bims-aukdir 2026-11-15 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2026–11–15
27 papers selected by
Mott Given

A Patch-Based Attention Framework for Automated Diabetic Retinopathy Severity Grading Using Fundus Images.
Benchmark Accuracy Does Not Transfer: Domain Adaptive Deep Learning for Diabetic Retinopathy Detection in a Bangladeshi Clinical Setting.
Machine Learning Models for Glycemic Prediction in Type 1 and Type 2 Diabetes: Systematic Review and Meta-Analysis.
Temporal stability of an electronic health record-driven machine learning model predicting type 2 diabetes among patients with mental illness.
Risk prediction models for macrovascular complications in type 2 diabetes mellitus: from traditional risk scores to multimodal intelligent prediction.
Development and temporal validation of an explainable machine learning model for predicting in-hospital diabetic foot ulcer in patients with type 2 diabetes using inflammatory-nutritional composite indices.
From task-specific networks to foundation models in diabetes care: a critical review of opportunities, evaluation gaps, and safety frontiers.
A Modular Diabetic Retinopathy Analysis Framework for Lesion Segmentation and Severity Grading: Cross-Dataset External Validation from APTOS to Messidor-2.
An explainable artificial intelligence system for prognostic assessment and decision support in diabetic foot ulcer care: a multicentre derivation and validation study with prospective randomised evaluation.
Per-Lesion Mask-Supervised Attention for Multi-Label Diabetic Retinopathy Lesion Detection.
Artificial Intelligence in Diabetes Mellitus Management: A Narrative Review of Advances, Architectures, and Translational Challenges.
Developing Machine Learning-Based Prognostic Models for Nonmelanoma Skin Cancer Patients With Comorbid Diabetes: A Retrospective Cohort Study.
Machine learning-based prediction of ventilator-associated pneumonia in ICU patients with type 2 diabetes.
Clinical readiness of first-trimester gestational diabetes prediction: From risk score to care pathway.
Exploring Predictive Modeling Algorithms for Good Glycemic Control in Patients With Type 2 Diabetes Mellitus: A Secondary Analysis of a Cluster Randomized Trial Cohort Further Explored by Sex.
Agentic LLM reasoning in pediatric diabetes outcome prediction from multimodal medical records.
Deep learning detection and biological interpretation of arterial stiffness from retinal fundus images in diabetes.
Global Trends and Exposure Profiling of Early-Onset Type 2 Diabetes: An Ecological Analysis Using Machine Learning.
Image-Based Dietary Assessment Using Multimodal Large Language Models: Performance, Limitations, and Clinical Applicability in Diabetes.
Design and Simulation of an AI-Powered Artificial Pancreas System With Personalized Glucose Sensing and Insulin Delivery.
Simulation-Based Development, Stress Testing, and External Physiological Benchmarking of a DKA Early-Warning Framework Intended for Future Nurse-Facing Evaluation.
Safety, accuracy, empathy and information quality of six large language models answering patient-facing questions about diabetic kidney disease: a cross-sectional comparative study.
A personalized glucose-aware meal guidance agent for type 2 diabetes self-management.
Targeted cystatin C testing for large negative eGFR discordance in adults with diabetes: development and external validation of a machine-learning model.
Identification of key indicators of colorectal cancer risk in patients with diabetes mellitus using machine learning and Mendelian randomization.
Structured multifrequency segmental bioimpedance descriptors for diabetes screening support.
Large Language Model-Assisted Discharge Instructions for Diabetic Kidney Disease: Comparative Evaluation Study.

Diagnostics (Basel). 2026 Nov 01. pii: 3601. [Epub ahead of print]16(21):

A Patch-Based Attention Framework for Automated Diabetic Retinopathy Severity Grading Using Fundus Images.

Abdulnasır Yıldız, Hasan Zan.

  Background/Objectives: Automated diabetic retinopathy (DR) severity grading from fundus images is a challenging task due to the high resolution of retinal images and the spatially sparse nature of pathological lesions. Methods: In this study, we propose a patch-based attention framework for DR severity grading that exploits localized information while avoiding aggressive global downsampling. Each fundus image is partitioned into non-overlapping patches, which are processed using a shared ResNet-18 encoder to extract patch-level feature representations. An attention-based aggregation mechanism then assigns adaptive importance weights to the extracted features and produces an image-level representation for severity classification. The proposed framework is evaluated on the APTOS 2019 dataset. Results: An analysis of different patch configurations identifies 25 patches as the best-performing setting, achieving a multi-class accuracy of 0.852, compared with 0.817 for the whole-image baseline. A controlled ablation using the same effective input resolution further shows that attention-based aggregation achieves the highest accuracy and F1-score among the evaluated whole-image and patch-based aggregation configurations. The model also achieves a binary DR/No-DR accuracy of 0.986. Attention visualizations provide qualitative insight into the regions contributing to model predictions. Conclusions: These results demonstrate the potential of attention-guided patch-based feature learning for accurate and interpretable diabetic retinopathy severity grading.

Keywords:  attention mechanism; diabetic retinopathy; fundus images; patch-based learning; severity grading

DOI:  https://doi.org/10.3390/diagnostics16213601
J Imaging Inform Med. 2026 Nov 12.

Benchmark Accuracy Does Not Transfer: Domain Adaptive Deep Learning for Diabetic Retinopathy Detection in a Bangladeshi Clinical Setting.

Riad Hossain, Ittekhar Mahin Abir, Ayesha Banu, Mohammed Nazim Uddin.

  Deep learning could make diabetic retinopathy (DR) screening more accessible where ophthalmologists are scarce, but high benchmark accuracy does not guarantee reliable clinical performance. We investigate this gap using 3662 images from APTOS 2019 for source training and a balanced evaluation corpus of 500 fundus images (250 DR, 250 No-DR) from a Bangladeshi clinic, graded by a single experienced ophthalmologist. Source-trained models achieved 94.26-96.80% accuracy under tenfold cross-validation on APTOS, yet achieved only 49.4-50.6% (mean 50.0%) when applied directly to the clinical images. The resulting 44.46-46.85 percentage-point losses across all six architectures provide consistent quantitative evidence of severe domain shift. A full-cohort feature comparison yielded a classical-feature Fréchet distance of 133.49 and an MMD 2 of 0.3967 (permutation p=0.0005 ). To recover target-domain performance, we evaluate an adaptation pipeline that integrates retinal-specific preprocessing, progressive fine-tuning with discriminative learning rates, soft-voting inference, and test-time augmentation. Aggregated out-of-fold evaluation produced 81.0% accuracy (95% Wilson CI 77.3-84.2%), 70.0% sensitivity, 92.0% specificity, 89.7% precision, and a 78.7% F1 score. Fine-tuning increased mean individual-model accuracy by 27.1 percentage points, whereas preprocessing contributed a descriptive 2.0-point increase. These findings quantify a substantial benchmark-to-clinic gap and show that target-domain adaptation can improve performance, while the remaining false-negative rate and single-clinic design preclude claims of clinical readiness.

Keywords:  Clinical deployment; Deep learning; Diabetic retinopathy; Domain adaptation; Domain shift; Ensemble learning; Test-time augmentation

DOI:  https://doi.org/10.1007/s10278-026-02454-5
J Med Internet Res. 2026 Nov 09. 28 e100653

Machine Learning Models for Glycemic Prediction in Type 1 and Type 2 Diabetes: Systematic Review and Meta-Analysis.

Haoran Yang, Diliyaer Abudukeremu, Yuhan Shi, Li Fang, Lei Li, Wenzheng Gu, Yunqi Jia, Rongfang He.

