bims-aukdir 2026-03-15 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2026–03–15
twenty-one papers selected by
Mott Given

Diagnostic Accuracy and Real-Life Advantages of the MONA.health Artificial Intelligence Software in Screening for Diabetic Retinopathy and Maculopathy.
Enhancing fundus image analysis for diabetic retinopathy using CheXNet with CBAM and Grad-CAM visualization.
BigEye: a clinically interpretable deep learning framework for diabetic retinopathy detection and stage prediction.
A multimodal retinal image dataset for diabetic retinopathy detection using foundation models.
Correction: Early prediction of gestational diabetes mellitus using machine learning-integrated metabolomic and clinical features.
Real world practice of Artificial intelligence Diagnostic System for Diabetic Retinopathy in Taiwan.
Validation of the Eyerobo FC Portable Fundus Camera for Diabetic Retinopathy Screening Using Public Datasets and Deep Learning.
AI-Driven Diagnostics vs. Clinician Assessment in Diabetic Retinopathy: A Comparative Analysis at a Secondary Eye Care Centre.
Development and External Validation of a Machine Learning Model for Predicting Wound Infection in Diabetic Foot Ulcers.
Training and validation of an automated algorithm to differentiate no and minimal diabetic retinopathy from more severe stages in wide-field images.
A study on the risk prediction of heart disease in diabetes patients based on machine learning.
Artificial intelligence in mobile health applications: A comprehensive review of its role in diabetes care.
Artificial Intelligence in Screening and Grading Diabetic Eye Diseases: A Systematic Review From Algorithms to Clinic.
Identifying Predictors of Early Treatment Intensification in Individuals With Type 2 Diabetes Treated With GLP-1 Receptor Agonists: A Machine Learning-Based Analysis of a Large Real World Cohort.
From raw data to actionable insights: preprocessing real-world data for machine learning in diabetes care.
Evaluating Large Language Models-Generated Health Education Materials for Discharged Patients with Diabetes: A Comparative Analysis.
Glucose Predictions Improve Glycemic Control: A Digital Twin Evaluation.
AI-Enabled Personalization of Semaglutide Therapy in Type 2 Diabetes: Systematic Review With an Integration Framework.
Machine learning and engagement insights for personalized blood glucose management.
Identification of Retinal Diseases Using Light Convolutional Neural Networks and Intrinsic Mode Function Technique.
Comparative assessment of large language models in diabetic foot infection management: alignment with IWGDF/IDSA guidelines.

Diagnostics (Basel). 2026 Mar 01. pii: 730. [Epub ahead of print]16(5):

Diagnostic Accuracy and Real-Life Advantages of the MONA.health Artificial Intelligence Software in Screening for Diabetic Retinopathy and Maculopathy.

Martina Tomić, Romano Vrabec, Toma Babić, Kristina Kljajić, Tomislav Bulum.

  Background/Objectives: We aimed to evaluate the diagnostic accuracy of the MONA.health artificial intelligence (AI) software (Version 1.0.0; MONA.health, Leuven, Belgium) and compare its advantages in screening for diabetic retinopathy (DR) and diabetic macular edema (DME) with standard fundus photography. Methods: This cross-sectional, real-life instrument validation study was conducted at the Vuk Vrhovac University Clinic in Zagreb during routine DR screening and included 296 patients (592 eyes) with diabetes. Following standard fundus photography using a 45° Zeiss VISUCAM NM/FA camera (Carl Zeiss Meditec AG, Jena, Germany), each patient also underwent imaging with an automated portable retinal camera (NFC-600, Crystalvue Ophthalmic Instruments, Taoyuan City, Taiwan). Two retina specialists independently graded images from the standard camera, while images from the NFC-600 were analyzed using the MONA.health AI software. Results: Among the 592 eyes, human grading identified 81 with any DR, including 17 with mild NPDR, 64 with referable DR (moderate/severe NPDR or PDR), and 13 with DME. The MONA.health AI software identified 65 eyes with referable DR and 19 with DME. For MONA DR screening compared to the standard fundus camera, the area under the curve, sensitivity, specificity, positive predictive value, negative predictive value, positive likelihood ratio, negative likelihood ratio, kappa agreement, diagnostic odds ratio, and diagnostic effectiveness were 99.74%, 100%, 99.81%, 99.33%, 100%, 528.00, 0.00, 0.99, infinity, and 99.85%, respectively. For MONA DME screening, these metrics were 97.97%, 100%, 98.95%, 85.93%, 100%, 95.67, 0.00, 0.81, infinity, and 99.02%, respectively. The MONA AI screening process required 1 day of training and approximately 5 min for image capture and analysis, compared to 7 days of training and 13 min for image acquisition and grading with the standard method. Conclusions: These findings demonstrate that the MONA.health AI software matches the accuracy of standard fundus photography for screening and early detection of referable DR and DME, while offering a faster, simpler, and more user-friendly workflow that significantly reduces the time to obtain screening results.

