bims-aukdir 2026-11-08 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2026–11–08
eighteen papers selected by
Mott Given

Learning to fuse and reconstruct multi-view graphs for diabetic retinopathy grading.
HalluciGuard-eye: knowledge graph-enhanced retrieval-augmented generation for hallucination detection and mitigation in ophthalmic multimodal large language models.
Explainable machine learning with wavelet-enhanced multi-scale texture analysis of diabetes-related heel pad ultrasound texture differences.
Personalized type-II diabetes diagnosis in smart healthcare using a neutrosophic collaborative learning approach.
Leveraging machine learning methods to predict post-transplant diabetes in solid organ transplant recipients.
A Robust and Trustable Approach to Incorporate Medical Domain Knowledge in Machine Learning Models for Diabetic Retinopathy Screening Using Routine Lab Results.
Development and validation of a machine learning-based model for predicting amputation risk in patients with lower extremity arterial disease and diabetes.
Cross-spatial attention guided multi-scale fusion network for precise microaneurysm segmentation.
Can artificial intelligence counsel expectant mothers? Expert evaluation of ChatGPT's educational performance in gestational diabetes mellitus.
Development and internal validation of machine learning models for predicting prediabetes in China.
Factors Influencing Diabetic Retinopathy Screening Attendance: Perspectives of People With Type 2 Diabetes and Healthcare Professionals.
Adaptive causal discovery with ordering-based reinforcement learning-approach for type 2 diabetes mellitus risk factors.
Serum Metabolic Differences Across Glucose Metabolic States Using Untargeted UPLC-HRMS Metabolomics.
Machine learning-based identification of depression in older patients with type 2 diabetes mellitus: A multicenter retrospective study.
Can Retrieval-Augmented Generation Improve the Potential of the LLMs in Diabetes Health Management? An Assessment Based on Health Management Tasks and Question Complexity.
AI-Delivered Short Video Cognitive Behavioral Therapy for Type 2 Diabetes Mellitus - A Randomized Controlled Trial Protocol.
Exploratory Evidence of a Gene-Smoking Interaction Involving MTR rs1805087 in Type 2 Diabetes Mellitus.
DiaBench: A Benchmarking Framework to Evaluate the Performance of LLMs in T2D Patient Education.

Med Image Anal. 2026 Oct 30. pii: S1361-8415(26)00436-6. [Epub ahead of print]116 104367

Learning to fuse and reconstruct multi-view graphs for diabetic retinopathy grading.

Haoran Li, Yuxin Lin, Huan Wang, Xiaoling Luo, Qi Zhu, Jiahua Shi, Huaming Chen, Bo Du, Johan Barthelemy, Zongyan Xue, Jun Shen, Yong Xu.

  Diabetic retinopathy (DR) is one of the leading causes of vision loss worldwide, making early and accurate DR grading critical for timely intervention. Recent clinical practices leverage multi-view fundus images for DR detection with a wide coverage of the field of view (FOV), motivating deep learning methods to explore the potential of multi-view learning for DR grading. However, existing methods often overlook the inter-view correlations when fusing multi-view fundus images, failing to fully exploit the inherent consistency across views originating from the same patient. In this work, we present MVGFDR, an end-to-end Multi-View Graph Fusion framework for DR grading. Different from existing methods that directly fuse visual features from multiple views, MVGFDR is equipped with a novel Multi-View Graph Fusion (MVGF) module to explicitly disentangle the shared and view-specific visual features. Specifically, MVGF comprises three key components: ① Multi-view Graph Initialization, which constructs visual graphs via residual-guided connections and employs Discrete Cosine Transform (DCT) coefficients as frequency-domain anchors; ② Multi-view Graph Fusion, which integrates selective nodes across multi-view graphs based on frequency-domain relevance to capture complementary view-specific information; and ③ Masked Cross-view Reconstruction, which leverages masked reconstruction of shared information across views to facilitate view-invariant representation learning. Extensive experimental results on MFIDDR, by far the largest multi-view fundus image dataset, demonstrate that the proposed approach achieves state-of-the-art overall performance for diabetic retinopathy grading. Code is available at https://github.com/HaoranLi525/MVGFDR.

Keywords:  Diabetic retinopathy detection; Feature fusion; Image classification; Medical imaging; Visual graph learning

DOI:  https://doi.org/10.1016/j.media.2026.104367
Front Endocrinol (Lausanne). 2026 ;17 1976190

HalluciGuard-eye: knowledge graph-enhanced retrieval-augmented generation for hallucination detection and mitigation in ophthalmic multimodal large language models.

Ge Wu, Chong Ai, Weiqi Wu.

