Diagnosing specific nutrition deficiencies based solely on clinical symptoms remains challenging in practice. The integration of Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) techniques offers new opportunities for improving nutrient deficiency diagnosis. Most existing RAG approaches retrieve relevant knowledge from a corpus based on semantic similarity between the user query and stored documents, and then guide the LLM to generate responses. However, such methods often overlook the semantic gap between user queries and their corresponding answers, which can limit diagnostic reliability. This study proposes a new RAG framework for LLMs in symptom-based nutrient deficiency diagnosis. Specifically, a Question-Answer (QA) pair knowledge base is constructed from relevant literature, where deficiency-related questions are used as retrieval targets and their corresponding answers are provided as contextual inputs to the LLM. By aligning retrieved content more closely with diagnostic intent, the proposed approach supplies more precise and relevant contextual information for generation. Experimental results show that the proposed QA pair-based RAG method improves diagnostic accuracy, demonstrating its potential for the application of RAG techniques in the field of nutritional diagnosis.

Early assessment of arterial stiffness (AS) is crucial for cardiovascular disease prevention, with photoplethysmography (PPG) emerging as a promising and convenient detection approach. While current research continuously explores novel PPG features and leverages advanced machine learning techniques to enhance AS detection accuracy, this proliferation of features introduces the challenge of the dimensionality curse. Moreover, the variation in robustness among PPG features is often overlooked in feature selection process. To address these challenges, this study proposes a robustness-aware multi-objective feature selection (RA-MOFS) framework that integrates feature robustness into a cost function for jointly optimizing prediction accuracy and cost efficiency. The method incorporates adaptive evolutionary strategies with hybrid search mechanisms to generate a spectrum of cost-accuracy trade-off options, catering to diverse clinical scenarios. Experimental results demonstrate that RA-MOFS achieves superior Pareto front positioning and enhanced solution diversity compared to conventional multi-objective algorithms, showing significant improvements in both convergence and distribution metrics. Notably, the framework identified a representative solution utilizing only six high-robustness features that achieved a satisfactory AS prediction performance (MAE = 0.59 m/s, r = 0.90) while meeting ARTERY Society validation standards (ME = -0.01 m/s, SDE = 0.74 m/s). Overall, this study provides an adaptable framework for PPG feature selection that enables viable AS assessment tools by balancing accuracy and practicality across diverse healthcare settings.

PurposesThis study aimed to develop a machine learning model to predict body mass index (BMI) in adolescents based on readily accessible daily information and to investigate the influence of modifiable factors on BMI changes through model interpretation techniques.MethodsThis study is a one-year prospective cohort study. Baseline data were collected through anthropometric measurements and questionnaires, and BMI were reassessed after 1 year. Six machine learning models were developed to predict BMI. Nested cross-validation (CV) was used for hyperparameter tuning and performance estimation. Predictors were prescreened on the inner-training folds of the nested CV using univariable analyses. Model performance was evaluated using Root Mean Squared Error (RMSE), Mean Squared Error (MSE), Mean Absolute Error (MAE), and coefficient of determination (R2). SHapley Additive exPlanations (SHAP) was used for global and local interpretations of the models.ResultsThe mean BMI of the 1,827 students included in the final analysis increased from 21.18 ± 3.63 kg/m2 at baseline to 21.54 ± 3.59 kg/m2 after 1 year, with an average change of 0.36 ± 1.40 kg/m2. The CatBoost (CB) model demonstrated the best predictive performance. After calibration, it achieved an RMSE of 1.200 [95% confidence interval (CI): 1.101–1.303], MSE of 1.440 (95% CI: 1.211–1.697), MAE of 0.895 (95% CI: 0.818–0.981) and R2 of 0.902 (95% CI: 0.882–0.918). In the SHAP analysis, the top 5 modifiable features at the population level were: level of health literacy, recognize self-weight status correctly, sedentariness duration on weekends, participation in professional sports training, frequency of staying up late.ConclusionThis study developed a BMI prediction model for adolescents using readily accessible daily information. The model accurately predicts BMI values 1 year later and provides both population-level and individual-level interpretability. Compared to existing studies, it offers key advantages, including independence from complex clinical data, the ability to predict continuous BMI values, and strong model interpretability. Our findings provide a promising research tool for screening high-risk adolescents, informing public health prevention and intervention strategies, and supporting personalized clinical interventions.

