bims-aukdir Biomed News
on Automated knowledge discovery in diabetes research
Issue of 2026–11–22
sixteen papers selected by
Mott Given



  1. Cureus. 2026 Oct;18(10): e117889
      Artificial intelligence (AI) is increasingly being investigated for glucose prediction and personalized glycemic management in diabetes, particularly through continuous glucose monitoring and multimodal patient data. However, strong predictive performance does not necessarily translate into clinically meaningful benefit. This systematic review evaluated the predictive performance and clinical outcomes of AI-based approaches for glucose prediction, prediction of glycemic events, and AI-enabled glycemic management in adults with diabetes mellitus. A systematic literature search was conducted in PubMed/MEDLINE, Scopus, Web of Science, the Cochrane Library, and Google Scholar from database inception through July 2026. Studies evaluating AI or machine-learning approaches for glucose prediction, prediction of hypoglycemia or hyperglycemia, or AI-enabled glycemic management were included. Data on study characteristics, diabetes type, AI architecture, input variables, prediction targets and horizons, validation strategies, predictive performance measures, clinical outcomes, and methodological quality were extracted. Risk of bias was assessed using design-appropriate validated tools. Thirty studies were included, comprising evidence from type 1 diabetes, type 2 diabetes, and mixed diabetes populations. Most studies investigated short-term glucose forecasting or hypoglycemia prediction, with prediction horizons ranging from approximately 15 minutes to 24 hours. Reported predictive performance was generally promising, although substantial heterogeneity in datasets, model architectures, prediction horizons, outcome definitions, and validation strategies limited direct comparisons. Increasingly, models incorporated insulin, meals, physical activity, demographic characteristics, and other clinical variables to support personalized and multimodal prediction. However, clinical evidence was considerably more limited than technical predictive evidence, with relatively few studies evaluating prospective clinical or decision-support interventions and clinically meaningful outcomes such as time in range or HbA1c. Overall, AI-based approaches demonstrated promising performance for glucose and glycemic event prediction, with a growing emphasis on personalized and multimodal modeling, but evidence for clinical effectiveness remains limited. Stronger prospective, externally validated, and randomized studies using standardized, clinically meaningful outcomes are needed to establish the clinical effectiveness, generalizability, and implementation of AI-based approaches in routine diabetes care.
    Keywords:  artificial intelligence; continuous glucose monitoring; diabetes mellitus; glucose prediction; glycemic management; type 1 diabetes; type 2 diabetes
    DOI:  https://doi.org/10.7759/cureus.117889
  2. BMC Med Inform Decis Mak. 2026 Aug 24. pii: 477. [Epub ahead of print]26(1):
       BACKGROUND: The global prevalence of diabetes is increasing, often leading to complications such as diabetic kidney disease (DKD). Uncontrolled blood glucose and hypertension are key risk factors that progressively impair kidney function, potentially resulting in kidney failure or end-stage renal disease (ESRD). Early detection of DKD is crucial but challenging due to its asymptomatic onset.
    METHODS: This study employs explainable artificial intelligence (XAI) to predict DKD risk in diabetic patients using tree-based ensemble models and Shapley Additive exPlanations (SHAP), leveraging the Medical Information Mart for Intensive Care (MIMIC)-IV dataset and a dataset from hospitals in Rwanda.
    RESULTS: Among the models used, Random Forest demonstrated superior performance, achieving accuracies of 87.97% on MIMIC-IV and 91.70% on the Rwandan dataset. XGBoost outperformed other models, achieving 79.32% accuracy when trained on the Rwanda dataset and evaluated on the MIMIC-IV dataset. Models were evaluated using multiple metrics, including ROC-AUC and calibration curves. SHAP provided both global and individual-level explanations, with predictions validated using estimated glomerular filtration rate (eGFR) values.
