the accuracy and reliability of depression diagnosis. Despite significant progress, several key challenges remain unresolved. These include the heterogeneity of data across modalities, variations in individual symptom expression, and the scarcity of datasets that incorporate more than three modalities.

Segmentation and classification of breast tumors are critical for computer-aided diagnosis (CAD) using breast ultrasound (BUS), a non-invasive modality for early cancer detection. However, BUS images present challenges such as low contrast, speckle noise, irregular morphology, and indistinct boundaries, which hinder automated analysis. Existing multi-task learning frameworks that utilize shared encoders often suffer from inadequate task-specific features and interference between objectives, resulting in suboptimal performance. To address this issue, we present MDFANet, a novel multi-task learning network designed to optimize tumor segmentation and classification in BUS images. The segmentation network incorporates a Feature Enhancement Aggregation (FEA) module, which enriches semantic representations by fusing multi-scale features from both encoder and decoder stages. Moreover, we introduce a distance auxiliary branch to facilitate accurate tumor boundary localization by integrating hierarchical features with spatial contextual cues. To balance the optimization of segmentation and classification objectives, we further employ a sample-adaptive joint loss function that dynamically adjusts task contributions based on the consistency of cross-task predictions. Experimental results on a BUS dataset demonstrate that our method significantly outperforms existing state-of-the-art approaches in both segmentation accuracy and classification reliability.

Helicobacter pylori infection remains a globally prevalent condition linked to peptic ulcers and gastric cancer, necessitating timely and accurate diagnosis. This review analyses a set of relevant studies that apply artificial intelligence methods to the detection and management of H. pylori. For each study, we examined key aspects including the AI technique employed, the type of input data used, reported performance, and stated limitations. We found that most approaches focused on early detection, particularly through medical imaging and deep learning models. While many systems achieved accuracy comparable to that of clinical experts, limitations were frequent, including small sample sizes, lack of external validation, and reduced effectiveness in post-eradication cases. Only a minority of works addressed explainability or assessed model performance across diverse populations. These findings reveal both the promise and the current barriers of AI-based tools for H. pylori, emphasizing the need for more robust, generalizable, and interpretable solutions.

The success rate of ablation in persistent atrial fibrillation (persAF) is still suboptimal. Machine learning (ML) with the help of wavelet scattering transform (WST) coefficients might enhance the ablation outcomes. 3206 non-contact electrograms (EGMs) were collected before and after ablation procedures using a balloon catheter from 10 patients with persAF. EGM signals were categorized into two classes: 1490 EGMs were labelled as positive ablation responses (AF termination or AF cycle length (AFCL) increased (≥10msec), and 1716 EGMs were labelled as negative responses (AFCL increase (<10msec)) to catheter ablation. WST coefficients were extracted after applying QRST subtraction from each EGM to remove the effect of far-field ventricular activity. These coefficients were fed into dimensionality reduction techniques. The original features (coefficients) and the transferred features were used to train and test 10 machine learning models using leave-one-patient-out 10-fold cross-validation (LOPOCV) method. The best scenario involved using the Morlet wavelet function in WST, PCA, and the decision tree (DT) classifier. The overall accuracy, sensitivity, specificity, precision, F1-score, AUROC, and balanced accuracy ranged between 75% and 82% across the 10-fold CV. We conclude that the WST coefficients, with the help of ML models, played an important role in predicting the responses of ablating the EGMs and their effect on AF termination and CL changes. A comparison was made between the proposed method and our previous work, showing the superiority of this approach.

Vedolizumab, a monoclonal antibody that targets the α4β7 integrin, is widely used to treat ulcerative colitis and Crohn's disease by blocking lymphocyte trafficking to the intestine. However, its precise mechanism of action in human tissues remains incompletely understood. A recent immunology study revealed that the efficacy of vedolizumab is not associated with T cell exclusion, but rather with the depletion of CD1c+ dendritic cells (cDC2) in the colon. We investigated whether machine learning models trained on bulk RNA-seq from colonic biopsies could independently rediscover this finding. Using the GSE234736 dataset (90 transcriptomes from IBD patients before and during treatment), we applied a Random Forest classifier to log-normalized gene expression to distinguish ON VEDO from PRE-treatment samples. The model achieved 72% accuracy and identified dendritic cell–associated genes, including CD207, among the top predictive features. This supports vedolizumab's role in blocking cDC2 gut trafficking and demonstrates that interpretable machine learning applied to public transcriptomic data can reveal therapeutic mechanisms in immune-mediated diseases.

The gut–brain axis represents a vital bidirectional communication network linking the gastrointestinal and central nervous systems. Growing evidence implicates disruptions in gut microbiota composition (dysbiosis) in the development of neurodegenerative conditions such as Alzheimer's disease, Parkinson's disease, autism spectrum disorder, and epilepsy. In this study, we introduce a computational agent-based model grounded in Lotka–Volterra dynamics to simulate interactions between key gut microbial populations and neuronal health. The model captures predator–prey-like dynamics, where microbiota act as microbial predators and neurons as vulnerable prey, while also incorporating the effects of metabolite accumulation and inflammatory signalling. Through sensitivity analysis, we identify biological parameters that most strongly influence system behaviour over time. Our findings provide mechanistic insights into how microbiome-driven inflammation may contribute to neurodegeneration and establish a systems-level framework for advancing gut–brain research in computational neuroscience and medical AI.

