bims-aukdir 2026-08-23 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2026–08–23
seventeen papers selected by
Mott Given

Artificial intelligence in diabetic retinopathy: from automated screening to risk-stratified care.
A comprehensive review on automated diabetic retinopathy detection and classification using fundus image.
Machine learning models for risk prediction of diabetic retinopathy in Fujian eye study.
ORDER-DR: external validation of severity grading and referable-risk stratification from fundus images.
Deep Learning of Fluorescence Lifetime Imaging Ophthalmoscopy for Type 2 Diabetes Classification.
A generalizability analysis framework for bias resilience in convolution neural network and transformer models for diabetic retinopathy screening.
Multimodal Deep Learning for Diabetic Retinopathy: A Survey.
Integrative Proteomics-Metabolomics Profiling of Umbilical Cord Plasma With Machine Learning Reveals Candidate Biomarkers for Gestational Diabetes Mellitus.
Subtyping and Prediction of the Glucose Based on the Hybrid Modeling of Deep Learning and Dietary Stimulated Glucose Metabolism Dynamical Model.
DSTransformer: A Dual-Scale Transformer Framework for Zero-Shot Blood Glucose Prediction via Domain Generalization.
Machine learning-based prediction of major amputation risk after initial limb-preserving surgery in diabetic foot.
Artificial Intelligence Analysis of Standard 12-Lead Electrocardiograms for Screening of Symptomatic Diabetic Peripheral Neuropathy.
SGLT2 inhibition, uric acid and inflammation in glycemic burden-related progression of cardiac autonomic neuropathy in type 2 diabetes mellitus: insights from machine learning and mediation analysis.
Evaluation of Differential Diagnosis of Odontogenic Lesions in Cone Beam Computed Tomography Images Using Radiomics-Based Machine Learning.
Machine learning-based prediction of prolonged length of stay in older patients with type 2 diabetes mellitus and cardiovascular disease.
Network Pharmacology, Machine Learning, and In Vivo Validation of Danzhi Jiangtang Capsule in Diabetic Nephropathy.
Assessing the information quality of AI-generated patient educational materials for diabetes: a scoping review.

Graefes Arch Clin Exp Ophthalmol. 2026 Aug 20.

Artificial intelligence in diabetic retinopathy: from automated screening to risk-stratified care.

Fenfen Zhang, Panpan Shao, Dongdong Wei, Chunhua Wang.

  Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide. The rising prevalence of diabetes has strained conventional screening pathways, which are labor-intensive and depend on specialist image interpretation, limiting access, particularly in resource-limited settings. Artificial intelligence (AI) offers a promising approach to improve the scalability, consistency, and reach of DR detection and management. This review examines AI applications in DR, spanning automated screening, grading, lesion segmentation, multimodal imaging integration, personalized risk prediction, and AI-driven metabolomics, while also addressing translational challenges and future directions. We conducted a narrative review using major biomedical databases and targeted searches of peer-reviewed literature, prioritizing recent systematic reviews, meta-analyses, clinical trials, implementation studies, and seminal technical reports. Current evidence shows that AI systems can detect referable DR with diagnostic performance approaching or exceeding expert human graders. Deep learning methods have improved lesion detection, severity grading, and risk assessment. Multimodal approaches integrating fundus photography with OCT, OCTA, and clinical data may enhance disease characterization and risk stratification, but most evidence remains developmental or early translational. Likewise, AI-guided personalized follow-up, treatment modeling, and metabolomics-based biomarker discovery are promising but less clinically validated than screening and grading applications. Real-world studies support implementation feasibility, with gains in screening uptake and workflow efficiency. Key barriers to broader adoption include external validation, heterogeneous reference standards, management of ungradable images, domain shift across devices and populations, workflow integration, and regulatory and ethical considerations. AI is poised to strengthen scalable DR care, but personalized applications require further validation before routine clinical use.

Keywords:  Artificial intelligence; Deep learning; Diabetic retinopathy; Machine learning; Retinal imaging; Risk stratification

DOI:  https://doi.org/10.1007/s00417-026-07390-2
Graefes Arch Clin Exp Ophthalmol. 2026 Aug 17.

A comprehensive review on automated diabetic retinopathy detection and classification using fundus image.

Ravi Bhushan Dixit, Chandan Kumar Jha.

   PURPOSE: Diabetic Retinopathy (DR) is a vision-threatening complication in diabetic patients. It harms retinal vessels and may lead to blindness. Detection at an early stage and its classification can prevent the risk of vision loss. However, fundus image-based manual screening of DR is a time-consuming and complex process.
METHODS: In recent years, many automated techniques for DR detection have been developed to screen and diagnose the disease condition at an early stage. These techniques are explored using the keywords diabetic retinopathy, fundus image, ophthalmology with machine learning (ML), and deep learning (DL). Search engines such as Google Scholar, PubMed, Medline, IEEE Explore, and Science Direct are utilised and explored to gather existing research papers.
RESULTS: This review systematically examines several techniques for DR detection and classification, using ML, DL, and hybrid approaches. This study analyses methodologies, datasets, pre-processing steps, performance evaluation metrics of existing techniques, and challenges associated with overfitting, model complexity, class imbalance, and deep feature extraction. Recent advancements in ensemble learning, transformer-based techniques, and attention mechanisms are also discussed for DR detection and classification.
CONCLUSION: The review explored and discussed the challenges of existing DR detection and classification methods. This paper suggests future research directions to improve the accuracy and robustness of DR detection systems.