    CONCLUSION: Our findings highlight the potential of combining predictive modeling with explainability to support transparent and trustworthy early detection of DKD in clinical practice. Compared with the current eGFR-based risk assessment, this approach incorporates additional predictors that improve risk identification. Blood urea nitrogen (Urea), Hemoglobin, Total cholesterol (TC), Blood glucose (BG), and Creatinine are the leading predictors of DKD risk. This study suggests that the XAI model can be used to assess DKD risk, while eGFR remains valuable for confirming the diagnosis and disease stage.
    Keywords:  Diabetes management; Diabetic kidney disease; Ensemble models; Explainable artificial intelligence; Machine learning; SHAP
    DOI:  https://doi.org/10.1186/s12911-026-03782-w
  3. Ophthalmol Sci. 2026 Dec;6(12): 101417
       Purpose: To use machine learning models to classify diabetic patients into low-risk (candidate for biennial screening) and high-risk (needing annual screening) categories based on their likelihood of worsening diabetic retinopathy (DR) only (excluding progression of diabetic macular edema) within 1 and 2 years from tabular data.
    Design: Retrospective cohort study.
    Participants: Deidentified electronic health record data of 32 542 patients with diabetes mellitus from the United States Veterans Health Administration Technology and Eye Care Services teleophthalmology program.
    Methods: Eight prediction approaches were evaluated, spanning rule-based methods, traditional machine learning, and deep learning architectures.
    Main Outcome Measures: Area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, expected calibration error, and Brier score, with 95% confidence intervals (CIs).
    Results: Sixteen thousand six hundred and six veterans with diabetes and no DR at baseline were included in the 1-year cohort and 15 397 in the 2-year cohort. Predictive models demonstrated clinically meaningful risk stratification across both prediction horizons. A deep learning model (Bidirectional Encoder Representations from Transformers for Healthcare) achieved the strongest discrimination, with an AUROC of 0.91 (95% CI, 0.85-0.96) at 1 year and 0.84 (95% CI, 0.76-0.91) at 2 years. When constraining sensitivity to 90% to capture ≥9 in 10 patients who will progress, Extreme Gradient Boosting correctly identified 53% of nonprogressors as low risk at 1 year and 52% at 2 years, suggesting that more than half of low-risk patients could be deferred from routine screening without substantially compromising the detection of true progressors. Performance across both horizons was achieved using only routinely collected tabular variables, without retinal imaging.
    Conclusions: Machine learning models can effectively identify low-risk diabetic patients suitable for biennial screening intervals without retinal imaging. These models may also help prioritize patients who have not yet been screened given that DR frequently remains undiagnosed. Together, our findings support the feasibility of transitioning a meaningful proportion of low-risk patients to biennial screening, with the potential to reduce unnecessary examinations and optimize resource allocation in high-burden health systems without compromising the detection of clinically significant disease progression.
    Financial Disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
    Keywords:  Diabetic retinopathy screening; Electronic health records.; Machine learning; Time series analysis
    DOI:  https://doi.org/10.1016/j.xops.2026.101417
  4. Pediatr Diabetes. 2026 ;2026 e011
      This paper evaluates deep learning models for blood glucose prediction in children and adolescents with type 1 diabetes using insulin pumps and investigates whether transfer learning from adult data improves prediction accuracy and safety. Models were trained using data from the Type 1 Diabetes Exercise Initiative Pediatric study (84 pediatric insulin pump users) and its adult counterpart (380 adult insulin pump users). Two transfer learning strategies were compared: one using pediatric data only and another pre-training the model on adult data before fine-tuning it on the pediatric cohort and each target patient. Prediction accuracy and clinical safety were assessed using root mean square error (RMSE) and the Clarke Error Grid (CEG) at 10-, 20-, and 30-min prediction horizons, with results analyzed separately for open-loop and closed-loop users. Pre-training on adult data yielded the best performance. At 30 min, mean RMSE was 19.61 mg/dL (1.09 mmol/L), with at least 92.64% of predictions falling within zone A of the CEG for 75% of patients. No predictions fell within the unsafe zone E. Delayed hypoglycemia detection occurred only in a small subset of patients and was mainly associated with missing contextual information or atypical clinical events. An additional evaluation on pediatric patients treated with multiple daily injections (MDI) demonstrated clinically acceptable prediction accuracy and safety, supporting the applicability of the proposed transfer learning approach to MDI users. These findings demonstrate that transfer learning from adult data effectively mitigates pediatric data scarcity and enables accurate, clinically safe blood glucose prediction under free-living conditions.