COVID-19 has exposed fundamental limitations of the capital-intensive, hospital-centred healthcare model. Ensuring equitable access to high-quality preventative and clinical care requires a collaborative approach enabled by digital technologies, artificial intelligence (AI), and data science. This paper examines the diverse ways in which these technologies are transforming healthcare and provides an overview of the current state of the art in their applications. The integration of AI and digital platforms is driving a paradigm shift in healthcare delivery and medicine. A cloud-based healthcare model is proposed, connecting distributed services across specialized centers and leveraging cloud computing to manage and deliver care efficiently, offering significant advantages over traditional hospital-centric systems. Key challenges in adopting AI and digital technologies, as well as implementing cloud-based hospitals, are discussed. The study also offers guidance for governments, policymakers, and investors in developing strategies and investments that support the digital transformation of healthcare, emphasizing that the successful evolution of healthcare service delivery requires the coordinated effort of all stakeholders. The cloud hospital model could promote the implementation of Sustainable Development Goal 3 (SDG 3): Good Health and Well-Being.

The international aviation network has facilitated global population movement. This has not only significantly contributed to the cross-border spread of infectious diseases, but has also had substantial economic consequences worldwide during the pandemic. Thus, understanding the dynamic relationship between mobility, epidemic spread, and economic trends is crucial for informing more balanced and adaptive public health and economic policies. Existing models have a tendency to prioritize infection trends, neglecting to account for the nonlinear interplay between population mobility, epidemic spread, and economic fluctuations. We propose a hybrid model architecture that combines the graph sample and aggregate (GraphSAGE) algorithm for modeling spatial dependencies driven by population flow and Long Short-Term Memory (LSTM) for capturing temporal patterns of infections and economic indicators. The simulation results demonstrate the model's capacity to predict infection dynamics and economic trends accurately. These findings underscore the value of data-driven strategies in enabling timely public health responses and in supporting the formulation of coordinated cross-border mobility policies during global health emergencies.

Document-level Biomedical Relation Extraction (BioRE) is a critical task in biomedical text mining, which automatically extracts relationships between entities to support downstream applications such as knowledge graph construction and question answering. Recent approaches often employ Bidirectional Encoder Representations from Transformers (BERT), which require fine-tuning on domain-specific annotated datasets and exhibit limited generalization across datasets and relation types. Advances in Large Language Models (LLMs) offer new opportunities for BioRE. This work evaluates multiple LLMs—both with and without fine-tuning—on contemporary BioRE datasets. Results indicate that LLMs prompted to directly output relations fail to match state-of-the-art (SOTA) BERT models, highlighting the need for more competitive and efficient LLM-based BioRE methodologies.

Ovarian cancer ranks as the third most common gynecological malignancy. It presents significant global health challenges due to its high mortality rates, mainly resulting from late-stage diagnoses and inadequate early detection methods. Existing diagnostic techniques primarily focus on identifying morphological changes in the ovaries, which often overlook early cellular-level alterations, resulting in delayed diagnoses. Consequently, early detection is crucial for preventing cancer progression and improving patient survival rates. Aptamers, small nucleotide sequences that specifically bind to target proteins, represent a novel approach for identifying tumor cells. Recent advances include the design of a bispecific aptamer capable of detecting exosomes released by ovarian cancer cells. We hypothesize that these aptamers can specifically bind to the receptor surface and can be used to detect ovarian cancer. The current research utilized computational methods to investigate the binding capabilities of aptamers to surface receptors (EpCAM and CD24) present on exosomes. By predicting the aptamers secondary and tertiary structures, we performed molecular docking simulations to assess their interactions with the target receptors. Our findings indicate that the aptamers Apt1 and Apt2 exhibit strong binding affinities, warranting their selection for further evaluation. The docked structures were further validated by using the GrASP web server, which is a machine learning based method to enhance the accuracy of our predictions. This research contributes to identifying the most effective aptamers for binding to exosome surfaces, which may potentially advance ovarian cancer screening methods.

Accurate prediction of virus–human protein–protein interactions (PPIs) is critical for understanding pathogen–host mechanisms, identifying therapeutic targets, and accelerating antiviral drug discovery. Recent advances in protein language models (PLMs), such as ESM2, have enabled powerful sequence-based representation learning for PPI prediction. However, these state-of-the-art PLMs are parameter-intensive, making full fine-tuning computationally expensive, time-consuming, and prone to catastrophic forgetting of general protein knowledge. To address these limitations, we explore parameter-efficient adaptation strategies, Low-Rank Adaptation(LoRA), Adapters, and IA3, for virus–human PPI prediction. We further propose Parameter-Efficient Protein Language Model(PE-PLM), a tailored approach that achieves superior predictive accuracy over supervised fine-tuning while significantly reducing trainable parameters and mitigating overfitting on small, class-imbalanced datasets. Compared to the state-of-the-art PPI prediction methods, PE-PLM consistently delivers the highest performance across virus-human PPI benchmark dataset. Comprehensive experiments demonstrate that our parameter-efficient adaptation framework offers a scalable, robust, and cost-effective solution for large-scale virus–human PPI prediction.

Our research introduces a health-aware dispatching framework for large-scale crowd events, focusing on safety and privacy. It combines (i) Disjoint Set Union (DSU) for efficient group management, (ii) Multi-Agent Systems (MAS) for decentralized coordination, and (iii) Ciphertext-Policy Attribute-Based Encryption (CP-ABE) for secure attribute protection. The DSU allows quick group updates, while the MAS policy forms temporary health-based subgroups and seamlessly reintegrates them. The framework balances throughput and safety with a bi-objective dispatch policy, minimizing risks and delays. A JADE simulation with 1,001 agents demonstrates scalability, showing stable monitoring and efficient DSU operations, effectively managing vulnerable subgroups without disrupting flow.