   Background: Diabetic retinopathy (DR) is the leading cause of preventable blindness in working-age adults, yet scalable early detection is limited by the unreliability of AI-assisted screening. Multimodal large language models can interpret fundus photographs and support DR grading, but they frequently hallucinate - fabricating lesions or misclassifying referable DR - which directly threatens timely treatment and prevention of vision loss. We developed HalluciGuard-Eye to make AI-assisted DR assessment trustworthy.
Methods: HalluciGuard-Eye combines knowledge graph-enhanced retrieval-augmented generation with hierarchical hallucination detection. An Ophthalmic Clinical Knowledge Graph containing 47,832 medical entities and 156,294 semantic relationships was constructed from clinical guidelines and published literature. A dual-stream architecture uses a Visual-Semantic Alignment Module to verify cross-modal consistency and a Clinical Factuality Grounding Network to anchor generated content to reranked evidence. The Hierarchical Factuality Score assesses anatomical, pathological, and clinical-recommendation accuracy. The framework was evaluated on four public ophthalmic imaging benchmarks.
Results: HalluciGuard-Eye reduced hallucination rates by 42.2% compared with state-of-the-art medical multimodal large language models and achieved 91.6% factual accuracy for clinical entity recognition on EyePACS. For diabetic retinopathy grading, sensitivity was 89.6% and specificity was 92.1%, representing improvements of 2.4% and 2.7%, respectively, over MedRAG. In the component ablation, removing knowledge-graph retrieval lowered EyePACS HFS by 9.4 percentage points. Additional analyses yielded 87.6% critical-evidence recall@20, 89.2% lesion-location accuracy, and 88.9% sensitivity in a patient-clustered cohort.
Conclusions: These results indicate that knowledge-grounded retrieval and hierarchical factuality assessment can reduce hallucinations while preserving diagnostic performance in ophthalmic multimodal language models.

Keywords:  AI screening; diabetic retinopathy; early detection; hallucination detection; knowledge graph; multimodal large language models; retrieval-augmented generation

DOI:  https://doi.org/10.3389/fendo.2026.1976190
J Appl Clin Med Phys. 2026 Nov;pii: e70888. [Epub ahead of print]27(11):

Explainable machine learning with wavelet-enhanced multi-scale texture analysis of diabetes-related heel pad ultrasound texture differences.

Xinlei Ji, Yiming Li, Jiachen Guo, Yitian Zhang, Liang Zhou, Tao Wu, Lei Yin.

   BACKGROUND: Diabetes is associated with structural and functional differences in plantar soft tissues, which may be reflected in ultrasound texture patterns. However, existing ultrasound texture studies are limited by single-scale feature extraction and limited feature-level interpretability.
PURPOSE: To develop an explainable machine learning framework incorporating a wavelet-enhanced multi-scale texture feature extraction module (WEMS-TFE), and to investigate whether heel pad ultrasound texture features differ between patients with diabetes without active foot ulcers and healthy controls.
METHODS: Heel pad ultrasonography was performed in 76 patients with diabetes (without active foot ulcers) and 47 healthy controls. Ultrasound images of the heel pad were analyzed at the microchamber and macrochamber levels. The WEMS-TFE module was proposed to extract multi-scale, multi-directional texture features via two-dimensional discrete wavelet decomposition combined with the gray-level co-occurrence matrix (GLCM). Features with significant between-group differences were selected and evaluated across seven machine learning classifiers using repeated cross-validation with participant-disjoint data partitions. SHapley Additive exPlanations (SHAP) analysis was used to interpret feature contributions in a representative XGBoost model.
RESULTS: The WEMS-TFE scheme achieved the highest cross-model composite performance and robustness among the five evaluated feature extraction schemes. Cross-classifier importance analysis identified 6 features that recurred across multiple classifiers. Consistently, most of these also ranked among the important features in the XGBoost SHAP analysis, and two representative features showed clear value-dependent contribution patterns. The identified features therefore reflected group-associated image-texture differences, although their associations may be affected by residual demographic confounding.
CONCLUSIONS: The proposed framework, with the WEMS-TFE module as its core component, enables quantitative characterization of group-associated differences in heel pad ultrasound texture features. The findings provide imaging-based evidence of group-level texture differences in the heel pad and support further investigation of plantar tissue characteristics in diabetes, while acknowledging that these differences may be affected by residual demographic confounding and do not establish diabetes-specific structural changes or causal tissue mechanisms.

Keywords:  diabetes mellitus; explainable artificial intelligence; heel pad ultrasonography; machine learning; ultrasound texture analysis

DOI:  https://doi.org/10.1002/acm2.70888
Digit Health. 2026 Jan-Dec;12:12 20552076261490345

Personalized type-II diabetes diagnosis in smart healthcare using a neutrosophic collaborative learning approach.

Tin-Chih Toly Chen.

   Objective: Machine learning (ML) and deep learning (DL) technologies have proven to be highly effective in diagnosing type-II diabetes. Despite this, their primary focus tends to be on maximizing overall diagnostic accuracy. Consequently, there may still be inaccuracies in diagnosing individual patients.
Methods: A neutrosophic collaborative learning (NCL) approach is proposed in this study for type-II diabetes diagnosis. In this approach, multiple local artificial intelligence models (LAMs) work simultaneously to predict the likelihood of a patient having type-II diabetes. This allows for an efficient and comprehensive diagnosis from various perspectives. Each LAM utilizes a neutrosophic deep neural network (NDNN) to estimate the probability of the condition. The NDNN model considers factors like local learning inadequacy and prediction inaccuracy when evaluating the indeterminacy of predictions. If the indeterminacy membership exceeds a certain threshold, the prediction may not be trustworthy. The results from these predictions are then sent to a global server, where they are aggregated using a neutrosophic weighted intersection (NWI).
Results: A benchmark case was employed to assess the effectiveness of the NCL approach in diagnosing type-II diabetes. The experimental results demonstrated that the NCL approach increased the hit rate for type-II diabetes diagnosis by an average of 9%. Furthermore, the collaboration of LAMs enhanced diagnostic accuracy by 5.6%.
Conclusions: Taking into account a variety of perspectives can enhance the effectiveness of diagnosing type-II diabetes. Converting diagnostic results into neutrosophic sets, which include three different memberships, also aids in evaluating the trustworthiness of each diagnostic outcome.

