bims-aukdir 2026-06-21 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2026–06–21
25 papers selected by
Mott Given

Real-World Performance of Artificial Intelligence in Diabetic Retinopathy Screening: A Systematic Review.
Comparative predictive accuracy of machine learning versus traditional statistical methods for diabetes-related complications: A systematic review and analysis.
Advanced CNN Deep Learning Model for Diabetic Retinopathy Classification.
An ensemble of deep learning models with falcon optimization assisted diabetic retinopathy diagnosis on retinal fundus images.
Explainable AI for diabetic retinopathy detection using vision transformers.
Diagnostic performance of an artificial intelligence software for diabetic retinopathy organised screening in a pilot study.
From discarding to leveraging: quality-aware collaborative learning for robust diabetic retinopathy grading.
Machine learning predicts diabetes risk in high-risk populations: analysis of National Health and Nutrition Examination Survey data.
An AI-driven framework for diabetic foot ulcer classification, segmentation, and depth estimation.
AI-driven cardiovascular risk prediction in patients with diabetes: bridging algorithmic innovation to equitable clinical application.
Ultra-widefield color fundus photography in diabetic retinopathy: from panretinal assessment to multimodal integration.
GAT-BiGRU: explainable multi-task temporal graph learning for glucose forecasting, hypoglycemia risk, and counterfactual insulin adjustment.
DFU-GCNet: a global context-enhanced inception network for robust and interpretable diabetic foot ulcer classification.
Digital Twin Applications in Diabetes Management: Scoping Review.
Application of Quality by Design (QbD) as Artificial Intelligence in Formulation for Diabetes: A Practical Review.
Development and validation of an explainable neural network model for predicting progression in type 2 diabetic kidney disease.
Machine Learning Based County Level Phenotypes Related to Diabetes Prevalence.
Predictive voting model for early diagnosis of diabetes dataset.
Machine learning-based predictive model for sleep disorders in diabetic patients: data analysis from CHARLS.
Enhancing diabetes risk stratification through natural language processing: a multimodal data integration approach.
Integrated Bioinformatics and Explainable Machine Learning Analyses Reveal Mesenchymal Stem Cell Exosome-Related Biomarkers for Diabetic Retinopathy.
An AI-Based OCT System to Detect Diabetic Macular Edema: A Prospective Validation and Noninferiority Randomized Clinical Trial.
Non-invasive opportunistic screening for diabetes mellitus: an interpretable stacking ensemble framework.
Predicting and explaining poor prognosis in diabetic kidney disease using SHAP-based interpretable machine learning.
The STEADY diabetic foot ulcer registry: methods, insights, and future directions.

Cureus. 2026 May;18(5): e108894

Real-World Performance of Artificial Intelligence in Diabetic Retinopathy Screening: A Systematic Review.

Saira Ahmed.

  Diabetic retinopathy (DR) remains a leading cause of preventable blindness worldwide, and timely screening is essential for early detection and intervention. Artificial intelligence (AI), particularly deep learning, has emerged as a promising tool for automated diabetic retinopathy screening. This systematic review evaluates the diagnostic performance and real-world applicability of AI-based systems across diverse clinical settings. A systematic search of PubMed, Excerpta Medica database (Embase), and the Cochrane Library was conducted, supplemented by screening of Google Scholar, with study selection performed in accordance with PRISMA 2020 guidelines. Studies were included if they assessed AI systems for diabetic retinopathy detection using fundus-based retinal imaging and reported diagnostic accuracy outcomes. A total of 30 studies published between 2016 and 2025 were included. Across studies, AI systems demonstrated consistently high diagnostic performance, with most reporting sensitivities above 85% and specificities above 80%. Large-scale and real-world studies confirmed the feasibility of implementing AI in national and community screening programmes. Additionally, smartphone-based and handheld imaging systems demonstrated promising potential for expanding screening access in resource-limited settings. Despite these encouraging findings, variability between AI systems and study designs highlights the need for external validation and standardisation prior to widespread clinical adoption. AI has significant potential to enhance screening efficiency and accessibility, but further research is required to evaluate long-term clinical outcomes and integration into healthcare systems.

Keywords:  artificial intelligence; deep learning; diabetic retinopathy; ophthalmology; retinal screening

DOI:  https://doi.org/10.7759/cureus.108894
Diabet Med. 2026 Jun 18. e70398

Comparative predictive accuracy of machine learning versus traditional statistical methods for diabetes-related complications: A systematic review and analysis.

Bismah Ghafoor, Francesco Zaccardi, Kamlesh Khunti, Sharmin Shabnam.

   AIMS: To compare the predictive accuracy of machine learning models versus traditional statistical models for predicting and detecting long-term complications among individuals with diabetes (PROSPERO: CRD420250629747).
METHODS: We systematically searched MEDLINE, PubMed, Cochrane and Scopus (2014-2025) for studies developing or validating prediction models in people with diabetes. Excluding case-control studies, we identified 36 eligible studies (280 model comparisons) from 18,237 records. We extracted study design, model details and performance metrics (primarily C-statistics). Risk of bias was assessed using PROBAST.
RESULTS: Across 280 comparisons, ensemble machine learning methods frequently outperformed logistic regression. Random forest models achieved higher discrimination in 63% (43/68) of comparisons, while extreme gradient boosting showed improvement in 58% (14/24). Support vector machines improved performance in only 44% (24/55). Generally, predictive accuracy gains were modest. Methodological quality was concerning as external validation was reported in only 8% (3/36) of studies, calibration in 13% (5/36), and 59% of studies demonstrated a high risk of bias.
CONCLUSIONS: Machine learning models, particularly ensemble methods, offer modest discrimination improvements over traditional statistics for predicting diabetes-related complications. However, widespread methodological limitations, specifically the lack of external validation, inconsistent calibration reporting and high bias, substantially limit our confidence and clinical readiness. Rigorous external validation and transparent reporting are needed before routine implementation.

Keywords:  ROC curve; diabetes mellitus; diabetic complications; machine learning; models, statistical; predictive value of tests; risk assessment

DOI:  https://doi.org/10.1111/dme.70398
J Biomed Phys Eng. 2026 Jun;16(3): 191-204

Advanced CNN Deep Learning Model for Diabetic Retinopathy Classification.

Noor Ali Sadek, Ziad Tarik Al-Dahan, Suzan Amana Rattan, Ahmed F Hussein, Brendan Geraghty, Ahmed Kazaili.