   Background: Machine learning models are increasingly used for glycemic event prediction and continuous glucose forecasting, but targets, populations, horizons, validation, and uncertainty reporting remain heterogeneous.
Objective: This study aimed to evaluate machine learning models for glycemic prediction in type 1, type 2, and mixed diabetes populations while distinguishing the full evidence map from the quantitatively analyzable subset.
Methods: We searched 9 English- and Chinese-language databases from inception to April 12, 2026, for studies predicting future glycemic events or glucose values in people with diabetes. Continuous forecasting and binary event prediction were analyzed separately. Root-mean-square error (RMSE) was the principal continuous end point; mean absolute error was supportive. Area under the receiver operating characteristic curve was synthesized using a Bayesian random-effects model because analyzable 2×2 data were sparse. Two reviewers (HY and DA) independently applied PROBAST (Prediction Model Risk of Bias Assessment Tool) to 19 quantitative-source reports (28 assessment units).
Results: We identified 787 records, assessed 310 full texts, and included 299 reports representing 270 studies; 19 reports contributed to quantitative synthesis. The review included 128 type 1, 26 type 2, 78 mixed, and 38 unclear-diabetes-type studies. Pooled RMSE was 1.18 mmol/L (95% CI 1.01-1.34; 37 effects), supportive mean absolute error was 0.90 mmol/L (95% CI 0.78-1.02; 20 effects), and Bayesian pooled area under the receiver operating characteristic curve was 0.926 (95% credible interval 0.877-0.957). Supportive sensitivity and specificity were 0.77 and 0.80. Of 28 PROBAST units, 24 were high risk, 2 unclear, and 2 low. Post hoc exploratory within-report analysis estimated an RMSE increase of 0.481 mmol/L per additional 30 minutes (95% CI 0.213-0.749), but overlap sensitivity reduced certainty. In the 270-study clinical readiness audit, external validation occurred in 4.0% (10/251), prospective evaluation in 0.7% (2/270), calibration in 4.8% (13/270), decision-curve or net-benefit analysis in 2.6% (7/270), and implementation or human-factors evaluation in 2.6% (7/270).
Conclusions: Glycemic prediction studies show forecasting and discrimination signals, but heterogeneity, high risk of bias, incomplete uncertainty reporting, and limited external and prospective validation preclude deployable benchmarks. Pooled estimates should be viewed as summaries of analyzable literature. Future studies need transparent horizon, threshold, calibration, validation, uncertainty, and implementation reporting.

Keywords:  artificial intelligence; blood glucose forecasting; continuous glucose monitoring; diabetes mellitus; hyperglycemia; hypoglycemia; machine learning; meta-analysis; systematic review; type 1 diabetes; type 2 diabetes

DOI:  https://doi.org/10.2196/100653
Acta Neuropsychiatr. 2026 Nov 13. 38 e62

Temporal stability of an electronic health record-driven machine learning model predicting type 2 diabetes among patients with mental illness.

Frida Hæstrup, Andreas Aalkjær Danielsen, Søren Dinesen Østergaard.

   OBJECTIVES: We previously trained and tested a machine learning model to predict incident type 2 diabetes (T2D) in patients with mental illness using electronic health record (EHR) data from routine clinical practice. This hyperparameter-tuned XGBoost model - based on information on sex, age and mean HbA1c within the past two years - predicts whether a patient will meet criteria for T2D in the two years following each contact to psychiatric services. The performance of such a model may drift over time due to changes in standards of care, baseline risk, registration procedures, etc. Therefore, we set out to examine its temporal stability.
METHODS: In the present study, we fitted the model described above on EHR data from January 1, 2013 to December 31, 2017, and examined its stability in predicting T2D two years into the future for patients with contacts in 2018, 2019, 2020, 2021, 2022, and 2023, respectively. Performance was quantified using the area under the receiver operating characteristic curve (AUROC).
RESULTS: The model showed good discriminative performance across all years (AUROC for 2018 = 0.84, 2019 = 0.83, 2020 = 0.87, 2021 = 0.80, 2022 = 0.84, 2023 = 0.80), with a linear model estimating a decrease in AUROC for T2D prediction of only 0.004 per year (95%CI: -0.024; 0.016). Secondary analyses suggested that models trained on recent data (past year) may achieve similar or better discrimination than models trained on longer historical periods.
CONCLUSIONS: The temporal stability of the T2D prediction model supports its potential for clinical implementation, facilitating early detection and tailored interventions, while highlighting the importance of continued evaluation.

Keywords:  Diabetes mellitus, Type 2; electronic health records; epidemiology; machine learning; mental disorders

DOI:  https://doi.org/10.1017/neu.2026.10105
Front Endocrinol (Lausanne). 2026 ;17 1980631

Risk prediction models for macrovascular complications in type 2 diabetes mellitus: from traditional risk scores to multimodal intelligent prediction.

Xuedong An, Jingru Cai, Xiaoting Cao, Jinglun Li.

  Macrovascular complications are the leading cause of morbidity and mortality in patients with type 2 diabetes mellitus (T2DM). Early risk identification and accurate risk stratification are essential for both primary and secondary cardiovascular prevention. With the rapid development of electronic health records, medical imaging, multi-omics technologies, and artificial intelligence (AI), risk prediction models for diabetic macrovascular complications have evolved from traditional statistical approaches to machine learning-, deep learning-, and multimodal AI-driven models. However, despite substantial improvements in predictive performance, current models remain limited by heterogeneous target populations, inconsistent outcome definitions, insufficient external validation, limited clinical interpretability, and inadequate real-world implementation, thereby restricting their clinical translation. This review summarizes the evolution, data sources, modeling methodologies, clinical applications, and future perspectives of risk prediction models for macrovascular complications in T2DM. We first describe the transition from conventional cardiovascular risk scores, including the Framingham Risk Score, to diabetes-specific models such as the UKPDS Risk Engine, ADVANCE, RECODe, and SCORE2-Diabetes, followed by the emergence of AI-based prediction models. We then review the major data sources used in contemporary prediction models, including clinical characteristics, laboratory biomarkers, vascular function and imaging, multi-omics profiles, and wearable device-derived data. Furthermore, we compare the strengths and limitations of traditional statistical methods, machine learning, deep learning, and multitask learning with respect to predictive performance, interpretability, and clinical applicability, and discuss the major challenges related to study design, external validation, data quality, algorithmic bias, and real-world implementation. Finally, we propose that future risk prediction models for diabetic macrovascular complications will evolve from single-source data to multimodal data integration, from static risk estimation to dynamic risk trajectories, from single outcomes to multiple interconnected cardiovascular outcomes, from prognosis prediction to individualized treatment benefit prediction, and from standalone algorithms to continuously learning clinical decision support systems. Emerging technologies, including foundation models, large language models, digital twins, causal inference, and continuous learning frameworks, are expected to drive the transition from Risk Prediction to Clinical Decision and ultimately Patient Benefit, thereby advancing precision prevention and individualized management of diabetes.

Keywords:  machine learning; macrovascular complications; precision medicine; risk prediction; type 2 diabetes mellitus

DOI:  https://doi.org/10.3389/fendo.2026.1980631
Front Med (Lausanne). 2026 ;13 1938429

Development and temporal validation of an explainable machine learning model for predicting in-hospital diabetic foot ulcer in patients with type 2 diabetes using inflammatory-nutritional composite indices.

Xue Wang, Chao Li, Yanan Tuo, Bo Zhao, Ayiding Xiahati, Lu Fan, Paerhati Wahafu.

   Objective: This study aimed to develop and temporally validate explainable machine learning models for predicting in-hospital new-onset DFU in hospitalized patients with type 2 diabetes mellitus (T2DM) using routinely available admission variables and inflammatory-nutritional composite indices.
Methods: This retrospective prediction-model study enrolled 550 consecutive hospitalized patients with T2DM admitted between January 2022 and December 2024. Patients admitted during 2022-2023 (n = 400) constituted the development cohort, whereas those admitted in 2024 (n = 150) served as an independent temporal validation cohort. Candidate predictors included demographic characteristics, diabetes duration, routine laboratory biomarkers, ankle-brachial index (ABI), and inflammatory-nutritional composite indices, including neutrophil percentage-to-albumin ratio (NPAR), albumin-to-neutrophil-to-lymphocyte ratio (ANLR), hemoglobin-albumin-lymphocyte-platelet score (HALP), systemic immune-inflammation index (SII), and C-reactive protein-to-albumin ratio (CRP/ALB). Candidate variables were screened using univariable logistic regression followed by least absolute shrinkage and selection operator (LASSO) regression. Logistic regression, random forest, LightGBM, and XGBoost models were developed and evaluated using the area under the receiver operating characteristic curve (AUC), average precision (AP), sensitivity, specificity, F1-score, Brier score, calibration analysis, and decision curve analysis (DCA). Model interpretability was assessed using SHapley Additive exPlanations (SHAP).
Results: During hospitalization, 66 patients (12.0%) developed DFU. Eleven predictors were retained after feature selection, including age, diabetes duration, neutrophil percentage, C-reactive protein, high-density lipoprotein cholesterol, glycated hemoglobin, ABI, NPAR, ANLR, HALP, and SII. In the temporal validation cohort, logistic regression had the highest observed discrimination (AUC = 0.976; bootstrap 95% CI, 0.936-0.992), with a sensitivity of 0.944, specificity of 0.932, and F1-score of 0.773 at the prespecified threshold; however, its Hosmer-Lemeshow test indicated evidence of lack of fit (P = 0.001). LightGBM had the lowest Brier score (0.064), whereas XGBoost achieved an AUC of 0.945, a Brier score of 0.067, and a Hosmer-Lemeshow P value of 0.350. At the prespecified thresholds, LightGBM and XGBoost sensitivities were only 0.056 and 0.389, respectively. SHAP analysis identified ANLR, glycated hemoglobin, HALP, age, and SII as the five most influential XGBoost predictors. Because the temporal validation cohort contained only 18 DFU events, these performance estimates should be interpreted cautiously.
Conclusion: Admission-based machine learning models showed high observed discrimination for in-hospital DFU risk stratification, but no model was uniformly superior across discrimination, calibration, and threshold-dependent sensitivity. The small number of validation events, possible redundancy among mathematically overlapping predictors, omitted established foot-risk variables, and the low sensitivity of LightGBM and XGBoost at their prespecified thresholds preclude direct use as screening tools. Prospective multicenter validation, formal assessment of predictor redundancy, threshold recalibration, and clinical-impact evaluation are required before implementation.