Keywords:  artificial intelligence; diabetic retinopathy and maculopathy screening; diagnostic accuracy; real-life advantages; standard fundus photography

DOI:  https://doi.org/10.3390/diagnostics16050730
Front Med (Lausanne). 2026 ;13 1732109

Enhancing fundus image analysis for diabetic retinopathy using CheXNet with CBAM and Grad-CAM visualization.

Wedad Al-Dolat, Salem Alhatamleh, Noor Alqudah, Amro Alhazimi, Mohammad Amin, Aseel Daamseh, Rola Madain, Raghad Malkawi, Rami Al-Omari, Faisal Almarek, Sarah Husam Aljefri.

   Introduction: Diabetic retinopathy (DR) is a leading cause of vision impairment among individuals with diabetes. Early detection and accurate grading are essential for timely clinical management. However, developing robust models for automated interpretation and grading of fundus images remains challenging due to variability in lesion appearance and image quality.
Methods: This study proposes a deep learning framework for DR classification from fundus images based on a DenseNet121 backbone initialized with CheXNet weights. A Convolutional Block Attention Module (CBAM) is integrated to enhance feature representation through channel and spatial attention mechanisms in a data-driven manner. In addition, Gradient-weighted Class Activation Mapping (Grad-CAM) is employed to provide post hoc visual explanations of model predictions. The proposed CheXNet_CBAM model is evaluated against several convolutional neural network architectures, including CheXNet, DenseNet121, MobileNetV2, VGG19, and ResNet50, using the APTOS 2019 and DDR datasets.
Results: On the APTOS 2019 dataset, the proposed model achieves an accuracy of 96.12%, while on the DDR dataset it attains 96.33%, outperforming the compared architectures on both benchmarks.
Discussion: The results indicate that incorporating CBAM improves discriminative feature learning within a DenseNet121-based framework. While the model demonstrates strong performance across two public datasets, further prospective evaluation and external validation are required to assess its clinical applicability in real-world settings.

Keywords:  Grad-CAM; deep learning; diabetic retinopathy; fundus imaging; image classification

DOI:  https://doi.org/10.3389/fmed.2026.1732109
Sci Rep. 2026 Mar 08.

BigEye: a clinically interpretable deep learning framework for diabetic retinopathy detection and stage prediction.

Hunter Mathias Gill, Doaa Hassan Salem, Okiemute Beatrice Omoru, Frank Dash Bogan, Jeffrey Xiao Liu, Michael Happe, Amir Reza Hajrasouliha, Sarath Chandra Janga.

  Diabetic Retinopathy (DR) is a major cause of vision loss and blindness in diabetic individuals. DR is conventionally diagnosed by assessing retinal lesion findings from fundus photographs taken during exams and applying a scale like International Classification of Diabetic Retinopathy (ICDR). The expected rise in future DR cases highlights the need for deep learning models capable of identifying relevant lesions and delivering explainable results. To this end we present BigEye, a novel framework that uses extracted lesion features to predict ICDR stage. A dataset of fundus images from a local hospital and a public dataset, annotated with segmentation masks and DR stages, is assembled to train a DeepLabV3 + model on six retinal lesions. Lesion quantities and pixel area features are integrated by a classifier model evaluated through 10-fold nested cross validation (0.77 ± 0.07 precision, 0.71 ± 0.06 recall, 0.72 ± 0.07 F1 score, 0.95 ± 0.02 ROC-AUC, 0.83 ± 0.03 accuracy). A Shapely Additive Explanations (SHAP) value analysis notably shows close alignment between discriminative lesions for each DR stage and corresponding ICDR stage criteria. These results demonstrate that BigEye is well suited for providing explainable ICDR stage predictions grounded in clinical knowledge.