Keywords:  deep learning; diagnosis; neutrosophic collaborative learning; neutrosophic weighted intersection; type-II diabetes

DOI:  https://doi.org/10.1177/20552076261490345
J Endocr Soc. 2026 Dec;10(12): bvag236

Leveraging machine learning methods to predict post-transplant diabetes in solid organ transplant recipients.

Rhea Varghese, Yingji Sun, Saba Maleki, Mamatha Bhat.

   Background: Post-transplant diabetes mellitus (PTDM) is a common complication following solid organ transplantation and is associated with increased morbidity and mortality. This study aimed to identify both generalizable and solid organ-specific predictors of PTDM with machine learning methods.
Materials & Methods: In this study, we used 2002-2023 data from the Scientific Registry of Transplant Recipients (SRTR) to examine the risk of PTDM development across the liver (n = 128 016), kidney (n = 360 562), heart (n = 45 505), and lung (n = 36 260) transplant recipients. We examined various features at baseline using logistic regression and random forest models. Performance was evaluated using area under the receiver operating characteristic curve (AUROC), Brier scores, sensitivity, and specificity. Shapley Additive exPlanations were used for model interpretation and to identify key relevant features.
Results: The incidence of PTDM ranged from 15.0% in kidney to 20.2% in lung recipients. Compared to logistic regression, random forest demonstrated modestly improved discrimination (AUROC: 0.62-0.63) and calibration (Brier: 0.124-0.156). It was found across all models that age and BMI were consistent predictors.Prednisone and cyclosporine maintenance were often found to be associated with higher PTDM risk while mycophenolate mofetil was seen as protective. SHAP analysis demonstrated complex, nonlinear interactions and supported known risk patterns such as increased age and male sex.
Conclusion: Our machine learning model identified risk factors common to post-transplant diabetes across organ transplant recipients, guiding more optimal patient survival post-transplant.

Keywords:  logistic regression; machine learning; post-transplant diabetes; random forest; solid organ transplant

DOI:  https://doi.org/10.1210/jendso/bvag236
Smart Health (Amst). 2026 Dec;pii: 100724. [Epub ahead of print]42

A Robust and Trustable Approach to Incorporate Medical Domain Knowledge in Machine Learning Models for Diabetic Retinopathy Screening Using Routine Lab Results.

Enrico Laoh, Oday Bani Ahmad, Tieming Liu.

  Many AI models for disease prediction remain difficult to integrate into clinical practice because model complexity can limit transparent interpretation. This study develops a domain-informed, low-dimensional machine-learning framework for diabetic retinopathy (DR) screening using routinely collected electronic health record and laboratory data. Medical-domain information was extracted from PubMed using a Term Frequency-Relative Citation Ratio (TF-RCR) procedure, and the selected feature set was constrained to remain cognitively manageable. The final model uses eight variables - age, glucose, nephropathy, neuropathy, HbA1c, hemoglobin, albumin, and creatinine - and its predictions are examined using post-hoc SHapley Additive exPlanations (SHAP) feature attribution. The data contained a 97,786-patient cohort, including 16,865 DR cases (17.25%). After a 64%/16%/20% training/validation/test split, SMOTE was applied only to the training subset, while validation and test sets retained the observed prevalence. On the untouched test set (n = 19,558), the eight-feature XGBoost model achieved 95.35% accuracy, 91.17% sensitivity, 96.22% specificity, an F1-score of 87.12%, AUROC of 97.24%, and AUPRC of 94.55%. Same-cohort comparisons with logistic regression, a 24-feature XGBoost model, a multilayer perceptron (MLP), and an FT-Transformer showed that the eight-feature model retained competitive discrimination while using a substantially smaller feature set. The framework provides a reproducible approach for combining literature-derived knowledge with empirical data while keeping the final model concise and its feature contributions available for post-hoc interpretation.

Keywords:  Diabetic Retinopathy Prediction; Elaborative Learning; Healthcare Analytics; PubMed; Trustable AI

DOI:  https://doi.org/10.1016/j.smhl.2026.100724
Front Cardiovasc Med. 2026 ;13 1893875

Development and validation of a machine learning-based model for predicting amputation risk in patients with lower extremity arterial disease and diabetes.

Lu Tang, Liang Wang, Jing Li, Xiaoyu Liu, Qin Chen, Pan Song, Xiaoyan Quan, Qingran Li, Gengliang Liu, Yang Liu, Chi Cui, Meihong Shi.