   Background: Diabetic Retinopathy (DR) is one of several retinal microvascular complications of Diabetes Mellitus (DM), a disease of increasing global prevalence. However, early detection and treatment can reduce or even prevent DR progression. In this work, Deep Learning (DL) techniques are used to grade DR from an early stage using either binary or multiclass classification as a clinical aid to help reduce the risk of patient vision loss.
Objective: The primary objective of this research is to develop a low-cost, fast, and accurate automated system using DL for the early detection and classification of DR from retina fundus images.
Material and Methods: This cross-sectional study employed three DL models, namely Convolutional Neural Networks (CNNs), decision tree, and logistic regression, to categorize three distinct clinically graded datasets, namely the Iraqi dataset, the Indian Diabetic Retinopathy Image Dataset (IDRiD) and the Eyepacs dataset, according to DR severity.
Results: Evaluation of the DL model results showed that logistic regression emerged as the most effective, where accuracies of 99%, 99.3%, and 99.4% were achieved for the Iraqi, IDRiD, and Eyepacs datasets, respectively. Conversely, the decision-tree model achieved the lowest accuracy across the three datasets with 95.2%, 95.9%, and 96.0%, respectively.
Conclusion: The logistic regression model demonstrated the highest overall accuracy of the three models for the classification of DR, with the Iraqi dataset with the highest accuracy of the three datasets.

Keywords:   CNN; Decision Trees; Deep Learning; Diabetic Retinopathy; EyePACS; IDRiD; Iraqi Dataset; Logistic Regression

DOI:  https://doi.org/10.31661/jbpe.v0i0.2406-1774
Sci Rep. 2026 Jun 14.

An ensemble of deep learning models with falcon optimization assisted diabetic retinopathy diagnosis on retinal fundus images.

Indresh Kumar Gupta, Shruti Patil, Ketan Kotecha, Swati Srivastava, Ahmed Farouk, Joel J P C Rodrigues.

  Diabetic Retinopathy (DR) is a prominent results of diabetes mellitus that causes abnormalities lesions in retina. If not identified at early, it may progress to complete loss of vision. Unfortunately, DR is an irreversible, and treatment only sustains existing vision. Timely detection and accurate treatment of DR can considerably decrease the chance of blindness. Manual diagnosis of DR in retinal fundus images (RFIs) by ophthalmologist is time consuming, costly and laborious tasks with a higher risk of misdiagnosis. Recently, Deep learning (DL) has gained popularity and shown remarkable performance particularly in medical image analysis and classification. Convolutional neural networks (CNNs) are increasingly being used as a DL approach in medical image analysis, and they are very efficient. This manuscript offers the design of Falcon Optimizer with Ensemble of Deep Learning Algorithm Assisted Diabetic Retinopathy Diagnosis Model (FOEDLA-DRDM) system on RFIs. The FOEDLA-DRDM system employs a Wiener filtering (WF) based preprocessing approach to eliminate noise from images. Following this, FOEDLA-DRDM system leverages the SE-DenseNet method to generate the feature vectors. For DR recognition FOEDLA-DRDM system applies an ensemble approach that combines - AutoEncoder, long short-term memory (LSTM), and deep belief network (DBN). Finally, Falcon Optimizer (FO) adjusts the hyperparameter values of the ensemble approach, giving rise to classification efficiency. The FOEDLA-DRDM system is validated by simulating it on a Kaggle DR dataset, with results being measured according to various criteria. The simulation findings showcase the effectiveness of the FOEDLA-DRDM system in diagnosis of DR.

Keywords:  Autoencoder; Deep belief network; Diabetic retinopathy; Falcon optimizer; Retinal fundus image

DOI:  https://doi.org/10.1038/s41598-025-11075-x
Sci Rep. 2026 Jun 15.

Explainable AI for diabetic retinopathy detection using vision transformers.

Mustafizur Rahaman, Masrufa Akter Muni, Saima Tasnim, Sanjida Shahid Juthi, Rakibul Islam, Khandakar Rabbi Ahmed.

  The heterogeneous acquisition, variability of orientation, and subtle lesions continue to challenge the screening of diabetic retinopathy through color fundus photographs. We formulate DR grading as a binary triage task (No-DR vs DR) and propose the All-ViT Hybrid framework, integrating complementary pretrained transformer backbones within a stability-oriented training schedule (head-only warm-up, partial unfreezing), optimized using AdamW with OneCycle scheduling, class-weighted cross-entropy, and mixed precision. Preprocessing includes luminance-space CLAHE and retinal field-of-view masking, and data splitting is performed using stratified sampling to preserve class balance. We perform post-hoc threshold tuning and test-time augmentation (optional), which is available for operating-point control and robustness. Over the competitive baselines (ConvNeXt, MobileNetV3-Large, EfficientNet-B0, DenseNet201), All-ViT Hybrid has an Accuracy of 0.9754, F1 of 0.9761, Precision of 0.9658, Recall of 0.9866, and measures of agreement κ of 0.9509, MCC of 0.9511, and Jaccard of 0.9532. Compared to the best baseline, F1 achieves a gain of +1.11 percentage points, an improvement in accuracy of +1.13, and a recall improvement of +0.81 percentage points without compromising controllable precision and these experiments were conducted on the APTOS 2019 Blindness Detection dataset. These visualizations provide qualitative interpretability and are not quantitatively validated against lesion-level annotations due to dataset limitations. These findings suggest that integrating complementary transformer representations within a unified fusion pipeline provides strong and threshold-adjustable performance for DR triage under real-world variability. The framework remains modular and extensible, with potential applicability to larger input resolutions, multi-class grading, and multi-modal clinical settings. The future research program will test external generalization and probability calibration across devices and centers.

Keywords:  BEiT); Deep learning; DeiT; Diabetic retinopathy screening; Feature-level fusion; Fundus photography; Threshold tuning; Vision Transformers (ViT

DOI:  https://doi.org/10.1038/s41598-026-56569-4
Sci Rep. 2026 Jun 16.

Diagnostic performance of an artificial intelligence software for diabetic retinopathy organised screening in a pilot study.

Emma Altobelli, Marco Baroni, Marco Ciancaglini, Paolo Carpineto, Francesco Masedu.

  Diabetic retinopathy (DR) is one of the most common complications of diabetes mellitus (DM) with a prevalence that varies from 10% to 61% in different countries. From healthcare to the precise prevention, diagnosis, and management of diseases, Artificial Intelligence (AI) is progressing rapidly in interdisciplinary fields, including ophthalmology. We aimed to explore the efficacy of artificial intelligence (AI)-based screening for diabetic retinopathy in type 2 DM patients. High-resolution colour fundus photographs were obtained for all patients using a non-mydriatic camera (Nidek AFC-330). DAIRET software uses a machine learning algorithm to identify the most important signs of DR. The software provides an output with a binary result: negative, in case of no referable disease or positive, when the presence of signs of DR is detected. The photographs were also manually graded negative or positive for DR by an expert ophthalmologist, masked to AI responses. Expert clinical opinion was used as the reference (gold standard), and the results were compared with those of the AI responses. A significant association was observed between DAIRET and expert diagnosis (χ2 = 18.6, p < 0.001). Disease prevalence, according to expert assessment, was 20.2% (33/163). Observed agreement between DAIRET and Expert diagnosis was 69.3%. Agreement metric was: Gwet's AC1 = 0.47 (95% CI: 0.33-0.61), indicating moderate agreement. Diagnostic performance of the DAIRET test indicates good sensitivity (73%), a specificity (68%) and high negative predictive value (NPV=90.8%), suggesting that DAIRET is more effective at excluding diabetic retinopathy than at confirming it. ROC analysis stratified by gender showed no difference in performance (p>0.05). Overall, these results suggest that DAIRET may be effectively integrated into screening workflows to rule out diabetic retinopathy, reduce specialist workload, and prioritize referrals. In particular, for daily practical use, DAIRET is managed within the MètaClinic electronic medical record. In-house data processing without going to an external cloud is a significant strength. Larger, prospective studies are needed to validate these findings, optimise thresholds, and assess performance across diverse populations and clinical settings.