Keywords:  DR; DR Staging; Diabetic Retinopathy; Explainable AI; Retinal Lesions

DOI:  https://doi.org/10.1038/s41598-026-43573-x
Sci Data. 2026 Mar 10.

A multimodal retinal image dataset for diabetic retinopathy detection using foundation models.

Zhenyu Tang, Lilong Wang, Zhen Guo, Qianqian Liang, Shuyue Xue, Chengcheng Feng, Lili Ran, Lingzhi Chen, Xiaosong Wang, Jun Li.

Diabetic Retinopathy (DR), a leading cause of preventable blindness worldwide, underscores the urgent need for robust AI-driven diagnostic tools. Although various deep learning models for retinal imaging have emerged, their evaluation remains constrained by limited public available datasets that lack both large-scale coverage and fine-grained annotations, compromising reliable assessments of model generalizability. To bridge this gap, we introduce a comprehensive multimodal dataset that includes three key retinal imaging modalities: color fundus photography (CFP), optical coherence tomography (OCT), and ultrawide-field fundus imaging (UWF). Our dataset is unprecedented in scale and modality diversity, provides detailed lesion-level annotations and severity grades for DR and Diabetic Macular Edema (DME). We benchmark a range of fundus foundation models and large vision-language models on this dataset, revealing critical performance gaps and domain-specific challenges. By unifying large-scale multimodal data with precisely annotated clinical labels, our work establishes a foundational benchmark to drive advances in AI reliability and real-world clinical utility.

DOI: https://doi.org/10.1038/s41597-026-07005-9
Front Endocrinol (Lausanne). 2026 ;17 1810108

Correction: Early prediction of gestational diabetes mellitus using machine learning-integrated metabolomic and clinical features.

Qun Ji, Lan Gao, Haiwei Liu, Xiaofang Chen, Boxia Fu, Yingbei Lin, Fei Wang.

  [This corrects the article DOI: 10.3389/fendo.2025.1687146.].

Keywords:  early prediction; gestational diabetes mellitus; machine learning; metabolites; metabolomic profiling

DOI:  https://doi.org/10.3389/fendo.2026.1810108
Retina. 2026 Mar 12.

Real world practice of Artificial intelligence Diagnostic System for Diabetic Retinopathy in Taiwan.

Ching-Chun Lin, Cheng-Kuo Cheng, Pai-Hui Peng, Sheng-Fu Cheng.

   PURPOSE: We evaluated the alterations of applying artificial intelligence (AI) diagnostic system for diabetic retinopathy (DR) screening in real-world practice.
METHODS: This retrospective study included 11,713 diabetic patients from the government-led Diabetes Shared Care Network. The AI system Verisee was integrated into the clinical workflow to identify referable diabetic retinopathy (RDR). Its performance was compared with ophthalmologist grading at the patient level using sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC). Subgroup analysis was performed by age and gender, with additional referral diseases identified by ophthalmologists.
RESULTS: Verisee achieved a sensitivity of 0.88, specificity of 0.86, accuracy of 0.86, PPV of 0.58, NPV of 0.97, and AUC of 0.87 in detecting RDR. Performance declined with increasing age, whereas sex distribution remained consistent across age groups. The AI system identified a higher proportion of RDR than ophthalmologists (27.45% vs. 18.15%). In addition to 1,818 patients with RDR, ophthalmologists identified other referral-warranted ocular conditions in 4.5% of cases. The AI system referred age-related macular degeneration (grades 2-4), whereas referral decisions for macular hole and macular edema (grades 1-2) varied; however, glaucoma (grades 0-1) identified by clinicians was not consistently referred.
CONCLUSION: Verisee demonstrated high accuracy in detecting RDR but exhibited reduced performance in older patients. It had a higher referral rate than ophthalmologists yet missed certain conditions such as glaucoma. Despite effectiveness in DR screening, further refinement is required to support broader ophthalmic disease detection.