    Keywords:  Blood glucose prediction; Deep learning; Pediatric type-1 diabetes; Transfer learning; Type 1 diabetes exercise initiative
    DOI:  https://doi.org/10.48130/pedi-0026-0010
  5. Healthc Inform Res. 2026 Oct;32(4): 354-364
       OBJECTIVES: Early detection of diabetic kidney disease (DKD) is difficult in semi-urban healthcare settings where diagnostic resources are limited and comorbidities overlap. This study developed and evaluated an interpretable ensemble learning pipeline that combined hybrid resampling and high-performance computing to predict DKD early in patients with type 2 diabetes mellitus (T2DM).
    METHODS: Data from a retrospective cohort of 210 adults with T2DM were analyzed using 35 demographic, lifestyle, and biochemical variables. Missing values (<7%) were imputed with the median or mode. Class imbalance (DKD:non-DKD = 1:2.6) was addressed by combining the synthetic minority oversampling technique with Tomek link cleaning and random undersampling. A stacking ensemble of logistic regression, random forest, and XGBoost was evaluated using accuracy, recall, specificity, F1-score, and area under the receiver operating characteristic curve (AUC). Shapley additive explanations (SHAP) were used to interpret the features.
    RESULTS: The ensemble achieved an AUC of 0.948 and an F1-score of 0.947. Data augmentation increased recall from 0.83 to 0.95 but reduced specificity from 1.00 to 0.89. SHAP identified urea, albumin, and diabetes duration as the most influential predictors, with better detection among patients with a longer diabetes duration. Using histogram-based XGBoost reduced training time by approximately 63%, and inference latency was below 0.02 seconds per patient record.
    CONCLUSIONS: The ensemble classifier pipeline discriminated DKD well and provided interpretable predictions in this moderate-sized cohort. Hybrid resampling improved sensitivity at the cost of lower specificity. External validation and longitudinal evaluation are needed before wider clinical use.
    Keywords:  Artificial Intelligence; Computing Methodologies; Diabetic Nephropathies; Early Diagnosis; Machine Learning
    DOI:  https://doi.org/10.4258/hir.2026.32.4.354
  6. J Patient Exp. 2026 ;13 23743735261493024
      Diabetic retinopathy (DR) remains a leading cause of preventable vision loss, yet the pathways patients navigate to be screened are fragmented and inconvenient, and structural barriers-not patient neglect-drive poor adherence. This commentary explores how artificial intelligence (AI)-enabled autonomous screening, embedded in primary care, can reshape the patient and family experience while improving clinical outcomes. Drawing on published evidence and early implementation experience from Saudi Arabia, where DR is common among people with diabetes, we present actionable insights for healthcare leaders. We emphasize experience-related gains-same-visit results that reduce waiting-related anxiety, fewer trips and unnecessary referrals, faster time to treatment, and clearer communication-alongside outcomes and cost. We address implementation challenges including clinician and patient trust, automation and confirmation bias, workflow integration, equity, and preserving human connection in AI-mediated care. Recommendations emphasize embedding screening in existing diabetes visits, pilot-to-enterprise scaling, transparent patient communication, and measuring patient-reported experience. AI-enabled DR screening offers a compelling, patient-centered model for specialty care transformation under value-based frameworks such as Saudi Vision 2030.