Keywords:  Deep learning; Diabetic retinopathy; Ensemble learning; Fundus image; Machine learning; Neural networks

DOI:  https://doi.org/10.1007/s00417-026-07435-6
Front Endocrinol (Lausanne). 2026 ;17 1834380

Machine learning models for risk prediction of diabetic retinopathy in Fujian eye study.

Yang Li, Qinrui Hu, Bin Wang, Xiangdong Luo, Mingqin Zhang, Xiaoxin Li.

   Background: Diabetic Retinopathy (DR) is among the most severe microvascular complications of diabetes, leading to visual impairment and diminished quality of life. This study developed and compared multiple machine learning models for DR risk prediction using population-based data from the Fujian Eye Study, aiming to identify the top five key predictors and establish a robust data-driven framework for early screening.
Methods: Data were obtained from the Fujian Eye Study, comprising 8211 participants and 51 variables. After data preprocessing, five machine learning models-Logistic Regression (LR), K-Nearest Neighbors (KNN), Support Vector Classifier (SVC), Decision Tree (DT), and Random Forest (RF)- were trained and optimized via cross-validation and grid search. Model performance was evaluated using multiple metrics, including accuracy, precision, recall, F1-score, and AUC. Feature importance was examined using SHAP (Shapley Additive Explanations) and validated through unsupervised and nonparametric approaches-Factor Analysis (FA), Highly Variable Feature Selection (HVGS), and Spearman's rank correlation.
Results: Among the five models, SVC model achieved the highest performance (F1-score 92.83%, AUC 0.99). SHAP analysis identified the top five predictors of DR risk: history of diabetes, age, pulse pressure difference (PPG), near visual acuity of the left eye, and height. Cross-method comparison confirmed high feature stability across models, indicating robust predictor reproducibility.
Conclusion: This study successfully established and validated a machine learning-based framework for predicting diabetic retinopathy risk using data from the Fujian Eye Study. The support vector machine (SVC) model demonstrated superior predictive capability. The identified key risk factors-diabetes history, age, pulse pressure difference, left eye near visual acuity, and height-provide actionable insights for early stratification. These findings provide a reliable, data-driven tool for early DR screening, which can facilitate population-level risk stratification and inform personalized preventive interventions in clinical and public health settings.

Keywords:  cross-sectional eye study; diabetic retinopathy; machine learning; model evaluation; risk prediction

DOI:  https://doi.org/10.3389/fendo.2026.1834380
Front Endocrinol (Lausanne). 2026 ;17 1923216

ORDER-DR: external validation of severity grading and referable-risk stratification from fundus images.

Quanwei Sheng, Qiyu Dong, Shuang Wu, Weijun Liang.

  Diabetic retinopathy (DR) is a common microvascular complication of diabetes, and fundus-image deep learning may support scalable screening and risk stratification. However, models trained on a single public dataset can show threshold shift when externally evaluated, limiting direct translation from internal accuracy to clinically interpretable risk. We developed ORDER-DR, a validation-calibrated dual-branch ordinal-risk framework for five-class DR severity grading and referable DR prediction from color fundus photographs. The final prespecified operating model used a high-resolution EfficientNet-B0 branch and a complementary lesion-order-sensitive risk (LORS) EfficientNet-B0 branch; checkpoints and decision thresholds were selected only from APTOS validation predictions. External validation was performed on 1,744 gradable Messidor-2 images. On held-out APTOS test splits, ORDER-DR achieved quadratic weighted kappa (QWK) 0.8960 ± 0.0049, macro-F1 0.6832 ± 0.0279, and accuracy 0.8352 ± 0.0118. On Messidor-2 with validation-calibrated thresholds, ORDER-DR achieved the strongest external ordinal agreement among the evaluated candidate models, with QWK 0.6423 ± 0.0364 and macro-F1 0.4578 ± 0.0332. The LORS branch retained higher external area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC), indicating a tradeoff between ordinal grade agreement and referable-risk ranking. Reliability analysis showed a higher referable-risk expected calibration error on Messidor-2 than on APTOS test (0.160 ± 0.008 vs. 0.049 ± 0.001). These findings indicate robust external ordinal consistency for ORDER-DR across datasets and show that operational threshold performance, threshold-independent risk ranking, and probability calibration should be evaluated as complementary dimensions. The primary contribution is a locked and reproducible empirical evaluation framework for DR ordinal grading and referable-risk stratification, integrating high-resolution class-balanced discrimination, ordinal-risk modeling, validation-derived threshold calibration, and external error analysis. At the validation-selected referable-risk threshold, Messidor-2 performance showed a high-specificity operating profile; high-sensitivity screening use can be adapted through operating-point selection and calibration for the target clinical setting.