Recent researches of large language models(LLM), which is pre-trained on massive general-purpose corpora, have achieved breakthroughs in responding human queries. However, these methods face challenges including limited data insufficiency to support extensive pre-training and can not align responses with users' instructions. To address these issues, we introduce a medical instruction dataset, CMedINS, containing six medical instructions derived from actual medical tasks, which effectively fine-tunes LLM in conjunction with other data. Subsequently, We launch our medical model, IIMedGPT, employing an efficient preference alignment method, Direct preference Optimization(DPO). The results show that our final model outperforms existing medical models in medical dialogue.Datsets, Code and model checkpoints will be released upon acceptance.

Diabetic foot (DF), a severe diabetes complication, remains a leading cause of lower-limb amputations, underscoring the urgent need for early-stage detection to enable timely intervention and improve outcomes. However, current approaches often fail to distinguish diabetic patients without foot complications (DM) from those with DF, as single-modality infrared thermography (IRT) struggles with subtle thermal cues and lacks localization of pathological regions. In this paper, we propose DFP-MMNet (Diabetic Foot Progression MultiModal Network), a novel two-stage multimodal framework that addresses this challenge. In the first stage, a deep registration model, guided by corresponding RGB images, accurately localizes lesion Regions of Interest (ROIs) on thermograms. In the second stage, a dual-branch network fuses deep visual features from the ROI (via RegNetY-16GF) with semantic embeddings derived from structured clinical data, encoded using CLIP text branch. Extensive experiments on the collected ITC dataset demonstrate that DFP-MMNet achieves an F1-score of 82.7%, outperforming unimodal baselines by over 15%, offering a robust and accurate solution for early DF recognition.

The convergence of artificial intelligence (AI) and wearable biosensors is revolutionizing personalized healthcare, enabling continuous monitoring, early detection of health issues, which enhances the efficiency of data processing and real-time decision-making. Multimodal Large Language Models (MLLMs) play a pivotal role in this ecosystem by offering advanced capabilities in analyzing complex health data, understanding nuanced health contexts, and generating tailored health recommendations instantaneously. This study provides insights into how machine learning, deep learning algorithms, and MLLM can work together to facilitate the analysis of physiologic data for real-time monitoring and early warning systems as well as complex decision support mechanisms. In addition, the emergence of clinical decision support systems (CDSS) driven by AI and MLLM provides comprehensive recommendations. Looking ahead, the potential convergence of digital people, meta-universes and world models with wearable biosensors presents an innovative vision for personalized health management.

Recent advancements in multimodal large language models (MLLMs) have opened new avenues for video understanding. However, achieving high fidelity in zero-shot video tasks remains challenging. Traditional video processing methods rely heavily on fine-tuning to capture nuanced spatial-temporal details, which incurs significant data and computation costs. In contrast, training-free approaches, though efficient, often lack robustness in preserving context-rich features across complex video content. To this end, we propose DYTO, a novel dynamic token merging framework for zero-shot video understanding that adaptively optimizes token efficiency while preserving crucial scene details. DYTO integrates a hierarchical frame selection and a bipartite token merging strategy to dynamically cluster key frames and selectively compress token sequences, striking a balance between computational efficiency with semantic richness. Extensive experiments across multiple benchmarks demonstrate the effectiveness of DYTO, achieving superior performance compared to both fine-tuned and training-free methods and setting a new state-of-the-art for zero-shot video understanding.