    Keywords:  access to care; artificial intelligence; diabetic retinopathy; patient experience; patient-reported experience measures; value-based care
    DOI:  https://doi.org/10.1177/23743735261493024
  7. Technol Health Care. 2026 Nov 18. 9287329261508376
      BackgroundDiabetes mellitus has become one of the prominent global health issues that demand early and precise predictions to avoid complications. Thus, an adaptive deep learning method for classifying diabetes is proposed in this work.ObjectiveThe proposed technique Progressive-Growth Deep Neural Network (PG-DNN) with Particle Swarm Optimization and the Genetic Algorithm (PSO-GA) to adjust the parameters. The experiments use PIMA, LMCH, and the large-scale diabetes prediction dataset.MethodsA comprehensive preprocessing technique is used to enhance data quality. Missing values are handled through imputation using polynomial regression models. Outliers are detected using the interquartile range and the Isolation Forest algorithm. For feature selection, Spearman correlation analysis is used. Data standardization and SMOTE-based imbalance handling are also considered. Stratification and repeated k-fold validation techniques have been used to compare the performance of the proposed PG-DNN with existing models, such as support vector machines, random forests, and standard deep neural networks. The PG-DNN architecture is incrementally expanded by learning deeper-level features; the GA-PSO optimization technique enables faster learning and improved convergence.ResultsExperiments demonstrate a consistent performance advantage for all datasets. The proposed approach achieves 95.6% accuracy on the primary dataset, 97.4% on the PIMA dataset, and 99.5% on the LMCH dataset. The interpretability analysis using LIME and SHAP identifies biologically significant features.ConclusionThe work presents a scalable and practical solution for predicting early diabetes.
    Keywords:  Diabetes mellitus; deep neural network; hyperglycemia; isolation forest; particle swarm optimization; polynomial regression; progressive-growth DNN; spearman correlation
    DOI:  https://doi.org/10.1177/09287329261508376
  8. Spectrochim Acta A Mol Biomol Spectrosc. 2026 Nov 15. pii: S1386-1425(26)01703-8. [Epub ahead of print]368(Pt 2): 129132
      The World Health Organization has labelled diabetes the 21st-century epidemic affecting hundreds of millions of people worldwide. It leads to significant complications affecting essential organs such as the kidneys, heart, and liver. This study investigated changes in phospholipid and lipid profiles in the heart and liver collected from control and type 2 diabetes mellitus (T2DM) model rats. Homogenized heart and liver samples were analyzed using Attenuated Total Reflectance- Fourier Transform Infrared Spectroscopy (ATR-FTIR) and Liquid Chromatography-Mass Spectrometry (LC-MS) metabolic profiling in conjunction with machine learning. We showed that in heart samples, diabetes induced a marked reduction in lipid content, disrupted phospholipid profiles, and altered amide I/II bands, suggesting protein glycation and membrane remodeling. In contrast, diabetic livers exhibited increased lipid accumulation and glycogen-related spectral features consistent with hepatic steatosis. PCA distinguished diabetic from control spectra with high confidence, particularly in the liver. Decision tree modelling identified the 1509 cm-1 band as a robust discriminant among the samples. Machine learning-guided analysis (AdaBoost) further identified key wavenumbers (e.g., 1779 cm-1 in heart, and 1021 cm-1 in liver) that robustly discriminate diabetic samples from control ones, achieving 100% classification accuracy for liver. In both heart and liver, lipidomic profiling revealed a predominant downregulation of phosphatidylcholines, lysophosphatidylcholines, and sphingomyelins, indicative of impaired membrane lipid metabolism, oxidative stress, and disrupted lipid signaling in the diabetic state. Notably, the liver exhibited a compensatory upregulation of PUFA-enriched phosphatidylcholines, suggesting an adaptive response to oxidative damage. Obtained results showed, that ATR-FTIR spectroscopy, combined with multivariate and machine learning approaches, effectively detects diabetes-related biochemical changes in cardiac and hepatic tissues.