Keywords:  calibration; deep learning; diabetic retinopathy; external validation; fundus image; ordinal classification; risk stratification

DOI:  https://doi.org/10.3389/fendo.2026.1923216
medRxiv. 2026 Aug 06. pii: 2026.08.04.26359728. [Epub ahead of print]

Deep Learning of Fluorescence Lifetime Imaging Ophthalmoscopy for Type 2 Diabetes Classification.

SeongJun Kwon, Cecilia S Lee, Aaron Y Lee, Linying Zhang.

Purpose: To evaluate whether fluorescence lifetime imaging ophthalmoscopy (FLIO) combined with deep learning can detect metabolic signatures for classification of type 2 diabetes mellitus (T2DM).
Design: Cross-sectional analysis of participants included AI-READI dataset (version 3) with FLIO imaging and and hemoglobin A1c (HbA1c) measurement.
Subjects: 1,783 participants from the AI-READI dataset (version 3) with HbA1c measurements and FLIO imaging scans (6,912 total): 671 normoglycemic, 726 prediabetic, and 386 diabetic.
Methods: Mean fluorescence lifetime maps were generated using a center-of-mass approach and used as inputs to AI models. We trained convolutional neural networks (CNNs), ResNet-18, and XGBoost under three-class (normal, prediabetic, diabetic) and two binary (normal vs. impaired; normal vs. diabetic) classification schemes, using nested 5-fold cross-validation with participant-level grouping.
Main Outcome Measures: Macro-averaged accuracy, F1 score, area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and positive predictive value (PPV).
Results: Group-averaged lifetime maps demonstrated consistent spatial differences across glycemic groups, with progressively longer lifetimes from normal to diabetic participants. The CNN achieved the best overall performance in the 3-class classification (accuracy 0.41 ± 0.03, F1 score 0.39 ± 0.02, AUROC 0.58 ± 0.02), compared to the random classifier for 3-class classification (AUROC = 0.50; accuracy = F1 = 0.33). ResNet-18 and XGBoost showed similar performance (AUROC 0.53-0.58). Confusion matrices revealed substantial overlap between classes, with frequent misclassification toward the prediabetes group. Binary reformulation (normal vs. diabetic) improved performance substantially, with the CNN resulting in AUROC 0.63 ± 0.02 and XGBoost 0.67 ± 0.07.
Conclusions: FLIO-derived lifetime maps capture metabolic signals associated with glycemic status but yield modest classification performance with current AI models. These findings highlight both the potential and the challenges of using FLIO for early metabolic screening and monitoring, informing future development of clinically applicable imaging biomarkers.

DOI: https://doi.org/10.64898/2026.08.04.26359728
PLoS One. 2026 ;21(8): e0351225

A generalizability analysis framework for bias resilience in convolution neural network and transformer models for diabetic retinopathy screening.

Majida Kazmi, Bisma Imran, Saad A Qazi, M A Rehman Siddiqui.

Automated diabetic retinopathy (DR) screening has seen extensive research progress, yet its clinical adoption remains limited, largely due to insufficient attention to model generalizability across diverse populations and imaging conditions. Existing studies often overlook the role of dataset biases, arising from demographic, acquisition, and preprocessing variations that undermine robustness in real-world settings. To address this critical gap, we propose a comprehensive generalizability analysis framework that systematically categorizes dataset biases, evaluates their impact on performance, and introduces robust inter-dataset comparison metrics. The framework is metric-agnostic and extensible to multi-class severity grading, with stage-wise BAG formulations introduced to support more detailed clinical evaluation. Using one primary (EyePACS) and nine secondary datasets, we assessed the generalizability of two distinct architectures: a convolutional neural network (MobileNetV2) and a transformer-based model (CvT). Performance was evaluated through intra-group measures (accuracy, sensitivity, specificity, and AUC) and a newly proposed Bias-Adjusted Generalization (BAG) index designed to quantify resilience to bias-induced domain shifts. Results demonstrate that CvT consistently outperformed MobileNetV2, achieving an average accuracy of 87% and a BAG index of 0.94, compared to MobileNetV2's 78% accuracy and BAG index of 0.91. Gradient-based explainability visualizations further confirm that both models attend to clinically relevant retinal regions, with architectural differences in activation patterns consistent with their respective generalizability profiles. These findings highlight the inherent bias resilience of transformer architectures, while highlighting their computational demands as a barrier to deployment in low-resource environments. Importantly, the study suggests that integrating transformer-inspired mechanisms into lightweight CNNs could yield clinically scalable models with both efficiency and robustness. Our framework offers a standardized approach for bias-aware evaluation, providing actionable insights for developing equitable, generalizable, and resource-adaptable AI solutions for DR screening.

