Volume 05 · Issue 04 · 2025

Large language models (LLMs), trained on vast amounts of textual data, have demonstrated strong capabilities in natural language understanding and generation. In the medical field, LLMs are increasingly applied across various domains such as disease screening, diagnostic assistance, and health management, playing a key role in advancing intelligent healthcare. In recent years, China has actively promoted the integration of artificial intelligence (AI) with healthcare through a series of policies that support enterprises in making breakthroughs in key technologies such as medical LLMs and multimodal data integration. Concurrently, efforts have accelerated the deployment of AI in applications such as health management and precision medicine to gradually establish a full-cycle intelligent healthcare system encompassing prevention, diagnosis, treatment, and rehabilitation. However, the rapid deployment of LLMs in healthcare has highlighted the lack of standardized evaluation criteria and consistent methodologies. To address this, this expert consensus focuses on establishing a retrospective evaluation framework tailored to medical applications. By integrating scientific evaluation metrics, standards, and procedures, the framework provides clear and actionable guidance for model evaluators, developers, and end users. It aims to unify assessment practices, enhance the scientific rigor and comparability of evaluations, and ensure the safe and effective use of LLMs in healthcare, ultimately supporting the high-quality development of AI-powered medical services.

Robotic surgery is transforming gynecologic care through continuous technological innovation. Its key advantages include three-dimensional visualization, enhanced dexterity, and improved precision. Furthermore, the integration of artificial intelligence (AI) enables tailored treatment strategies. Robotic platforms have evolved from multi-port systems to minimally invasive single-port designs, expanding indications and improving patient outcomes. Gynecologic applications, particularly those using the vaginal natural orifice (e.g., vNOTES), have broadened owing to the anatomical benefits of this approach. However, challenges such as high costs, limited insurance coverage, and a steep learning curve hinder its widespread adoption. Addressing these barriers requires domestic technological advancement, standardized training, and policy support. Interdisciplinary integration is another major frontier with 5G-based telesurgery, AI-assisted decision-making, and multidisciplinary collaboration enhancing surgical planning and execution. Continued innovation is essential to reduce costs, extend access, and achieve the goals of precision, minimal invasiveness, and equitable care—ultimately providing safer, more efficient gynecologic surgery for diverse populations.

Evidence-based medicine (EBM) faces inherent challenges in bridging population-based evidence with personalized medical needs. The rapid advancement in artificial intelligence (AI) offers unprecedented opportunities to transform this paradigm. However, applications without theoretical guidance pose risks to the application, regulation, and orderly development of AI technologies such as large language models (LLMs). This study proposes a novel L0-L5 evolutionary framework to systematically guide the integration of LLMs into evidence-based clinical decision-making. The framework delineates a progressive path from current EBM practices (L0) through AI-assisted evidence retrieval (L1), accelerated evidence synthesis (L2), real-world data analysis (L3), and digital twin-based personalized evidence (L4), to generative model-driven virtual evidence creation (L5). Each level represents increasing capabilities in addressing the core tensions between evidence timeliness, personalization resolution, and decision transparency. This framework offers a structured approach to harness the transformative potential of LLMs while preserving the fundamental principles of EBM, ultimately enabling truly personalized precision medicine grounded in robust evidence.

China’s healthcare reform faces significant hurdles like inefficient medical insurance fund utilization, imbalanced medical resource distribution, and limited innovation in biopharmaceuticals, necessitating smarter technological interventions. This article assesses the impact of smart healthcare in China’s reform agenda. Innovative payment methods like Diagnosis-Related Group (DRG) and Disease Group Payment (DIP), bolstered by big data, have reduced patient burdens. Digitization in medical services has streamlined processes, improved patient experiences, and tackled regional resource disparities. Technologies such as artificial intelligence have accelerated drug development, boosting efficiency and precision. Yet, smart healthcare encounters challenges. To address these, the article suggests enhancing top-level design for technology standards, ensuring secure data sharing, advancing health technology assessments, and nurturing skilled personnel in smart technology.

Cardiac magnetic resonance imaging (CMR) has become a routine and primary tool for the clinical assessment of cardiovascular system function and structure. Segmentation of the left ventricle (LV) in CMR images is an important step in calculating clinical indicators such as ventricular volume, LV mass, LV wall thickness, and ejection fraction, and analyzing abnormalities in LV wall motion.

