bims-aukdir Biomed News
on Automated knowledge discovery in diabetes research
Issue of 2026–08–30
thirty-two papers selected by
Mott Given
  1. Correction: Machine learning versus traditional regression models for predicting diabetic retinopathy screening adherence among community-dwelling older adults with diabetes.
  2. Machine learning based prediction of gestational diabetes mellitus using early pregnancy biomarkers and clinical data.
  3. Uncertainty-Aware AI-Assisted Diabetic Retinopathy Grading from Fundus Images with Ordinal Conformal Prediction.
  4. An Optimized and Explainable Machine Learning Framework for Diabetes Prediction Using Marine Predators Algorithm and SHAP.
  5. Evaluation and Comparison of Machine Learning Methods for Type 2 Diabetes Classification and Associated Factors.
  6. Machine learning versus traditional regression models for predicting diabetic retinopathy screening adherence among community-dwelling older adults with diabetes.
  7. Multidimensional feature fusion in longitudinal physical examination data: a machine learning framework for classification of type 2 diabetes.
  8. Artificial intelligence in prediabetes care: applications in screening, risk prediction, and lifestyle intervention.
  9. REFA-DINO: A Relative-Efficient Fusion Attention Framework for Binary and Multiclass Diabetic Retinopathy Classification.
  10. Enhanced diabetic retinopathy detection via multimodal fusion of fundus imaging features and clinical demographic information using verifiable convolutional neural network.
  11. Grading-labour cost analysis of AI-assisted versus human-only diabetic retinopathy screening in two Danish healthcare settings.
  12. Dual-modality ultrasound radiomics model for classifying diabetic peripheral neuropathy in type 2 diabetes: a multicenter prospective study.
  13. Correction: Artificial intelligence for early detection of diabetic retinopathy: A vision transformer-based approach.
  14. Reliability-aware fusion of ocular thermal imaging and radiometric features for proof-of-concept classification of diabetic retinopathy vs. healthy controls.
  15. A multi-level machine learning pipeline for prediction and prioritization of hybrid insulin peptides in type 1 diabetes.
  16. Correction: Associations of dietary patterns and obesity with type 2 diabetes in elderly Chinese men: a machine learning approach.
  17. Quantum machine learning in diabetes and metabolism: Applications and translational challenges.
  18. Development, validation, and visualization of a machine learning-based predictive model for sarcopenia risk in patients with diabetes mellitus.
  19. Identification of Progressive Islet Biomarkers for Type 2 Diabetes by Integrated Transcriptomics and Mendelian Randomization: Alterations Spanning the Normal Glucose Tolerance-Impaired Glucose Tolerance-Type 2 Diabetes Continuum.
  20. Weight-to-waist circumference index enhances interpretable machine learning models for screening depressive symptoms in middle-aged and elderly adults comorbid with hypertension and diabetes.
  21. Quantifying the imputation paradox and XAI inconsistency in multi-source diabetes prediction: a 353,680-record leakage-free stacking ensemble with dynamic routing architecture.
  22. Artificial intelligence-integrated multimodal retinal imaging for early detection and risk stratification of systemic vascular and neurodegenerative diseases.
  23. Development and Validation of an Interpretable Machine Learning Model for Predicting In-Hospital Mortality in Diabetic Patients with Sepsis-Associated Acute Kidney Injury.
  24. Reflection- and Curriculum-Based Instruction Tuning for Diabetes-Specialized Large Language Models: Model Development and Evaluation Study.
  25. Applying Artificial Intelligence to Childhood Obesity: T2DM and MASLD Risk Predictive Models.
  26. Merging artificial intelligence and retinal vascular geometric parameters: A novel tool for diagnosis and prognosis prediction of diabetic nephropathy.
  27. Machine learning prediction models for diabetic kidney disease: A systematic review and meta-analysis.
  28. FLNA promotes the progression of diabetic retinopathy by increasing M1 macrophages.
  29. Glycosylation-Associated Macrophage Signatures Define a Diagnostic Model for Diabetic Retinopathy via Single-Cell Analysis.
  30. Development of a Shapley Additive Explanations (SHAP)-Interpretable Model for Cardiovascular Risk Stratification in Type 2 Diabetes Mellitus: Implications of MDA-Modified Protein Adducts and Autoantibodies.
  31. The reliability-clarity paradox in vascular digital health: A nurse-led evaluation of ChatGPT-4's accuracy in diabetic foot ulcer management.
  32. Intramuscular Fat Assessed by AI-Assisted Muscle Ultrasound: Association With Cardiometabolic Risk Factors and Diabetic Nephropathy in Diabetes Mellitus.
  1. Front Endocrinol (Lausanne). 2026 ;17 1958085
    Correction: Machine learning versus traditional regression models for predicting diabetic retinopathy screening adherence among community-dwelling older adults with diabetes.
    Frontiers Production Office
      [This corrects the article DOI: 10.3389/fendo.2026.1909594.].
    Keywords:  community-dwelling older adults; diabetic retinopathy screening adherence; logistic regression; machine learning; protection motivation theory; random forest
    DOI:  https://doi.org/10.3389/fendo.2026.1958085
  2. Bioinformation. 2026 ;22(5): 2838-2842
    Machine learning based prediction of gestational diabetes mellitus using early pregnancy biomarkers and clinical data.
    Karnaditya Rana, Bikramaditya Mukherjee, Ajith Antony.
      Gestational diabetes mellitus (GDM) is a common pregnancy-related condition that can lead to significant maternal and neonatal complications, but conventional screening methods often delay diagnosis. Therefore, it is of interest to develop a machine learning model for early prediction of GDM using first-trimester biomarkers and clinical data. Hence, a prospective study of 100 pregnant women was conducted and various machine learning algorithms were trained to predict GDM. The random forest model showed the best performance with an accuracy of 86% and an AUC of 0.90. Thus, we show the potential of machine learning in enabling early prediction and timely intervention for GDM, improving maternal and neonatal outcomes.
    Keywords:  Gestational diabetes mellitus (GDM); early pregnancy biomarkers; insulin resistance; machine learning (ML); predictive modeling
    DOI:  https://doi.org/10.6026/973206300222838
  3. Diagnostics (Basel). 2026 Aug 19. pii: 2645. [Epub ahead of print]16(16):
    Uncertainty-Aware AI-Assisted Diabetic Retinopathy Grading from Fundus Images with Ordinal Conformal Prediction.
    Umar Hasan, Muhammad Ali Nayeem, Turki G Alghamdi.
      Background: Artificial intelligence (AI) systems for diabetic retinopathy (DR) grading require reliable uncertainty estimates when applied to fundus images outside the development dataset. We evaluated whether conformal prediction can provide structured set-valued outputs and whether internal uncertainty calibration remains reliable during external evaluation. Methods: EfficientNet-B0, ResNet-50, and Vision Transformer (ViT-Base) classifiers were trained on APTOS 2019. Split-conformal predictors were calibrated exclusively on held-out APTOS images using three categorical scores, LAC, APS, and RAPS, and an ordinal score restricted to adjacent severity grades. Performance was assessed internally on APTOS and externally on IDRiD, with an independent five-seed ViT replication extending evaluation to Messidor-2, at target coverages of 90% and 95%. Results: Coverage was approximately nominal internally but decreased on both external datasets. On IDRiD, the largest deficit occurred for severe DR (grade 3). The ordinal method produced contiguous intervals in 100% of cases and achieved the highest grade-3 coverage in every tested backbone-risk configuration. For ResNet-50 at 95% target coverage, grade-3 coverage increased from 0.750 with APS to 0.945 with the ordinal method, while average set size increased from 2.89 to 3.03. In the independent ViT replication on Messidor-2, ordinal marginal coverage exceeded APS at both targets (0.676 versus 0.632 and 0.741 versus 0.714). Conclusions: Internal calibration did not ensure reliable class-specific uncertainty during external evaluation. Ordinal prediction sets improved structural coherence and mitigated severe-grade undercoverage, but did not restore formal coverage guarantees after dataset shift.
    Keywords:  artificial intelligence; conformal prediction; dataset shift; deep learning; diabetic retinopathy; diagnostic support; external validation; fundus photography; ordinal classification; uncertainty quantification
    DOI:  https://doi.org/10.3390/diagnostics16162645
  4. J Pers Med. 2026 Aug 10. pii: 423. [Epub ahead of print]16(8):
    An Optimized and Explainable Machine Learning Framework for Diabetes Prediction Using Marine Predators Algorithm and SHAP.
    Alifa Nasrin, Muhammad Bin Asif, Ramasamy Naidu, Afzal Haq Asif, Muhammad Shahzad Chohan, Gausal Azam Khan, Md Arifuzzaman, Akm Azad, Muhammad Ali Martuza.