Lower limb isometric strength is crucial for predicting diseases, monitoring rehabilitation, and assessing activity levels. Manual testing lacks quantitative evaluation, while handheld dynamometers (HHDs) require skilled raters and isokinetic dynamometers are expensive and complex. Existing devices often focus on single-joint measurements for specific populations. To address the need for multi-joint quantitative muscle strength assessment, along with portability, affordability, and ease of use, this study developed the isometric dynamometer for the lower extremity (IDLE) to measure hip flexion, knee extension, knee flexion, and ankle dorsiflexion strength. Its validity and reliability were evaluated in 20 healthy adults (50% female). The IDLE demonstrated excellent validity compared to a strap-fixed HHD (Pearson's r ≥ 0.907, ICC ≥ 0.908, P < 0.01). Intra-rater reliability was excellent (ICC ≥ 0.926) for male knee extension (bilateral), left knee flexion, and right ankle dorsiflexion; as well as for female right hip flexion, knee extension (bilateral), and right knee flexion, and good (ICC ≥ 0.808) for other measurements. Inter-rater reliability was excellent (ICC ≥ 0.901) for all measurements except male right ankle dorsiflexion. The IDLE is a valid and reliable device for measuring lower extremity isometric strength in healthy adults, with further validation in clinical populations recommended.

The trophoblast cell-surface antigen 2 (Trop2) is markedly overexpressed in breast cancers, with a particularly high incidence in triple-negative breast cancer. The therapeutic relevance of Trop2 expression is underscored by the approval of an antibody-drug conjugate for triple-negative breast cancer treatment. However, there is no a predictive technique for accurate whole-body mapping of Trop2 expression in patients. In this study, we developed a novel Trop2-specific molecular probe, [99mTc]Tc-MY6349, and evaluated its safety and feasibility for detecting Trop2 expression in breast cancer using SPECT/CT imaging. Methods: Trop2 expression in different breast cancer cell lines was assessed via immunofluorescence and flow cytometry. The Trop2-specific nanobody MY6349 was site-specifically labeled with 99mTc via a C-terminal GGGC tag, and its binding affinity to the Trop2 receptor was tested in vitro. The in vivo tumor uptake and distribution of [99mTc]Tc-MY6349 were examined through SPECT imaging and biodistribution studies. Furthermore, a pilot clinical study of [99mTc]Tc-MY6349 SPECT/CT was conducted in 8 patients with breast cancer, and the results were compared with [18F]FDG PET/CT. Results: [99mTc]Tc-MY6349 achieved a greater than 95% radiochemical purity after purification. In vitro and in vivo experiments demonstrated the binding specificity of [99mTc]Tc-MY6349 to the Trop2 receptor. In vivo imaging and biodistribution studies revealed a significant correlation between tumor uptake and Trop2 expression levels. In the pilot clinical study, SPECT imaging with [99mTc]Tc-MY6349 successfully detected Trop2-positive tumors 15 min after tracer injection. Delayed imaging showed reduced uptake in normal organs but sustained retention of [99mTc]Tc-MY6349 in tumors. Importantly, [99mTc]Tc-MY6349 SPECT/CT imaging highlighted Trop2 expression heterogeneity and visualized primary and metastatic lesions with a favorable tumor-to-background ratio in breast cancer. Conclusion: [99mTc]Tc-MY6349 was successfully prepared and exhibited a high binding affinity and Trop2 specificity. The pilot clinical study validated the safety and feasibility of [99mTc]Tc-MY6349 SPECT/CT for detecting Trop2 expression in vivo in patients with breast cancer. This imaging modality could complement existing methods, aiding in the guidance of Trop2-targeted therapies and advancing personalized treatment while also promoting the application of SPECT/CT nuclear medicine imaging technology.