   Background: Patients with lower extremity arterial disease (LEAD) complicated by diabetes face substantial amputation risk. Pre-amputation death acts as a critical competing event that distorts risk estimation. Validated personalized prediction tools for this population remain scarce and hinder precise clinical interventions. This study aimed to develop and validate two complementary interpretable models for 12-month amputation risk in patients with LEAD and diabetes to support clinical decision-making.
Methods: This retrospective multicenter study enrolled 1,919 patients with LEAD and diabetes from four Chinese hospitals (January 2018-September 2024), with follow-up ending in September 2025. The primary outcome was 12-month amputation (major or minor). Prior to model construction, candidate predictors were screened via LASSO regression, random forest classifier and Boruta algorithm. Ten machine learning algorithms were then trained and evaluated by discrimination and calibration metrics. SHapley Additive Explanations (SHAP) was applied for interpretability. A Fine-Gray competing risk regression model was built using the same screened predictor set.
Results: Patients were split into training (n = 967), internal validation (n = 414), and external validation (n = 538) cohorts. The 12-month amputation rate was 14.0%. Median follow-up was 12 months; mean age was 72.7 ± 9.5 years and 65.7% were male. The optimal GraBoost model achieved internal validation AUC = 0.860 and external validation AUC = 0.828, incorporating 14 clinical predictors, including: Rutherford classification, financial burden, age, Barthel Index, diabetic foot, statin, fibrinogen (FIB), neutrophil (NEUT), residence, prealbumin (PALB), C-reactive protein (CRP), activated partial thromboplastin time (APTT), body mass index (BMI) classification and wound. SHAP revealed higher grade Rutherford classification, lower Barthel Index, younger age, wounds, diabetic foot, and elevated FIB, APTT, CRP, NEUT correlated with higher amputation risk. Urban residence, normal BMI classification, lower financial burden, statin use and higher PALB were protective. Shared predictors showed consistent directional effects across GraBoost and Fine-Gray models. A dual-model web-based calculator was constructed for risk assessment.
Conclusion: We developed and validated complementary interpretable GraBoost machine learning and Fine-Gray competing risk models for 12-month amputation risk. The dual-model web calculator may help clinicians identify high-risk patients and guide clinical management.

Keywords:  amputation; clinical prediction model; diabetes; lower extremity arterial disease; machine learning

DOI:  https://doi.org/10.3389/fcvm.2026.1893875
Biomed Tech (Berl). 2026 Nov 05.

Cross-spatial attention guided multi-scale fusion network for precise microaneurysm segmentation.

Zuxiang Wang, Weiwei Gao, Yu Fang, Nan Song, Feng Xiao.

   OBJECTIVES: Accurate detection of Microaneurysms (MAs) is crucial for the early diagnosis of diabetic retinopathy (DR). However, due to their small size and low contrast with the background, precisely segmenting them from the background poses significant challenges.
METHODS: The Cross-Spatial Attention Embedded Multi-Scale Fusion Network (CSAFNet) is proposed to enhance the segmentation accuracy of MAs in fundus images. Based on the U-Net architecture, an improved ResNet50 is used as the backbone for feature extraction. An Efficient Multi-scale Attention (EMA) module is embedded to capture key features across different spatial locations and scales. Simultaneously, a Multi-Scale Deep Fusion (MSDF) module is designed to integrate contextual features from multiple scales. Additionally, the proposed Balanced Focusing Boundary (BFB) loss function improves segmentation by enhancing geometric accuracy at object boundaries.
RESULTS: The method is evaluated on four datasets (E-ophtha, TongjiU-DROD, IDRiD, Messidor) with various advanced methods. On IDRiD, its F1-score and IoU improved by 6.16 and 5.63 % respectively, compared to the state-of-the-art (SOTA). On the Messidor and TongjiU-DROD datasets, it demonstrated comprehensive superiority in sensitivity, specificity, F1-score, and IoU.
CONCLUSIONS: The proposed CSAFNet can accurately extract MAs from fundus images, thereby effectively assisting in the precise diagnosis of DR.

Keywords:  deep feature fusion; diabetic retinopathy; microaneurysms; multi-scale attention

DOI:  https://doi.org/10.1515/bmt-2025-0495
BMC Endocr Disord. 2026 Aug 14. pii: 357. [Epub ahead of print]26(1):

Can artificial intelligence counsel expectant mothers? Expert evaluation of ChatGPT's educational performance in gestational diabetes mellitus.

Mücahit Furkan Balcı, Celal Akdemi̇r, Fatih Yildirim.