      Background: Diabetes mellitus affects over 500 million people worldwide, yet many machine learning prediction models remain difficult to interpret, limiting their clinical applicability. This study proposes an explainable machine learning framework integrating the Marine Predators Algorithm (MPA) for hyperparameter optimization with SHAP-based explainability to diabetes prediction. Methods: Logistic Regression (LR), Random Forest (RF), and MPA-optimized XGBoost were evaluated using a publicly available Kaggle diabetes dataset of approximately 100,000 records. Statistical significance was assessed using Wilcoxon signed-rank tests with Bonferroni correction for fold-wise cross-validation results, while McNemar's and DeLong's tests were employed for paired comparison of independent test-set predictions and ROC-AUC values, respectively. Performance was assessed using accuracy, precision, recall, F1-score, specificity, ROC-AUC, and Brier score. SHAP was used to provide global and local model explanations. Results: The MPA-optimized XGBoost model achieved the highest performance, with 96.72% accuracy, 97.70% precision, 95.70% recall, 96.71% F1-score, and 99.56% ROC-AUC, significantly outperforming LR and RF (p < 0.001). The model demonstrated good calibration with a Brier score of 0.0262. SHAP analysis identified HbA1c level, blood glucose level, and age as the most influential predictors, while interaction analysis indicated a synergistic relationship between HbA1c and blood glucose. Conclusions: The proposed framework demonstrated strong predictive performance and interpretable model behavior on the publicly available diabetes dataset used in this study. These findings indicate the potential of MPA-based optimization combined with SHAP explainability for supporting transparent machine learning research in diabetes prediction. However, additional external validation using independent clinical datasets is required before considering the framework for clinical decision support or real-world deployment.
    Keywords:  SHAP; SMOTE; XGBoost; class imbalance; clinical decision support; diabetes prediction; explainable artificial intelligence; hyperparameter optimization; machine learning; marine predators algorithm
    DOI:  https://doi.org/10.3390/jpm16080423
  5. Healthc Inform Res. 2026 Jul;32(3): 214-223
    Evaluation and Comparison of Machine Learning Methods for Type 2 Diabetes Classification and Associated Factors.
    Masoumeh Dadashpour, Mohammad Yousefi, Sohrab Effati, Somayeh Ghiasi Hafezi.
       OBJECTIVES: Type 2 diabetes mellitus (T2DM) is a prevalent chronic metabolic disorder associated with serious complications, including nephropathy, cardiovascular disease, retinopathy, and neuropathy. Given its increasing incidence and the complexity of associated factors-such as obesity, metabolic syndrome, and sedentary lifestyle-accurate identification is essential. This study aimed to evaluate and compare the performance of several machine learning algorithms to identify key associated factors and detect individuals with T2DM within this dataset.
    METHODS: A publicly available dataset from Kaggle, comprising health records of 99,982 individuals, was used. Five supervised machine learning models were evaluated: Bayesian ridge regression, logistic regression, extreme gradient boosting (XGBoost), artificial neural networks, and random forest. Each model was trained and evaluated to assess classification performance. Performance was measured using the area under the receiver operating characteristic curve (AUC-ROC) and accuracy. SHapley Additive Explanations (SHAP) values were used to interpret model outputs and identify the most influential features.
    RESULTS: Among the five models, XGBoost demonstrated the highest performance, achieving an accuracy of 96% and an AUC-ROC of 0.98. SHAP analysis identified hemoglobin A1c, blood glucose, age, body mass index, and sex as the most influential predictors of T2DM.
    CONCLUSION: XGBoost was the most effective algorithm for identifying individuals with T2DM in this dataset. It also provided insights into the relative importance of clinical features, supporting more precise classification. However, results should be interpreted with caution until validated in independent cohorts.
    Keywords:  Biomarkers; Machine Learning; Predictive Value of Tests; Risk Factors; Type 2 Diabetes Mellitus
    DOI:  https://doi.org/10.4258/hir.2026.32.3.214
  6. Front Endocrinol (Lausanne). 2026 ;17 1909594
    Machine learning versus traditional regression models for predicting diabetic retinopathy screening adherence among community-dwelling older adults with diabetes.
    Yawen Wang, Lanting Xia, Guiling Geng, Chunxing Ge, Xiangyun Qian, Haiyan Shen.
       Background: Diabetic retinopathy (DR) is a leading cause of preventable vision loss among patients with diabetes, yet screening adherence remains suboptimal. Existing studies have mainly focused on clinical or sociodemographic determinants, with limited evidence integrating psychological mechanisms. The application of Protection Motivation Theory (PMT)-based constructs combined with machine learning for predicting DR screening adherence remains underexplored in community populations.
    Objective: This study compared machine learning models (decision tree and random forest) with traditional logistic regression for predicting DR screening adherence among community-dwelling older adults with diabetes, incorporating a validated PMT-based questionnaire as a key psychological predictor.
    Methods: A cluster random sampling design recruited 1,021 older adults with diabetes from four community health centers in Nantong, China (March-October 2025). Data included sociodemographic characteristics, clinical indicators, health behaviors, and PMT-based constructs. Participants were classified as good (n = 159) or poor adherence (n = 862). SMOTE and class weighting addressed class imbalance. Models were evaluated using 10-fold cross-validation. Logistic regression, decision tree, and random forest models were built using identical predictors. Decision curve analysis assessed clinical utility.
    Results: The random forest model achieved a marginally higher AUC (0.774, 95% CI: 0.678-0.869) compared with logistic regression (AUC = 0.751, 95% CI: 0.678-0.824) and decision tree (AUC = 0.709, 95% CI: 0.596-0.821); however, DeLong's tests indicated no statistically significant differences (all p > 0.05). The decision tree exhibited the best calibration (lowest Brier score = 0.1142). DCA indicated that random forest provided the highest net benefit across most threshold probabilities. Multivariable analysis identified history of ocular disease (OR = 3.529, 95% CI: 2.430-5.139) as the strongest positive predictor, while higher HbA1c, lower self-efficacy, higher perceived severity, absence of exercise therapy, and smoking were associated with poorer adherence.
    Conclusion: Machine learning models demonstrated comparable discriminative performance to traditional logistic regression for predicting DR screening adherence, while offering distinct profiles in calibration and clinical net benefit. Integration of PMT-based psychological constructs with clinical and behavioral factors provides a multidimensional framework for understanding screening behavior and supports risk stratification for precision diabetes care.
    Keywords:  community-dwelling older adults; diabetic retinopathy screening adherence; logistic regression; machine learning; protection motivation theory; random forest
    DOI:  https://doi.org/10.3389/fendo.2026.1909594
  7. Front Public Health. 2026 ;14 1841719
    Multidimensional feature fusion in longitudinal physical examination data: a machine learning framework for classification of type 2 diabetes.
    Yan Zhou, Bijun Sun, Xiang Cai, Jiandong Qi.
       Background: Machine learning-based risk classification of type 2 diabetes mellitus (T2DM) has become a prevailing research direction in intelligent healthcare. Nevertheless, current studies suffer from two key limitations. First, most existing models are built on cross-sectional data and cannot capture the dynamic progression characteristics of T2DM. Second, traditional longitudinal time-series models are commonly adopted without sufficiently exploring the deep interactive relationships among multi-temporal physical examination features, limiting model predictive performance.
    Methods: To tackle the above drawbacks, this study proposes a novel multi-dimensional feature extraction and fusion network (MFFNet) for T2DM early risk prediction based on sequential annual physical examination data. The proposed architecture comprehensively mines latent feature information from three complementary dimensions: the horizontal interaction of multiple examination indicators within a single period, the longitudinal dynamic evolution pattern of individual indicators across consecutive years, and the cross-temporal synergistic correlations among different indicators across multiple periods. Experiments were conducted on a real-world longitudinal dataset consisting of three consecutive years of physical examination records from the Health Management Center of Beijing Hospital, which presents an extremely imbalanced sample distribution with a positive-negative ratio of 1:8.34.
    Results: Comparative experiments with state-of-the-art time-series models (RNN, LSTM, GRU, and Transformer) demonstrate the superiority of MFFNet. The proposed model achieved a sensitivity of 0.8766, a specificity of 0.7082, a negative predictive value (NPV) of 0.9799, and a PR-AUC of 0.4322. Furthermore, SHAP interpretability analysis identified core predictive features for T2DM risk assessment, including age, creatinine (Cr), waist circumference (WC), alanine aminotransferase (ALT), triglycerides (TG), and systolic blood pressure (SBP).
    Conclusion: The proposed MFFNet can effectively implement accurate T2DM classification and early risk prediction using longitudinal physical examination sequences. This multi-dimensional feature fusion strategy substantially improves the mining capability of temporal healthcare data. The lightweight and cost-effective MFFNet provides a reliable artificial intelligence-assisted solution for large-scale early pre-screening and risk intervention of T2DM in primary clinical healthcare.
    Keywords:  intelligent medicine; longitudinal study; multi-dimensional feature extraction and fusion; physical examination data series; type 2 diabetes
    DOI:  https://doi.org/10.3389/fpubh.2026.1841719
  8. Front Endocrinol (Lausanne). 2026 ;17 1885905
    Artificial intelligence in prediabetes care: applications in screening, risk prediction, and lifestyle intervention.
    Runxuan Chen, Min Song, Tong Xiang.