DOI: https://doi.org/10.1371/journal.pone.0351225
J Ophthalmol. 2026 ;2026 4612172

Multimodal Deep Learning for Diabetic Retinopathy: A Survey.

Genyan Qin, Qinghua Peng, Yasha Zhou, Jun Peng, Qing Liu, Jian Xu.

  Diabetic retinopathy (DR) remains a leading cause of blindness globally, driving the rapid development of automated diagnostic systems leveraging deep learning. Recent research demonstrates that multimodal deep learning models that integrate retinal images, electronic health records (EHR), and clinical text data can surpass single-modality approaches in DR screening accuracy. In this survey, we systematically review advances from the last 3 years, highlighting key models and methodologies, as well as datasets and evaluation metrics. We find multimodal approaches consistently outperform single-modality models, and often by a clear margin, and identify data scarcity, domain adaptation, and interpretability as the three main hurdles ahead.

Keywords:  diabetic retinopathy; medical image analysis; multimodal deep learning; retinal imaging; vision language models

DOI:  https://doi.org/10.1155/joph/4612172
Diabetes Obes Metab. 2026 Aug 17.

Integrative Proteomics-Metabolomics Profiling of Umbilical Cord Plasma With Machine Learning Reveals Candidate Biomarkers for Gestational Diabetes Mellitus.

Rui Zhao, Yidan Zhu, Yuanjie Liu, Zhaoqi Yuan, Haozheng Lu, Aimin Jiang, Tingting Chai, Meiying Guo, Cuiping Xu.

   BACKGROUND: Gestational diabetes mellitus (GDM) is a common pregnancy complication, yet its molecular mechanisms remain incompletely understood. This study applied a multi-omics strategy combined with machine learning to identify candidate biomarkers and dysregulated molecular networks at the maternal-fetal interface.
METHODS: Umbilical cord plasma from 61 pregnant women, including 34 with GDM and 27 normoglycemic controls, was analysed using high-resolution proteomics and untargeted metabolomics. Hub candidate biomarkers were prioritised using LASSO, random forest and XGBoost; their discriminatory performance was assessed by ROC analysis and internally validated through bootstrap resampling and repeated cross-validation. Proteome-metabolome integration was performed with mixOmics. Immunohistochemistry (IHC) of placental tissues and RT-qPCR in HTR8/SVneo cells provided tissue-level and transcriptional validation, respectively.
RESULTS: Proteomics identified 37 differentially expressed proteins and selected five hub proteins-AFP, ORM1, PRSS2, LACRT and LCN1. The composite protein score demonstrated strong discriminatory performance, with an AUC of 0.984 (95% CI: 0.958-1.000). IHC confirmed the dysregulation of these hub proteins in GDM placentas and RT-qPCR revealed that the mRNA levels of AFP, ORM1 and PRSS2 were consistently upregulated under high-glucose conditions. Metabolomics identified 185 differential metabolites and three hub metabolites, including C30H38O7, nicotine N-oxide and 7(1)-hydroxychlorophyll. The composite metabolite score showed an AUC of 0.991 (95% CI: 0.977-1.000). Cross-omics integration revealed an exploratory proteome-metabolome network associated with cornified envelope formation, keratinisation, humoral immunity and vitamin/nucleoside transport.
CONCLUSIONS: Integrative proteomic-metabolomic profiling of umbilical cord plasma revealed coordinated immune, barrier and metabolic dysregulation in GDM, supporting exploratory multi-omics biomarker development and future mechanistic studies.

Keywords:  biomarkers; gestational diabetes mellitus; machine learning; metabolomics; proteomics; umbilical cord plasma

DOI:  https://doi.org/10.1111/dom.71241
Bull Math Biol. 2026 Aug 18. pii: 158. [Epub ahead of print]88(9):

Subtyping and Prediction of the Glucose Based on the Hybrid Modeling of Deep Learning and Dietary Stimulated Glucose Metabolism Dynamical Model.

Yuhao Zhou, Xianglong Liu, Shi Zhao, Yu Zhao.

  Diabetes represents a significant global health challenge, underscoring the need for enhanced methodologies in glycemic monitoring and risk assessment. Static biomarkers, including fasting plasma glucose and glycated hemoglobin, may have limitations in capturing the temporal variations and individual heterogeneities inherent in glucose regulation. Continuous glucose monitoring (CGM) offers a high-resolution approach that may facilitate timely metabolic categorization. However, there is a notable scarcity of modeling frameworks that integrate CGM data while maintaining both interpretability and predictive accuracy. This study proposes an integrated analytical framework that amalgamates CGM data-driven classification of glucose response with a hybrid modeling approach that combines dynamical systems and deep learning methodologies for personalized glucose regulation assessments. A cohort of 44 adults was enrolled and followed up on an individual basis for 7 to 14 days, and typical daily glucose profiles were derived through dynamic time warping. K-shape clustering analysis was employed to identify significant glucose response subtypes, differentiating participants into three categories: health, prediabetes, and diabetes based on CGM-derived indicators. We subsequently developed an integrated hybrid model that synergizes a dietary stimulation glucose-insulin dynamical model with a long short-term memory residual network. This hybrid model demonstrated substantially improved accuracy, achieving a root mean square error (RMSE) of 6.61 and a coefficient of determination ( R2 ) of 0.91. The glucose-insulin dynamical model effectively elucidates the physiological mechanisms underlying postprandial glucose-insulin regulation. The findings of this study emphasize the utility of CGM-derived glucose phenotyping and hybrid predictive models as viable tools for individualized metabolic risk assessment. Additionally, they contribute to the early identification of dysglycemia and establish a practical framework for precision glycemic management.