Keywords:  Artificial intelligence; Diabetes; Retinopathy; Screening

DOI:  https://doi.org/10.1038/s41598-026-58158-x
Front Med (Lausanne). 2026 ;13 1795484

From discarding to leveraging: quality-aware collaborative learning for robust diabetic retinopathy grading.

Yuan Pan, Xiangwen Cai, Pan Xiong, Fang Mei, Zhengfei Wang, Jiaming Hong.

  Real-world diabetic retinopathy (DR) screening faces a paradox: the most diagnostically critical images are often the lowest in quality, because advanced disease itself produces vitreous hemorrhage, proliferative tissue, and media opacities that degrade fundus imagery. We characterize this quality-severity coupling quantitatively (Spearman ρ = 0.420, odds ratio 4.17 for referable DR in Reject vs. Good strata on DDR, p < 0.001) and show that conventional pipelines work against it: filtering low-quality images discards the most severe cases, while uniform processing leads to misclassification. Both behaviors stem from treating image quality assessment (IQA) as a binary preprocessing decision. We argue that quality should serve as a continuous guidance signal that conditions the diagnostic process, and propose QGDR, a quality-guided dynamic routing framework realizing this paradigm through three coordinated mechanisms: (i) a multi-level IQA module that extracts hierarchical quality features across backbone stages; (ii) a quality-conditioned context gating mechanism that modulates spatial attention according to the predicted quality state; and (iii) an adaptive gated fusion mechanism that routes inputs to scale-specialized experts, with high-quality images preferentially activating fine-scale experts for subtle lesions and degraded images relying on coarse-scale experts for robust global pattern recognition. On EyeQ and DDR, QGDR attains 78.32% accuracy with 0.6863 QWK and 80.85% accuracy with 0.8231 QWK respectively, outperforming representative CNN, transformer, and foundation-model baselines while remaining within the compute envelope of standard single-stream backbones. Counter-intuitively, performance is preserved or improved on the lowest-quality stratum (77.61% on EyeQ-Reject; 82.21% on DDR-Reject, exceeding the Good-quality accuracy on the same dataset), and a Shuffled-IQA counterfactual confirms that QGDR exploits the semantic content of quality information rather than a generic auxiliary signal. Test-only evaluation on the external IDRiD and DeepDRiD cohorts confirms cross-dataset generalization. By treating image quality as guidance rather than as a filter, QGDR preserves screening coverage without sacrificing diagnostic reliability.

Keywords:  collaborative learning; diabetic retinopathy grading; dynamic expert routing; fundus image quality assessment; multi-scale feature learning

DOI:  https://doi.org/10.3389/fmed.2026.1795484
Arch Med Sci. 2026 ;22(2): 708-720

Machine learning predicts diabetes risk in high-risk populations: analysis of National Health and Nutrition Examination Survey data.

Xiaohua Yang, Meiqi Yao, Jia Huang, Zhuojing Cheng, Ting Sun.

   Introduction: This project intended to develop and validate a diabetes prediction model for high-risk populations based on machine learning algorithms.
Material and methods: A total of 2,355 samples from the National Health and Nutrition Examination Survey (NHANES) database covering three cycles from 2013 to 2018 were included. The data were divided into training and testing sets in a 7 : 3 ratio. Nineteen risk prediction factors were selected as feature variables, including demographic baseline data, measurement data, medical history, and psychological health. Five machine learning models - decision tree, random forest (RF), multilayer perceptron (MLP), Adaboost, and Extreme Gradient Boosting (XGBoost) - were developed based on the data and variables mentioned above. Model performance was evaluated using accuracy, sensitivity, specificity, the area under curve (AUC) values of receiver operating characteristic (ROC) curves, and Matthews Correlation Coefficient (MCC) scores. Finally, the Shapley feature importance measurement tool was employed to select features in the optimal model.
Results: The present work ultimately included 2,355 individuals at high risk of diabetes for analysis, with 260 cases of diabetes and 2,095 cases without diabetes. Among the five machine learning models established in this project., the RF and XGBoost models exhibited better overall performance compared to other models. In the test set, the RF model had an AUC of 0.896, accuracy of 0.784, sensitivity of 0.739, specificity of 0.849, and MCC of 0.418. The XGBoost model had corresponding values of AUC as 0.903, accuracy of 0.815, sensitivity of 0.962, and MCC of 0.443. According to the importance analysis of features in these two optimal models, waist circumference, age, BMI, gender, systolic blood pressure (SBP), diastolic blood pressure (DBP), education level, poverty income ratio (PIR), Patient Health Questionnaire (PHQ)-9 score, and race were the top ten key risk factors for diabetes in the high-risk population.
Conclusions: The RF and XGBoost machine learning models demonstrated strong performance in predicting the occurrence of diabetes in high-risk populations. These models can aid in developing more precise intervention measures and personalized treatment plans to effectively reduce the incidence of diabetes and related risks in this population.

Keywords:  National Health and Nutrition Examination Survey; diabetes; machine learning; prediction model

DOI:  https://doi.org/10.5114/aoms/209547
Sci Rep. 2026 06 17. pii: 18861. [Epub ahead of print]16(1):

An AI-driven framework for diabetic foot ulcer classification, segmentation, and depth estimation.

Saswati Debnath, Daniel Willson, Renukadevi M N, Shreenidhi H S, Jaydeep Kishore, Rohit Singh.

  Diabetic Foot Ulcers are a severe complication of diabetes that can lead to infection, amputation, and increased mortality, making timely and objective assessment essential. Conventional evaluation relies primarily on visual inspection and manual measurements, which are often subjective and inconsistent. This paper presents a comprehensive automated deep learning-based framework for DFU assessment that integrates classification, segmentation, relative depth estimation, and explainability. An EfficientNet-B0 model is employed for binary classification of ulcerated versus healthy skin. At the same time, a U-Net with an EfficientNet-B0 encoder is used for precise delineation of ulcer boundaries, enabling quantitative morphometric analysis such as area and width estimation. Model interpretability is incorporated through Gradient-weighted Class Activation Mapping and Local Interpretable Model-Agnostic Explanations, providing visual insights into the decision-making process. To characterize wound topology, relative pseudo-depth maps are generated from single RGB images using a MiDaS-based monocular depth estimation approach. Finally, an automated reporting module synthesizes outputs from all components into structured, clinician-friendly summaries. Experimental results demonstrate strong classification performance, accurate segmentation, and meaningful characterization of relative depth, highlighting the potential of the proposed framework as a research-oriented AI-assisted decision-support tool for DFU assessment. However, further clinical validation involving expert clinicians and prospective studies is required before real-world clinical deployment.