      Prediabetes is a highly prevalent intermediate metabolic state and a major public health target for preventing type 2 diabetes mellitus (T2DM). However, current approaches to prediabetes screening, risk stratification, and lifestyle management remain limited by inconsistent diagnostic definitions, incomplete case detection, heterogeneous progression risk, and the resource-intensive nature of conventional face-to-face prevention programmes. Artificial intelligence (AI), including machine learning, deep learning, explainable AI, and algorithm-driven digital interventions, is increasingly being explored as a tool to address these gaps. This narrative review used a structured search of PubMed, Web of Science Core Collection, and Embase for studies published from January 2010 to April 2026, selecting articles that evaluated AI-assisted or algorithm-driven approaches for prediabetes screening, progression risk prediction, or lifestyle intervention and reported relevant model performance, validation, or intervention outcomes. Current evidence indicates that AI-based models can improve discrimination beyond traditional risk scores by integrating routine clinical data, longitudinal electronic health records, continuous glucose monitoring profiles, wearable-derived behavioural signals, and emerging molecular biomarkers. Some externally validated models have shown clinically relevant performance for identifying individuals at high risk of progression and for guiding more targeted preventive strategies. In parallel, fully or semi-automated digital programmes delivered through mobile applications, web platforms, connected scales, and sensor-based feedback systems have demonstrated potential to support lifestyle change, improve engagement, and reduce reliance on labour-intensive counselling. Nevertheless, translation into routine care remains constrained by heterogeneity in prediabetes definitions, limited external validation across diverse populations, uncertain long-term effectiveness, geographical imbalance in evidence, privacy concerns, and the need for stronger governance frameworks. Overall, AI should be viewed as an assistive technology that may support earlier detection, more precise risk stratification, and scalable lifestyle management in prediabetes. Further multi-centre, prospective, and implementation-focused studies are needed to establish clinical utility, equity, safety, and cost-effectiveness.
    Keywords:  artificial intelligence; diabetes prevention; lifestyle intervention; machine learning; prediabetes; risk stratification; type 2 diabetes
    DOI:  https://doi.org/10.3389/fendo.2026.1885905
  9. Int J Biomed Imaging. 2026 ;2026 4374759
    REFA-DINO: A Relative-Efficient Fusion Attention Framework for Binary and Multiclass Diabetic Retinopathy Classification.
    Hussain Dawood, Muniba Noreen, Ali Javed, Hafsa Ilyas, Usman Muhammad, Hassan Dawood.
      Diabetic retinopathy (DR) has become a major cause of preventable blindness, and its early detection is crucial for maintaining vision in diabetic patients. Existing computer-aided diagnosis (CAD) systems still face challenges in detecting early-stage lesions, integrating local and global features, and maintaining consistent performance in binary and multiclass settings. To address these limitations, this study proposes a relative-efficient fusion attention framework (REFA-DINO) that integrates a convolutional local-feature learning module with a transformer-based global contextual method. The architecture employs a relative-efficient fusion adapter (REFA) block comprising two branches, EfficientNet and attention-enhanced patch embedding. EfficientNet extracts lesion-sensitive retinal features, and an attention-enhanced patch-embedding module is used to generate token representations. An adaptive relative position-aware cross-attention mechanism is introduced to enable effective interaction among heterogeneous features, which are further processed through the DINO-based transformer module. The extracted global representations are passed to the linear classification head for DR classification. REFA-DINO is evaluated on two diverse and large-scale fundus imaging datasets, EyePACS and APTOS, for binary and multiclass classification. The proposed framework achieved accuracies of 98.60% and 93.77% for binary classification and 85.52% and 87.51% for multiclass classification on the APTOS and EyePACS datasets, respectively. A comparative analysis against state-of-the-art methods, cross-corpora evaluation, and an ablation study are also conducted to demonstrate the effectiveness of the proposed REFA-DINO. Compared with state-of-the-art methods, the REFA-DINO framework enhances binary classification on both the EyePACS and APTOS datasets, with significant improvements over the strongest baselines. The proposed approach attains up to 6.19% and 9.39% improvement in the binary and multiclass classification accuracy on EyePACS. Likewise, it achieves an improvement of 2.64% on the binary classification score on APTOS with competitive multiclass classification performance. Experimental findings highlight the robustness and generalizability of the REFA-DINO framework in classifying DR.
    Keywords:  DINOv2; EfficientNet; computer-aided diagnosis; diabetic retinopathy; medical imaging classification; vision transformers
    DOI:  https://doi.org/10.1155/ijbi/4374759
  10. Comput Biol Med. 2026 Aug 27. pii: S0010-4825(26)00456-7. [Epub ahead of print]214 111892
    Enhanced diabetic retinopathy detection via multimodal fusion of fundus imaging features and clinical demographic information using verifiable convolutional neural network.
    Amanulla Mohammad, N Ramakrishnaiah.
      Traditional systems of ocular disease diagnosis and many deep learning-based systems are limited in their practice to analyzing fundus images from a single eye, often using raw, unprocessed data. This fails to include the critical correlation between bilateral eyes, has several issues such as low image contrast, noise, and class imbalance, and generally concentrates on the detection of no more than a single disease. In order to bridge the gap, Enhanced Diabetic Retinopathy Detection through Multimodal Integration of Fundus Imaging Features and Clinical Demographic Information utilizing Verifiable Convolutional Neural Network (DD-FIF-CDI-VCNN) is proposed. In the initial phase, the fundus images of the eyes along with their corresponding demographic information are sourced from the Ocular Disease Recognition, which acts as the main feed for the proposed framework. During the pre-processing stage, it uses Robust Consensus Tobit Kalman Filtering (RCTKF) to resize the fundus images, normalizes the pixel intensities, and reduces noise in the process of quality improvement for subsequent feature extraction. The pre-processed fundus images are provided to the ResNet-fused External Attention Network (ResfEANet) based feature extraction that extracts rich visual and texture features from the retinal images. In parallel to this, the demographic attributes are fed to a feature extraction module using TabNet which converts the structured affected person facts into a meaningful function vector. This aims at making sure that relevant patient-specific styles are encoded and prepared for integration with visual functions. Then, the visible and texture capabilities extracted from the fundus photos are combined with the demographic attributes the usage of Hierarchical Multi-Scale Feature Fusion (HMSFF). This creates a complete representation that leverages both imaging and structured affected person statistics. This blended feature set is the input to the categorization model for correct diabetic retinopathy detection. Next, this fused characteristic vector, combining the features extracted from the fundus images with their corresponding demographic attributes, is passed to a Verifiable Convolutional Neural Network (VCNN)-based classifier model. This comprehensive multi-modal representation ensues, rendering the VCNN model capable of accurate detection of diabetic retinopathy classifying normal and diabetes, thereby enhancing predictability and robustness compared with the unimodal approach. The VCNN parameters are fine-tuned by the Warthog Optimization Algorithm (WOA) to ensure improved convergence and better detection performance. The proposed DD-FIF-CDI-VCNN method is analyzed under performance metrics: accuracy, precision, recall, F1-Score, Area under Curve (AUC) and Error rate when compared with existing models.
    Keywords:  Fundus image; Hierarchical multi-scale feature fusion; Ocular disease recognition; ResNet-fused external attention network and verifiable convolutional neural network; Robust consensus tobit kalman filtering
    DOI:  https://doi.org/10.1016/j.compbiomed.2026.111892
  11. Acta Ophthalmol. 2026 Aug 22.
    Grading-labour cost analysis of AI-assisted versus human-only diabetic retinopathy screening in two Danish healthcare settings.
    Lars Grønlykke, Sebastian Wiberg, Alex Carter, Peter Heilbo Ratgen, Simon Joel Lowater, Jakob Kristian Holm Andersen, Thiusius Rajeeth Savarimuthu, Jakob Grauslund.
       PURPOSE: Diabetic retinopathy (DR) screening is essential to prevent vision loss, yet rising diabetes prevalence threatens to outpace ophthalmology capacity. Artificial intelligence (AI) systems can triage retinal images to reduce clinician workload, but economic evidence from high-income, tax-funded health systems remains limited. This study presents a cost-minimisation analysis (CMA) of clinician grading labour comparing deep-learning models (DLM)-assisted screening versus human-only DR screening.
    METHODS: This CMA evaluates two Danish healthcare settings differing in DR prevalence: In tertiary diabetes centres with DR grade 2-4 prevalence of 41% and in publicly contracted private ophthalmology practices with DR grade 2-4 prevalence of 6%. Costs were estimated from the healthcare-system perspective over a one-year horizon, including clinician time and wages for image assessment and verification. Sensitivity analyses explored variations in staff time, DR prevalence, DLM false-positive rate, and reimbursement effect.
    RESULTS: In the tertiary-centre setting (23 696 screens), total annual grading-labour costs were €282 192 for AI-assisted versus €421 425 for human-only screening (33% reduction). In private practice (84 494 screens), the AI-assisted pathway costs were €569 692 versus €1 502 698 (62% reduction). Savings were driven by reduced ophthalmologist time on cases the DLM classified as DR negative. Across all sensitivity analyses, the AI-assisted strategy remained cost-saving.
    CONCLUSION: AI-assisted DR screening was cost-minimising in both high- and low-prevalence settings within Denmark's tax-funded health system, with greater savings in lower-prevalence populations. Tariff structures strongly influence payer impact, suggesting that reimbursement models must align with efficiency gains to realise health-system savings from AI triage in routine DR screening.
    Keywords:  artificial intelligence; cost‐minimisation analysis; deep learning; diabetic retinopathy
    DOI:  https://doi.org/10.1111/aos.70225
  12. Front Endocrinol (Lausanne). 2026 ;17 1900881
    Dual-modality ultrasound radiomics model for classifying diabetic peripheral neuropathy in type 2 diabetes: a multicenter prospective study.