Keywords:  Deep learning; Dynamical models; Glucose subtyping; Glycemic prediction; Hybrid modeling

DOI:  https://doi.org/10.1007/s11538-026-01678-4
IEEE J Biomed Health Inform. 2026 Aug 18. PP

DSTransformer: A Dual-Scale Transformer Framework for Zero-Shot Blood Glucose Prediction via Domain Generalization.

Yi Wu, Yuanyuan Zou, Jie Cheng, Kan Wang, Bin Dai, Jinhong Guo.

Diabetes is a global chronic metabolic disorder that requires precise blood glucose control to delay complications. Multi-step glucose prediction can provide early warnings of abnormal fluctuations and create a time window for proactive intervention. However, existing methods still face two major challenges: limited zero-shot generalization caused by patient heterogeneity, and difficulty in jointly capturing local fine-scale fluctuations and global long-term trends within a single-scale framework. To address these issues, this paper proposes DSTransformer, a dual-branch multi-resolution framework for multi-step blood glucose prediction. By combining parallel multi-scale feature extraction with layer-wise learnable routers, DSTransformer dynamically fuses local and global temporal information for more balanced glucose sequence modeling. It also introduces a Fishr-inspired patient-level gradient regularization strategy, treating each patient as an independent domain and aligning the loss landscape by matching batch-wise gradient variance across domains, which significantly enhances zero-shot generalization performance. Relying solely on univariate CGMS time series to keep clinical deployment costs low, the framework delivers strong results in patient-level zero-shot evaluations across three independent datasets (ShanghaiT2DM, OhioT1DM, and REPLACE-BG), predicting future blood glucose levels over 30-120-minute horizons. At the best performing 30-minute horizon, the average performance across the three datasets reached 13.335 mg/dL root mean square error (RMSE), 8.606 mg/dL mean absolute error (MAE), and 93.807% accuracy. Even for the more challenging 120 minute long-horizon prediction, the framework still achieved an average RMSE of 25.793 mg/dL, MAE of 16.995 mg/dL, and accuracy of 86.538% across the three datasets. These results outperform existing state-of-the-art (SOTA) methods across all key metrics, fully demonstrating the framework's excellent generalization capability and its potential to support future clinical decision-making after prospective clinical validation.

DOI: https://doi.org/10.1109/JBHI.2026.3724928
Front Endocrinol (Lausanne). 2026 ;17 1889329

Machine learning-based prediction of major amputation risk after initial limb-preserving surgery in diabetic foot.

Kerim Bora Yilmaz, Gokalp Tulum, Ferhat Cuce, Zehra Simsek, Necibe Sare Mert, Sema Horasan Hatipoglu, Cansu Bozkurt, Muhammet Ikbal Isik, Ali Murat Basak, Onur Osman.

   Background: Accurate preoperative prediction of whether an initially limb-preserving strategy in diabetic foot management will culminate in minor or major amputation remains a clinical challenge. This study aimed to develop and evaluate using two temporally separated cohorts a machine-learning framework using routinely available baseline clinical, laboratory, and selected imaging and vascular variables.
Methods: Two temporally separated cohorts were used, with Dataset 1 for model development and Dataset 2 for temporally separated evaluation. A 20-repetition stratified outer-split workflow was implemented, incorporating two-step feature selection, Optuna-based hyperparameter optimization, training-only SMOTE, and threshold tuning to maximize the F2-score under a recall constraint of ≥0.70. Six classifiers were evaluated using average precision (AP), ROC-AUC, recall, precision, specificity, accuracy, and Brier score.
Results: The major-amputation group exhibited a more severe baseline phenotype, including higher inflammatory burden, worse neuropathy and wound severity, and a higher prevalence of necrotizing fasciitis. Internally, multilayer perceptron achieved the highest AP (55.9% ± 13.6%). In external evaluation, k-nearest neighbors achieved the highest AP (65.1% ± 10.2%) and recall (72.8% ± 19.6%), whereas multilayer perceptron showed higher precision and specificity. Key contributors included necrotizing fasciitis, neuropathy severity, hemoglobin, PEDIS classification, and inflammatory indices.
Conclusion: These findings suggest that prediction of amputation level is feasible, validated in a temporally separated cohort, and clinically interpretable, and may support future decision-support applications, although further validation is required before clinical implementation.