This paper proposed a level set method with a curvature converging mechanism to ensure the convexity of the segmented left ventricle, which is called Curvature Converging Active Contour (CCAC) model. By utilizing the curvature of the level set contour, the method controled and maintained its convexity, resulting in a final segmented contour that is convex in shape. This ensured that the segmentation results conform to the anatomical structure of the left ventricle, providing strong support for accurate assessment of cardiac structures, detection of myocardial lesions, and other clinical applications.

In the experimental section, we conducted a detailed comparison with other methods. Using the Dice coefficient as the evaluation metric, batch data results were compared and analyzed using box plots. The results show that the CCAC model outperforms other models. Compared to the RSF model, it achieves a higher Dice coefficient, with significantly improved segmentation accuracy and better alignment with anatomical structures. Compared to the DRLSE model, it effectively avoids under-segmentation. Additionally, it further enhances accuracy based on U-Net segmentation results, maintains result convexity, and still delivers good segmentation performance even when deep learning results are suboptimal.

The mean curvature and curvature convergence mechanism may effectively address the issue of maintaining convexity in left ventricular segmentation. This feature could be used for left ventricular segmentation, aiding doctors in evaluating cardiac structures, predicting disease progression, and assessing potential risks.

Magnetic resonance imaging (MRI) brain tissue segmentation is essential for the diagnosis and treatment of neurological diseases; however, intensity inhomogeneity poses a significant challenge, particularly in data-limited scenarios. This study aimed to explore an algorithm designed to robustly address intensity inhomogeneity for brain MRI tissue segmentation under dataset constraints.

We propose a two-stage framework that leverages data-driven and knowledge-driven approaches. Initially, a data-driven model was employed for skull-stripping, where a lightweight module, named multi-view dilated convolution attention (MDCA), is integrated into skip connections. The MDCA module eliminates the effect of intensity inhomogeneity enriched in shallow features at multiple scales, thus avoiding the negative impact on deeper abstract features. Furthermore, we introduced multiplicative and additive intrinsic components optimization (MAICO) algorithm, which decomposes MRI images into their real anatomical structures, multiplicative and additive bias fields, and zero-mean Gaussian noise, thus enabling precise anatomical segmentation. Experiments on MRBrainS13 and MRBrainS18 public datasets involved the random introduction of intensity inhomogeneity to generate training, validation, and testing sets with 60%, 20%, and 20% splits, respectively. Segmentation performance was measured using Dice coefficients and compared to methods such as MICO, FSL, and UNet. An ablation study further validated the efficacy of the MDCA module.

Our approach improved MRI brain tissue segmentation accuracy, achieving a mean Dice coefficient of 0.7733 across tissue types. With MDCA and MAICO, it reached 0.8163 for white matter, 0.7402 for gray matter, and 0.7634 for cerebrospinal fluid, outperforming other algorithms. Additionally, MDCA module integration in skip connections yielded a 5% average accuracy boost.

This study effectively combined knowledge-driven and data-driven techniques to enhance MRI brain segmentation stability and accuracy, thereby demonstrating strong potential for clinical application in managing intensity inhomogeneity in data-constrained settings.

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that leads to cognitive decline and memory impairment, posing a public health concern in aging populations. Early and accurate detection of AD using non-invasive imaging biomarkers remains a critical clinical need for timely intervention and disease management. This study aimed to develop an advanced artificial intelligence (AI)-based diagnostic framework, ViTranZheimer, that leverages video vision transformers to analyze magnetic resonance imaging (MRI) and improve AD classification accuracy.

This study presents "ViTranZheimer," an AD diagnosis approach that leverages video transformers to analyze MRI volumes. Our proposed deep learning framework aimed to improve the accuracy and sensitivity of AD diagnosis, thereby equipping clinicians with a tool for early detection and intervention. We exploited the temporal dependencies between slices by treating the MRI volumes as videos to capture intricate structural relationships. We evaluated ViTranZheimer on the publicly available Alzheimer’s Disease Neuroimaging Initiative: complete 3Yr 3T data collection, which includes 351 T1-weighted MRI scans categorized into normal controls (NC = 129), mild cognitive impairment (MCI = 145), and AD = 77 groups. Each MRI volume was preprocessed using spatial normalization and skull stripping, and modeled as a video sequence for input to a video vision transformer. The model was trained from scratch using 10-fold stratified cross-validation and optimized with the Adam optimizer over 500 epochs. Classification performance was evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Statistical comparison was conducted using the Wilcoxon signed-rank test against 2 baseline models: a convolutional neural network with bidirectional long short-term memory (CNN-BiLSTM) and vision transformer with bidirectional long short-term memory (ViT-BiLSTM).