    Rong Xiao, Wanyan Li, Wenqian Qiu, Chenglong Zhang, Enfa Zhao, Tingting Xie, Danqing He.
       Objective: Diabetic peripheral neuropathy (DPN) is a prevalent and disabling complication of type 2 diabetes mellitus (T2DM), yet detection remains constrained by limited accessibility of nerve conduction studies. This study developed and validated a dual-modality ultrasound radiomics model for individualized DPN classification.
    Methods: In total, 253 T2DM patients from three centers were prospectively enrolled between June 2025 and February 2026, allocated to training (n = 122), internal test (n = 53), and external validation (n = 78) cohorts. Radiomics features were extracted from longitudinal B-mode ultrasound and shear wave elastography images of the tibial nerve. After reproducibility filtering, batch-effect correction, and elastic-net feature selection, four machine learning algorithms were compared and the best-performing was used to construct modality-specific radiomics scores (Rad-scores). A combined model integrating Rad-scores with independently associated clinical factors was developed. Discrimination, calibration, clinical usefulness, and SHapley Additive exPlanations (SHAP) interpretability were evaluated.
    Results: Multivariable analysis identified diabetes duration and minimum elastic modulus as independently associated factors, with both Rad-scores significantly associated with DPN status. The combined model demonstrated good discrimination across all cohorts (AUC 0.892 [95% CI 0.828-0.950], 0.824 [0.696-0.925], and 0.838 [0.735-0.926]), outperforming all other models (p < 0.05), with adequate calibration in both training and internal test cohorts (p > 0.05). Decision curve analysis confirmed clinical benefit. SHAP analysis identified diabetes duration as the most influential variable, followed by the shear wave elastography Rad-score.
    Conclusions: The combined model demonstrates favorable diagnostic performance for individualized DPN risk stratification and holds promise as a noninvasive complement to nerve conduction.
    Keywords:  SHapley additive explanations; diabetic peripheral neuropathy; machine learning; radiomics; shear wave elastography
    DOI:  https://doi.org/10.3389/fendo.2026.1900881
  13. PLoS One. 2026 ;21(8): e0357114
    Correction: Artificial intelligence for early detection of diabetic retinopathy: A vision transformer-based approach.
    PLOS One Staff
      [This corrects the article DOI: 10.1371/journal.pone.0350854.].
    DOI:  https://doi.org/10.1371/journal.pone.0357114
  14. Front Med (Lausanne). 2026 ;13 1888280
    Reliability-aware fusion of ocular thermal imaging and radiometric features for proof-of-concept classification of diabetic retinopathy vs. healthy controls.
    Persiya J, Sasithradevi A, Anand Rajendran.
       Introduction: Early detection of diabetic retinopathy (DR) is crucial for preventing permanent vision impairment, yet widely used screening techniques such as fundus photography and optical coherence tomography (OCT) depend on expensive equipment and expert clinical interpretation, restricting their use in large-scale and low-resource environments. To overcome these challenges, this study introduces ReliaFusion-Net, a reliability-aware multimodal framework for proof-of-concept binary classification of eyes with DR vs. healthy controls, using thermal eye images and radiometric temperature features.
    Methods: The proposed ReliaFusion-Net architecture combines bidirectional co-attention with feature-wise linear modulation to effectively model complementary relationships between thermal image features and physiological temperature descriptors. A clinically curated dataset containing 558 thermal eye images, including 278 normal and 280 diabetic samples, was collected with ophthalmologist support. The framework was trained and evaluated for binary classification, and cross-validation experiments were conducted to assess model generalization.
    Results: Using an EdgeNeXt backbone, the proposed model achieved 93.18% accuracy, 93.26% F1-score, and 0.9785 AUC on the test set. Cross-validation results further confirmed strong generalization performance. These findings indicate that the reliability-aware multimodal fusion strategy effectively discriminates eyes with DR from healthy controls using non-invasive thermal imaging data.
    Discussion: By dynamically adapting modality contributions based on reliability, ReliaFusion-Net overcomes limitations associated with traditional fixed fusion strategies and improves classification robustness. Overall, the findings demonstrate proof-of-concept of reliability-aware multimodal learning to distinguish eyes with DR from healthy controls using non-invasive thermal imaging, with potential applicability in tele-ophthalmology settings. Future work focuses on multi-stage disease classification and longitudinal disease progression analysis.
    Keywords:  diabetic retinopathy affected eye; multimodal deep learning; ocular thermography; radiometric temperature features; thermal imaging
    DOI:  https://doi.org/10.3389/fmed.2026.1888280
  15. Front Immunol. 2026 ;17 1917213
    A multi-level machine learning pipeline for prediction and prioritization of hybrid insulin peptides in type 1 diabetes.
    Ilya Kandinov, Dmitry Trukhin, Dmitry Gryadunov, Elena Savvateeva.
      Hybrid insulin peptides (HIPs) are neoepitopes involved in type 1 diabetes (T1D), but their complete repertoire remains unknown. The vast combinatorial space makes experimental screening unfeasible and requires bioinformatics-based prioritization. We developed a multi-level machine learning pipeline for ranking HIP candidates. First, 36 physicochemical and junction-specific features were computed for a reference library of 240 HIPs with known enzyme-linked immunospot (ELISPOT) reactivity, and a baseline Ridge regression model was trained. Next, all possible HIP candidates with 7-9 amino acid residues per fragment were generated from eight pancreatic β-cell secretory granule source proteins, including insulin chains and C-peptide, islet amyloid polypeptide, chromogranin A, neuropeptide Y, and two secretogranins, yielding 1,057,374 candidates. For each source protein, a local weighted XGBoost (Extreme Gradient Boosting) model was trained using Ridge-score-derived pseudo-labels together with weighted ELISPOT-derived and literature-derived reference HIPs. Finally, anchor-calibrated re-ranking was performed in the global model using cosine similarity to positive anchors (n = 46) and negative anchors (n = 210). The Ridge model achieved 5-fold out-of-fold R² = 0.711 and an area under the receiver operating characteristic curve (AUC) of 0.967. The global model produced a prioritized list of 40 HIP candidates, five per source protein. The highest ranks were observed for candidates with right fragments from neuropeptide Y, secretogranins 1 and 2, islet amyloid polypeptide, and chromogranin A. Candidates carrying the insulin fragment on the right side were systematically down-ranked, suggesting asymmetry in HIP formation. The proposed pipeline reduces the HIP search space from more than one million sequences to a limited set of candidates for experimental validation and provides a framework adaptable to other chimeric neoepitopes in autoimmunity.
    Keywords:  XGBoost; bioinformatics; hybrid insulin peptides; machine learning; neoepitopes; peptide prioritization; regression model; type 1 diabetes
    DOI:  https://doi.org/10.3389/fimmu.2026.1917213
  16. Front Nutr. 2026 ;13 1882361
    Correction: Associations of dietary patterns and obesity with type 2 diabetes in elderly Chinese men: a machine learning approach.
    Haowei Sun, Lijin Zhu, Peng Wang, Keqing Yuan, Saida Salima Nawrin, Yufei Cui, Longfei Li.
      [This corrects the article DOI: 10.3389/fnut.2025.1705683.].
    Keywords:  SHAP analysis; dietary patterns; obesity; type 2 diabetes mellitus; unsupervised machine learning
    DOI:  https://doi.org/10.3389/fnut.2026.1882361
  17. Comput Methods Programs Biomed. 2026 Aug 16. pii: S0169-2607(26)00358-5. [Epub ahead of print]286 109609
    Quantum machine learning in diabetes and metabolism: Applications and translational challenges.
    Alessia Riente, Cassandra Serantoni, Michele Maria De Giulio, Stefano Capezzone, Rosita Esposito, Marco De Spirito, Giuseppe Maulucci.
       BACKGROUND AND OBJECTIVE: Quantum machine learning is emerging as a promising extension of artificial intelligence, with potential advantages over classical approaches in handling complex biomedical data. This review aims to evaluate quantum machine learning applications in the detection, prediction, and personalized management of metabolic syndrome and type 2 diabetes mellitus.
    METHODS: We reviewed literature published between 1994 and 6 July 2026, with peer-reviewed journal articles and conference proceedings as the principal evidence base, complemented by selected preprints and technical sources. Quantum machine learning approaches, including quantum support vector machines, quantum neural networks, and quantum echo state networks, were classified and compared with classical counterparts across obesity and early metabolic dysregulation, diabetes diagnosis and glycemic management, and chronic complications.
    RESULTS: Quantum machine learning and hybrid quantum-classical systems demonstrated potential benefits in small-sample and noisy environments typical of wearable and biomedical sensor data. Reported performance gains included improvements in accuracy, robustness, and scalability, though interpretability and reproducibility remain challenges. Hardware limitations associated with noisy intermediate-scale quantum devices, data encoding, and privacy considerations emerged as key barriers to clinical translation.
    CONCLUSIONS: Preliminary studies highlight the promise of quantum machine learning for predictive and personalized management of metabolic syndrome and type 2 diabetes mellitus. However, successful clinical adoption will require robust validation pipelines, regulatory sandboxes, and harmonized compliance frameworks. Domain-specific evaluation metrics and transparent conformity assessments are essential to ensure trustworthy, scalable, and equitable deployment. A staged roadmap is proposed to bridge experimental progress with ethical and regulatory readiness.