Keywords:  diabetic foot ulcer; explainable artificial intelligence; inflammatory biomarkers; machine learning; nutritional status; shapley additive explanations; temporal external validation; type 2 diabetes mellitus

DOI:  https://doi.org/10.3389/fmed.2026.1938429
Front Artif Intell. 2026 ;9 1983619

From task-specific networks to foundation models in diabetes care: a critical review of opportunities, evaluation gaps, and safety frontiers.

M Nanda Kishore, C Navaneethan.

  Artificial intelligence in diabetes care has historically relied on task-specific supervised learning, such as bespoke networks for retinopathy grading, kidney disease prediction, and glucose forecasting. The emergence of foundation models (FMs: large models pretrained on broad data, typically with self-supervision, and adapted to many downstream tasks) is displacing this paradigm across diabetes-relevant modalities: ophthalmic models such as RETFound, pretrained on 1.6 million retinal photographs, have been reported to outperform task-specific networks on retinopathy grading and systemic disease prediction, especially when labeled data are scarce; transformer-based time-series models are replacing long short-term memory (LSTM) pipelines for continuous glucose monitoring (CGM) forecasting; clinical large language models (LLMs) such as Med-PaLM and GPT-4 attain expert-level scores on medical question-answering benchmarks; and generative models supply synthetic CGM and electronic health record (EHR) data. This review provides a critical, modality-organized synthesis of this transition. We (i) formalize a three-era taxonomy: rule-based risk scores (era I), task-specific deep learning (era II, 2016-2020), and foundation models (era III, from 2021), and map the diabetes literature onto it; (ii) survey foundation-model architectures, pretraining objectives, and adaptation strategies across retinal imaging, CGM, longitudinal EHR, and multimodal data; (iii) expose an evaluation crisis: benchmark contamination of LLM medical evaluations, a near-universal absence in the literature we reviewed of external and prospective validation for retinal and CGM foundation models, miscalibration under dataset shift, and, to our knowledge and within the reviewed literature (January 2016-December 2025; negative findings are time-sensitive), the absence of any randomized evaluation of a foundation model in a diabetes workflow; and (iv) analyze safety risks: hallucinated insulin-dosing advice, automation bias, inherited algorithmic inequity, and privacy exposure during adaptation. We conclude that the decisive open problem is no longer representation learning but grounded, calibrated, causally informed deployment, requiring contamination-resistant evaluation, prospective trials, and hybrid architectures coupling foundation-model representations with treatment-effect estimation.

Keywords:  clinical evaluation; continuous glucose monitoring; diabetic retinopathy; electronic health records; foundation models; large language models; model calibration; self-supervised learning

DOI:  https://doi.org/10.3389/frai.2026.1983619
Diagnostics (Basel). 2026 Oct 31. pii: 3596. [Epub ahead of print]16(21):

A Modular Diabetic Retinopathy Analysis Framework for Lesion Segmentation and Severity Grading: Cross-Dataset External Validation from APTOS to Messidor-2.

Sofiane Boumeshed, Yazid Cherfa, Assia Cherfa, Noureddine Belkhamsa.

  Background/Objectives: High in-domain accuracy of diabetic retinopathy (DR) grading models does not guarantee a similar performance for a new population. We quantify how an APTOS-trained grader degrades on an independent cohort and examine why. Methods: A modular framework combines a ResNet-34 U-Net for lesion segmentation (IDRiD) with an EfficientNet-B0 severity grader (APTOS 2019), optionally fused with lesion-derived and frequency features. The grader was evaluated zero-shot on Messidor-2 (n=1744) with bootstrap confidence intervals, four training seeds, referable-DR analysis, a discrimination-versus-calibration decomposition, a paired preprocessing ablation, and a duplicate audit. Results: Quadratic weighted kappa (QWK) fell from 0.873 (95% CI 0.831-0.909) on the APTOS test set to 0.455 (0.400-0.507) on Messidor-2; across four seeds the external QWK was 0.410±0.057. Referable-DR sensitivity fell from 92.7% to 35.0% (specificity 93.9% and 96.6%) and the AUC from 0.984 to 0.795, indicating both a loss of discrimination and a calibration shift; recalibrating thresholds with target labels (cross-validated) raised QWK only to 0.583. Omitting CLAHE reduced QWK to 0.326 (paired difference 0.129, 0.089-0.171). Thirty-three APTOS test images duplicated in the training or validation data changed QWK marginally (0.866). Lesion-feature fusion did not significantly improve grading. The segmentation Dice on lesion-containing patches was 0.51-0.68, whereas the per-image Dice was 0.21-0.61. Conclusions: In-domain agreement is a poor proxy for external performance; multi-seed reporting, explicit calibration analysis, and preprocessing control are needed before clinical use.

Keywords:  automated grading; calibration shift; computer-aided diagnosis; cross-dataset external validation; deep learning; diabetic retinopathy; fundus imaging; lesion segmentation

DOI:  https://doi.org/10.3390/diagnostics16213596
EClinicalMedicine. 2026 Nov;101 104284

An explainable artificial intelligence system for prognostic assessment and decision support in diabetic foot ulcer care: a multicentre derivation and validation study with prospective randomised evaluation.

Jia-Yu Zhang, Yangyang Zhuang, Qingping Cao, Chenxi Song, Yanjun Feng, Gaosheng Zhuang, Zhe Ji, Xueyang Li, Qiang Li, Peisheng Jin, Aijun Zhang, Chi Zhang.

   Background: Diabetic foot ulcers (DFUs) carry risks of recurrence, amputation, and all-cause mortality, but most prediction models address single outcomes and have limited interpretability. We developed an interpretable multi-outcome machine-learning framework for individualised DFU risk assessment and prospective clinical evaluation.
Methods: We conducted a development and validation study across three hospitals in China from January 1, 2018, to April 30, 2025. We used electronic medical records and follow-up data from 763 patients at Centres A and B to develop and internally validate an ensemble of single-task extreme gradient boosting (ST-XGBoost) and single-task multilayer perceptron (ST-MLP) models to predict cure, recurrence, amputation, and all-cause mortality. The locked ensemble was externally validated in 100 patients at Centre C. Model performance was assessed using discrimination, calibration, and decision curve analysis, with interpretability evaluated using Shapley additive explanations (SHAP). From January 5 to March 5, 2026, we performed a prospective comparative evaluation. We randomised 202 participants (1:1) to usual care or artificial intelligence (AI)-assisted care, of whom 200 were analysed. We compared AI-assisted physician decision-making via an evolvable web-based online prediction platform versus usual care. The outcome was overall quality score of the final clinical recommendation. The platform supports follow-up data collection for future offline model updating.
Findings: In the training dataset, cure, recurrence, amputation, and all-cause mortality occurred in 638, 208, 226, and 167 patients, respectively; the corresponding numbers in the external validation dataset were 80, 34, 24, and 25. In independent external validation, the ST-XGBoost + ST-MLP ensemble outperformed six comparators, with a macro-averaged AUC of 0.901, accuracy of 0.875, and F1 score of 0.830. Outcome-specific discrimination was highest for amputation (AUC 0.953) and all-cause mortality (AUC 0.951), followed by cure (AUC 0.880) and recurrence (AUC 0.820). In the prospective evaluation, AI-assisted recommendations received higher scores than corresponding clinician recommendations (7.77 vs 6.47; mean difference 1.30, 95% confidence interval (CI) 1.08-1.52; p < 0.0001), with the largest gain in completeness. The AI-assisted group showed better short-term follow-up engagement and glucose-management adherence.
Interpretation: We demonstrate that integrating gradient-boosted trees and neural networks enables explainable prediction of complex DFU trajectories and may support prioritised, individualised clinical decision-making. Nevertheless, recurrence estimates should be interpreted cautiously.
Funding: Xuzhou Municipal Health Commission Science and Technology Project (XWYHT20240128).