Keywords:  Computer-aided diagnosis; Deep learning; Diabetic foot ulcer; Explainable AI; Hybrid model; Medical imaging; Multi-class classification; Transformer

DOI:  https://doi.org/10.1038/s41598-026-52864-2
Front Med (Lausanne). 2026 ;13 1831220

AI-driven cardiovascular risk prediction in patients with diabetes: bridging algorithmic innovation to equitable clinical application.

Hongxuan Li, Zheyi Xu, Yanhui Cen, Xin Liu.

  Machine learning models hold promise to revolutionize cardiovascular disease (CVD) prediction in patients with type 2 diabetes, with algorithms such as neural networks demonstrating superior discriminative performance in internal validations. However, a systematic review has revealed that existing models generally carry a high risk of bias and exhibit poor adherence to transparent reporting standards, severely hindering their clinical translation and real-world application. Furthermore, current models are predominantly developed using populations from Europe and North America, resulting in a critical lack of representativeness for Asian populations, where the burden of cardiovascular disease is particularly heavy. This article argues that the field is undergoing a pivotal transition-from an exclusive focus on algorithmic performance to ensuring clinical equity and fairness. Future advancements should prioritize external validation, calibration-aware assessment, subgroup-specific performance reporting, and cautious integration of biologically plausible biomarkers rather than relying on discrimination alone. Only through this approach can machine learning-driven predictive tools truly bridge the gap between innovation and equitable clinical implementation, ultimately alleviating the global burden of diabetes-related cardiovascular complications.

Keywords:  cardiovascular disease; lipid peroxidation; machine learning; population fairness; prediction model; risk of bias; type 2 diabetes

DOI:  https://doi.org/10.3389/fmed.2026.1831220
Front Med (Lausanne). 2026 ;13 1845156

Ultra-widefield color fundus photography in diabetic retinopathy: from panretinal assessment to multimodal integration.

Ruolin Fang, Ruixuan Xu, Zhipeng Wu, Yuhua Tong.

  Diabetic retinopathy (DR) is the leading cause of vision loss among working-age adults. Early screening and precise staging are crucial for delaying disease progression. Although the traditional Early Treatment of Diabetic Retinopathy Study (ETDRS) 7-field method is the gold standard for grading, it covers only 35% of the retinal area, carrying a risk of missing peripheral lesions. Ultra-widefield color fundus photography (UWF-CFP) overcomes this limitation by capturing up to 200° of the retina in a single image, enabling comprehensive visualization of peripheral ischemia, neovascularization, and peripheral-posterior pole asymmetry lesions. In recent years, combining UWF-CFP with deep learning algorithms has achieved robust performance in automated DR screening, grading, and quantitative vascular analysis. When further integrated with ultra-widefield swept-source optical coherence tomography angiography (UWF SS-OCTA) or multimodal clinical data, these AI-driven methods can improve diagnostic consistency and staging accuracy. Moreover, because the retinal microvasculature mirrors systemic microcirculation, vascular parameters from UWF-CFP have shown correlations with diabetic nephropathy, coronary artery calcification, and stroke risk, highlighting a potential role in non-invasive assessment of systemic complication risks in diabetic patients. UWF-CFP is facilitating a shift in the DR assessment paradigm from ETDRS 7-field to full retinal assessment. Combined with artificial intelligence and multimodal data integration, UWF-CFP has the potential to contribute to a future care model that integrates ocular and systemic risk assessment, which we tentatively term "eye-system collaborative management," though this paradigm remains to be validated in prospective studies.

Keywords:  artificial intelligence; color fundus photography; diabetic retinopathy; multimodal integration; panretinal assessment; ultra-widefield imaging

DOI:  https://doi.org/10.3389/fmed.2026.1845156
J Am Med Inform Assoc. 2026 Jun 18. pii: ocag104. [Epub ahead of print]

GAT-BiGRU: explainable multi-task temporal graph learning for glucose forecasting, hypoglycemia risk, and counterfactual insulin adjustment.

Muhammad Abdullah Sarwar, Sarmad Maqsood, Egle Belousovienė, Rytis Maskeliūnas.

   BACKGROUND: People with type 1 diabetes mellitus (T1DM) show glucose variability driven by insulin dosing, meals, activity, and circadian rhythms. Many deep learning approaches treat glucose forecasting and hypoglycemia detection as separate tasks and provide limited transparency.
OBJECTIVE: We developed an explainable, multi-task temporal graph framework that jointly predicts glucose trajectories and hypoglycemia risk at 30 and 60 minute horizons, and provides bounded, patient-specific insulin adjustment recommendations.
METHODS: Temporal GAT-BiGRU transforms multimodal continuous glucose monitoring (CGM) time series into a temporal k-neighborhood graph, with each time point represented as a feature-enriched node. A graph-attention encoder performs multi-head message passing over history edges, while an attention-based BiGRU captures longer dependencies. We evaluated OhioT1DM and BrisT1D using hypoglycemia and predicted Time in Range (TIR) metrics. A prediction-driven counterfactual module retrospectively generates bounded basal/bolus adjustments using patient-specific Insulin-to-Carbohydrate Ratio (ICR) and Insulin Sensitivity Factor (ISF); interpretability is supported via GNNExplainer.
RESULTS: On OhioT1DM, Temporal GAT-BiGRU achieved hypoglycemia precision-recall area under the curve (PR-AUC) 0.93, mean absolute error (MAE) 9.40 mg/dL, root mean square error (RMSE) 15.8 mg/dL, mean absolute relative difference (MARD) 6.01%, and predicted TIR 71.39%. On BrisT1D, performance remained strong with PR-AUC 0.98, MAE 9.36 mg/dL, RMSE 15.3 mg/dL, MARD 7.55%, and predicted TIR 64.78%. The insulin module generated bounded, subject-specific recommendations, typically suggesting ∼10% basal increases with individualized meal-bolus updates.
CONCLUSIONS: Temporal GAT-BiGRU provides accurate glucose prediction through temporal graph reasoning, sequence modeling, and interpretable explanations. It supports personalized decision support and closed-loop glucose management systems.

Keywords:  GNNExplainer; blood glucose forecasting; hypoglycemia prediction; insulin dose adjustment; multi-task learning

DOI:  https://doi.org/10.1093/jamia/ocag104
Front Digit Health. 2026 ;8 1811311

DFU-GCNet: a global context-enhanced inception network for robust and interpretable diabetic foot ulcer classification.

Md Tofael Ahmed Bhuiyan, Md Abdur Rahman, Farzan Majeed Noori, Md Zia Uddin, Abdul Kadar Muhammad Masum.