   BACKGROUND: Gestational diabetes mellitus (GDM) is an increasingly common pregnancy complication associated with adverse maternal-fetal outcomes and long-term risk of type 2 diabetes. Artificial intelligence (AI) tools such as ChatGPT have emerged as potential instruments for patient education, yet their reliability and safety in obstetric counseling remain underexplored. This study aimed to evaluate the educational quality of ChatGPT-5 responses to frequently asked questions about GDM.
METHODS: This cross-sectional study was conducted between January and June 2025 at Torbalı State Hospital, İzmir, Türkiye, following STROBE guidelines. Twenty frequently asked questions about gestational diabetes mellitus, identified from routine antenatal counseling through clinician observation and consensus, were entered into ChatGPT-5 using a standardized prompt. Thirty physicians specializing in obstetrics and gynecology from Türkiye, the United Kingdom, Canada, and Germany independently rated the AI-generated answers on a 5-point Likert scale across five domains: accuracy, comprehensiveness, clarity, safety, and appropriateness. Reliability was analyzed using Cronbach's α and intraclass correlation coefficient (ICC). Comparative analyses were performed using Friedman and Dunn-Bonferroni tests.
RESULTS: Overall, ChatGPT-5 achieved high ratings for accuracy (4.23 ± 0.32), clarity (4.18 ± 0.27), and appropriateness (4.09 ± 0.31), while comprehensiveness (3.92 ± 0.29) and safety (3.88 ± 0.34) scored lower. Internal consistency was strong (Cronbach's α = 0.87) and interrater reliability was good (ICC = 0.79). Among thematic domains, "Definition, causes, and risk factors" received the highest scores (4.20 ± 0.30), whereas "Follow-up and treatment" scored the lowest (3.96 ± 0.33). Safety ratings were significantly lower than accuracy and clarity (p = 0.018).
CONCLUSION: ChatGPT-5 demonstrated high accuracy and clarity in GDM counseling, suggesting its potential use as a physician-supervised educational support tool. However, limitations in comprehensiveness and safety highlight the need for guideline-based refinement and human oversight to ensure ethical and reliable integration of AI in obstetric patient education.
TRIAL REGISTRATION: Not applicable. This study is an observational cross-sectional design and did not involve patient enrollment or intervention.

Keywords:  Artificial intelligence; ChatGPT; Digital health; Gestational diabetes mellitus; Patient education; Pregnancy

DOI:  https://doi.org/10.1186/s12902-026-02490-w
PLoS One. 2026 ;21(11): e0361704

Development and internal validation of machine learning models for predicting prediabetes in China.

Zhengxin Jiang, Wen Qian, Haifang Yang, Yifang Luo, Wen Peng, Qiong Wu, Xiaoxing Wei.

BACKGROUND: Diabetes imposes a growing health burden among Chinese adults, necessitating scalable tools for early identification of high-risk individuals. We aimed to develop and internally validate survival machine-learning models for prediabetes risk stratification using routinely available clinical indicators.
METHODS: This retrospective cohort study included 77,478 Chinese adults from the Dryad repository (2010-2016). Candidate predictors were selected via LASSO regression and random survival forests (RSF). Six survival models, Cox proportional hazards (Cox-PH), CoxBoost, DeepSurv, gradient boosting machine (GBM), random survival forests (RSF), and survival tree, were developed and evaluated using the concordance index (C-index), time-dependent ROC curves, calibration curves, and decision curve analysis (DCA). The best-performing model was deployed as a web-based application.
RESULTS: Eight predictors were consistently selected across methods: age, BMI, fasting plasma glucose, total cholesterol, triglycerides, HDL-C, LDL-C, and creatinine clearance rate. In the test set, DeepSurv and GBM achieved comparable discrimination (C-index: 0.756 vs. 0.751; adjusted P = 0.694). GBM was preferred due to better calibration and greater net benefit across clinically relevant thresholds, leading to its selection as the candidate model.
CONCLUSION: The GBM-based prediabetes risk model, built on eight inexpensive and routinely available clinical measures, showed promising internal discrimination and calibration in a large Chinese cohort. An interactive web tool was developed as a research prototype to facilitate future external validation and early risk identification. External validation in independent populations is warranted before clinical use.

DOI: https://doi.org/10.1371/journal.pone.0361704
J Eval Clin Pract. 2026 Dec;pii: e70723. [Epub ahead of print]32(8):

Factors Influencing Diabetic Retinopathy Screening Attendance: Perspectives of People With Type 2 Diabetes and Healthcare Professionals.

Gabriela B Petersen, Bryan Cleal, Jette K Kristensen, Henrik Vorum, Stine Byberg, Mia V Fangel, Lene E Joensen.

   RATIONALE: Diabetic retinopathy screening is an established component of effective type 2 diabetes management, yet participation remains suboptimal. Understanding why individuals do not attend screening requires consideration of interactions between patients, healthcare providers, and the organisation of care.
AIM: To explore patient-, provider-, and system-level factors influencing diabetic retinopathy screening attendance within routine type 2 diabetes care, from the perspectives of individuals with type 2 diabetes, general practitioners and eye clinic staff.
METHODS: Semi-structured interviews were conducted with people with type 2 diabetes (n = 20), general practitioners (n = 10) and eye clinic staff (n = 8). Data were analysed using systematic text condensation. Findings were mapped onto the Chronic Care Model to categorize system, provider, and patient-level components.
RESULTS: Non-attendance was influenced by multiple interacting factors, including transportation difficulties, limited awareness of diabetic retinopathy and the purpose of screening, competing health priorities, and challenges integrating diabetes care into everyday life. Healthcare professionals identified time constraints, unclear responsibilities, and insufficient communication between primary and specialist care as important organisational barriers. Facilitators included family support, encouragement from healthcare providers, and awareness of the potential consequences of diabetes and untreated diabetic retinopathy.
CONCLUSIONS: Non-attendance at diabetic retinopathy screening is rarely an explicit or informed decision to decline care. Rather, participation is shaped by interacting structural, informational, contextual, and relational factors across the healthcare system. Improving screening attendance requires approaches that extend beyond increasing patient awareness, including systematic support within routine diabetes care, tailored communication, improved coordination between primary and specialist care, and proactive follow-up strategies.