Keywords:  Artificial intelligence; Clinical decision support systems; Diabetic foot ulcers; Machine learning; Prognosis

DOI:  https://doi.org/10.1016/j.eclinm.2026.104284
Diagnostics (Basel). 2026 Nov 05. pii: 3662. [Epub ahead of print]16(21):

Per-Lesion Mask-Supervised Attention for Multi-Label Diabetic Retinopathy Lesion Detection.

Wael Alghamdi, Fawaz Alsolami, Mostafa Saleh.

  Background/Objectives: Deep learning models for diabetic retinopathy (DR) usually report a severity grade, and their attention maps explain that grade rather than individual lesions. We aimed to detect microaneurysms, hemorrhages, hard exudates and soft exudates (cotton-wool spots) separately, each with an attention map trained to fall on that lesion. Methods: An EfficientNet-B4 and Transformer encoder, pretrained on severity grades, feeds four lesion queries that each return a score and an attention map; a localization loss trains each map toward its lesion mask. We trained it on DDR and TJDR, tested it on their test sets and on two external sets (IDRiD and e-ophtha), and compared it with a U-Net. Results: The mean areas under the curve (AUCs) were 0.93-0.94 for hemorrhages, 0.90-0.96 for hard exudates and 0.81-0.90 for soft exudates; microaneurysms reached 0.91-0.96, except on DDR (0.58). We analyze this drop in detail. Attention ranked lesion pixels above the rest of the retina with an AUC-Judd of 0.948-0.996. Without the localization loss, detection was slightly lower, and attention moved to regions outside the lesions that still raised the prediction. The model detected lesions better than U-Net (mean AUC of 0.876 against 0.853) and, unlike the latter, also learned from images that have labels but no masks: with masks on half of the training images, detection and AUC-Judd were nearly unchanged. Conclusions: Supervising each lesion's attention with its own mask gives a per-lesion score and a map that falls on that lesion, without reducing detection.

Keywords:  attention supervision; cross-attention; deep learning; diabetic retinopathy; explainable AI; fundus photography; hard exudates; microaneurysms; retinal hemorrhages; soft exudates

DOI:  https://doi.org/10.3390/diagnostics16213662
Curr Diabetes Rev. 2026 Nov 10.

Artificial Intelligence in Diabetes Mellitus Management: A Narrative Review of Advances, Architectures, and Translational Challenges.

Uriti Sri Venkatesh, Ankita Wal, Kumar Vimal, Sakshi Srivastava, Tushar Lokhande, Akanksha Singh, Mohd Javed Naim, Krishana Kumar Sharma, Amin Gasmi.

   INTRODUCTION: Diabetes Mellitus (DM), a chronic disorder caused by insulin deficiency, dysfunction, or both, is projected to affect 640 million people by 2040. Traditional management is limited by drug resistance, poor adherence, and insufficient personalization. This review examines Artificial Intelligence (AI) applications addressing these gaps in DM care.
METHODS: PubMed, Scopus, Web of Science, IEEE Xplore, and Google Scholar were searched (January 2020-December 2025) using combinations of diabetes mellitus, artificial intelligence, machine learning, deep learning, federated learning, explainable AI, and blockchain. Studies addressing AI-based risk prediction, glucose forecasting, complication detection, and secure or transparent frameworks were included, reported by diabetes subtype where relevant.
RESULTS: Convolutional neural networks, long short-term memory networks, support vector machines, and transformer architectures have improved glucose forecasting, complication screening, and risk prediction. Federated learning enables privacy-preserving multi-institutional training, while SHAP and LIME improve interpretability, and blockchain offers a tamper-proof audit layer. However, most high-performing models remain internally validated, limiting generalizability.
DISCUSSION: Unlike prior reviews that address a single application domain or treat T1DM and T2DM as a single population, this review is organized around their distinct AI design needs and classifies applications as experimental, commercial, or regulator-approved. AI can extend beyond technical accuracy by integrating diverse clinical data, though explainability and validation remain central to adoption.
CONCLUSION: AI has the potential to transform DM care through more precise forecasting, earlier complication detection, and individualized treatment. Realizing this requires stronger regulatory frameworks, external validation, reduced demographic bias, and improved interpretability, prioritized through prospective multicenter studies, standardized datasets, and explainability standards embedded in clinical workflows.

Keywords:  Diabetes mellitus; artificial intelligence; clinical translation; explainable AI; federated learning; glucose forecasting; machine learning; narrative review.

DOI:  https://doi.org/10.2174/0115733998507302261019052129
Cancer Med. 2026 Nov;pii: e72050. [Epub ahead of print]15(11):

Developing Machine Learning-Based Prognostic Models for Nonmelanoma Skin Cancer Patients With Comorbid Diabetes: A Retrospective Cohort Study.

Junjie Huang, Claire Chenwen Zhong, Zhaojun Li, Zehuan Yang, Yu Jiang, Jinqiu Yuan, Jonathan Poon, Qi Dou, Martin C S Wong.

   BACKGROUND: Nonmelanoma skin cancer (NMSC) is one of the most common malignancies worldwide, with increasing incidence. Type 2 diabetes mellitus (T2DM) is a prevalent comorbidity that may adversely affect cancer prognosis through shared metabolic and inflammatory pathways. This study aims to identify risk factors influencing the survival of NMSC patients with T2DM, construct and evaluate the performance of different predictive models, and develop a comprehensive risk score to accurately predict the survival probability of these patients.
METHODS: We collected data from the Hong Kong Hospital Authority Data Collaboration Laboratory, including 1325 NMSC patients with T2DM between 2000 and 2020. Cox proportional hazards (CoxPH), least absolute shrinkage and selection operator-Cox model (LASSO-Cox), random survival forest (RSF), and survival tree were employed to construct prediction models. SHapley Additive exPlanations (SHAP) analysis was conducted to identify key risk factors and attribute model outputs. A risk score system was developed using the AutoScore-Survival package.
RESULTS: We identified seven significant predictors for poorer overall survival among patients with T2DM, including higher age at cancer diagnosis, male sex, smoking, longer duration of T2DM, elevated creatinine and glycated hemoglobin (HbA1c), and lower fasting glucose. The RSF model showed the highest predictive performance among the evaluated machine learning models. The tuned AutoScore-Survival risk score system demonstrated good predictive performance (concordance index (C-index) = 0.712 in the test set), and was based on the following variables: age at cancer diagnosis, duration of T2DM, creatinine, insulin treatment, stroke, and HbA1c. The areas under the curve (AUCs) at 1, 3, and 5 years in the test set were 0.677, 0.697, and 0.729, respectively.
CONCLUSION: Older age at cancer diagnosis, male sex, longer diabetes duration, smoking, elevated creatinine and HbA1c, and lower fasting glucose were identified as adverse prognostic factors for NMSC patients with T2DM. The RSF model outperformed other algorithms in predictive accuracy. The proposed risk score offers a clinically interpretable and effective means of predicting survival among NMSC patients with T2DM.

Keywords:  diabetes mellitus; machine learning; nonmelanoma skin cancer; random survival forest; risk score; survival analysis

DOI:  https://doi.org/10.1002/cam4.72050
BMC Pulm Med. 2026 Aug 18. pii: 562. [Epub ahead of print]26(1):

Machine learning-based prediction of ventilator-associated pneumonia in ICU patients with type 2 diabetes.

Aoxing Shi, Meixia Song, Zhongqiang Guo, Lunmeng Cui, Guangxin Chen, Ying Cai, Zhaoyu Gao.