   Introduction: Diabetic foot ulcers (DFUs) are severe complications that cause frequent lower extremity amputations. Timely diagnosis is crucial for effective clinical management. Although deep learning approaches improve detection, the models often struggle to capture different lesion scales. Furthermore, opaque algorithmic decisions often lower medical trust. Therefore, this study introduces DFU-GCNet for robust and interpretable ulcer classification.
Methods: The proposed architecture merges inception modules with global context blocks. This combination extracts multi-scale features from different wound sizes and simultaneously models broad spatial dependencies across tissue regions. Thus, it effectively distinguishes pathology from surrounding healthy skin. We evaluate this framework using the Kaggle DFU dataset. We integrate explainable AI techniques to ensure clinical transparency. GradCAM++, Local Interpretable Model-Agnostic Explanations, and SHapley Additive exPlanations are used to provide high-resolution diagnostic heatmaps and confirm that the network prioritizes clinically relevant wound boundaries.
Results: The model achieved a superior classification accuracy of 97.16%, with an F1-score of 0.9715 and a Matthews correlation coefficient of 0.9437. DFU-GCNet demonstrated decisive superiority compared with standardized modern baselines such as VGG16 and EfficientNet.
Discussion: The findings indicate that DFU-GCNet is a highly reliable automated screening instrument.

Keywords:  DFU-GCNet; deep learning; diabetic foot ulcer; explainable AI (XAI); global context attention; medical image classification

DOI:  https://doi.org/10.3389/fdgth.2026.1811311
JMIR Diabetes. 2026 Jun 18. 11 e83059

Digital Twin Applications in Diabetes Management: Scoping Review.

Fatemeh Sarani Rad, Maryam Jafarpour, Ehsan Bitaraf, Katayoon Khaleghdadi, Juan Li.

   Background: Digital twin (DT) systems have emerged as a promising approach in health care, enabling real-time, patient-specific virtual modeling and personalized interventions. In diabetes care, DTs offer the potential to revolutionize glucose management, decision support, and therapy personalization through integration of real-time and longitudinal patient data.
Objective: This scoping review mapped the current landscape of DT applications in diabetes and synthesized evidence across 13 research questions organized into 7 thematic domains: system design, target conditions, data sources, personalization strategies, intelligence and adaptability, validation methods, and implementation considerations.
Methods: This scoping review was conducted in accordance with the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) and JBI methodological guidance for scoping reviews. A literature search was performed in PubMed, IEEE Xplore, Scopus, and Web of Science for studies published up to April 2025; all databases were last searched on June 23, 2025. Eligible studies were original empirical articles in English that described patient-specific DT systems or closely related individualized virtual models applied to diabetes diagnosis, monitoring, management, treatment, or complication-related care. Reviews, editorials, commentaries, theoretical papers without original data, and studies not focused on diabetes were excluded. Furthermore, FSR, MJ, and KK independently screened records and assessed full texts, with disagreements resolved through discussion and, when needed, by EB. Data were charted using a structured framework based on 13 predefined research questions, and were synthesized descriptively and thematically.
Results: Of 208 records identified, 123 underwent title and abstract screening, 39 full texts were assessed for eligibility, and 28 studies were included. Most studies focused on type 1 or type 2 diabetes and used data-driven, hybrid, or simulation-based DT approaches. Common clinical applications included therapeutic control, glucose prediction, decision support, and disease management. Lifestyle data, wearables, continuous glucose monitoring, and electronic health records were the dominant inputs, while personalization relied on adaptive feedback, insulin optimization, and behavior-driven tools. Intelligent features, such as adaptive learning, explainable artificial intelligence, and real-time synchronization, enhanced adaptability, although human oversight was rare. Validation was mainly retrospective or simulation-based, with few clinical trials; reported outcomes included improved hemoglobin A1c, time-in-range, and reduced hypoglycemia. Ethical discussions focused on data privacy, while implementation barriers centered on validation gaps, data quality, and workflow integration.
Conclusions: DT research in diabetes is expanding and shows strong potential for personalized and data-driven care; however, the evidence base remains heterogeneous, inconsistently reported, and limited in prospective clinical validation. Key gaps include standardized definitions, robust real-world evaluation, fairness and governance considerations, and integration into clinical workflows. Future work should prioritize clinically grounded validation, regulatory readiness, and interoperable architectures to support safe, equitable, and scalable implementation.

Keywords:  automated insulin delivery; clinical decision support; continuous glucose monitoring; diabetes mellitus; digital twin; ethics; machine learning

DOI:  https://doi.org/10.2196/83059
Recent Adv Drug Deliv Formul. 2026 Jun 16.

Application of Quality by Design (QbD) as Artificial Intelligence in Formulation for Diabetes: A Practical Review.

Gaurav Kumar Chaurasia, Ritesh Kumar Tiwari, Shashi Verma.

   INTRODUCTION: The increasing incidence of diabetes has made it essential to create more efficient, customized, and reproducible drug delivery systems. Quality by Design (QbD) is viewed as a science-driven, disciplined approach to drug development that relies more on ensuring consistent product quality through the determination of the Critical Quality Attributes (CQA), Critical Process Parameters (CPP), and development of a Design Space. Nonetheless, the sophistication of contemporary drugs and volumes of data involved tend to confine the independent effectiveness of QbD. Drug development has been transformed by Quality by Design (QbD), which has replaced reactive quality testing with proactive, scientifically based approaches. With its roots in ICH Q8-Q11 principles, QbD places a strong emphasis on defining Critical Quality Attributes (CQAs), creating Design Space, and incorporating risk management to improve the regulatory flexibility and resilience of products.
METHODS: The review gives a regulatory perspective on QbD and covers important tools from AI such as artificial neural networks (ANN), support vector machines (SVM), and response surface methodology (RSM). In particular, these AI tools offer predictive modeling, pattern recognition, and optimization in support of formulation design. A comprehensive search of PubMed, Google Scholar, and ScienceDirect identified relevant articles focused on the use of AI and machine learning (ML) in diabetes care. Case studies are provided to demonstrate applications in practice, including oral insulin nanoparticles, extended-release metformin, and oral peptide formulations.
RESULTS: By integrating AI with QbD, an ideal environment is created, one that will enhance formulation accuracy, reduce development times, and increase the possibility of regulatory acceptance.
DISCUSSION: This review investigates the convergence of Artificial Intelligence (AI) technologies and QbD principles in advanced diabetes therapy formulation development. It evaluates how AI tools improve the efficiency, precision, and regulatory acceptability of QbD-guided pharmaceutical design, especially for diabetes therapy.
CONCLUSION: Although challenges exist with data integrity, regulatory interpretation, and the necessity for a combination of scientific themes, the AI-QbD framework is a transformational journey toward intelligent, patient-focused drug design. Future directions include advances like real-time release testing (RTRT), AI-enabled personalized medicine, and the integration with digital health.

Keywords:  QbD; artificial intelligence; diabetes; drug design; pharmaceutical formulation; predictive modeling.