Objective:To accurately measure human energy metabolism with high temporal resolution, a respiratory gas analysis system was designed using a breath-by-breath approach. Methods:Firstly, indirect calorimetry was employed in respiratory gas analysis to measure the respiratory flow and concentration signals in real-time. Secondly, oxygen consumption QO2 and carbon dioxide production QCO2 were calculated through respiratory characteristic alignment and respiratory signal segmentation. Finally, metabolic indexes were calculated according to the Weir formula. Furthermore, a controlled trial was formulated for validation comparisons with the MGC ULTIMA SYSTEM PFX CARDIO2. Results:The results indicate that the data points of metabolic indexes measured by this system all fall within the confidence interval when compared with those of the MGC. Conclusion:The system has high consistency with the MGC measurement results. Thus, the system can accurately measure metabolism in real-time and provide data support for clinical nutrition assessment and therapy.

Among the ever-evolving development of vision-language models, contrastive language-image pretraining (CLIP) has set new benchmarks in many downstream tasks such as zero-shot classifications by leveraging self-supervised contrastive learning on large amounts of text-image pairs. However, its dependency on rigid one-to-one mappings overlooks the complex and often multifaceted relationships between and within texts and images. To this end, we introduce RankCLIP, a novel pretraining method that extends beyond the rigid one-to-one matching framework of CLIP and its variants. By leveraging both in-modal and cross-modal ranking consistency, RankCLIP improves the alignment process, enabling it to capture the nuanced many-to-many relationships between and within each modality. Through comprehensive experiments, we demonstrate the enhanced capability of RankCLIP to effectively improve performance across various downstream tasks, notably achieving significant gains in zero-shot classifications over state-of-the-art methods, underscoring the potential of RankCLIP in further advancing vision-language pretraining.

Sleep disturbance is common in alcohol dependence (AD). Far-infrared ray (FIR) therapy has anti-inflammatory properties and may improve sleep, offering a potential non-pharmacological treatment during withdrawal. The primary objective of this study was to compare sleep quality and serum pro-inflammatory cytokine levels between AD patients who received FIR therapy and those who received standard care only. A total of 114 male AD patients who completed a two-week withdrawal program were randomized to a control group (n = 61) receiving supportive care, or a treatment group (n = 53) receiving supportive care plus five FIR sessions. Sleep quality was assessed using cardiopulmonary coupling (CPC), and serum levels of 10 cytokines were measured via Luminex assay before and after the intervention. Analysis of covariance (ANCOVA) compared post-treatment outcomes between groups, controlling for baseline values. Pearson correlation analysis examined relationships between changes in sleep parameters and cytokine levels. Compared to controls, the treatment group showed significantly increased total sleep time and sleep efficiency. Moreover, serum levels of IL-1β, IL-6, and TNF-α were significantly reduced in the treatment group. Addtionally, improvements in total sleep time and sleep efficiency were negatively correlated with changes in IL-1β and TNF-α levels. FIR therapy may improve sleep quality and reduce pro-inflammatory cytokines in AD patients during withdrawal, suggesting its potential as an alternative to hypnotic medications for sleep management.

The automation of scientific research through large language models (LLMs) presents significant opportunities but faces critical challenges in knowledge synthesis and quality assurance. We introduce Feedback-Refined Agent Methodology (FRAME), a novel framework that enhances medical paper generation through iterative refinement and structured feedback. Our approach comprises three key innovations: (1) A structured dataset construction method that decomposes 4,287 medical papers into essential research components through iterative refinement; (2) A tripartite architecture integrating Generator, Evaluator, and Reflector agents that progressively improve content quality through metric-driven feedback; and (3) A comprehensive evaluation framework that combines statistical metrics with human-grounded benchmarks. Experimental results demonstrate FRAME's effectiveness, achieving significant improvements over conventional approaches across multiple models (9.91