   BACKGROUND: Ventilator-associated pneumonia (VAP) remains one of the most common and serious complications among mechanically ventilated patients in the ICU. Patients with type 2 diabetes mellitus (T2DM) are considered a vulnerable population for VAP because of impaired immune function and metabolic dysregulation. This study aimed to develop and validate machine learning (ML) models for the early prediction of VAP in mechanically ventilated ICU patients with T2DM.
METHODS: A total of 876 mechanically ventilated ICU patients with T2DM from the MIMIC-IV database were included and randomly divided into training and testing sets at a ratio of 7:3. Feature selection was performed using the Boruta algorithm, Least Absolute Shrinkage and Selection Operator (LASSO) regression, and Spearman correlation analysis. The Synthetic Minority Over-sampling Technique (SMOTE) was applied to address class imbalance. Four ML models, including logistic regression (LR), random forest (RF), gradient boosting machine (GBM), and eXtreme Gradient Boosting (XGBoost), were developed. Model performance was comprehensively evaluated using receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis (DCA), and multiple performance metrics, including recall, specificity, and F1-score. Shapley Additive Explanations (SHAP) were applied to interpret the optimal model, and a web-based risk calculator was further developed. In addition, sensitivity analyses were performed by retaining the Glasgow Coma Scale (GCS) score and using the original dataset without SMOTE to evaluate the robustness of the model.
RESULTS: A total of 876 patients were included, of whom 73 developed VAP and 803 did not. XGBoost demonstrated the best predictive performance, with an area under the receiver operating characteristic curve (AUC) of 0.891 (95% CI 0.862-0.920), a recall of 0.832, a specificity of 0.877, and an F1-score of 0.851 in the testing set. Calibration curves showed good agreement between predicted and observed outcomes, and DCA indicated favorable potential clinical net benefit. SHAP analysis identified the seven most influential predictive features in the model, including Glucose_max, SOFA-Renal score, hypertension, Temp_min, Scr_max, WBC_min, and SOFA-CNS score.
CONCLUSION: Four ML models were developed and compared for the early prediction of VAP in mechanically ventilated ICU patients with T2DM, among which XGBoost demonstrated the best predictive performance. SHAP improved the interpretability of the model and identified the seven predictive features that contributed most to model predictions. Future prospective multicenter external validation studies are needed to further evaluate the model's generalizability and provide additional evidence for the early risk assessment of VAP in patients with T2DM.

Keywords:  ICU; MIMIC-IV database; Machine learning; Predictive model; Type 2 diabetes mellitus; Ventilator-associated pneumonia

DOI:  https://doi.org/10.1186/s12890-026-04609-0
Diabetes Res Clin Pract. 2026 Nov 13. pii: S0168-8227(26)00587-5. [Epub ahead of print] 113667

Clinical readiness of first-trimester gestational diabetes prediction: From risk score to care pathway.

Eduard Cristian Mihoci, Demetra Socolov, Răzvan Vladimir Socolov, Ana Maria Mihoci, Dragoş Nemescu.

   PURPOSE: To assess whether first-trimester gestational diabetes mellitus (GDM) prediction can improve clinical decisions, rather than merely predict a later abnormal glucose test.
METHODS: We critically reviewed targeted PubMed searches, independent validations, prospective implementation studies, randomized trials and guidelines. Searches were non-systematic and selection limitations are reported.
RESULTS: Routine clinical and glycemic predictors are feasible, but transportability, calibration, and strategy-level benefit are not established consistently. RESPECT 2 demonstrated implementation feasibility; a trial of nutrition support in 519 model-selected women did not significantly reduce GDM. TOBOGM investigated treatment after early biochemical diagnosis, not care guided by a risk score. Direct trials of the complete prediction-guided strategy remain limited.
CONCLUSION: First-trimester scores should not replace recommended diagnostic testing. Clinical adoption requires local validation, a prespecified beneficial action, evaluation of costs and harms, and comparative trials measuring maternal and neonatal outcomes.

Keywords:  Clinical utility; External validation; First trimester; Gestational diabetes mellitus; Machine learning; Prediction model

DOI:  https://doi.org/10.1016/j.diabres.2026.113667
Health Sci Rep. 2026 Nov;9(11): e73465

Exploring Predictive Modeling Algorithms for Good Glycemic Control in Patients With Type 2 Diabetes Mellitus: A Secondary Analysis of a Cluster Randomized Trial Cohort Further Explored by Sex.

INTEGRATE team

   Background and Aims: Predictive modeling algorithms show promise in identifying risk factors influencing clinical outcomes. This study aims to evaluate different machine learning algorithms to identify an accurate predictive model of glycemic control at 18-months follow-up in type 2 diabetes mellitus (T2DM) patients and explore sex-specific risk factors.
Methods: Data of T2DM patients from the INTEGRATE dataset (2016-2019) were used. Baseline demographic, anthropometric, clinical biomarkers, and Charlson Comorbidity Index (CCI) were included to develop predictive models using LASSO logistic regression, CART, XGBoost, and Random Forest. Model discrimination was assessed using the area under the curve (AUC), calibration using calibration intercept, slope, and Brier score, clinical utility using decision-curve analysis (DCA), and model interpretability using SHapley Additive exPlanations (SHAP). Sex-specific predictors were identified by training separate LASSO models independently within each sex stratum.
Results: Participants (n = 4821) had a mean age of 65.4±12.8 years, BMI of 31.4 ± 6.7 kg/m2, and 46.6% were female. The predictive models had an AUC of 0.68-0.71, with broadly acceptable calibration across all models. Decision-curve analysis demonstrated net benefit over treat-all and treat-none strategies across clinically plausible threshold probabilities (10%-50%). Insulin use was the strongest predictor of poor glycemic control (HbA1c ≥ 7.0%) in both sexes (males: -2.46, 95% CI: -2.91 to -1.96; females: -2.10, 95% CI: -2.67 to -1.64). SHAP analysis confirmed diabetes medication intensity as the dominant predictor. Sex-specific predictors included higher HDL in males, and severe comorbidity index in females, although estimates were imprecise.
Conclusion: LASSO logistic regression and XGBoost showed the highest discrimination (AUC: 0.71) with broadly acceptable calibration. Decision-curve analysis supported clinical utility across plausible threshold probabilities. SHAP analysis confirmed diabetes medication intensity as the dominant predictor, consistent with sex-stratified findings. External validation in independent cohorts remains necessary before clinical implementation.

Keywords:  glycemic control; machine learning; predictive modeling; sex differences; type 2 diabetes

DOI:  https://doi.org/10.1002/hsr2.73465
ArXiv. 2026 Oct 29. pii: arXiv:2610.40379v1. [Epub ahead of print]

Agentic LLM reasoning in pediatric diabetes outcome prediction from multimodal medical records.

Daniel Palacios, Guantong Qi, Anna Kadi, Natali Borrego, Sukru Aras, Michael L Ferm, Elizabeth Kubota-Mishra, Maria J Redondo, Zhandong Liu, Mustafa Tosur, Hu Chen.

Large language model (LLM) agents are proposed for clinical prediction but are untested in pediatrics, where cohorts are small and outcomes are uncommon. Using records from 2,489 youth with type 2 diabetes (T2D) and 4,018 with type 1 diabetes (T1D), spanning structured data, note-derived social determinants and, in T1D, continuous glucose monitoring data, we predicted seven T2D outcomes, four also in T1D, at 24 months from 18 months of follow-up. CEDAR (Clinical Evidence-Driven Agent for Risk-prediction), an evidence-citing, self-auditing LLM pipeline, outperformed trained machine learning in both cohorts (pooled area under the receiver operating characteristic curve 0.815 versus 0.656 to 0.700 in T2D and 0.756 versus 0.603 to 0.631 in T1D), most on the sparsely screened outcomes. Only 4 of 70,246 values it cited did not match the chart. Reading only the evidence and reasoning cited by CEDAR, two pediatric endocrinologists reached the same prediction as CEDAR in about nine of ten cases, agreeing with each other on 90.5%.
iScience. 2026 Nov 20. 29(11): 118046

Deep learning detection and biological interpretation of arterial stiffness from retinal fundus images in diabetes.

Zhiyun Zhao, Kang Dang, Ziheng Zhou, Ying Peng, Yan Liu, Shiyuan Zhan, Shengyin Jiao, Zilong Wang, Jingqi Niu, Qidong Zheng, Dong Zhao, Liujun Fu, Yuancheng Dai, Ping Zhang, Pei Gu, Yan Wang, Qin Wan, Yufan Wang, Yikui Lv, Weijiang Chu, Yanmei Yu, Xingjian Zhou, Hongmei Qiu, Zhiqiang Kang, Mingfu Ma, Yifei Zhang, Guang Ning, Xiaowei Ding, Weiqing Wang.