Keywords:  artificial intelligence; diabetic retinopathy; diagnostic system retina; real world practice

DOI:  https://doi.org/10.1097/IAE.0000000000004826
Ophthalmol Ther. 2026 Mar 11.

Validation of the Eyerobo FC Portable Fundus Camera for Diabetic Retinopathy Screening Using Public Datasets and Deep Learning.

Emmanuel Eric Pazo, Xiangying Liu, Shoukuan Liu, Qinru Zhang, Mingxuan Zhang, Zhihui Zhang, Salissou Moutari, Muhammad Usama, Xinjun Ren.

   INTRODUCTION: To validate the diagnostic performance of the Eyerobo FC, a new portable non-mydriatic fundus camera for diabetic retinopathy (DR) screening, against an established desktop fundus camera benchmark using a transfer-learning approach in which artificial intelligence (AI)-based detection algorithms trained on desktop images were applied to Eyerobo FC images.
METHODS: This prospective validation study employed a three-tier experimental design. Tier 1 involved training a deep learning model (EfficientNet-B4) on standard desktop camera images from EyePACS and APTOS 2019 datasets. Tier 2 established the reference standard by evaluating the trained model on the Messidor-2 dataset (N = 1748 eyes) captured with a Topcon TRC NW6 desktop camera (sensitivity 92.7%, 95% CI 91.2-94.2%; AUC 0.952, 95% CI 0.943-0.961). Tier 3 validated the same AI model (without retraining) on images from the Eyerobo FC in a prospective cohort (N = 104 eyes: 52 referable DR, 52 non-referable). The primary outcome was noninferiority of sensitivity and specificity (margin Δ = 10%) compared with the desktop benchmark. Statistical analysis included bootstrap resampling (1000 iterations) for confidence intervals and a one-sided Z-test for the difference of proportions to assess noninferiority.
RESULTS: In a balanced cohort of 104 eyes (52 referable DR, 52 non-referable), the Eyerobo FC achieved sensitivity of 92.3% (95% CI 84.4-98.2%) and specificity of 94.2% (95% CI 87.0-100%), demonstrating noninferior performance compared with the desktop benchmark (sensitivity 92.7%, specificity 94.3%). The sensitivity difference of -0.4 percentage points and the specificity difference of -0.1 percentage points were both within the noninferiority margin. AUC was 0.977 (95% CI 0.945-0.997) versus 0.952 for the desktop benchmark. The AI model correctly classified 97 of 104 eyes (93.3% accuracy, 95% CI 88.5-98.1%), with 4 false negatives and 3 false positives. Noninferiority was statistically confirmed for both sensitivity and specificity (P < 0.05). Inter-grader agreement was excellent (Cohen's kappa = 0.917). Nonmydriatic image gradability rate was 94.4%. Grad-CAM visualization confirmed appropriate model attention to hemorrhages, exudates, and microaneurysms rather than artifacts.
CONCLUSIONS: The Eyerobo fundus camera demonstrates noninferior diagnostic performance (sensitivity 92.3%, specificity 94.2%, AUC 0.977) compared with desktop systems when evaluated with AI algorithms trained exclusively on desktop images. These findings support deploying portable AI-assisted screening in resource-constrained and point-of-care settings, with successful cross-domain transfer learning enabling algorithmic generalizability across imaging platforms.

Keywords:  Artificial intelligence; Deep learning; Diabetic retinopathy; Handheld fundus camera; Point-of-care diagnostics; Telemedicine; Validation study

DOI:  https://doi.org/10.1007/s40123-026-01353-w
Ophthalmic Epidemiol. 2026 Mar 08. 1-9

AI-Driven Diagnostics vs. Clinician Assessment in Diabetic Retinopathy: A Comparative Analysis at a Secondary Eye Care Centre.