DOI:  https://doi.org/10.2174/0126673878428147260531183116
Front Endocrinol (Lausanne). 2026 ;17 1858808

Development and validation of an explainable neural network model for predicting progression in type 2 diabetic kidney disease.

Binfeng Xiong, Chengzheng Duan, Keying Lin, Sheng Xu.

   Background: Type 2 diabetic kidney disease (T2DKD) affects 20-40% of patients with type 2 diabetes mellitus and has become the leading cause of end-stage renal disease globally. Early identification of patients at risk of rapid progression remains challenging, as existing prediction models often rely on complex indicators unsuitable for primary care settings. This study aimed to develop and validate a machine learning model using routine clinical parameters to predict T2DKD progression.
Methods: This single-center retrospective cohort study enrolled 349 patients diagnosed with T2DKD according to clinical criteria at Quzhou People's Hospital in China between June 2022 and June 2025. From 36 baseline characteristics, four core predictors were identified through least absolute shrinkage and selection operator (LASSO) regression and multivariable logistic regression. Six machine learning models were constructed, and model performance was evaluated by discrimination, calibration, and clinical net benefit. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) analysis.
Results: Metabolic dysfunction-associated steatotic liver disease (MASLD), aspartate aminotransferase (AST), diabetic peripheral neuropathy (DPN), and age were identified as independent core predictors. The neural network (NN) model achieved optimal performance in the test dataset, with an area under the curve (AUC) of 0.742, satisfactory calibration (Hosmer-Lemeshow P = 0.2020), the lowest Brier score (0.2105), and superior clinical net benefit across risk thresholds of 0.2-0.6. SHAP analysis confirmed stable feature importance rankings between training and test datasets (Pearson r = 0.976) and revealed synergistic interactions among MASLD, AST, and DPN.
Conclusion: A NN model incorporating four routine clinical indicators effectively predicts T2DKD progression risk. This cost-effective tool is suitable for clinical practice and community health services, offering a scalable solution for early intervention and prognosis improvement.

Keywords:  machine learning; neural network; predictive model; serum creatinine; type 2 diabetic kidney disease; urinary albumin-to-creatinine ratio

DOI:  https://doi.org/10.3389/fendo.2026.1858808
AMIA Jt Summits Transl Sci Proc. 2026 ;2026 193-201

Machine Learning Based County Level Phenotypes Related to Diabetes Prevalence.

Md Fitrat Hossain, Fadia T Shaya.

In the US, diabetes prevalence rates continue to rise. However little focus is given on the association of diabetes with the social determinants of health (SDoH). This study focuses on developing phenotypes based on county-level SDoH which are related to diabetes prevalence. Machine learning algorithms such as classification and regression tree (CART) model were used to define phenotypes based on county-level SDoH. Random forest was also used to identify additional risk factors. Five different phenotypes identified by the CART model divided the counties into five groups. Counties with high food insecurity rates (more than 16%) and high poverty rates (more than 24%) were found to have higher mean prevalence rate of diabetes (17.64%, SD 2.42). These can help individuals involved in healthcare and policy makers to make tailored region-based interventions to reduce diabetes prevalence and improve living conditions for people with diabetes.
Front Cardiovasc Med. 2026 ;13 1780009

Predictive voting model for early diagnosis of diabetes dataset.

Talaa Al-Bimani, Jamil AlShaqsi, Rami Alkhawaldeh, Adil S Al Busaidi, Ali Tarhini.

  This research aimed to construct predictive voting models (hard vote and soft vote) to improve the diabetes diagnosis system at the initial pre-diabetes stage using several risk factors retrieved from a dataset collected from a government hospital in Oman. The study focused on identifying significant predictors of diabetes and enhancing the accuracy of early diagnosis. The Knowledge Discovery in Database (KDD) model was utilized to conduct the experiments. A 33-month historical dataset comprising N = 4104 registered patients and 14 variables was analyzed. The features used for diabetes classification included age, height, weight, gender, diastolic and systolic blood pressure, cholesterol level, blood glucose level, and haemoglobin level. Five supervised classification algorithms were applied to construct the voting models: Decision Tree (J48), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest, and Naïve Bayes. The findings revealed that the hard-vote model achieved the highest predictive accuracy of 84.7% compared with the soft-vote model. Additionally, the haemoglobin A1C test (HbA1c), Fasting Plasma Glucose (FPG), and age were identified as the most significant factors for predicting diabetes. The extracted rules indicated that HbA1c served as the initial criterion for diabetes diagnosis, with a threshold value of 6.3. The study demonstrated the effectiveness of ensemble voting models in improving diabetes prediction during the pre-diabetes stage. The identified predictors and extracted rules may support healthcare professionals in making earlier and more accurate diagnostic decisions. Furthermore, the involvement of domain experts and the validation of rules using classified patient cases strengthened the reliability and practical applicability of the proposed models.

Keywords:  classification; data mining; diabetes in Oman; knowledge discovery in database (KDD); prediction algorithms; voting ensemble algorithms

DOI:  https://doi.org/10.3389/fcvm.2026.1780009
Sci Rep. 2026 Jun 18.

Machine learning-based predictive model for sleep disorders in diabetic patients: data analysis from CHARLS.

Maoqin Tian, Shunli Zuo, Yongping Sun, Xue Gong, Ning Liu.

  Sleep disorders are prevalent and constitute a major concern in patients with diabetes mellitus. Therefore, the aim of this study was to investigate the applicability of machine learning methods in predicting sleep disorders among diabetic patients. Six relevant features were selected using single-factor correlation analysis and the LASSO algorithm. We developed and evaluated five ML models: logistic regression, decision tree, extreme gradient boosting, support vector machine, and light gradient boosting machine. Data from the China Health and Retirement Longitudinal Study database were utilized, with a total of 60,308 elderly individuals screened, of which 1276 diabetic patients were included in the analysis. Of these, 777 did not develop sleep disorders, while 499 did. Fifteen statistically significant predictors were identified through single-factor analysis, and six relevant variables were determined via LASSO regression, including family history of diabetes, education, marital status, chronic diseases, chronic pain, and depression. Based on these six variables, five ML models were constructed to predict the risk of sleep disorders in diabetic patients. Among these, the XGB model demonstrated superior performance, with an area under the curve of 0.850. The calibration curve indicated a good fit of the model on the development set, and decision curve analysis further confirmed the model's excellent net benefit and prediction accuracy. The overall performance of the XGB model was the best. Our findings suggest that ML models, particularly extreme gradient boosting, offer the most effective approach for predicting the risk of sleep disorders in diabetic patients.

Keywords:  CHARLS; Diabetes; Machine learning; Prediction model; Sleep disorders

DOI:  https://doi.org/10.1038/s41598-026-53312-x
Front Public Health. 2026 ;14 1793361

Enhancing diabetes risk stratification through natural language processing: a multimodal data integration approach.

Yaoyan Lu, Tingting Li, Liyuan Ge, Yujun Wei, Yujing Li, Shangyuhui Huang, Tao Jiang, Yanrui Chen, Zhen Xiang.