  Pulse wave velocity is the gold standard for assessing arterial stiffness (AS) but is costly and not widely accessible. This multicenter study developed a deep learning model to predict AS from fundus images. Trained on 64,440 images from 36,412 adults with diabetes at China's National Metabolic Management Centers and externally validated on 6,217 images from two independent hospitals, the model achieved an AUC of 0.780 (95% CI: 0.769-0.792) and 0.772 (95% CI: 0.754-0.790) in internal and external validation, respectively, outperforming models using classical risk factors alone. The fusion model integrating fundus images with clinical factors significantly improved test-set performance (AUC = 0.853, 95% CI: 0.844-0.863) and demonstrated good external generalizability (AUC = 0.801, 95% CI: 0.784-0.816). Ablation studies indicated retinal arteries were associated with AS, with other retinal structures also contributing. These findings support deep-learning-based fundus-image analysis for AS screening in real-world clinical settings, and retinal-structure analyses for exploring AS pathophysiology.

Keywords:  arterial stiffness; deep learning; diabetes mellitus; retinal fundus

DOI:  https://doi.org/10.1016/j.isci.2026.118046
Diabetes Obes Metab. 2026 Nov 09.

Global Trends and Exposure Profiling of Early-Onset Type 2 Diabetes: An Ecological Analysis Using Machine Learning.

Wude Guo, Yang Liu, Jing Wang, Yi Ma, Jiaxin Chen, Chengyao Sun, Guoxiang Liu.

   BACKGROUND: Early-onset Type 2 diabetes (EOT2D) is an increasing global health concern. We examined global trends in EOT2D incidence and identified robust population-level exposure correlates.
METHODS: Using the data from the GBD 2021, we analysed EOT2D incidence among individuals aged 15-39 years across 204 countries and territories from 1990 to 2021. Temporal trends were assessed using estimated annual percentage change. Candidate exposures were evaluated using Elastic Net, Random Forest and XGBoost with temporally separated training, validation and test periods. Country-level bootstrap resampling, robust rank aggregation (RRA) and SHapley Additive exPlanations (SHAP) analyses assessed stability and cross-model consensus. Selected exposures were further examined using year fixed-effects and two-way fixed-effects (TWFE) models. Composite exposure indices were constructed to characterize geographic patterns of joint exposure.
RESULTS: EOT2D incidence increased globally, with persistently high rates in Pacific Island countries and rapid increases in Egypt, Greenland and Afghanistan. XGBoost showed the best out-of-time predictive performance. High body mass index and diet low in whole grains were the only cross-model consensus predictive exposures retained by RRA after multiple-testing correction and consistently ranked highest across models and temporal periods. In the year fixed-effects model, both exposures were positively associated with EOT2D incidence. After adjustment for country fixed effects, high body mass index remained significant. Composite exposure indices showed marked geographic heterogeneity but broadly similar patterns across weighting strategies.
CONCLUSIONS: EOT2D incidence is rising globally. Population adiposity showed the most consistent association across analytical approaches, while the association with EOT2D incidence of low whole-grain intake was more context dependent.

Keywords:  Type 2 diabetes; global burden of disease; machine learning; risk factors

DOI:  https://doi.org/10.1111/dom.71495
medRxiv. 2026 Oct 24. pii: 2026.10.21.26365949. [Epub ahead of print]

Image-Based Dietary Assessment Using Multimodal Large Language Models: Performance, Limitations, and Clinical Applicability in Diabetes.

Lila Gravellier, Chuyi Zhang, Yue Wu, Vijaya Parameswaran, Agatha Scheideman, Mandy Shao, Emma Zhang, Dalia Perelman, Christopher Gardner, David C Klonoff, Helge Ræder, Michael P Snyder, Tao Wang.

Background: Image-based dietary assessment is valuable for precision diabetes care. Multimodal large language models (mLLMs) may substantially facilitate dietary self-monitoring. However, comprehensive evaluation of their real-world accuracy remains limited.
Methods: We developed SeeFood, an open-source benchmark, and evaluated eight commercial and open-source mLLMs on 80 real-world meal photographs from the SNAPMe dataset for nutrient and energy estimation (calories, carbohydrates, protein, fat, fiber), food-item identification, and day-level diet quality (Healthy Eating Index, HEI-2020). Three finalists were evaluated on the complete dataset (1680 photographs, 275 participant-days), and the best model was analyzed by meal type, portion size, sizing-marker use, ingredient visibility and meal complexity.
Results: GPT-4o mini achieved the lowest errors, with a significant advantage for carbohydrates; the three finalists performed similarly on other outcomes and on HEI-2020. Mean absolute percentage errors (MAPE) remained high (52%-90%), and the direction of bias was model-specific. The carbohydrate mean absolute error was 13.3 g (95% CI 12.4-14.3), and only 57% of meals were within ±10 g, a clinically relevant threshold for insulin dosing in type 1 diabetes. Calorie errors for lunch and dinner were in the range reported for human estimation on other images (∼55% MAPE), whereas drink-only images had the highest relative errors (312% MAPE) because of dissolved sugars. Food items were reliably identified when fully visible (Dice = 0.74), and HEI-2020 estimates showed moderate agreement with reference values (RMSE = 12.5 points; r = 0.61). Larger portions were increasingly underestimated, and sizing references did not improve accuracy. Newer or larger models and chain-of-thought prompting did not consistently improve performance.
Conclusions: Current mLLMs can extract meaningful dietary information from real-world meal photographs, but substantial errors in nutrient quantification and portion-size estimation limit their use for precision diabetes management. Integrating visual assessment with additional contextual information may enable more reliable, lower-burden dietary monitoring.

DOI: https://doi.org/10.64898/2026.10.21.26365949
Artif Organs. 2026 Nov 13.

Design and Simulation of an AI-Powered Artificial Pancreas System With Personalized Glucose Sensing and Insulin Delivery.

G Kalyani, T Pratap, Sateesh Jasti.

   BACKGROUND: Achieving tight glycemic control in Type 1 diabetes is challenged by non-linear metabolic dynamics and insulin-stacking risk. Artificial intelligence models offer improved personalization, but their opacity raises safety concerns, particularly for basal insulin dosing, which is often treated as static in existing decision-support systems.
METHODS: We propose a hybrid closed-loop system combining a data-driven LightGBM regressor for basal dose prediction, trained on the T1D-UOM dataset, with a physiologically grounded, rule-based safety layer at the edge. Basal insulin is predicted from cyclical time encoding and glucose rate-of-change; bolus dosing follows carbohydrate intake and correction requirements. A time-bounded insulin-on-board model, an uncertainty-adaptive post-bolus damping factor, and a predictive low-glucose-suspend rule reproducing the Tandem Basal-IQ protocol together suspend or reduce delivery when residual insulin is high or hypoglycemia is projected. The system was evaluated in silico over 60 continuous days under shadow-mode replay and closed-loop reduced-order simulation, including an ablation with bolus delivery suppressed.
RESULTS: Under chronological, within-subject validation, the offline basal estimator achieved R2=0.68 and MAE=0.21 U; under strict leave-one-subject-out validation on the same real cohort, performance did not generalize to unseen subjects ( R2=-0.49 , MAE=0.49 U). The hybrid controller maintained Time-in-Range at 62.8% within the reduced-order closed-loop simulation, with no violations of the insulin-on-board suspension or damping rules; suppressing bolus delivery in an ablation within this simulation dropped Time-in-Range to 26.1%, indicating that, within the reduced-order model, bolus delivery is a determinant of simulated glycemic outcome.
CONCLUSIONS: The proposed architecture demonstrates a safety-audited template for combining machine-learning basal prediction with clinically grounded dosing rules. Validation on real outcome data and physical hardware is identified as the next steps toward clinical translation.

Keywords:  LightGBM; Type 1 diabetes; artificial intelligence; artificial pancreas; closed‐loop control system; insulin dose prediction

DOI:  https://doi.org/10.1111/aor.70278
Healthcare (Basel). 2026 Nov 06. pii: 3790. [Epub ahead of print]14(21):

Simulation-Based Development, Stress Testing, and External Physiological Benchmarking of a DKA Early-Warning Framework Intended for Future Nurse-Facing Evaluation.

Muneera Alharbi.