Ram Sudarshan Ravindran, Syed Mohideen Abdul Khadar, Kim Ramasamy, Aarif Khan, Nikhil Gopalakrishnan, Mohammed Sithiq Uduman, Harinikrishna Balakrishnan, Meenakshi Ravindran, Rengappa Ramakrishnan.

   PURPOSE: To determine the diagnostic accuracy and reliability of artificial intelligence (AI) in identifying diabetic retinopathy (DR) and macular oedema (ME) compared to ophthalmologists.
METHODS: This prospective study included 294 patients (576 eyes). Fundus images obtained using a non-mydriatic Topcon NW400 fundus camera were analyzed by an AI tool (Google ARDA (Automated Retinal Disease Assessment). Clinical grading was performed by a retina specialist using the International Clinical DR Severity Scale and considered the reference standard. Sensitivity, specificity, predictive values, and inter-grader agreement (κ statistics) were calculated.
RESULTS: The AI tool identified 69.8% of the eyes as DR, compared to 75.2% by the retina specialist, with an 83.3% accuracy rate, specificity of 90.9%, sensitivity of 97.1%, and Kappa = 0.77. For DME, AI classified 15.3% of eyes, compared to 5.9% by ophthalmologists, with an 89.9% diagnostic efficiency and Kappa = 0.48.
CONCLUSION: AI tools show high sensitivity and substantial agreement with ophthalmologists in diagnosing DR and DME, indicating their potential to enhance diagnostic accuracy and efficiency in retinal health screening.

Keywords:  Artificial intelligence; diabetic macular oedema; diabetic retinopathy; diabetic retinopathy screening

DOI:  https://doi.org/10.1080/09286586.2026.2641114
Diabetes Metab Syndr Obes. 2026 ;19 586810

Development and External Validation of a Machine Learning Model for Predicting Wound Infection in Diabetic Foot Ulcers.

Xingguo Nie, Yang Jiang, Xiangyan Meng, Ju Liu, Haijian Zhao, Yundong Chen, Junbo Wang, Tan Lu.

   Purpose: Wound infection is a major determinant of poor prognosis in patients with diabetic foot ulcers (DFUs). This study aimed to develop, compare, and externally validate multiple machine learning (ML) models for predicting wound infection in DFUs using routinely collected clinical indicators.
Methods: A total of 800 patients with DFUs were retrospectively enrolled. The primary cohort (n=500) was randomly divided into training (70%, n=350) and internal testing (30%, n=150) sets, while an independent cohort (n=300) was used for external validation. Eight ML algorithms were constructed and compared, including logistic regression, decision tree, random forest, support vector machine, k-nearest neighbor, naive Bayes, extreme gradient boosting, and light gradient boosting machine. Model performance was evaluated using area under the curve (AUC), accuracy, sensitivity, specificity, and other metrics in internal cross-validation and external validation. SHapley Additive exPlanations (SHAP) were applied for feature interpretability.
Results: The RF model demonstrated the best performance, with an AUC of 0.937 (95% CI 0.906 to 0.969) in training, 0.853 (95% CI 0.804 to 0.901) in internal testing, and 0.832 (95% CI 0.779 to 0.885) in external validation. Six key variables (age, duration of diabetes, ankle brachial index, ulcer area, vascular complications, and osteomyelitis) were identified as the most influential predictors. SHAP analysis provided interpretable insights into their contributions to infection risk.
Conclusion: The RF model showed robust predictive performance and generalizability for wound infection in DFUs. Its integration into clinical practice could enable early risk stratification and personalized interventions, potentially reducing amputations and improving outcomes. Future prospective studies are needed for further validation.

Keywords:  diabetic foot; external validation; machine learning; predictive modeling; ulcer infection

DOI:  https://doi.org/10.2147/DMSO.S586810
Acta Ophthalmol. 2026 Mar 09.

Training and validation of an automated algorithm to differentiate no and minimal diabetic retinopathy from more severe stages in wide-field images.

Lars Morten Skollerud, Jakob Kristian Andersen, Sebastian Dinesen, Thiusius Rajeeth Savarimuthu, Jakob Grauslund.