The proposed ViTranZheimer model achieved 98.6% accuracy in classifying NC, MCI, and AD cases, outperforming CNN-BiLSTM (96.5%) and ViT-BiLSTM (97.5%). It also attained superior precision, recall, F1-score (all 0.97), and an AUC of 0.99. Performance differences were statistically significant based on the Wilcoxon signed-rank test (P < 0.05).

ViTranZheimer demonstrated strong potential for accurate and early AD diagnosis using non-invasive MRI data. By leveraging video vision transformers, the model provides a promising tool for clinical decision support in neurodegenerative disease detection.

Patients who experience acute hospitalization face a risk of suffering adverse events, such as delirium, pressure ulcers, or pain. This risk gets aggravated in individuals with multimorbidity. Furthermore, the prevalence of multimorbidity is notably high, and gets even higher for elder people. In addition, the interaction between multiple adverse events can significantly impact mortality. Previous efforts to predict this kind of events have not produced satisfactory results, particularly for older patients with multimorbidity in emergency room settings. Having a clinical prediction rule (CPR) that can accurately predict adverse events in this population is crucial to prevent these events and improve patient outcomes.

This study enrolled patients with multimorbidity who were admitted to an acute care unit from December 2021 to June 2023. The dimensionality of this dataset was reduced from 43 to 10 features through the implementation of a normalization-based ensemble technique, integrating feature selection methods from different categories: filter methods, wrapper methods, and embedded models to ensure robust validation. A stratified k-fold cross-validation was applied to reduce the risk of overfitting caused by the imbalanced distribution of the data set. Once the relevant predictors were identified, the sequential forward selection (SFS) technique was used to determine the optimal subsets of predictors that maximize model accuracy.

The evaluation of the performance of these subsets using different classification algorithms led to the development of a CPR using only the three most relevant predictors. The metrics of different models were compared, and the support vector machine (SVM) model was selected due to its superior area under curve (AUC)-receiver operator characteristic (ROC) (0.93) and better handling of class unbalancing and rest of parameters (accuracy 0.91, precision and recall 0.83, and specificity 0.94). To facilitate the application of this prediction rule, a web application that streamlines the detection, classification, and prediction processes of these outcomes was developed.

The proposed model may achieve high accuracy and stability by requiring fever events to predictadverse outcomes in patients with multimorbidity in emergency settings compared with conventional methods.

To address the limitations of conventional surveillance systems in providing real-time predictions, we aimed to develop and validate a multi-output least absolute shrinkage and selection operator (LASSO) model integrating web-based search data with traditional surveillance for high-resolution influenza forecasting across China.

We constructed a multi-output LASSO regression model by incorporating Baidu search (8 keywords) and regional influenza surveillance (2012-2023) data covering 31 provinces and 27 cities in China. The model was trained using 2013-2022 data (n = 30,160) and validated using 2023 data (n = 3,074). Comprehensive feature engineering incorporated temporal offsets, trend slopes, seasonal components, and geographical neighborhood effects. Model performance was assessed using R², root mean squared error, and mean absolute error.

For one-week forecasts, the model achieved an R² of 0.967 for influenza-positive rates and maintained robust performance across viral subtypes (R² = 0.953 for influenza B, 0.929 for H1N1, and 0.918 for H3N2). Two-week forecasts retained substantial accuracy (R² = 0.752-0.892) for viral indicators. Influenza-like illness predictions showed moderate accuracy (R² = 0.799) for 1-week forecasts. Regional validation in Chongqing demonstrated consistent performance (R² ≥ 0.850) across all indicators for one-week predictions.

The multi-output LASSO model integrating web-based search data with traditional surveillance demonstrated satisfactory performance for influenza forecasting across diverse geographical regions in China. This methodological framework may contribute to the advancement of evidence-based approaches for influenza monitoring and epidemic preparedness.

对小编来说，

对小编来说，