Keywords:  barriers; diabetic retinopathy; facilitators; qualitative research; screening attendance; type 2 diabetes

DOI:  https://doi.org/10.1111/jep.70723
BMC Bioinformatics. 2026 Aug 05. pii: 260. [Epub ahead of print]27(1):

Adaptive causal discovery with ordering-based reinforcement learning-approach for type 2 diabetes mellitus risk factors.

Bo Chen, Ziyi Yuan, Yufeng Chen, Wenxue Xie, Xiue Gao, Jiangang Hu, Yuan Sang.

   BACKGROUND: Causal discovery of type 2 diabetes mellitus (T2DM) risk factors is crucial for revealing underlying disease mechanisms and supporting prevention and treatment. However, clinical risk-factor data are typically high-dimensional, limiting the performance of many conventional causal discovery methods. Causal discovery with ordering-based reinforcement learning (CORL) reduces the search space via variable orderings, but it uniformly treats all variable states during initial state construction, failing to account for the heterogeneous importance of individual risk factors.
METHODS: To address these issues, this study proposes an adaptive CORL framework (ACORL) tailored for T2DM risk factors. First, the search for a causal graph is reformulated into a variable ordering task, where feature contributions are estimated via an ensemble Random Forest classifier to construct a weighted initial state. Next, the variable ordering is optimized through a reinforcement learning reward mechanism to generate a directed acyclic graph. Subsequently, ordinary least squares linear regression pruning and empirical information entropy estimation are applied to eliminate redundant relations and quantify causal strengths. Finally, comparative experiments were conducted on both synthetic and empirical clinical datasets to evaluate structural accuracy and medical plausibility.
RESULTS: Simulations on a synthetic dataset demonstrated that ACORL reduced the structural Hamming distance from 3 to 1 and enhanced the F1 score from 50% to 85.71% relative to CORL. On empirical clinical data, ACORL identified causal relations highly consistent with established medical knowledge, achieving a correct-edge rate of 83.3% on the 13-feature NHANES dataset and 91.6% on the 29-feature NHANES dataset. Furthermore, it matched the baseline performance on the lower-dimensional Pima datasets while requiring significantly fewer training iterations.
CONCLUSIONS: This study introduces an adaptive reinforcement learning framework for T2DM causal discovery. The proposed ACORL model effectively leverages feature heterogeneity and statistical refinement to discover medically plausible causal relations among risk factors. Consequently, it provides an efficient and reliable tool for epidemiological researchers investigating risk-factor structures in large-scale clinical studies.

Keywords:  Causal discovery; Reinforcement learning; Risk factors; Type 2 diabetes mellitus (T2DM); Variable ordering

DOI:  https://doi.org/10.1186/s12859-026-06578-x
J Proteome Res. 2026 Nov 06. 25(11): 5952-5967

Serum Metabolic Differences Across Glucose Metabolic States Using Untargeted UPLC-HRMS Metabolomics.

Zhulidezi Bahati, Runze Zou, Zhongda Zeng, Wei Duan.

  Insulin resistance (IR) and impaired glucose regulation (IGR) involve metabolic disturbances beyond changes in blood glucose. This cross-sectional study applied untargeted UPLC-HRMS metabolomics to serum from 109 participants across normal glucose tolerance (NGT), IR, IGR, and type 2 diabetes mellitus (T2DM). An integrated analysis of 457 putatively annotated metabolites combined multivariate statistics, clustering, and machine learning to characterize differences among glycometabolic groups and assess the discriminatory information. Pairwise comparisons respectively identified 60, 42, and 59 differential metabolites for NGT vs IR, IR vs IGR, and IGR vs T2DM, including five shared across all comparisons. LASSO and random forest evaluated fixed differential-metabolite panels through stratified 5-fold cross-validation repeated ten times and 1,000 label permutations. The corresponding pooled AUC ranges were 0.918-0.923, 0.995-0.996, and 0.828-0.950. Because panel selection preceded cross-validation, these estimates describe within-cohort discrimination rather than generalizable predictive performance. Pathway analysis of the union of differential metabolites indicated nominal enrichment of lysine degradation and nicotinate and nicotinamide metabolism. Shared metabolites also showed associations with clinical measures of glucose homeostasis and insulin response. These findings characterize distinct yet partly shared metabolic features across glycometabolic states and provide candidates for targeted confirmation and independent longitudinal validation.

Keywords:  differential metabolites; insulin resistance; machine learning; type 2 diabetes mellitus; untargeted metabolomics

DOI:  https://doi.org/10.1021/acs.jproteome.6c00446
Sci Prog. 2026 Oct-Dec;109(4):109(4): 368504261507162

Machine learning-based identification of depression in older patients with type 2 diabetes mellitus: A multicenter retrospective study.

Xiao Zhang, Juntao Tan, Manni Xiao, Yuxin He, Si Tang.