    Keywords:  Chronic complications; Digital health; Hybrid quantum-classical models; Metabolic syndrome; Quantum machine learning; Type 2 diabetes mellitus
    DOI:  https://doi.org/10.1016/j.cmpb.2026.109609
  18. Front Med (Lausanne). 2026 ;13 1904187
    Development, validation, and visualization of a machine learning-based predictive model for sarcopenia risk in patients with diabetes mellitus.
    Mingjie Tang, Shiyi Zheng, Hang Xu, Yinghong Li, Chaoming Chen.
       Background: Sarcopenia is a common complication in patients with diabetes mellitus (DM), yet early screening remains challenging due to the lack of accessible tools. This study aimed to develop and externally validate a machine learning (ML) -based predictive model for sarcopenia risk in DM patients, with enhanced interpretability.
    Methods: Data from 1,074 DM patients at a single center (training and internal validation) and two external datasets (NHANES, n = 1,078; CHARLS, n = 547) were used. Seven ML methods were evaluated. Feature selection used LASSO and Boruta. Model performance was assessed by AUC, calibration (Brier score), and decision curve analysis. SHAP was applied to explain the final model.
    Results: Six predictors (age, sex, BMI, hemoglobin, uric acid, creatinine) were identified. In internal validation, LightGBM achieved the highest AUC (0.973). However, on both external datasets, logistic regression showed superior and stable performance (NHANES AUC = 0.967, CHARLS AUC = 0.949), with lower Brier scores (0.042 and 0.061) and favorable net benefit. LightGBM's performance decreased externally (AUC 0.873 and 0.868). Therefore, logistic regression was selected as the final model. SHAP analysis revealed that low BMI, older age, and low creatinine were the strongest predictors of sarcopenia.
    Conclusion: Using six readily available clinical features, the logistic regression model shows robust generalizability across populations. This clinically practical tool supports early risk identification in diabetic patients, enabling timely interventions to mitigate sarcopenia-related complications.
    Keywords:  SHAP; diabetes mellitus; machine learning; predictive model; sarcopenia
    DOI:  https://doi.org/10.3389/fmed.2026.1904187
  19. Biomedicines. 2026 Aug 18. pii: 1850. [Epub ahead of print]14(8):
    Identification of Progressive Islet Biomarkers for Type 2 Diabetes by Integrated Transcriptomics and Mendelian Randomization: Alterations Spanning the Normal Glucose Tolerance-Impaired Glucose Tolerance-Type 2 Diabetes Continuum.
    Nilupaer Aisikaer, Zaoling Liu, Shuya Cai.
      Background: Impaired glucose tolerance (IGT) is the principal prediabetic stage preceding type 2 diabetes mellitus (T2D), yet islet-specific biomarkers capable of tracking progressive molecular changes from normal glucose tolerance (NGT) through IGT to T2D remain unestablished. We sought to identify a multi-gene biomarker panel with monotonically increasing expression and causal support across the full glycemic continuum. Methods: Two human islet transcriptomic datasets (GSE76895, GSE164416; n = 181: NGT 50, IGT 56, T2D 75) were integrated following ComBat batch correction. Candidate biomarkers were identified at the intersection of limma differential expression and weighted gene co-expression network analysis (WGCNA), then refined through a five-algorithm machine learning consensus (LASSO, random forest, XGBoost, SVM-RFE, elastic net). Progressive expression was assessed by the Jonckheere-Terpstra (J-T) trend test. Two-sample Mendelian randomization (MR) with eQTLGen cis-eQTL instruments, Steiger directionality testing, and Bayesian colocalization provided causal inference. A diagnostic model was internally validated via 1000-iteration bootstrap resampling. Cell-type specificity was verified using single-cell RNA sequencing (GSE200044; 127,919 cells). Results: A 12-gene biomarker panel (ALDOB, DKK3, PCOLCE2, KCNE4, INHBA, IRF8, ITGB2, LAPTM5, MYOF, RAMP3, RUNX2, S100A4) was identified, with all members passing Bonferroni-corrected J-T trend testing across the NGT-IGT-T2D axis (p ≤ 2.4 × 10-3). Notably, a direct IGT-versus-NGT transcriptome-wide comparison (11,948 genes) yielded no significant DEGs after FDR correction, indicating that prediabetic islet signals are subtle and detectable only through progressive trend analysis on preselected candidates. Nevertheless, the mean IGT-stage effect size of the 12 hub genes reached 41.9% of the T2D value, with KCNE4 achieving 94.0% (nominal p = 1.75 × 10-4), identifying it as the earliest-altered biomarker. Two-sample MR using whole-blood eQTLs suggested protective effects of genetically proxied MYOF (OR 0.999, FDR = 1.68 × 10-4) and RUNX2 (OR 0.998, FDR = 1.68 × 10-4) on T2D risk, with Steiger testing supporting an expression-to-disease direction (p < 10-36). However, Bayesian colocalization indicated independent causal variants at both loci (PP.H3 > 0.76, PP.H4 < 0.001), substantially weakening the causal interpretation and suggesting that the MR associations may reflect linkage disequilibrium rather than shared causal biology. The panel achieved a bootstrap-corrected AUC of 0.833 (apparent 0.879) with PR-AUC of 0.951. Single-cell validation confirmed upregulation of 7 hub genes in β cells and revealed cell-type-specific patterns invisible in bulk data, including bidirectional INHBA regulation between β and α cells and progressive α-cell proportion expansion (28.97% → 46.41%). Pathway enrichment converged on three mechanistic axes: extracellular matrix remodeling, immune activation, and autoimmune-like responses, with direct enrichment of the type 1 diabetes pathway (hsa04940). Conclusions: This study establishes a 12-gene progressive islet biomarker panel spanning the NGT-IGT-T2D continuum, supported by machine learning robustness, genetic causal evidence, diagnostic modeling, and single-cell biological validation. KCNE4 emerges as a candidate early-warning biomarker for prediabetes, while MYOF and RUNX2 represent causally supported compensatory targets, collectively providing a multilayered foundation for T2D risk stratification and precision intervention. From a clinical perspective, the identification of progressive islet biomarkers at the prediabetic stage provides molecular support for early lifestyle intervention, reinforcing that timely detection and behavioral modification remain the most effective strategies to prevent T2D progression.
    Keywords:  biomarker panel; diagnostic model; impaired glucose tolerance; machine learning; mendelian randomization; single-cell RNA sequencing; type 2 diabetes mellitus
    DOI:  https://doi.org/10.3390/biomedicines14081850
  20. Nutr Hosp. 2026 Aug 07.
    Weight-to-waist circumference index enhances interpretable machine learning models for screening depressive symptoms in middle-aged and elderly adults comorbid with hypertension and diabetes.
    Dandan Xu, Li Jiang, Qizhi Yang.
       OBJECTIVE: patients with comorbid hypertension and diabetes are at elevated risk of depressive symptoms, but simple metabolic predictors are insufficient. This study aimed to develop an interpretable machine learning model incorporating clinical data and novel composite indices for early screening of depressive symptoms.
    METHODS: a total of 546 participants ≥45 years with the comorbidities were enrolled from CHARLS (2011-2020). Eight composite indices (TyG, WHtR, WHT.5R, WWI, ABSI, CTI, hsCRP/HDL, log_tg) were included. Restricted cubic splines explored dose-response relationships. Ten machine learning models were compared via AUC, calibration curves, and DCA; SHAP interpreted the optimal model.
    RESULTS: the Gradient Boosting model performed best (validation AUC = 0.899, good calibration). SHAP identified frailty score and WWI as top predictors, with ABSI and TyG also contributing significantly. Subgroup analyses confirmed robust generalizability across gender and age in screening for depressive symptoms.
    CONCLUSION: novel composite indices (especially WWI, ABSI, TyG) have substantial clinical value for depression symptom screening in this population. The interpretable model provides a convenient, effective screening tool for clinical practice.
    DOI:  https://doi.org/10.20960/nh.06815
  21. Front Artif Intell. 2026 ;9 1898153
    Quantifying the imputation paradox and XAI inconsistency in multi-source diabetes prediction: a 353,680-record leakage-free stacking ensemble with dynamic routing architecture.
    M Nanda Kishore, C Navaneethan.