   Background: This study investigates whether integrating Natural Language Processing (NLP) with traditional clinical data improves type 2 diabetes risk prediction by extracting latent risk factors from unstructured medical text.
Methods: We analyzed a public dataset of 1,879 individuals. Structured variables (BMI, HbA1c, blood pressure) were normalized, while unstructured textual entries (symptom descriptions, lifestyle notes) were processed using a BERT-based NLP pipeline for feature extraction (risk behaviors, family history). A hybrid model integrated these NLP-derived features with traditional variables using logistic regression. Performance was evaluated via accuracy, precision, recall, F1-score, and AUC-ROC. Robustness was assessed through bootstrap confidence intervals, sensitivity analysis, optimism adjustment, and temporal validation on a post-2020 cohort (n = 939). Generalizability across learners was tested using random forest, XGBoost, and neural networks.
Results: NLP identified non-traditional risk indicators including sedentary occupation, poor dietary adherence, and chronic stress. The integrated model significantly outperformed the structured-only baseline (Accuracy: 88.2 vs. 76.5%; AUC-ROC: 0.92 vs. 0.83). Key independent predictors included NLP-identified sedentary behavior (OR = 1.80, p = 0.003), hypertension (OR = 2.12, p < 0.001), elevated HbA1c (OR = 2.38, p < 0.001), and high BMI (OR = 1.67, p < 0.001). Validation confirmed stability: optimism-adjusted AUC was 0.91, temporal validation yielded AUC 0.89 (performance decay <3%), and all four classifiers showed consistent improvement with NLP features (ΔAUC: +0.07 to +0.09).
Conclusion: NLP effectively unlocks latent risk information from unstructured clinical text, significantly enhancing diabetes risk prediction. This framework enables more holistic patient risk assessment for personalized prevention. Future work requires external validation, privacy-preserving methods, and enhanced interpretability via explainable AI.

Keywords:  BERT model; diabetes mellitus; machine learning; natural language processing; public health informatics; risk assessment; type 2; unstructured data

DOI:  https://doi.org/10.3389/fpubh.2026.1793361
J Ocul Pharmacol Ther. 2026 Jun 18. 10807683261460674

Integrated Bioinformatics and Explainable Machine Learning Analyses Reveal Mesenchymal Stem Cell Exosome-Related Biomarkers for Diabetic Retinopathy.

Liu Zheng, Jiaojiao Jiang, Jinqing Li, Xingyu Su, Shixiong Peng, Zhixiang Ding.

   PURPOSE: This study aims to identify potential biomarkers for diabetic retinopathy (DR) by focusing on genes associated with mesenchymal stem cell-derived exosomes (MSCs-Exo).
METHODS: By integrating DR transcriptome data with the protein dataset of MSCs-Exo, we utilized a comprehensive array of bioinformatics techniques, including weighted gene coexpression network analysis, Mfuzz clustering, and machine learning algorithms such as least absolute shrinkage and selection operator regression and random forest to pinpoint key genes. Functional mechanisms were explored through functional enrichment analysis, immune infiltration, and single-cell RNA sequencing. The immunohistochemistry and Western blotting were used for validation on DR mice models.
RESULTS: Our comprehensive analysis identified 16 hub genes associated with MSCs-Exo. Through the application of interpretable machine learning techniques, YBX1 and PSMA7 were further identified as central genes within this network. A predictive diagnostic model for DR was developed and validated using receiver operating characteristic curve analysis, which demonstrated modest diagnostic efficacy, as indicated by an area under the curve exceeding 0.7. Importantly, experimental validation showed that the protein expression levels of YBX1 and PSMA7 were significantly reduced in the retinal tissues of DR mice compared with the control group (P < 0.05). Functional enrichment analysis suggested that YBX1 and PSMA7 are involved in critical biological processes, specifically the regulation of protein and amino acid metabolism. In addition, immune infiltration results show that they are significantly associated with the immune dysregulation of DR, especially in CD4T memory cells. Single-cell analysis also supported the above finding.
CONCLUSION: These findings suggest that YBX1 and PSMA7, derived from MSCs-Exo, may serve as potential biomarkers for DR. Further studies are needed to confirm their clinical utility and therapeutic relevance.

Keywords:  PSMA7; YBX1; diabetic retinopathy (DR); machine learning; mesenchymal stem cell-derived exosomes (MSCs-Exo); multiomics

DOI:  https://doi.org/10.1177/10807683261460674
JAMA. 2026 Jun 15.

An AI-Based OCT System to Detect Diabetic Macular Edema: A Prospective Validation and Noninferiority Randomized Clinical Trial.

Shuyi Zhang, Anran Ran, Jiaying Zhou, Anni Ling, Kaiser Sham, Yu Zhang, Ziqi Tang, Truong X Nguyen, Dawei Yang, Naiman Lam, Hunter K L Yuen, Victor T T Chan, Mary Ho, Julia Y Y Chan, Thomas C H Lam, Charlene C L Yim, Claire W Y Chow, Stephanie S L Cheung, Matthew C W Lam, Anthony C H Lai, Cherie Y K Wong, Amy H Y Yu, Vivian W K Hui, Andrew C Y Mak, Sophia Ling Li, Noel C Y Chan, Wilson W K Yip, Alvin L Young, Risa Ozaki, Kit-Man Loo, Andrea O Y Luk, Juliana C N Chan, Pangfai Chan, Matthew M H Luk, Loretta K P Lai, Daniel H T Wong, Jasmine Shiu, Xi Wang, Pheng-Ann Heng, Haotian Lin, Sobha Sivaprasad, Tien Yin Wong, Kenneth K W Li, Carmen K M Chan, Chi Pui Pang, Clement C Tham, Timothy Y Y Lai, Simon K H Szeto, Carol Y Cheung.