  ObjectiveType 2 diabetes mellitus (T2DM) and depression frequently co-occur in older adults, creating a bidirectional relationship that complicates disease management and worsens clinical outcomes. Early identification of depression in this vulnerable population remains challenging due to the absence of practical, externally validated screening tools tailored to routine clinical settings in China.MethodsIn this multicenter retrospective study, we developed and validated nine machine learning classifiers to identify clinically documented depression in older patients (aged ≥ 65 years) with T2DM during hospitalization, using electronic health records from 31,980 individuals across six centers. Candidate predictors encompassed demographic characteristics, comorbidities, lifestyle factors, and 15 composite laboratory biomarkers. Least absolute shrinkage and selection operator (LASSO) regression was applied for feature selection, and model performance was evaluated via discrimination (area under the receiver operating characteristic curve (AUC)), calibration, and decision curve analysis across training, internal validation, and institutionally independent external validation sets.ResultsSeven predictors-age, sex, hypertension, cerebrovascular disease (CeVD), drinking history, surgical history, and neutrophil-to-albumin ratio (NAR)-were retained as selected predictors associated with model-predicted depression risk. All nine algorithms except SVM showed comparable discriminative performance in the external validation set (AUCs 0.769-0.772). XGBoost achieved AUCs of 0.836 (95% CI: 0.829-0.843), 0.784 (95% CI: 0.771-0.797), and 0.772 (95% CI: 0.756-0.787) in the training, internal validation, and external validation sets, respectively, which was comparable to logistic regression (LR) (AUC = 0.772) in the external validation set. XGBoost was therefore selected as a representative model for SHAP-based interpretability analysis.ConclusionsThe machine learning model, integrating routinely available clinical and laboratory variables, may offer an interpretable and institutionally externally validated tool for depression identification in older Chinese patients with T2DM. Its discriminative performance was moderate (external validation AUC = 0.772) and comparable to that of conventional LR, with no clear advantage of more complex algorithms and only minimal differences among most models. Validation in other geographic regions and healthcare systems is warranted.

Keywords:  SHAP; depression; machine learning; multicenter; prediction model; type 2 diabetes mellitus

DOI:  https://doi.org/10.1177/00368504261507162
J Med Syst. 2026 Nov 05. pii: 159. [Epub ahead of print]50(1):

Can Retrieval-Augmented Generation Improve the Potential of the LLMs in Diabetes Health Management? An Assessment Based on Health Management Tasks and Question Complexity.

Shanshan Guo, Kuiyou Chen, Hong Wu, Qing Ye.

  Retrieval-Augmented Generation (RAG) has shown significant potential in enhancing the performance of large language models (LLMs) in knowledge-intensive NLP tasks. However, in medical Q&A, particularly in post-diagnosis diabetes management, it remains unclear whether RAG can improve the quality of responses. We developed the DiabMedQA dataset to evaluate the optimization performance of retrieval-augmented LLMs (RA-LLMs) in various Q&A scenarios, encompassing three levels of question complexity and four health management tasks. Our experiments revealed that RA-LLMs achieved higher semantic similarity scores and were generally preferred by clinicians over standard LLMs. In the semantic evaluation, Qwen2.5-14B performed the best, with its BERTScore improving by an average of 10.56% after applying RAG. Notably, the combined analysis showed significant BERTScore gains at all complexity levels, with significantly smaller gains at higher complexity. High-complexity improvements were significant for only one of the four models. In diet advice tasks, the RA-LLM's BERTScore was below 0.8, the lowest among all tasks. Human evaluation showed that Clinicians preferred retrieval-augmented Qwen2.5-14B responses over its baseline responses. Our findings support the value of RAG for post-diagnosis diabetes Q&A and highlight question complexity as an important consideration when applying retrieval augmentation.

Keywords:  Diabetes health management; Health management tasks; Large language models; Question complexity; Retrieval-augmented generation

DOI:  https://doi.org/10.1007/s10916-026-02485-z
Diabetes Metab Syndr Obes. 2026 ;19 616432

AI-Delivered Short Video Cognitive Behavioral Therapy for Type 2 Diabetes Mellitus - A Randomized Controlled Trial Protocol.

Yijing Zhao, Su Wang, Yu Yang, Wen Cao, Kun Chen, Jie Liu, Chao Liu, Kun Wang.

   Background: This article presents a protocol for a randomized, controlled, multicenter trial that aims to evaluate whether the cognitive behavioral therapy (CBT) based on artificial intelligence (AI) short video push technology can improve glycemic control in patients with type 2 diabetes mellitus (T2DM) more effectively than traditional standard care. CBT has been shown to improve glycemic control in patients with T2DM. To help T2DM patients have and preserve behavioral changes for a long time, we have developed a self-management application TANKON. Patients can receive CBT of diabetes-related knowledge and skills by watching different short videos automatically played in the application, and the orders and content of short videos are precisely calculated based on AI technology.
Methods: This is a non-blinded, multicenter, parallel-group, randomized controlled trial with 1:1 equal randomization. A total of 298 participants (149 per group) will be enrolled from seven endocrinology departments across general hospitals in China. The intervention group (n = 149) will receive CBT via the TANKON app, an AI-powered platform that automatically delivers short educational videos. The control group (n = 149) will receive care-as-usual. In the following 12-week follow-up, the patients in the intervention group will choose to prolong or discontinue the use of the application. The primary endpoint is the difference in hemoglobin A1c (HbA1c) change from baseline to 12 weeks for the intervention group compared with those assigned to the control group. Safety assessment includes hypoglycemia events and other adverse events.
Conclusion: This trial evaluates CBT based on AI short video push technology which has the potential to provide a scalable treatment option for patients with T2DM.
Trial Registration: Trial-ID: ChiCTR2400080050, registration date: 18 January 2024 (https://www.chictr.org.cn/showproj.html?proj=216374).