This study addresses the challenge of achieving precise endpoint positioning for robotic arms in high-precision grinding, often compromised by processing and installation errors, as well as non-geometric parameters. Focusing on a five-degree-of-freedom grinding robot, we present an error calibration and compensation method utilizing grating ruler measurement for the robotic arm. Firstly, an improved D-H parameter method establishes the forward kinematic model, coupled with a mathematical model for compensating error parameters in combination with the robotic arm’s geometric configuration. Then, based on the measurement data of the grating ruler, the fusion algorithm is used to solve the error parameter values of the lengths and angles of each joint and compensate them into the kinematic model. Finally, the error calibration results are compared with those of the laser tracker to verify the positioning accuracy of the end of the robotic arm after the grating ruler calibration. The experimental results show that the fluctuation range of the repeat positioning accuracy of the grating ruler calibration and the laser tracker calibration is basically the same, and the stability is good. Compared with the uncalibrated robotic arm, the absolute positioning accuracy has increased by 69.95%, and the difference from the calibration of the laser tracker is 0.069 mm. Its positioning accuracy basically meets the requirements of high-precision grinding. Meanwhile, the cost of the grating ruler equipment is much lower than that of the laser tracker, and it can replace the laser tracker to calibrate the error of the mechanical arm.

Flexible sweat sensors play a crucial role in health monitoring and disease prevention by enabling real-time, non-invasive assessment of human physiological conditions. Sweat contains a variety of biomarkers, offering valuable insights into an individual's health status. In this study, we developed an advanced flexible electrochemical sensor featuring reduced graphene oxide (rGO)-based electrodes, modified with a composite material comprising nitrogen and sulfur co-doped holey graphene (HG) and MXene, with in-situ-grown TiO2 nanoparticles on the MXene. The sensor design leverages the synergistic properties of its components: MXene provides a conductive scaffold, TiO2 enhances electrocatalytic activity, and the porous HG network facilitates efficient ion and electron transfer, with doping increasing the number of active sites. This configuration enables sensitive and simultaneous detection of ascorbic acid (AA), uric acid (UA), and dopamine (DA). Additionally, a potassium-selective (K⁺) film applied to rGO-enhanced electrodes supports concurrent detection of K⁺ in sweat. The sensor array demonstrates a broad detection range, low detection limits, and high sensitivity, while maintaining mechanical flexibility, anti-interference capabilities, and repeatability. Validated through in-situ sweat biomarker detection during exercise, the sensor effectively tracked fluctuations in K⁺, AA, and UA levels in a volunteer. This practical application underscores the sensor's potential for continuous health monitoring, and early disease detection, establishing it as a promising tool in modern healthcare.

Early screening of arterial stiffness plays a crucial role in reducing cardiovascular risk and mortality, as well as promoting healthy aging. However, asymptomatic patients are frequently overlooked due to the lack of specialized interpretation and the costly hospital-based equipment required by traditional clinical methods, potentially leading to adverse consequences. This study presents IF-ASS, an interpretable framework designed to facilitate widespread screening of arterial stiffness at an earlier stage by utilizing routinely acquired photoplethysmography (PPG) signals. The data was collected from 205 participants (aged 21-82 years) through two recruitment phases. A novel vascular volume analysis (VVA) of PPG conduction was proposed to expanded the methods for extracting general PPG features. And subsequently 216 features were derived. The features, characterized by their meaningful functionalities, were combined using the proposed feature pooling (FP) algorithm. This method attained the highest dimensionality reduction rate (DRR = 94.9 %, 15.17 % improvement), while offering scalable combinations for model enhancement. The optimal trade-off between effectiveness and interpretability was achieved by modeling an improved probability risk explanation method based on eXtreme Gradient Boosting and Shapley Additive exPlanations. Among multidimensional comparisons and validations, IF- ASS demonstrated superior stability, delivering the highest area under the receiver operating characteristic curve (AUC) of 0.924, and accuracy (ACC) of 0.850 (3.25 % mean improvement). Attributively, inspired explorations significantly boosted the trustworthiness of IF-ASS. In conclusion, the IF-ASS could be considered a reliable and cost-effective tool for screening latent arterial stiffness in ubiquitous settings, particularly in the trends of disease onset at younger ages and an aging population.