Importance: Screening for diabetic retinopathy using fundus photographs is the global standard of care but results in high false-positive referrals to evaluate diabetic macular edema (DME), placing a substantial burden on specialist eye clinics. Integrating an AI-based optical coherence tomography (AI-OCT) system into screening pathways may reduce potentially unnecessary referrals.
Objective: To evaluate the diagnostic and referral performance of an AI-OCT system for DME detection within a diabetic retinopathy screening pathway in clinical settings.
Design, Setting, and Participants: Stepwise evaluation conducted in Hong Kong Special Administrative Region: a prospective silent-mode validation (February 2020 to July 2023) recruiting 603 patients with diabetes at a tertiary hospital triage unit, followed by a multicenter noninferiority RCT (September 2023 to April 2025), with follow-up completed in May 2025, recruiting 276 patients with suspected DME referred from a territory-wide diabetic retinopathy screening program.
Interventions: RCT participants were randomized to intervention (referral for DME evaluation based on both fundus photograph-based screening reports and AI-OCT reports [n = 137]) or control (automatic referral based solely on fundus photograph-based screening reports [n = 139]) groups. The AI-OCT system incorporated image-quality assessment, DME detection, and uncertainty flagging. Study outcomes focused on referral rates under the 2 pathways; for ethical reasons, all participants ultimately underwent specialist evaluation.
Main Outcomes and Measures: The primary outcome was false-positive DME referral rate, with a prespecified noninferiority margin of 20%. The secondary outcomes included sensitivity and specificity for DME detection and DME referral.
Results: In prospective silent-mode validation (mean age, 64.7 [SD, 9.4] years; 56.2% male), 86 of 1200 scans (7.2%) were identified as ungradable and 49 of 1114 gradable scans (4.4%) were classified as uncertain. The system achieved 98.8% (95% CI, 94.5%-100.0%) sensitivity and 90.7% (95% CI, 88.7%-92.4%) specificity for DME detection. In the RCT (mean age, 63.9 [SD, 10.9] years; 54.7% male), DME prevalence was similar in the intervention and control groups (30.9% vs 29.9%). The false-positive DME referral rate was 24.1% (95% CI, 14.6%-37.0%) and 69.1% (95% CI, 61.0%-76.1%), respectively (absolute difference, -45% [95% CI, -58.2% to -31.9%; P < .001 for noninferiority]; upper bound of the CI below the prespecified noninferiority margin of 20%). Sensitivity for DME referral was 100.0% (95% CI, 100.0%-100.0%) in both groups. Specificity for DME referral was 86.5% (95% CI, 79.3%-92.9%) in the intervention group and 0.0% (95% CI, 0.0%-0.0%) in the control group. No cases of DME occurred among nonreferred participants in the intervention group.
Conclusions and Relevance: Compared with standard practice, incorporation of the AI-OCT system as a secondary screening tool was noninferior with respect to false-positive referral rates and was associated with a substantial reduction in potentially unnecessary DME referrals without compromising sensitivity.
Trial Registration: Chinese Clinical Trial Registry: ChiCTR2300075087.

DOI: https://doi.org/10.1001/jama.2026.7025
Sci Rep. 2026 Jun 16.

Non-invasive opportunistic screening for diabetes mellitus: an interpretable stacking ensemble framework.

Lin Zhang, Junming Xu, Weigang Wang, Rui Chang, Linghua Wang.

  Opportunistic screening for type 2 diabetes offers a potentially accessible approach to preliminary case detection without relying on invasive testing. In this study, we developed a heterogeneous Stacking ensemble model (Task C) using exclusively non-invasive demographic, lifestyle, medical-history, and symptom-based features. The model prioritized sensitivity, achieving a Recall of 0.9267, while showing modest discriminative performance (AUC = 0.5515), low specificity (0.1106), and moderate probability calibration (Brier Score = 0.2482). Targeted simulation analyses revealed that adjusting the top three modifiable behavioral factors captured approximately 85.5% of the reduction in model-estimated screening probability observed under the all-six-factor adjustment. Individual-level case simulation illustrated a stepwise reduction in model-estimated screening probability under increasingly comprehensive hypothetical adjustments. Decision curve analysis suggested potential screening utility mainly within the lower-threshold range. These findings suggest that the proposed ensemble may serve as a technically feasible and interpretable tool for preliminary non-invasive diabetes case-finding, while providing hypothesis-generating insights into modifiable factors for future validation.

Keywords:  Diabetes mellitus; In Silico simulation; Interpretable stacking; Opportunistic screening; SMOTENC

DOI:  https://doi.org/10.1038/s41598-026-57865-9
iScience. 2026 Jun 19. 29(6): 116280

Predicting and explaining poor prognosis in diabetic kidney disease using SHAP-based interpretable machine learning.

Man Qian, Lin Li, Yanli Cheng, Yue Hou, Shaojie Fu, Zhonggao Xu.

  Prognostic assessment of diabetic kidney disease (DKD) is essential for personalized management. This study developed eight machine learning models using data from 180 biopsy-proven patients with DKD to predict a composite endpoint of all-cause mortality, dialysis initiation, or renal transplantation. Internally, the Naive Bayes (NB) model achieved the highest accuracy of 82.3%, while the logistic regression (LR), support vector machine (SVM), and NB models shared the highest AUC of 0.788. An independent external validation confirmed robust generalizability, yielding an AUC of 0.834. SHAP analysis identified eGFR, serum albumin, C3, serum creatinine, and urinary red blood cell count (URBC) as the most impactful features. Feature stability was confirmed via a "leave-top1-out" sensitivity analysis. The models highlighted the predictive value of C3 and URBC by capturing non-linear patterns often missed by traditional linear methods, providing granular insights for personalized prognosis evaluation.

Keywords:  Artificial intelligence; Health sciences; Machine learning; Nephrology

DOI:  https://doi.org/10.1016/j.isci.2026.116280
Wounds. 2026 Apr;38(4): 97-106

The STEADY diabetic foot ulcer registry: methods, insights, and future directions.

Jared R Adams, Jennifer N Lannon, Vickie R Driver, David Alper, Cyaandi R Dove, John C Lantis, Joseph Rolley, Robert Snyder.

   BACKGROUND: Diabetic foot ulcers (DFUs) are a major cause of morbidity, amputation, and mortality among individuals with diabetes, with disproportionate impact on underserved populations. Comprehensive real-world data on DFU management and outcomes are lacking.
OBJECTIVE: To describe the design and methodology of the STEADY (Structured Evaluation and Analysis of Diabetic Foot Ulcers in the US) registry, a national prospective cohort study of patients with DFUs in the United States whose objective is to evaluate DFU treatment patterns, outcomes, and health care resource utilization in real-world settings, to assess comparative effectiveness, cost effectiveness, and safety of DFU therapies and therapy combinations, and to advance disease management through risk- and site-stratified treatment optimization models.
METHODS: STEADY is a 10-year prospective multicenter observational study with an aim of enrolling 5000 adults with active DFUs in the United States. Data sources include electronic case report forms, electronic medical records (EMRs), patient-reported outcomes via mobile app, and optional insurance claims. Primary and secondary end points will include time and incidence of partial and complete wound closure; wound and disease characteristics; rates of recurrence, infection, ischemic events, and amputation; health care utilization, including surgical procedures; health-related quality of life; work productivity; and additional patient reported outcomes. Descriptive, survival, and comparative effectiveness analyses will be performed. Data governance ensures full regulatory compliance and robust data security and integrity, supporting the potential use of the registry dataset as a synthetic control arm in future clinical research.
CONCLUSION: STEADY leverages an artificial intelligence (AI)-enabled platform to integrate multisource data, including wound photography, social determinants of health, patient reported outcomes and caregiver information. The platform uses AI for transcription and interpretation of patient and provider dictation, supports patient-controlled EMR synchronization for comprehensive longitudinal tracking across providers, offers participant incentives to enhance engagement, and ensures rigorous, automated data quality assurance at all stages.

Keywords:  artificial intelligence; diabetic foot ulcer; machine learning; registry; wound management

DOI:  https://doi.org/10.25270/wnds/26010