    Keywords:  ATR-FTIR; Diabetes; Machine learning; Metabolic profiling; Phospholipid-lipid balance
    DOI:  https://doi.org/10.1016/j.saa.2026.129132
  9. Int J Low Extrem Wounds. 2026 Nov 17. 15347346261507599
      Purposeof the research: Diabetic peripheral neuropathy induces progressive gait deterioration, yet most Machine Learning (ML)-based gait studies rely on binary classification frameworks that fail to capture the stage-wise biomechanical changes occurring across the diabetes neuropathy continuum. To identify stage-consistent and stage-specific gait biomarkers across three clinical groups healthy controls (N), diabetes without neuropathy (D), and diabetes with peripheral neuropathy (DPN) using an interpretable machine learning framework.MethodsGait data were collected from 394 participants (N = 109, D = 163, DPN = 122) using the Zebris FDM platform. Spatiotemporal, plantar pressure, and symmetry parameters were extracted. Four pairwise classification tasks (N vs D, D vs DPN, N vs DPN, Healthy vs Diseased) were evaluated using Results: Support Vector Machine (SVM), Random Forest, Extreme Gradient Boosting (XGBoost), and Logistic Regression under stratified five-fold cross-validation. Cross-validated SHapley Additive exPlanations (SHAP)-based feature selection and a final XGBoost model were used for biomarker identification and interpretation. SVM achieved the highest area under the receiver operating characteristic curve (AUC) across all tasks (N vs D: 0.80, D vs DPN: 0.75, N vs DPN: 0.91, Healthy vs Diseased: 0.85). SHAP analysis identified bilateral heel contact time and step length as stage-consistent core biomarkers, present across all four classification tasks. Stage-specific discriminators included forefoot load-timing disruption in early diabetes, midfoot unloading and lateral symmetry reversal at the neuropathic transition, and posterior Centre of Pressure (CoP) displacement in established neuropathy.ConclusionsDiabetic gait pathology progresses in a measurable, stage-ordered sequence from early forefoot load-timing disruption through midfoot redistribution and symmetry reversal at the neuropathic transition, to global spatiotemporal deterioration in established neuropathy. The stage-wise biomarker framework presented here provides an interpretable, clinically grounded basis for gait-based screening of diabetic neuropathic progression, warranting prospective validation.
    Keywords:  diabetic peripheral neuropathy; gait analysis; gait biomarkers; gait variability; machine learning
    DOI:  https://doi.org/10.1177/15347346261507599
  10. IEEE Open J Eng Med Biol. 2026 ;7 241-249
      Glucose dysregulation in non-diabetic individuals can precede metabolic disease onset by years, yet most glucose prediction models target diabetic populations with structured insulin records. This article presents Patch Time-Aware Cross-Attention (Patch-TACA), a multimodal transformer-based framework for long-term blood glucose forecasting in healthy individuals. The approach combines patch-based self-supervised pretraining with a time-aware cross-attention mechanism to integrate continuous glucose monitoring (CGM) data with physiological and behavioral measures, such as accelerometry, heart rate, electrodermal activity, and meal macronutrients, without requiring explicit resampling of irregularly sampled signals. Patch-TACA is assessed using data from twelve healthy participants across three laboratory sessions. It achieves an RMSE of 14.26 ± 3.48 mg/dL at the 90-minute prediction horizon, outperforming GlySim and Gluformer baselines. Moreover, employing a self-supervised method for pre-training reduced RMSE by 7.9 compared to training with random weight initialization, and hyperglycemia prediction accuracy reached 93.9. These results demonstrate that multimodal sensor fusion with self-supervised learning enables accurate long-horizon glucose forecasting in non-diabetic populations, supporting proactive metabolic health monitoring before clinical disease onset.
    Keywords:  Blood glucose forecasting; continuous glucose monitoring; cross-attention; self-supervised learning; wearable sensors
    DOI:  https://doi.org/10.1109/OJEMB.2026.3705652
  11. Int J Med Inform. 2026 Nov 10. pii: S1386-5056(26)00551-4. [Epub ahead of print]224 106811
       BACKGROUND: Machine-learning models for undiagnosed diabetes are often trained on self-reported diagnosis, a label that depends on access to testing, and few are compared with guideline rules or deployed clinically.
    OBJECTIVE: To quantify how the training label changes whom a screening model prioritizes, to correct models trained on self-reported diagnosis, and to develop, validate and deploy an interpretable screener anchored to glycated hemoglobin (HbA1c).