Keywords:  amputation prediction; clinical decision support; diabetic foot; explainable artificial intelligence; limb-preserving surgery; machine learning; temporal validation; web-based deployment

DOI:  https://doi.org/10.3389/fendo.2026.1889329
Diabetes Obes Metab. 2026 Aug 19.

Artificial Intelligence Analysis of Standard 12-Lead Electrocardiograms for Screening of Symptomatic Diabetic Peripheral Neuropathy.

Paweł Ignacy, Hanadi Aldosari, Krzysztof Irlik, Martyna Gut-Misiaga, Aleksandra Gil, Hanna Kwiendacz, Janusz Gumprecht, Gregory Y H Lip, Uazman Alam, Katarzyna Nabrdalik.

   AIM: To investigate whether artificial intelligence (AI) models trained on standard 12-lead electrocardiograms (ECG) can identify symptom-defined diabetic peripheral neuropathy (DPN) as assessed by the Michigan Neuropathy Screening Instrument Questionnaire (MNSI-Q).
MATERIALS: This was an observational study of people with diabetes enrolled in the Silesia-Diabetes Heart Project. DPN was assessed using the MNSI-Q. We used an original questionnaire cut-off score of ≥ 7 and revised cut-off score of ≥ 4 to diagnose DPN and classified DPN as highly symptomatic and moderately symptomatic accordingly. Feature extraction from 10-s raw ECG recordings utilised algorithms to identify recurring signal segments (motifs) and anomalies (discords). These features were used to train XGBoost, Support Vector Machine (SVM) and Ridge classifiers to differentiate between DPN-positive and DPN-negative people.
RESULTS: A total of 640 participants (mean age 54 ± 17; 52% female) were included. Of these, 95 (15%) had highly symptomatic DPN (MNSI-Q ≥ 7) and 281 (44%) had moderately symptomatic DPN (MNSI-Q ≥ 4). For the highly symptomatic DPN, the XGBoost classifier utilizing a combination of motifs and discords demonstrated the highest performance, achieving an area under the receiver-operating characteristic curve (AUC) of 0.89 (95% CI 0.88-0.91), an accuracy of 88.5% and a sensitivity of 93.4%. The model showed substantially lower predictive capability when tested against the broader screening threshold of MNSI ≥ 4 (AUC 0.64).
CONCLUSION: AI analysis of the standard ECG demonstrated a strong association with the presence of symptom-defined peripheral neuropathy but lacked sensitivity for milder presentation. With further validation, this method could serve as an accessible, supplementary screening aid to help identify high-risk patients during routine cardiovascular assessment.

Keywords:  Michigan neuropathy screening instrument; artificial intelligence; diabetic peripheral neuropathy; electrocardiography; machine learning

DOI:  https://doi.org/10.1111/dom.71240
Front Endocrinol (Lausanne). 2026 ;17 1901683

SGLT2 inhibition, uric acid and inflammation in glycemic burden-related progression of cardiac autonomic neuropathy in type 2 diabetes mellitus: insights from machine learning and mediation analysis.

Jing Jing Wang, Jie Hu, Wu Dai, Yan Liu, Pan Pan Jiang, Jun Ye, Yong Hong Cao.

   Objective: To identify independent predictors of diabetic cardiac autonomic neuropathy (DCAN) deterioration in patients with type 2 diabetes mellitus (T2DM), develop machine learning (ML)-based risk prediction models, and explore the underlying mediation of glycemic burden.
Methods: This prospective cohort study included 293 T2DM patients who underwent standardized Ewing testing at baseline and follow-up. Univariable and multivariable logistic regression were utilized to identify predictors. Restricted cubic splines (RCS) were applied to evaluate non-linear relationships. Nine ML algorithms were developed and evaluated using AUC and calibration metrics, with SHAP values illustrating feature importance. Mediation analysis was performed to investigate whether longitudinal biomarker changes accounted for the association between HbA1c and DCAN deterioration.
Results: During follow-up, 80 patients (27.3%) experienced DCAN deterioration. Multivariable logistic regression identified elevated HbA1c at baseline as an independent risk factor, while SGLT2 inhibitor use was significantly associated with a lower risk of deterioration. RCS analysis revealed a continuous linear risk increase for HbA1c at baseline, whereas the platelet-to-lymphocyte ratio (PLR) at baseline exhibited an inverted U-shaped relationship. Among the ML algorithms, KNN achieved the highest AUC among models meeting the prespecified calibration criterion (AUC = 0.913; Brier score = 0.084; calibration slope = 0.701; calibration intercept = -0.053). SHAP analysis confirmed SGLT2 inhibitor use, HbA1c, and PLR as the top three predictors. Mediation analysis demonstrated that longitudinal increases in uric acid (ΔUA) and white blood cell count (ΔWBC) accounted for 20.6% and 12.9% of the association between HbA1c and DCAN progression, respectively.
Conclusions: SGLT2 inhibitor use is significantly associated with a lower risk of DCAN deterioration in T2DM patients, whereas elevated HbA1c levels correlate with disease progression, potentially involving pathways of exacerbated uric acid metabolism and systemic inflammation. Furthermore, the KNN-based ML model serves as a promising proof-of-concept tool for clinical risk stratification that warrants future external validation.