      Diabetes affects 537 million adults globally, a figure projected to reach 783 million by 2045. Despite over 4,200 ML prediction studies, clinical translation remains hindered by an over-reliance on benchmark datasets, unmeasured information costs of multi-source fusion, and untested XAI convergence assumptions. We address these issues by evaluating 353,680 records (from 455,446 candidates) across a Clinical-Biomarker Set (CBS) and a Lifestyle-Survey Set (LSS) using a strict leakage-free protocol. Under this protocol, normalisation, imputation, and SMOTE synthesis were fitted exclusively on training data. We formally define the Imputation Paradox as the AUC loss that occurs when critical diagnostic biomarkers are median-imputed under structural MNAR missingness. An Optuna-tuned stacking ensemble (Random Forest, XGBoost, LightGBM) achieves an AUC-ROC of 0.8481 on a held-out 70,736-sample test set. While a simple average baseline marginally exceeded stacking on pure discrimination (AUC 0.8521), stacking was retained for its superior precision (0.3439) and interpretable meta-learner trust weights, yielding a cross-validation AUC of 0.8491 ± 0.0014 and a Wilcoxon-significant improvement over baselines (p < 0.001, n = 1,000 replicates). A CBS sub-model trained on complete HbA1c and glucose data achieves an AUC of 0.9781. The resulting 0.130-unit gap quantifies this fusion-induced information cost. To recover this, our end-to-end dynamic routing pipeline routes 28.1% of patients to the CBS sub-model and 71.9% to the stacking ensemble, achieving a blended AUC of 0.8496. At a 0.30 mass screening threshold, the model identifies ~16,600 additional diabetic patients per million screened, with decision curve analysis confirming net benefit across the 5-30% threshold range. XAI conflict analysis yields Spearman correlations of r = -0.4561 (SHAP/permutation, p = 0.141), r = +0.2417 (SHAP/LIME, p = 0.737), and r = +0.1494 (permutation/LIME, p = 0.350). While non-significant at n = 12 features (critical r = ±0.576, α = 0.05), this provides preliminary directional evidence that these methods capture structurally different aspects of feature relevance.
    Keywords:  LIME; clinical decision support; decision curve analysis; diabetes prediction; dynamic routing pipeline; explainable AI; imputation paradox; leakage-free machine learning
    DOI:  https://doi.org/10.3389/frai.2026.1898153
  22. Front Neurol. 2026 ;17 1885303
    Artificial intelligence-integrated multimodal retinal imaging for early detection and risk stratification of systemic vascular and neurodegenerative diseases.
    Kaili Du, Yuqiang Zheng, Zhuoshi Wang.
       Background: We developed and validated RetinalVNG-Net, a multimodal deep learning framework for simultaneous risk stratification of hypertensive retinopathy, diabetic retinopathy, and neurodegenerative-associated retinal changes from integrated fundus photography, optical coherence tomography, and clinical metadata.
    Methods: This retrospective multi-center study included 2,740 subjects from three independent ophthalmology centers (January 2019-December 2023). Centers A and B (n = 2,220) constituted the development set, with a stratified 15% subset (n = 333) reserved for hyperparameter tuning and the remainder (n = 1,887) used for five-fold cross-validation. Center C (n = 520) served as a geographically distinct, device-heterogeneous external test set. RetinalVNG-Net employs a RETFound ViT-Large fundus encoder, a dual-stream Optical Coherence Tomography (OCT) branch (ResNet-3D-18 for volumetric B-scans and 2D-CNN for layer thickness maps), and a tabular transformer for metadata, fused via cross-modal attention. An auxiliary regression head outputs a continuous Retinal Biological Age Gap (RBAG) score as an interpretable severity biomarker.
    Results: Internal cross-validation yielded macro-averaged AUC-ROC 0.957 (±0.008), sensitivity 0.913, specificity 0.941, and F1 0.908 across four classes. On the external test set, macro-averaged AUC-ROC reached 0.944 (95% CI 0.922-0.958), with per-class AUCs of 0.963 (hypertensive retinopathy), 0.951 (diabetic retinopathy), 0.924 (neurodegenerative changes), and 0.938 (controls). RetinalVNG-Net significantly outperformed the best single-modality fundus model (macro-AUC 0.944 vs. 0.901; p < 0.001).
    Conclusion: RetinalVNG-Net demonstrates promising robustness and generalizability for simultaneous multimodal risk stratification of vascular and neurodegenerative retinal changes across diverse devices and settings, supporting further evaluation as a tool for risk stratification of retinal manifestations associated with systemic vascular and neurodegenerative disease. Prospective longitudinal studies would be required to establish value for early or predictive detection.
    Keywords:  artificial intelligence; diabetic retinopathy; hypertensive retinopathy; multimodal deep learning; neurodegeneration; retinal imaging
    DOI:  https://doi.org/10.3389/fneur.2026.1885303
  23. J Intensive Care Med. 2026 Aug 27. 8850666261479295
    Development and Validation of an Interpretable Machine Learning Model for Predicting In-Hospital Mortality in Diabetic Patients with Sepsis-Associated Acute Kidney Injury.
    Yanni Wang, Hongjie Shen, Shengze Wu, Suibi Yang, Feng Guo, Min Yang, Zhongheng Zhang.
      BackgroundSepsis-associated acute kidney injury (SA-AKI) is a common and severe complication in critically ill patients, with poor prognosis. Diabetes may further increase adverse outcomes through infection susceptibility, immune dysfunction, and renal vulnerability. However, mortality prediction models for patients with diabetes complicated by SA-AKI remain limited. This study aimed to develop and validate a machine learning-based model for early in-hospital mortality prediction in this population.MethodsA total of 6929 patients with SA-AKI and diabetes were identified from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database and randomly divided into training and validation sets at a ratio of 7:3. Ninety-four variables, including demographics, diagnoses, clinical parameters, and medication records within the first 24 h after ICU admission, were extracted. Twelve machine learning algorithms were developed and compared, and the optimal model was selected. Recursive feature elimination was used to identify key predictors, while SHapley Additive exPlanations were applied for model interpretation. The final model was deployed as a web-based tool and externally tested using the eICU Collaborative Research Database.ResultsThirty-two key predictors were ultimately selected, including urine output rate, platelet count, lactate, weight, blood glucose, SOFA score, pH, blood urea nitrogen, vital signs, coagulation indices, vasopressor use, and other clinically relevant variables. The categorical boosting algorithm model presented better predictive performance [receiver operating characteristic (AUC): 0.828] than other models [accuracy (ACC): 70.9%, sensitivity: 78.7%, specificity: 69%, F1 score: 0.509, positive predictive value (PPV): 33.7%, and negative predictive value (NPV): 93.1%]. External testing using data from the eICU database was also well validated (AUC: 0.793).ConclusionsA CatBoost-based machine learning model incorporating 32 clinically accessible variables showed good predictive performance for in-hospital mortality in patients with diabetes and SA-AKI, supporting early risk stratification and clinical decision-making.
    Keywords:  acute kidney injury; diabetes; machine learning; mortality; predictive model; sepsis
    DOI:  https://doi.org/10.1177/08850666261479295
  24. JMIR AI. 2026 Aug 25. 5 e92843
    Reflection- and Curriculum-Based Instruction Tuning for Diabetes-Specialized Large Language Models: Model Development and Evaluation Study.
    Jaesung Hwang, Deniise Liz Namayanja, Donghyeon Park.
       Background: Effective diabetes management requires continuous interpretation of glycemic trends, personalized dietary guidance, and sustained patient education. Although large language models (LLMs) are increasingly being explored for health-related applications, existing general-purpose and biomedical models often struggle with diabetes-specific reasoning and instruction-following, limiting their reliability for domain-focused tasks such as clinical question answering and dietary recommendation tasks.
    Objective: This study aimed to develop and evaluate a diabetes-specialized LLM optimized for diabetes-specific reasoning, instruction-following, and dietary recommendation tasks.
    Methods: This was a model development and benchmark evaluation study. We developed a model-centric instruction refinement framework using instruction-following difficulty and reversed instruction-following difficulty to identify and replace suboptimal instruction-response pairs during instruction tuning. Curriculum-based instruction tuning was applied by sequencing instructions from lower to higher difficulty. The resulting diabetes-specialized model was evaluated across diabetes-related question answering, natural language inference, information extraction, summarization, clinical answer generation, and dietary recommendation tasks. Performance was compared with biomedical LLMs and general-purpose baselines using diabetes-specific benchmark subsets and simulation-based glycemic evaluation.
    Results: Across diabetes-related benchmark tasks, the proposed model demonstrated improved performance in question answering, information extraction, and generative tasks. Ablation experiments showed that the full reflection- and curriculum-based instruction tuning strategy improved overall performance by 59.03% relative to the zero-shot LLaMA3.1 8B baseline. For dietary recommendation tasks, the proposed model achieved a 0.66% improvement in the Diet Quality Index-International score compared with ChatGPT (GPT-4). Simulation-based evaluation using the SimGlucose simulator further showed that meal plans generated by the proposed model resulted in a reduced postprandial glycemic burden, as measured by a lower incremental area under the curve, compared with GPT-4-generated meal plans.
    Conclusions: This study demonstrates that domain-specific instruction tuning can effectively adapt general-purpose LLMs for diabetes management. By combining reflection-based instruction replacement with curriculum-based instruction tuning, the proposed approach enhances instruction-following, reasoning capability, and dietary guidance for diabetes. The results highlight the potential of specialized LLMs to provide more reliable and clinically aligned support for diabetes-related decision-making and self-management, offering a promising direction for safe and effective AI deployment in chronic disease care.
    Keywords:  AI in health care; curriculum learning; diabetes mellitus; dietary recommendation; glycemic simulation; instruction tuning; instruction-following difficulty; large language models
    DOI:  https://doi.org/10.2196/92843
  25. Diagnostics (Basel). 2026 Aug 11. pii: 2533. [Epub ahead of print]16(16):
    Applying Artificial Intelligence to Childhood Obesity: T2DM and MASLD Risk Predictive Models.
    Marianna Amitrano, Gianluca Mondillo, Mario Emiliano, Umberto Paolo Santoro.