    METHODS: This retrospective cross-sectional study used the National Health and Nutrition Examination Survey (NHANES) 2013-2014 (5,328 adults with HbA1c), the Behavioral Risk Factor Surveillance System (BRFSS) 2015 (253,680 adults) and an external cohort (388 adults) for a measurement-only model. An Explainable Boosting Machine (EBM) and three guideline rules were compared at equal referral rates in survey-weighted cross-validation. Paired models differing only in their label were compared by Shapley attribution and risk-rank shift. A diagnosis-propensity model learned in NHANES corrected a BRFSS model by positive-unlabeled learning.
    RESULTS: Of all diabetes, 16.4 % was undiagnosed. At the US Preventive Services Task Force referral rate (44 % of adults), the 13-item EBM found 85 % (95 % CI 75-94) of HbA1c-defined undiagnosed cases versus 65 % (55-74) for the age-limited rule. Access and socioeconomic items carried 23.8 % of attribution under the self-reported label and 17.7 % under the HbA1c label. The correction raised a BRFSS model's area under the receiver operating characteristic curve from 0.775 to 0.816, a gain robust to moderate violations of its assumptions; externally, a measurement-only EBM reached 0.788 but needed recalibration. The offline tool reproduced every tested prediction.
    CONCLUSIONS: Self-reported diagnosis labels partly encode access to care. In retrospective data from one survey cycle, an interpretable HbA1c-anchored screener was more sensitive than the age-limited rule at equal testing; with only internal and small measurement-only external validation, prospective evaluation in primary care must precede routine use.
    Keywords:  Clinical decision support; Explainable artificial intelligence; Health equity; Label bias; Machine learning; Screening; Undiagnosed diabetes
    DOI:  https://doi.org/10.1016/j.ijmedinf.2026.106811
  12. J Clin Nurs. 2026 Nov 14.
       AIMS: This study aimed to develop a predictive model for diabetes self-management (DSM) using a fully connected neural network (FCNN) architecture. The model's performance was rigorously evaluated using key metrics, including accuracy, precision, recall, and the area under the receiver operating characteristic curve (AUC-ROC).
    DESIGN: A quantitative, multicenter cross-sectional design.
    METHODS: A systematic literature review identified key psychosocial factors and corresponding validated scales, which formed the initial item pool. A multicenter cross-sectional study was then conducted to collect data on these items from 2005 participants across five research centers in Hainan, Zhejiang, Jiangsu, and Henan provinces. We chose to develop an FCNN model over traditional approaches including logistic regression and random forest, as it could automatically capture complex, non-linear relationships among predictors without manual feature engineering. Using this model for gradient-based feature importance ranking, we selected the most predictive items from the initial pool following preliminary screening of overlapping items. After feature selection, the model was retrained using the selected features, with 80% of the data allocated for training and 20% for testing. Finally, model performance was validated through a two-week prospective observational study.
    RESULTS: Twenty-four out of 99 candidate variables were identified as model characteristics for predicting DSM behaviours. The model demonstrated strong performance across all four self-management domains. For healthy eating, the accuracy, precision, recall, and AUC were 83.34%, 88.56%, 90.05%, and 0.84%, respectively. For medication adherence, these values were 82.14%, 86.60%, 90.27%, and 0.84%, respectively. For regular exercise, they were 82.39%, 82.89%, 86.09%, and 0.89%, respectively. For blood glucose monitoring, they were 83.39%, 85.45%, 89.72%, and 0.87%, respectively. The prospective validation cohort (n = 50) further confirmed the model's robust performance, yielding AUCs of 0.84, 0.82, 0.83, and 0.89 for diet, exercise, blood glucose monitoring, and medication adherence, respectively.
    CONCLUSION: The FCNN-based predictive model demonstrated promising performance in forecasting DSM behaviours, with particularly strong discriminative ability as evidenced by high accuracy and AUC values. These findings suggest that deep learning approaches incorporating psychosocial and clinical variables may help predict DSM behaviours, although the model requires further validation before it can be considered a clinical tool for personalized diabetes care. Although the prospective validation results are encouraging, they should be interpreted as preliminary given the limited sample size (n = 50); larger-scale, multicenter external validation studies are warranted to further confirm the model's generalizability and clinical applicability.