Keywords:  causal mediation analysis; diabetic cardiac autonomic neuropathy; machine learning; sodium-glucose cotransporter 2 inhibitors; uric acid

DOI:  https://doi.org/10.3389/fendo.2026.1901683
Dentomaxillofac Radiol. 2026 Aug 21. pii: twag062. [Epub ahead of print]

Evaluation of Differential Diagnosis of Odontogenic Lesions in Cone Beam Computed Tomography Images Using Radiomics-Based Machine Learning.

Zarif Ece Hammudıoglu, Merve Kucuk Kurtgoz, Elif Çeçen Erol, Ceren Aktuna Belgın, Kaan Orhan.

   OBJECTIVE: This study aimed to evaluate the structural characteristics of mandibular alveolar bone in patients with Type 1 diabetes mellitus (T1DM), Type 2 diabetes mellitus (T2DM), and systemically healthy controls using panoramic radiography-based radiomic analysis combined with machine learning algorithms.
MATERIALS AND METHODS: A total of 225 panoramic radiographs (75 T1DM, 75 T2DM, 75 healthy controls) were retrospectively analyzed. ROIs were segmented from eight anatomical mandibular segments per subject, and 107 radiomic features were extracted using PyRadiomics. Interobserver reliability was confirmed by two-way random-effects ICC (≥0.85). A leakage-free pipeline was applied. Four machine learning algorithms were evaluated: Random Forest, ExtraTrees, SVM-RBF, and Logistic Regression.
RESULTS: Significant differences were identified in age (Kruskal-Wallis p = 0.0003) and sex (χ²=17.857, p = 0.0001). The best single-segment performance was achieved in the left mandibular corpus with Logistic Regression (Accuracy=0.833, F1=0.832, AUC=0.958). All segments showed significant radiomic differences (FDR q < 0.001). Sensitivity of 1.000 was achieved for T1DM and AUC=1.000 for T2DM. The feature glszm_SizeZoneNonUniformityNormalized showed the strongest discriminative power (H = 86.928, ε²=0.578).
CONCLUSION: Panoramic radiography-based radiomic analysis demonstrates high diagnostic performance in non-invasively distinguishing mandibular bone alterations among T1DM, T2DM, and healthy individuals, with potential as a clinical bone monitoring tool.

Keywords:  diabetes mellitus; machine learning; mandibular bone; panoramic radiography; radiomics; trabecular bone

DOI:  https://doi.org/10.1093/dmfr/twag062
Front Cardiovasc Med. 2026 ;13 1903579

Machine learning-based prediction of prolonged length of stay in older patients with type 2 diabetes mellitus and cardiovascular disease.

Yixia Zuo, Jianfei Chen, Jie Wang, Zheng Zhu, Longao Huang, Qi Deng, Yan Li, Liping You, Jun Gu.

   Background: Older patients with type 2 diabetes mellitus (T2DM) and cardiovascular disease (CVD) frequently experience prolonged length of stay (PLOS). This condition increases healthcare burden and worsens prognosis. However, no predictive model specifically addresses PLOS in this high-risk multimorbid population.
Methods: This single-center retrospective study included hospitalized older T2DM-CVD patients. PLOS was defined as hospital stay exceeding the 75th percentile of the training set population. Potential predictors were selected via LASSO regression. Eight machine learning (ML) models were developed to predict PLOS risk. Model performance was evaluated using receiver operating characteristic curves, calibration curves, and decision curve analysis. SHAP analysis was employed for model interpretability.
Results: A total of 27,629 patients were included. The XGBoost model achieved the highest training AUC (0.819) and demonstrated competitive predictive performance in both the internal (AUC = 0.753) and time-based external (AUC = 0.728) validation sets. However, its performance advantage over simpler models such as logistic regression was modest in validation, and XGBoost showed some degree of overfitting (AUC drop of 0.066 from training to validation). Although logistic regression showed comparable validation performance with less overfitting, XGBoost was selected as the final model for its ability to capture complex nonlinear interactions and provide SHAP-based interpretability, with the understanding that further external validation is needed. Key predictors included cerebral infarction, white blood cell count, anemia, pulse rate, the glycated hemoglobin to high-density lipoprotein cholesterol ratio (GHR), and osteoporosis. Most continuous variables showed nonlinear associations with PLOS risk.
Conclusions: The XGBoost-based model effectively predicts PLOS risk in older T2DM-CVD patients. This tool shows promise for early identification of high-risk individuals and optimization of medical resource allocation within our institutional setting. However, further prospective and multi-center validation studies are required before clinical adoption.