   BACKGROUND/OBJECTIVES: We evaluated a simulation-based diabetic ketoacidosis (DKA) early-warning framework and examined whether machine learning advantages reflected simulator assumptions.
METHODS: Twenty thousand synthetic 15 h episodes were generated at 5 min intervals. Validation-frozen beta-hydroxybutyrate (BHB) rules and XGBoost models were compared. Post hoc analyses assessed phenotype-specific glucose, counterfactual vital signs, comparator thresholds, and calibration. Public IOBP2 data were reconstructed for physiological benchmarking.
RESULTS: Among 449 reference-DKA test episodes, the original BHB+trend rule achieved 98.9% sensitivity and 1.060 false alerts/day; multimodal XGBoost achieved 89.5% and 0.456/day. At BHB ≥3.0 mmol/L, median glucose was 152.9 mg/dL in classical-DKA versus 304.5 mg/dL in non-DKA hyperglycemia, demonstrating a generator shortcut. With phenotype-identical vital signs, dual/multimodal AUPRC were 0.6044/0.6036; removing glucose and vital-sign phenotype cues yielded 0.5142/0.5164. An extended validation grid changed the BHB+trend cutoff and achieved 91.5% sensitivity at 0.761 false alerts/day. The adult IOBP2 benchmark contained 854 paired observations from 165 adults. Observed glucose medians exceeded synthetic medians by 108.7-243.4 mg/dL across five BHB strata.
CONCLUSIONS: Conditional XGBoost discrimination depends on simulator construction and does not establish physiological superiority. IOBP2 exposes substantial glucose-distribution mismatch and does not validate DKA prediction or kinetics. Prospective clinical and nurse-facing assessment remains necessary.

Keywords:  alert fatigue; beta-hydroxybutyrate; clinical decision support; continuous ketone monitoring; diabetic ketoacidosis; early warning system; machine learning; nursing; real-world data; synthetic data

DOI:  https://doi.org/10.3390/healthcare14213790
Front Med (Lausanne). 2026 ;13 2007153

Safety, accuracy, empathy and information quality of six large language models answering patient-facing questions about diabetic kidney disease: a cross-sectional comparative study.

Shuai Yu, Ruirui Huang, Wenli Zhou, Ning Duan, Wenjie Huang.

   Background: Large language models (LLMs) may support patient education, but their suitability for diabetic kidney disease (DKD) remains uncertain. We compared six publicly accessible LLM chatbots-ChatGPT, Gemini, DeepSeek, Doubao, Kimi, and Qwen-across safety, accuracy, empathy, information quality, and readability, with safety as the primary outcome.
Methods: In this CHART-guided cross-sectional study, 42 standardized English DKD questions were submitted once to each chatbot in independent single-turn sessions, generating 252 responses. Five source-masked clinical raters assessed safety, accuracy, empathy, DISCERN, EQIP, JAMA benchmarks, and the Global Quality Scale; readability was assessed automatically. Agreement was evaluated using Fleiss'κ and ICC (2,1)/ICC (2,5). Model differences were tested using Cochran's Q or Friedman tests with Kendall's W and Benjamini-Hochberg correction. Potentially unsafe responses underwent qualitative review.
Results: Safety showed substantial agreement (κ = 0.854); ICC (2,1) ranged from 0.552 to 0.855 and ICC (2,5) from 0.860 to 0.967. Safe-response rates ranged from 83.3% to 92.9% without significant between-model differences (P = 0.726). Accuracy and empathy differed significantly (both P < 0.001), with ChatGPT highest for accuracy and Gemini highest for clinician-rated empathy. Information quality also differed significantly, with ChatGPT performing best overall and Doubao lowest. All six readability indices differed across models (all P < 0.001; Kendall's W = 0.139-0.255). DeepSeek had the lowest mean scores on all five grade-level indices and the highest mean FRES, but mean grade-level estimates exceeded the sixth-grade reference level for every model. Thirty of 252 responses (11.9%) were potentially unsafe, most often because of incomplete medication-related information or insufficiently individualized fluid advice.
Conclusions: Publicly accessible LLMs can provide plausible DKD information but show variable performance and safety risks. They may support, but should not replace, professional patient education; guideline anchoring, source verification, improved readability, and clinical oversight remain essential.

Keywords:  artificial intelligence; diabetic kidney disease; large language models; patient education; safety

DOI:  https://doi.org/10.3389/fmed.2026.2007153
Int J Med Inform. 2026 Nov 09. pii: S1386-5056(26)00548-4. [Epub ahead of print]224 106808

A personalized glucose-aware meal guidance agent for type 2 diabetes self-management.

Seongjae Lee, Kawon Jeong, Wonjun Hwang, Gwanwoo Kim, Hyokyoung Jeon, Nan Hee Kim, Sun-Joon Moon, Da Young Lee, Sung-Min Park.

   BACKGROUND: Type 2 diabetes self-management requires individuals to coordinate meals, glucose monitoring, medications, insulin use, and daily routines outside the clinic. Although personalized postprandial glucose response (PPGR) prediction has advanced, numerical predictions alone do not provide the explanations and practical meal-level guidance needed for everyday decision-making. We developed a Personalized Glucose-aware Meal guidance Agent (PiGMA), to translate individualized PPGR predictions into prediction-linked explanations and meal-level guidance.
METHODS: This retrospective proof-of-concept study used 3857 meal records from 140 individuals with type 2 diabetes across two clinical centers, comprising a development cohort of 130 participants with 3580 meals and a held-out cohort of 10 participants with 277 meals. PiGMA combined a personalized PPGR predictor with an instruction-tuned LLaMA3.1-based guidance module trained for meal scoring, dietary advising, and PPGR interpretation. Evaluation included task-specific quantitative benchmarks, five-run G-Eval assessment, and blinded review by two endocrinologists.
RESULTS: Participant-specific fine-tuning reduced mean absolute error in seven of ten held-out participants, with significant improvement using five or more days of personalization data. PiGMA outperformed the base model in meal scoring, achieved higher scores for PPGR interpretation, and outperformed medical-domain language models in dietary advising. In blinded expert evaluation, PiGMA received higher PPGR-interpretation ratings and was preferred in 33 of 40 endocrinologist rater-case judgments. However, character-span-level unsupported content was higher for PiGMA than for the base model in PPGR interpretation (12.29 % vs. 3.60 %).
CONCLUSIONS: PiGMA provides retrospective proof-of-concept evidence that individualized PPGR predictions can be translated into prediction-linked meal guidance. However, the higher hallucination or unsupported-content rate and absence of prospective patient evaluation indicate a safety-reliability trade-off and preclude conclusions regarding clinical effectiveness, safety, or deployment readiness.

Keywords:  Diabetes self-management; Digital health; Large language model; Personalized nutrition; Postprandial glucose response; Type 2 diabetes

DOI:  https://doi.org/10.1016/j.ijmedinf.2026.106808
Front Endocrinol (Lausanne). 2026 ;17 2002540

Targeted cystatin C testing for large negative eGFR discordance in adults with diabetes: development and external validation of a machine-learning model.

Xuanxuan Wang, Mengyu Zhang, Ke Zhang, Yuanhong Xu, Yingying Chen, Yuanyuan Hu.

   Background: Large negative discordance between cystatin C- and creatinine-based estimated glomerular filtration rate (eGFR) may reveal clinically relevant renal-function differences that are not apparent from creatinine alone. When universal cystatin C testing is impractical, a method to prioritize patients for additional testing may improve testing efficiency.
Methods: We conducted a retrospective two-center study of adults with diabetes. Large negative eGFR discordance (D2) was defined as eGFRcys/eGFRcr <0.70. Seven machine-learning algorithms were evaluated in a development cohort of 1,170 participants (214 D2-positive), followed by SHAP-guided feature reduction. The final model was applied unchanged to an independent external validation cohort of 1,619 participants (112 D2-positive). Performance was compared with a prespecified clinical core (age, sex, and eGFRcr); additional analyses evaluated logistic regression using eGFRcr alone or with blood urea nitrogen and albumin. Prespecified cystatin C testing allocations of 10%-50% were evaluated using model-based risk ranking.
Results: The final nine-predictor XGBoost model achieved an AUROC of 0.863 (95% CI, 0.833-0.889) and an AUPRC of 0.600 (95% CI, 0.532-0.668) in development. In independent external validation, AUROC was 0.806 (95% CI, 0.769-0.840) and AUPRC was 0.258 (95% CI, 0.193-0.337), compared with 0.661 and 0.126, respectively, for the clinical core. At the 20% testing allocation, risk-guided testing identified 61.2% of D2 cases in development and 55.4% externally, with 3.06-fold and 2.77-fold enrichment over random testing. At this allocation in external validation, XGBoost identified 62 D2 cases versus 43 with the three-variable model (recall difference, 17.0 percentage points; 95% CI, 8.6-27.2). However, raw external probabilities overestimated D2 risk (mean predicted probability, 29.9% vs. observed prevalence, 6.9%), and the Brier score was higher than that of the clinical core (0.144 vs. 0.130).
Conclusions: In two cohorts of adults with diabetes who underwent paired testing, routine clinical laboratory data supported prioritization for large negative eGFR discordance. Model-based ranking may help target cystatin C testing in similar settings, but raw probabilities require local calibration before threshold-based decisions and prospective evaluation is needed before clinical implementation.