      Pediatric obesity is a complex, multifactorial pandemic with serious early-onset comorbidities, including prediabetes, type 2 diabetes, metabolic dysfunction-associated steatotic liver disease (MASLD), and cardiovascular disorders. While lifestyle modifications and the Mediterranean diet remain primary interventions, artificial intelligence (AI) is emerging as a critical tool for early diagnosis and personalized management. This review evaluates the current role of AI in predicting and treating childhood obesity and its complications. A literature search was conducted on PubMed and Google Scholar for English-language articles published from 2015 onward. Search terms included combinations of keywords related to "obesity", "pediatric", "comorbidities" (e.g., MASLD, diabetes), and "artificial intelligence" (e.g., machine learning, deep learning, multi-omics). Eligible study types ranged from original articles to systematic reviews and clinical guidelines. By integrating multi-omic data (genome, epigenome, transcriptome, metabolome, microbiota) with socio-psychological metrics, AI can predict obesity risk and early complications. Machine learning (ML) and deep learning have successfully identified specific metabolites, gut flora alterations, neurological pathways, and metabolic SNPs linked to obesity susceptibility. Furthermore, ML-driven prognostic models enable risk assessment for MASLD or diabetes progression, while specialized software supports remote lifestyle monitoring and tailored dietary interventions. AI has the potential to revolutionize pediatric obesity management through precision medicine. However, challenges regarding data privacy, digital literacy, and equitable access persist. Because current evidence relies heavily on limited and heterogeneous pediatric datasets, large-scale, well-characterized, and externally validated cohorts are essential to establish the clinical applicability of AI models before routine implementation.
    Keywords:  MASLD; artificial intelligence; childhood obesity; type 2 diabetes
    DOI:  https://doi.org/10.3390/diagnostics16162533
  26. Chin Med J (Engl). 2026 Aug 24.
    Merging artificial intelligence and retinal vascular geometric parameters: A novel tool for diagnosis and prognosis prediction of diabetic nephropathy.
    Sai Pan, Xiaoniao Chen, Qian Wang, Zhaohui Li, Fang Liu, Liqiang Wang, Yong Wang, Dajiang Wang, Li Zhang, Feng Liu, Ying Li, Jie Wu, Pu Chen, Xueying Cao, Guangyan Cai, Zheyi Dong.
       BACKGROUND: Diabetic nephropathy (DN) is the leading cause of end-stage renal disease. The retinal microvasculature, as the only directly observable microvasculature, may reflect DN progression. This study aims to construct a non-invasive diagnostic and prognostic prediction model using the mixed effects of retinal vascular geometric parameters and clinical data.
    METHODS: We constructed a multimodal database including 397 patients with type 2 diabetes and chronic kidney disease from multiple centers in China. The primary cohort (374 patients) was recruited from the Department of Nephrology at the First Medical Center of the Chinese People's Liberation Army General Hospital in Beijing between 2017 and 2022, while an external validation cohort (23 patients) was collected from five other hospitals across China from September 2022 to March 2023. Fundus images, clinical characteristics, renal biopsy diagnoses, and follow-up data were collected. Unsupervised learning and Resnet neural networks were used to segment and calculate retinal vascular geometric parameters. Weighted quantile regression (WQS), Lasso, and COX univariable regressions were employed to assess the mixed effects of retinal vascular geometric parameters and select relevant clinical characteristics. Logistic regression and Cox regression with random forest (COX-RF) were used for model construction.
    RESULTS: A multimodal database of 397 patients was constructed. A diagnostic model combining retinal parameters (WQS-diagnosis) and seven clinical characteristics achieved superior performance: AUC 0.98, accuracy 0.92 on the test set, outperforming models using only clinical data (AUC 0.93) or two retinal parameters plus clinical data (AUC 0.94). On the multi-centre validation set, the model maintained accuracy 0.91 and AUC 0.95. For prognosis, a mixed-effects parameter (WQS-prognosis) was derived; the COX-RF model achieved an AUC of 0.88.
    CONCLUSIONS: Retinal microvasculature are effective biomarkers for DN. The proposed non-invasive model demonstrated high accuracy and generalizability, offering a valuable tool for optimising DN management.
    Keywords:  Diabetic nephropathy; Diagnosis; Machine learning; Prognosis; Retinal vascular geometry
    DOI:  https://doi.org/10.1097/CM9.0000000000004211
  27. Exp Gerontol. 2026 Aug 27. pii: S0531-5565(26)00283-4. [Epub ahead of print] 113304
    Machine learning prediction models for diabetic kidney disease: A systematic review and meta-analysis.
    Chenglong Zhou, Xiaochu Wu, Zhili Tian, Li Mo, Fei Liu, Jun Li.
       BACKGROUND: Machine learning (ML) has emerged as a promising tool for predicting diabetic kidney disease (DKD), yet the performance and clinical utility of these models remain unclear. We conducted a systematic review and meta-analysis to evaluate ML models for DKD prediction.
    METHODS: We systematically searched seven databases from inception to May 16, 2026. Studies developing or validating ML models for DKD diagnosis were included. Quality was assessed using the Prediction model Risk of Bias Assessment Tool (PROBAST) and its AI extension (PROBAST+AI), and reporting completeness was evaluated using the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) statement and its AI extension (TRIPOD+AI). Pooled sensitivity, specificity, and area under the curve (AUC) were estimated using bivariate random-effects models.
    RESULTS: Sixty-six models were included. The pooled AUC was 0.926 (95% CI: 0.898-0.947), with sensitivity of 0.83 (95% CI: 0.79-0.86) and specificity of 0.90 (95% CI: 0.87-0.93). Neural networks achieved the highest AUC (0.973), followed by random forest (0.910). However, heterogeneity was substantial (I2 > 85%). Meta-regression identified model type (P = 0.019) and study design (P = 0.037) as significant effect modifiers.
    CONCLUSIONS: ML models show promising discriminative performance for DKD prediction, but the evidence is severely limited by high heterogeneity, lack of external validation and reporting of calibration metrics, and poor methodological transparency. Future research must prioritize external validation, calibration assessment, standardized reporting, and prospective impact studies.
    Keywords:  Diabetic kidney disease; Machine learning; Meta-analysis; Prediction model; Predictors; Systematic review
    DOI:  https://doi.org/10.1016/j.exger.2026.113304
  28. Int J Ophthalmol. 2026 ;19(9): 1687-1699
    FLNA promotes the progression of diabetic retinopathy by increasing M1 macrophages.
    Chun-Hua Li, Mao-Fei Feng, Xue Zhang, Lei Yu.
       AIM: To clarify the role of filamin A (FLNA) in promoting diabetic retinopathy (DR) progression and explore its underlying mechanism via modulating M1 macrophage inflammation and adenosine 5'-monophosphate activated protein kinase (AMPK) phosphorylation, as well as its potential as a diagnostic biomarker and therapeutic target.
    METHODS: Gene Expression Omnibus (GEO) dataset and Weighted Gene Co-Expression Network Analysis (WGCNA) were used to identify differentially expressed genes across DR, diabetes mellitus (DM), and control groups. Gene Set Enrichment Analysis (GSEA), Gene Ontology (GO), and Kyoto Encyclopedia of Genes and Genomes (KEGG) were employed to analyze the enrichment of intersecting genes. Machine learning models assessed the predictive performance of these genes, while single-cell sequencing and cell chat examined their expression in M1 macrophages and interactions with other cell types. Quantitative PCR (q-PCR) measured mRNA levels of interleukin-6 (IL-6), interleukin-1β (IL-1β), and tumor necrosis factor-α (TNF-α), and Western blotting assessed protein levels of FLNA, AMPK, and phosphorylated-AMPK (p-AMPK).
    RESULTS: Key findings revealed FLNA, plectin (PLEC), and Tweety family member 3 (TTYH3) as critical genes in DR, with FLNA showing the highest area under the curve (AUC) value and importance, and the random forest model performing best. The AMPK signaling pathway was enriched for FLNA, which was most significantly expressed in M1 macrophages. M1 macrophages promote IL-6, IL-1β, and TNF-α expression. FLNA increases the effect of M1 macrophages on DR progression and AMPK phosphorylation.
    CONCLUSION: Our study identifies FLNA as a key regulator in DR progression, primarily by enhancing M1 macrophage-mediated inflammation and AMPK phosphorylation. These findings highlight FLNA's critical role in DR pathogenesis and its potential as a biomarker for early diagnosis and a therapeutic target for intervening in macrophage-driven inflammatory responses.
    Keywords:  FLNA; M1 macrophages; diabetic retinopathy; machine learning; phosphorylated-AMPK
    DOI:  https://doi.org/10.18240/ijo.2026.09.04
  29. J Ophthalmol. 2026 ;2026 7199483
    Glycosylation-Associated Macrophage Signatures Define a Diagnostic Model for Diabetic Retinopathy via Single-Cell Analysis.
    Zhujuan Pan, Shaobo Zhou, Wenjuan Qi, Jing Wang, Yan Luo, Feihong Fan.
       Purpose: Diabetic retinopathy (DR) remains a leading cause of vision loss, with macrophages and glycosylation dysregulation implicated in DR pathogenesis. However, the potential as diagnostic biomarkers has been rarely investigated.
    Methods: We integrated single-cell RNA sequencing (scRNA-seq) datasets to profile DR cell landscapes. The immune cell heterogeneity was dissected, and glycosylation-related transcriptional programs were delineated in DR. Machine learning-based diagnostic modeling was also conducted to identify macrophage differentiation-related glycosylation genes (MDRGGs).