Keywords:  XGBoost; cardiovascular disease; machine learning; prolonged length of stay; type 2 diabetes mellitus

DOI:  https://doi.org/10.3389/fcvm.2026.1903579
J Vis Exp. 2026 Jul 28.

Network Pharmacology, Machine Learning, and In Vivo Validation of Danzhi Jiangtang Capsule in Diabetic Nephropathy.

Ziqiang Li, Ying Yuan, Yueyue Zhao, Yifang Deng, Gang Yin, Anlu Shen.

Although Danzhi Jiangtang Capsule (DJC) is a traditional Chinese herbal preparation used clinically for diabetes, how it may protect the kidney during diabetic nephropathy (DN) has not been fully clarified. This investigation was designed to explore potential mechanisms by which DJC affects DN, with a focus on the NLR family pyrin domain-containing 3 (NLRP3)/Caspase-1/Gasdermin D (GSDMD) pyroptosis-related signaling cascade. An integrated strategy combining network pharmacology and machine learning was employed. The Traditional Chinese Medicine Systems Pharmacology Database and Analysis Platform (TCMSP) and the Bioinformatics Analysis Tool for Molecular Mechanism of Traditional Chinese Medicine (BATMAN-TCM) were used to screen bioactive ingredients and their corresponding protein targets of DJC. DN-associated genes were retrieved from Gene Expression Omnibus (GEO), GeneCards, and Online Mendelian Inheritance in Man (OMIM). Key candidate targets were screened and ranked using multiple machine learning algorithms. The binding affinity between DJC's active ingredients and core targets was assessed via molecular docking. Finally, the therapeutic efficacy and predicted mechanisms were evaluated in db/db diabetic mice. Network pharmacology analysis identified 599 DJC targets and 68 overlapping genes shared with DN. Using machine learning algorithms, C-C motif chemokine ligand 2 (CCL2) and CASP1 were identified as prioritized candidate targets. Molecular docking predicted possible strong binding affinities between DJC active ingredients and these core proteins. Functional enrichment analyses (GO/KEGG) suggested that DJC modulation is associated with inflammatory responses and the MAPK pathway. In vivo validation showed that DJC treatment attenuated renal injury and fibrosis markers. These findings suggest that DJC may attenuate DN in part through CCL2/C-C motif chemokine receptor 2 (CCR2)-related pyroptosis signaling changes.

DOI: https://doi.org/10.3791/71328
Front Public Health. 2026 ;14 1902424

Assessing the information quality of AI-generated patient educational materials for diabetes: a scoping review.

Jingwen Song, Norafisyah Makhdzir, Zarina Haron, Chunping Qin, Li Wang.

   Objective: To map the evidence on artificial intelligence (AI)-generated diabetes-related patient education materials and patient-facing health information, with particular attention to AI models, prompting approaches, evaluation methods, and information-quality outcomes.
Methods: This scoping review was conducted in accordance with the JBI methodology for scoping reviews and reported following the PRISMA-ScR checklist. The review was registered on the Open Science Framework (doi: 10.17605/OSF.IO/U4FAE) PubMed, Web of Science, Embase, Scopus, Cochrane CENTRAL, CNKI, WanFang Data, and SinoMed were searched from inception to May 1, 2026. Chinese- and English-language literature was searched. Two reviewers independently screened studies, charted data, and mapped reported outcomes to Wang and Strong's information quality framework. Outcomes not adequately represented by the framework were retained as additional dimensions. Descriptive statistics and narrative synthesis were used.
Results: Of 6,049 records identified, 24 studies from 11 countries or regions were included. All studies evaluated ChatGPT or another GPT-family model; 21 used zero-shot or direct prompting, three used role prompting, and two implemented retrieval-augmented generation. Eleven indicators were mapped to the information quality framework, with ease of understanding (n = 14), accuracy (n = 13), and believability (n = 9) assessed most frequently. Six additional outcomes were identified: clinical safety (n = 5), actionability (n = 3), response efficiency (n = 1), personalization (n = 1), transparency (n = 1), and empathy (n = 1). Most studies reported reading demands above those generally recommended for patient education, although findings varied by language, material type, and assessment method. Study-specific instruments were used in 17 studies (70.8%), whereas 10 (41.7%) used structured or established tools. Only six studies reported full source or model blinding, 10 reported quantitative inter-rater agreement, and three involved patients or members of the public.
Conclusion: Research on AI-generated diabetes education is expanding, but substantial heterogeneity in prompts, evaluators, tools, and outcome definitions limits comparison across studies. Future research should prioritize validated, multilingual, and patient-centered evaluation tools that integrate conventional information-quality attributes with clinically relevant dimensions such as safety, actionability, personalization, transparency, empathy, and response efficiency.

Keywords:  artificial intelligence; diabetes; information quality; patient education materials; scoping review

DOI:  https://doi.org/10.3389/fpubh.2026.1902424