Identifying travel modes is essential for modern urban transportation planning and management. Recent advancements in data collection, especially those involving Global Positioning System (GPS) technology, offer promising opportunities for rapidly and accurately inferring users' travel modes. This study presents an innovative method for inferring travel modes from GPS trajectory data. The method utilizes multi-scale convolutional techniques to capture and analyze both temporal and spatial information of the data, thereby revealing the underlying spatiotemporal relationships inherent in user movement and behavior patterns. In addition, an attention mechanism is integrated into the model to enable autonomous learning. This mechanism enhances the model's capacity to identify and emphasize key information across different time periods and spatial locations, thus improving the accuracy of travel mode inference. Evaluation on the open-source GPS trajectory dataset, GeoLife, demonstrates that the proposed method attained an accuracy of 83.3%. This result highlights the effectiveness of the method, demonstrating that the model can more accurately understand and predict user travel modes through the integration of multi-scale convolutional technologies and attention mechanisms.

Herein, we aimed to develop an easily available and efficient screening method for diabetic peripheral neuropathy (DPN) suitable for primary care settings, emphasizing simplicity, speed, and accuracy. Nerve conduction studies were conducted on 214 patients with diabetes, encompassing the outcomes of five distinct assessments: diabetic neuropathy symptom (DNS), vibration perception threshold (VPT), and nerve screening. The diagnostic accuracy of the VPT and nerve screening was evaluated by comparing them with that of the nerve conduction study. To assess diagnostic efficacy, various combinations were examined, including DNS combined with VPT, pain, temperature, touch, and ankle reflex. The diagnostic performance of DNS was superior to that of the five neurological screening items and VPT, with sensitivity, specificity, and accuracy of 0.68, 0.81, and 0.73, respectively. Among the two combined methods, "DNS + ankle reflex" was identified as having the highest diagnostic value, with an area under the curve, a sensitivity, a specificity, and an accuracy of 0.81, 0.89, 0.70, and 0.80, respectively. Furthermore, a combination of "DNS + ankle reflex + touch + pain + VPT" achieved the best performance among the five combinations, with an area under the curve, sensitivity, specificity, and accuracy of 0.85, 0.93, 0.68, and 0.81, respectively. The combination of DNS, ankle reflex, touch, pain, and VPT methods showed the highest diagnostic value for DPN. However, considering factors including accuracy, time, and economic cost, we recommend using a simpler combination of DNS and ankle reflex for large-scale screening of patients with DPN.

Intratumoral immune status influences tumor therapeutic response, but it remains largely unclear how the status determines therapies for patients with intrahepatic cholangiocarcinoma. Here, we examine the single-cell transcriptional and TCR profiles of 18 tumor tissues pre- and post- therapy of gemcitabine plus oxaliplatin, in combination with lenvatinib and anti-PD1 antibody for intrahepatic cholangiocarcinoma. We find that high CD8 GZMB+ and CD8 proliferating proportions and a low Macro CD5L+ proportion predict good response to the therapy. In patients with a poor response, the CD8 GZMB+ and CD8 proliferating proportions are increased, but the CD8 GZMK+ proportion is decreased after the therapy. Transition of CD8 proliferating and CD8 GZMB+ to CD8 GZMK+ facilitates good response to the therapy, while Macro CD5L+-CD8 GZMB+ crosstalk impairs the response by increasing CTLA4 in CD8 GZMB+. Anti-CTLA4 antibody reverses resistance of the therapy in intrahepatic cholangiocarcinoma. Our data provide a resource for predicting response of the combination therapy and highlight the importance of CD8+T-cell status conversion and exhaustion induced by Macro CD5L+ in influencing the response, suggesting future avenues for cancer treatment optimization.