    IMPACT: Predicting core diabetes self-management behaviours using an FCNN may provide useful insights into individualized adherence patterns and could inform targeted, stepped-care interventions, although further validation is needed before clinical application.
    REPORTING METHOD: The reporting of this cross-sectional study followed the STROBE guideline.
    PATIENT OR PUBLIC CONTRIBUTION: Patients with Type 2 diabetes mellitus contributed to this study by completing questionnaires during the model development phase (n = 2005) and by participating in the prospective validation phase (n = 50). No patients or public members were involved in study design, data analysis, or interpretation of results. Study findings will be disseminated to participants through their respective endocrinology departments.
    Keywords:  Type 2 diabetes; deep learning; diabetes selfmanagement; prediction model
    DOI:  https://doi.org/10.1111/jocn.70608
  13. J Diabetes Res. 2026 ;pii: e3351404. [Epub ahead of print]2026(1):
       OBJECTIVE: The aim of this study is to systematically investigate molecular mechanisms associated with gut microbial metabolite effects in gestational diabetes mellitus (GDM), and to screen and validate key host targets to provide foundations for early diagnosis and targeted intervention.
    METHODS: Gut microbial metabolite targets from the gutMGene database were integrated with GDM patient differentially expressed genes from the GEO database (GSE103552, 11 GDM vs. 8 controls). Core targets were screened via protein-protein interaction network and machine learning, then validated by Western blot (n = 5 per group). Functional enrichment analysis was used to elucidate biological pathways; single-cell transcriptomic data (GSE173193: 2 GDM vs. 2 controls; GSE212309: 4 GDM vs. 3 controls) validated cell-specific target expression; immune signature analysis explored microenvironment correlations. Drug-likeness and toxicity were assessed for candidate metabolites.
    RESULTS: Seventeen intersection targets between gut microbiota metabolites and GDM were identified, primarily enriched in AGE-RAGE, PI3K-Akt, and cell cycle signaling pathways. Machine learning consistently screened GSK3B and CDK4 as core targets. Western blot analysis confirmed significant upregulation of GSK3B in GDM placental tissues. Single-cell analysis revealed high GSK3B expression in extravillous trophoblasts and CDK4 downregulation in villous cytotrophoblasts. Both genes showed associations with immune cell signature scores. The constructed metabolite-target network revealed complex interaction relationships involving 26 metabolites and 51 gut microbes.
    CONCLUSION: GSK3B and CDK4 were identified as candidate host targets associated with gut microbiota metabolites in GDM, potentially contributing to pathogenesis through placental cell function and immune microenvironment changes, providing targets for mechanistic investigation and therapeutic exploration.
    Keywords:  CDK4; GSK3B; gestational diabetes mellitus; gut microbiota metabolites; machine learning; network pharmacology
    DOI:  https://doi.org/10.1155/jdr/3351404
  14. Healthc Inform Res. 2026 Oct;32(4): 365-375
       OBJECTIVES: Type 2 diabetes mellitus (T2DM) arises from complex interactions between genetic and environmental factors, creating a need for approaches that integrate inherited susceptibility and cumulative environmental exposure. Previous models have focused on clinical or genetic data and provided limited mechanistic insight. This study developed a machine-learning model that integrated a polygenic risk score (PRS) and DNA methylation to classify prevalent T2DM and assessed their individual and combined contributions to classification performance.
    METHODS: Models were developed and validated using data from the Ansan/Ansung cohort (n = 1,318) and the independent Health Examinee cohort (n = 358). PRS was calculated using external genome-wide association study summary statistics, and DNA methylation features were represented by β-values at T2DM-related cytosine-phosphate-guanine sites. Three feature sets were compared: Base, Base + PRS, and Base + PRS + DNA methylation. Performance was evaluated using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). Decision curve analysis assessed clinical utility, and Shapley additive explanations (SHAP) were used to interpret the models.