Keywords:  cystatin C; diabetes; eGFR discordance; estimated glomerular filtration rate; external validation; machine learning; targeted testing

DOI:  https://doi.org/10.3389/fendo.2026.2002540
Endokrynol Pol. 2026 Nov 10.

Identification of key indicators of colorectal cancer risk in patients with diabetes mellitus using machine learning and Mendelian randomization.

Haiyan Shangguan, Jie Li, Ruoyu Sun, Zuoling Xie, Chenchen Wang, Zheng Wang, Xi Huang, Shaohua Wang.

   INTRODUCTION: Previous studies have indicated that diabetes mellitus may increase the risk of colorectal cancer (CRC), yet the precise associationremains unclear. To identify critical clinical predictors of CRC risk in patients with type 2 diabetes mellitus (T2DM) using machinelearning (ML) and to subsequently evaluate their causal relationship with CRC via Mendelian randomization (MR) analysis.
MATERIAL AND METHODS: A retrospective cohort design was adopted, analyzing clinical records from patients with T2DM and CRC and from those with T2DM alone. Significant predictors were determined using univariate and multivariate Cox regression analyses. Four MLmodels were developed, with essential indicators defined by the overlap of top-ranked clinical variables and risk factors. Mendelian randomizationanalysis was then applied to assess causality between the key indicators (exposures) and CRC (outcome), supplemented bysensitivity tests and Steiger analysis. Reverse MR analysis was conducted to examine potential reverse causality between CRC (exposure)and the identified indicators (outcomes).
RESULTS: Univariate and multivariate Cox regression revealed that age [p < 0.05, hazard ratio (HR) = 1.041, 95% confidence interval(CI) = 1.022-1.062] and hemoglobin A1c (HbA1c, p < 0.05, HR = 1.176, 95% CI = 1.094-1.264) were significant risk factors. Logistic regressionwas ultimately identified as the most suitable model, from which two key indicators (age and HbA1c) were derived. Mendelianrandomization analysis demonstrated a causal effect of HbA1c on CRC, confirming HbA1c as a risk factor [odds ratio (OR) = 1.14, 95%CI = 1.01-1.28, p = 0.0336]. Sensitivity analyses and Steiger testing supported the robustness and directionality of these findings. Notably,reverse MR analysis indicated no significant causal effect of CRC on HbA1c [p inverse variance weighted = 0.07853).
CONCLUSIONS: Age and HbA1c were identified as pivotal indicators, with HbA1c exhibiting a unidirectional causal association with CRC, thereby offering insights for potential therapeutic strategies.

Keywords:  Mendelian randomization; age; colorectal cancer; hemoglobin A1c; machine learning; type 2 diabetes mellitus

DOI:  https://doi.org/10.5603/ep.113150
Front Physiol. 2026 ;17 1958735

Structured multifrequency segmental bioimpedance descriptors for diabetes screening support.

Minho Choi, Kun Ho Lee, Jaeuk U Kim.

   Background: Blood-based measurements remain the standard for identifying diabetes mellitus (DM), but repeatable non-invasive approaches may have value for screening support. Bioelectrical impedance analysis (BIA) captures body-composition, hydration, and bioelectrical properties associated with metabolic health. However, previous studies have largely focused on analyzer-derived body-composition variables or individual parameters such as phase angle, whereas multifrequency segmental impedance responses have not been systematically characterized for screening-support applications. This study aimed to computationally characterize multifrequency segmental BIA responses using structured descriptors and to evaluate their potential for identifying previously diagnosed DM status within a proof-of-concept screening-support framework.
Methods: An ExtraTrees-based classification framework was developed to discriminate previously diagnosed DM status. Feature selection, input dimensionality determination, and model optimization were performed within a nested cross-validation framework, and performance was evaluated using pooled out-of-fold predictions. Alongside demographic and device-derived body-composition variables, we formulated 440 base BIA descriptors capturing amplitude and dispersion, phase angle, Cole and impedance geometry, curve shape and coupling, and regional-pattern information. We additionally developed frequency-interval and reliability-weighted interval descriptors to characterize interval-specific variations in multifrequency impedance-magnitude responses and their local agreement with the fitted frequency-response trend. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC) and average precision (AP), together with calibration and threshold-dependent metrics.
Results: Evaluated in 2,745 participants aged 47-90 years from a predominantly older-adult cohort, the ExtraTrees model achieved an AUROC of 0.8336 (95% CI, 0.8138-0.8523) and an AP of 0.6754 (95% CI, 0.6395-0.7092). The full descriptor configuration showed the highest overall discrimination, although the incremental gains associated with the newly developed frequency-interval and reliability-weighted descriptors were modest. These two descriptor groups together accounted for approximately 35.1% of the pooled out-of-fold SHAP contribution, consistent with a complementary rather than dominant predictive role.
Conclusion: This study demonstrates the feasibility of using structured multifrequency segmental BIA-derived descriptors to discriminate previously diagnosed DM status. The newly developed descriptors provided complementary predictive information, although their incremental improvement in discrimination was limited. These findings provide a methodological basis for future evaluation of BIA-assisted screening using contemporaneous blood-test-defined outcomes in appropriately recruited populations.

Keywords:  bioelectrical impedance analysis; diabetes mellitus; diabetes screening; frequency-interval descriptor; machine learning; multifrequency bioimpedance; reliability-weighted descriptor

DOI:  https://doi.org/10.3389/fphys.2026.1958735
JMIR Med Inform. 2026 Nov 13. 14 e100792

Large Language Model-Assisted Discharge Instructions for Diabetic Kidney Disease: Comparative Evaluation Study.

Miaomiao Cheng, Huiru Wang, Qi Zhang, Jun Qin, Lei Gong, Ling Guo, Xiaoyan Xiao.

   Background: Discharge instructions are an important component of the transition from hospital to home for patients with diabetic kidney disease (DKD), whose care often involves multiple comorbidities, medications, and stage-specific management needs. However, producing complete, accurate, clear, and personalized instructions can be challenging for junior physicians.
Objective: This study aimed to evaluate whether large language model (LLM) assistance improves the quality of DKD discharge instructions produced by junior physicians and whether its benefits vary according to chronic kidney disease (CKD) stage.
Methods: The study included 101 hospitalized patients with DKD. Discharge instructions for each patient were initially drafted by a group of junior residents (resident group); subsequently, another group of residents at the same level, with the assistance of DeepSeek-V3.2 (AI-assisted group), generated a separate set of discharge instructions for the same 101 patients, resulting in a total of 202 documents. Four clinical experts, including 2 nephrologists and 2 endocrinologists, independently scored both versions of the discharge instructions using a 5-point Likert scale across 5 dimensions: completeness, accuracy, safety, clarity, and personalization. Interrater reliability was assessed using the intraclass correlation coefficient (ICC).
Results: Interrater reliability was excellent in both the resident group (ICC=0.817, 95% CI 0.763-0.864; P<.001) and the AI-assisted group (ICC=0.860, 95% CI 0.816-0.897; P<.001). Compared with resident-only instructions, AI-assisted instructions received significantly higher scores across 5 evaluation dimensions and the overall evaluation (all Bonferroni-adjusted P<.001), with large effect sizes (r=0.741-0.927). In the resident-only group, CKD stage was negatively correlated with accuracy, safety, clarity, and overall quality (Spearman ρ=-0.33 to -0.37; all P<.001), whereas no significant correlations were observed for most dimensions in the AI-assisted group (|ρ|<0.15; P>.05). The proportion of cases achieving a ≥1-point improvement in overall quality was 99% (100/101; 95% CI 94.6%-100.0%).
Conclusions: LLM assistance improved the completeness, clarity, and personalization of DKD discharge instructions generated by junior physicians, with advantages greatest in complex cases. A human-AI collaborative model is a promising approach to reduce documentation burden.

Keywords:  artificial intelligence; diabetic kidney disease; discharge instructions; large language model; patient education

DOI:  https://doi.org/10.2196/100792