    Results: Macrophages exhibited elevated abundance in proliferative DR and showed intense interactions with other monocytes. Endothelial cells were subdivided into four subtypes, with Endo_KCNQ3 representing a dominant proliferative and highly glycosylated phenotype. Monocytes were clustered into three subtypes; Mono_RGS1 emerged as a transitional phenotype in the monocyte-to-macrophage trajectory. Macro_MIR181A1HG was identified as a proliferative and glycosylation-active macrophage subset. A total of 100 MDRGGs were identified. Among them, seven hub genes (AKAP13, SRGAP2, AFF1, ARHGAP24, RNF149, PTK2B, ATP1B3) were incorporated into a diagnostic model. The model achieved high predictive accuracy in both training and external validation cohorts (AUC > 0.85) and was further validated via nomogram and decision curve analyses.
    Conclusion: Glycosylation is closely associated with macrophage heterogeneity in DR. A seven-gene MDRGG-based diagnostic model demonstrated robust diagnostic performance in DR.
    Keywords:  diabetic retinopathy; diagnostic model; glycosylation; machine learning; macrophage
    DOI:  https://doi.org/10.1155/joph/7199483
  30. ACS Omega. 2026 Aug 25. 11(33): 50168-50184
    Development of a Shapley Additive Explanations (SHAP)-Interpretable Model for Cardiovascular Risk Stratification in Type 2 Diabetes Mellitus: Implications of MDA-Modified Protein Adducts and Autoantibodies.
    Chien-Yi Hsu, Shou-Cheng Lu, Liang-Wei Lin, Chu-Heng Yen, Hung-Tse Lin, Shao-Chun Lin, Ching-Yu Lin, Chao-Lien Liu.
      Coronary artery disease (CAD) is the leading cause of mortality in type 2 diabetes mellitus (T2DM). This study explores the two-hit oxidative-immune hypothesis by evaluating malondialdehyde (MDA)-modified protein (MDA-protein) adducts and their corresponding autoantibodies for precise and fair CAD severity stratification and classification. Novel MDA-peptide epitopes (apolipoprotein B-100 (ApoB-100), fibronectin (FINC), and complement C4A/C4B (C4A/C4B)) were identified via proteomics, and plasma levels of MDA, adducts, and autoantibodies were measured in 165 Taiwanese T2DM patients. Multivariate logistic regression estimated odds ratios (ORs) per 1-SD increase. Machine learning with Shapley additive explanations (SHAP)-interpretable analysis and nested cross validation distinguished obstructive from nonobstructive CAD. Fairness was assessed across age and sex subgroups. MDA (OR 3.148) and MDA-protein adducts (OR 2.090) were independent markers associated with advanced CAD severity, while IgG anti-ApoB-1001662-1683 MDA (OR 0.481) was independently associated with a lower risk. In differentiating obstructive CAD in T2DM patients, random forest achieved a pooled AUC of 0.958, a Brier score of 0.089, and a net benefit of 0.433 at the 20% threshold, outperforming conventional markers. SHAP analysis highlighted IgG anti-C4A/B167-187 as the dominant positive contributor within the machine learning framework, suggesting a possible proinflammatory signature, underscoring nonlinear push-pull dynamics between oxidative-immune signatures. As an exploratory finding, fairness analysis demonstrated consistent performance across sexes and a notable benefit in older patients, while revealing age-related imbalance in calibration and decision utility. Integrating MDA-related immune signatures into interpretable, fairness-aware categorical boosting and logistic regression models provides a robust, noninvasive framework for CAD severity stratification in T2DM, clarifying immunometabolic interactions and supporting equitable clinical decision-making.
    DOI:  https://doi.org/10.1021/acsomega.6c05628
  31. J Vasc Nurs. 2026 Sep;pii: S1062-0303(26)00042-7. [Epub ahead of print]44(3): 195-200
    The reliability-clarity paradox in vascular digital health: A nurse-led evaluation of ChatGPT-4's accuracy in diabetic foot ulcer management.
    Pınar Tunç Tuna, Hamide Şişman.
       BACKGROUND: Diabetic foot ulcers remain a leading cause of non-traumatic amputations worldwide, necessitating precise patient education and clinical management. As large language models become increasingly integrated into patient-facing platforms, evaluating their clinical safety within specialized nursing contexts is imperative. This study aimed to examine the performance of ChatGPT-4 in DFU management, focusing particularly on linguistic clarity and scientific accuracy.
    METHODS: A cross-sectional evaluative design was employed in April 2025. An expert panel of six certified wound care nurses, each possessing over a decade of clinical experience, assessed AI-generated responses to seven core clinical inquiries (encompassing 31 subtopics). Evaluation was conducted using a 5-point Likert scale. Methodological rigor was established through Content Validity Index (CVI) and Intraclass Correlation Coefficients (ICC) to determine expert consensus.
    RESULTS: ChatGPT-4 exhibited exceptional performance in linguistic clarity, scoring 152.0 out of 155 based on expert evaluation. This high clarity refers to the model's ability to translate complex medical data into accessible language. However, a significant decline in scientific accuracy was observed when addressing formal clinical protocols (6.67/10), particularly regarding the International Working Group on the Diabetic Foot (IWGDF) 2023 updates. Despite these inaccuracies, the expert panel demonstrated a high level of consensus (ICC = 0.937) regarding the model's limitations.
    CONCLUSIONS: The study identifies a 'reliability-clarity paradox' where ChatGPT-4's high linguistic fluency (152/155) creates an authoritative tone that effectively masks underlying clinical inaccuracies. This is evidenced by a significant drop in scientific precision (6.67/10), particularly regarding the integration of IWGDF 2023 updates. The exceptional inter-rater consensus (ICC = 0.937) further confirms that these clinical risks are systematic rather than subjective. Consequently, while AI can assist in administrative drafting, the specialized nurse's role as a 'clinical gatekeeper' remains vital to provide a necessary safety filter and ensure evidence-based accuracy in DFU care.
    Keywords:  Artificial intelligence; ChatGPT-4; Diabetic foot ulcer; Health communication; Patient safety; Vascular nursing
    DOI:  https://doi.org/10.1016/j.jvn.2026.06.002
  32. Diabetes Obes Metab. 2026 Aug 24.
    Intramuscular Fat Assessed by AI-Assisted Muscle Ultrasound: Association With Cardiometabolic Risk Factors and Diabetic Nephropathy in Diabetes Mellitus.
    Paloma Pérez-López, Juan José López-Gómez, Álvaro de la Calzada-Martínez, Jaime González-Gutiérrez, Beatriz Ramos-Bachiller, Esther Delgado-García, Emilia Gómez-Hoyos, Ana Ortolá-Buigues, Gonzalo Díaz-Soto, Rebeca Jiménez-Sahagún, Mario Alfredo Saavedra-Vásquez, Pablo Fernández-Velasco, Daniel De Luis-Román.
       AIMS: Intramuscular fat (IMF) is increasingly recognized as a marker of ectopic adiposity and adverse cardiometabolic outcomes. Artificial intelligence (AI)-assisted ultrasound of the rectus femoris (RF) offers a non-invasive approach for quantifying IMF. This study evaluated the association of IMF with diabetes-related complications (particularly diabetic nephropathy) and metabolic risk factors in patients with diabetes mellitus (DM).
    MATERIALS AND METHODS: In this cross-sectional study, outpatients from a tertiary Endocrinology and Nutrition Department underwent anthropometric assessment, bioimpedance analysis, and muscular ultrasound. Ultrasound images were analysed using the PIIXMED AI-system (DAWAKO MedTech; Valencia, Spain) to quantify muscle (Mi) and fat (FATi) percentages (the latter indicating the IMF). Associations between IMF, clinical characteristics, metabolic biomarkers, and vascular complications were examined.
    RESULTS: A total of 120 patients were included (57.5% men), with a mean age of 70.8 ± 10.3 years and diabetes duration of 13.8 ± 9.6 years. Most participants had type 2 DM (79.2%), with suboptimal glycaemic control (HbA1c 8.1% ± 1.5%). Individuals in the highest FATi quartile (> 43%) showed higher prevalence of microvascular complications, particularly diabetic nephropathy (44.0 vs. 14.7%, p = 0.001), and lower estimated glomerular filtration rate (63.6 ± 21.3 vs. 72.8 ± 20.1 mL/min/1.73m2, p = 0.045). Multivariate analysis showed that FATi remained independently associated with nephropathy (OR 6.01, 95% CI 1.99-18.14; p < 0.01). ROC analysis identified a FATi threshold of 43.5% with modest discriminative ability (AUC = 0.668).
    CONCLUSIONS: AI-assisted ultrasound-derived FATi is associated with adverse metabolic profiles and diabetic nephropathy. Assessment of IMF may provide an accessible biomarker related to microvascular complications. Longitudinal and multicentre studies are needed to determine the role of IMF in diabetes-related complications.
    Keywords:  artificial intelligence; diabetes mellitus; diabetic nephropathy; intramuscular fat; muscle ultrasound
    DOI:  https://doi.org/10.1111/dom.71249
This page is being maintained by Thomas Krichel. It was last updated on 2026‒08‒30 at 0:12.