    RESULTS: Adding PRS and then DNA methylation features progressively increased AUROC across algorithms. The linear support vector machine with the fully integrated feature set achieved the highest discrimination (AUROC = 0.896, AUPRC = 0.801). PRS improved accuracy, and DNA methylation data produced further gains, particularly in recall and F1-score. Decision curve analysis showed that the integrated model provided greater net benefit across clinically relevant threshold probabilities. SHAP analysis identified PRS as the most influential individual feature and DNA methylation features as the largest contributors at the group level.
    CONCLUSIONS: Integrating PRS and DNA methylation improves the performance of machine-learning models for T2DM classification. Genetic and epigenetic factors thus provide complementary information for characterizing prevalent T2DM.
    Keywords:  Cohort Studies; DNA Methylation; Genetic Risk Score; Machine Learning; Type 2 Diabetes Mellitus
    DOI:  https://doi.org/10.4258/hir.2026.32.4.365
  15. Front Endocrinol (Lausanne). 2026 ;17 2030428
      [This corrects the article DOI: 10.3389/fendo.2026.1936566.].
    Keywords:  disease risk prediction; hyperspectral tongue imaging; machine learning algorithms; tongue coating microbiota; type 2 diabetes mellitus
    DOI:  https://doi.org/10.3389/fendo.2026.2030428
  16. Front Med (Lausanne). 2026 ;13 1893262
       Objective: To analyze the factors associated with sarcopenic obesity (SO) in patients with type 2 diabetes (T2DM), and to develop and internally validate an SO risk stratification model based on interpretable machine learning, thereby providing guidance for the clinical screening of high-risk patients and the implementation of personalized risk assessment.
    Methods: A single-center cross-sectional study design was adopted, enrolling 378 patients with T2DM, who were divided into an SO group and a non-SO group based on SO diagnostic results. Demographic characteristics, disease-related data, anthropometric measurements, and laboratory test results were collected. The cohort was randomly divided into a training set (n = 265) and a validation set (n = 113) in a 7:3 ratio. Univariate analysis, LASSO regression, and multivariate logistic regression were used to identify factors associated with SO, and based on these results, five risk-stratification models were constructed: logistic regression, random forest, SVM, LightGBM, and XGBoost. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), accuracy, sensitivity, recall, precision, F1 score, and Brier score. Model calibration performance and clinical applicability were assessed using calibration curves, calibration slope, calibration intercept, calibration-in-the-large (CITL), and decision curve analysis (DCA). Supplementary internal validation of the LightGBM model was performed using 1,000 bootstrap resamples, and interpretability analysis was conducted using the SHAP method.
    Results: In this study, the detection rate of SO in patients with T2DM was 36.2%, age, ADL, waist circumference, HDL, and 25(OH)D were the primary factors influencing SO. The AUC values of the five models on the validation set ranged from 0.806 to 0.821. Among them, LightGBM achieved the numerically highest AUC in this sample, at 0.821 (95% CI: 0.744-0.897), with an AUPRC of 0.712; however, the DeLong test indicated that the differences in AUC among the models were not statistically significant (all P > 0.05). Overall, the models demonstrated good calibration performance and potential clinical net benefit; however, bootstrap internal validation indicated that LightGBM still exhibited a certain degree of optimism bias and under-calibration. SHAP analysis revealed that age, ADL, waist circumference, HDL, and 25(OH)D contributed significantly to model predictions, with age being the most important predictive feature.
    Conclusion: The SO risk stratification models developed using machine learning methods demonstrated good performance and potential clinical utility, with LightGBM achieving the highest numerical performance. SHAP-based interpretability analysis enhanced model transparency and clinical credibility, providing a reference for assisted screening, risk stratification, and individualized assessment of SO in patients with T2DM.
    Keywords:  internal validation; machine learning; risk stratification model; sarcopenic obesity; type 2 diabetes mellitus
    DOI:  https://doi.org/10.3389/fmed.2026.1893262