Keywords:  artificial intelligence; cognitive behavioral therapy; self-management; smartphone application; type 2 diabetes mellitus

DOI:  https://doi.org/10.2147/DMSO.S616432
Electrophoresis. 2026 Nov 03.

Exploratory Evidence of a Gene-Smoking Interaction Involving MTR rs1805087 in Type 2 Diabetes Mellitus.

João Pedro Costa Roriz, Leandro Prado Assunção, Nayane Soares de Lima, Diolina Gonçalves da Silva, Rodrigo da Silva Santos, Angela Adamski da Silva Reis.

  Type 2 diabetes mellitus (T2DM) is the most common form of diabetes worldwide. In Brazil, diabetes affects 7.7% of adults and represents a substantial public health and economic burden. Variants in the MTR gene may disrupt one-carbon metabolism, potentially contributing to hyperhomocysteinemia and increased susceptibility to T2DM. A case-control study involving 219 age- and sex-matched individuals was conducted to evaluate the association between the MTR rs1805087 variant and susceptibility to T2DM. Associations under different genetic inheritance models were examined using χ2 tests and odds ratios (95% confidence interval [CI]). In parallel, machine-learning (ML) algorithms were applied, with predictive performance assessed using accuracy, precision, sensitivity, and receiver operating characteristic-area under the curve (ROC-AUC). The variant was not independently associated with T2DM. However, among individuals with T2DM, genotype distributions differed according to smoking status, with significant associations under the codominant and recessive genetic models. Exploratory ML analyses showed limited discriminative ability in the independent test set, with ROC-AUC values of 0.555 for K-nearest neighbors (KNN), 0.524 for support vector machine (SVM), and 0.503 for random forest (RF). All models exhibited substantial declines in performance compared with cross-validation results obtained in the training set. MTR rs1805087 was not independently associated with T2DM in this studied population. However, differences in genotype distributions according to smoking status among individuals with T2DM may reflect context-dependent genetic effects. Exploratory ML analyses showed limited predictive performance but identified smoking as one of the most influential predictors across all evaluated algorithms.

Keywords:  Type 2 diabetes mellitus; machine learning; single nucleotide variant; smoking

DOI:  https://doi.org/10.1002/elps.70184
medRxiv. 2026 Oct 29. pii: 2026.10.26.26366562. [Epub ahead of print]

DiaBench: A Benchmarking Framework to Evaluate the Performance of LLMs in T2D Patient Education.

Yang Li, Syna Syna, Chuyi Zhang, Bhavana Kunisetty, Elizabeth Healey, Agatha F Scheideman, Mandy M Shao, Han Meng, Yu Kuei Lin, Jacob Kohlenberg, Yangfu Zhang, David C Klonoff, Anna Simos, Marina Basina, Michael P Snyder, Haipeng Chen, Tao Wang.

Background: Effective type 2 diabetes (T2D) education improves patient outcomes, but clinician time constraints limit what can be covered during encounters. Large language models (LLMs) offer a scalable way to answer patient questions, yet their accuracy and reliability in T2D education across health literacy levels remain poorly characterized.
Methods: We developed DiaBench, a standardized benchmarking framework built from 96 question- and-answer pairs collected by clinicians from representative real-world patient encounters. GPT-5 transformed each pair into multiple-choice questions (MCQs) under three configurations (1-3) simulating different question formats and levels of user background. A five-member panel of certified diabetes care and education specialists (CDCESs), medical doctors (MDs), and endocrinologists validated the transformations, rating question rephrasing, answer rephrasing, and answer correctness. We then evaluated seven LLMs (GPT-5 1 , Gemini-2.5-Flash 2 , Cohere Command 3 , Llama-4 4 , DeepSeek-R1 5 , Perplexity Sonar 6 , Mistral-Large-2.0 7 ) for accuracy and self-consistency across three runs.
Results: (i) Expert validation showed a consensus over MCQ transformation. For configuration 1, 98.1% (SD 0.01) of rephrased answers were appropriate, 98.1% (SD 0.01) of MCQ answers were correct, and 95.0% (SD 0.05) of rephrased questions were appropriate; corresponding values for configuration 2 were 96.3% (SD 0.05), 97.5% (SD 0.02), and 89.8% (SD 0.14). (ii) across benchmarking for configurations 1 and 2, most models achieved high accuracy (≥95%), but Mistral-Large-2.0, Llama 4, and Perplexity Sonar showed substantially lower self-consistency (60.4%-91.7%) than other models (≥95%). All models degraded in both accuracy and self-consistency on configuration 3.
Conclusions: DiaBench quantitatively differentiates LLM capability in T2D patient education, with Gemini, GPT-5, and DeepSeek-R1 performing strongly and other models suboptimally. DiaBench provides a transferable platform for evaluating existing and newly developed LLMs, addressing the need for standardized benchmarking before clinical deployment of LLM-based systems in patient education.

DOI: https://doi.org/10.64898/2026.10.26.26366562