   PURPOSE: Diabetic retinopathy (DR) is a leading cause of blindness in the working-age population. Screening is essential to identify and treat sight-threatening stages prior to irreversible visual loss. This study aimed to train and validate an automated algorithm to identify no or minimal DR, potentially saving resources for specialist evaluation.
METHODS: We included 307 wide-field images (Optomap, Dunfermline, Scotland) classified by a certified retinal expert according to the International Clinical Diabetic retinopathy (ICDR) scale. The expert manually annotated 26 995 diabetic-related retinal lesions using the Computer Vision Annotation Tool (CVAT). A segmentation model was trained to detect lesions using a 70/15/15 split for training, tuning, and validation. A classification model used the outputs of the segmentation model as input features for a decision tree classifier, categorizing patients into Group 0 (DR level 0-1) or Group 1 (DR level 2-4). Classifier hyperparameters and input feature selection were optimized based on binary classification performance using cross-validation. The classifier was evaluated on 48 validation images and further validated with 200 images graded by up to four independent certified graders.
RESULTS: The area under the curve was 0.94 for the 48 validation images, with specificity, sensitivity, and kappa values of 0.89, 0.93, and 0.83, respectively. For the 200 expert-validated images, the values were 0.91, 0.98, 0.82, and 0.79, respectively.
CONCLUSION: The combined method of segmentation followed by feature count analysis shows promising results for binary DR classification in Optomap wide-field images without requiring a large dataset for model development.

Keywords:  Optomap; artificial intelligence; automated segmentation; retinopathy; wide‐field

DOI:  https://doi.org/10.1111/aos.70122
Sci Prog. 2026 Jan-Mar;109(1):109(1): 368504261424391

A study on the risk prediction of heart disease in diabetes patients based on machine learning.

Tinghua Zhang, Huan Liu.

  ObjectiveThis study aims to identify the most suitable machine-learning model for early heart disease risk screening in diabetic populations.MethodsThis retrospective cohort study utilized data from the China Health and Retirement Longitudinal Study, with baseline data from 2011 and follow-up data from 2020. Using features selected by Least Absolute Shrinkage and Selection Operator (LASSO) regression, we systematically constructed 16 distinct machine-learning models. Model performance was evaluated using a comprehensive set of metrics, including the area under the receiver operating characteristic curve, F1-score, sensitivity, specificity, precision, accuracy, and balanced accuracy. To interpret the decision-making process of the best-performing model, we conducted Shapley additive explanations (SHAP) analysis.ResultsAfter the 9-year follow-up period concluding in 2020, 157 of the 819 patients with diabetes at baseline (2011) developed heart disease. From the available features, LASSO regression selected 19 core features for model construction. Among the models developed, the K-Nearest Neighbors (KNN) model demonstrated optimal performance across key metrics, achieving the highest F1 score, balanced accuracy, and precision. The SHAP analysis identified body mass index, systolic blood pressure, and waist circumference as the three most important predictive features within the diabetic cohort. The contribution patterns of these features in the KNN model align closely with clinical expertise, achieving a strong balance between predictive power and interpretability.ConclusionThis study developed a machine-learning model to predict heart disease risk in patients with diabetes. Although the model exhibited only modest predictive performance, it provides a valuable empirical foundation and clear direction for constructing more reliable and clinically useful prediction tools in this field.

Keywords:  Diabetes; heart disease; machine learning; prediction; risk

DOI:  https://doi.org/10.1177/00368504261424391
World J Methodol. 2026 Mar 20. 16(1): 107488

Artificial intelligence in mobile health applications: A comprehensive review of its role in diabetes care.

Wen-Jie Li, Lin-Ze Li.

  This review explores the integration of artificial intelligence (AI) in mobile health applications for diabetes care. It focuses on key AI methodologies - machine learning, deep learning, and natural language processing - and their roles in glucose monitoring, personalized self-management, risk prediction, and clinical decision support. Drawing on recent literature (2018-2024), the study outlines the benefits of AI in improving accuracy, engagement, and precision in diabetes treatment. Challenges such as data privacy, algorithmic bias, and regulatory barriers are also examined. A new section discusses when AI technologies may become burdensome, especially in low-resource settings or for users with limited digital literacy. The review concludes with directions for enhancing model explainability and integrating AI with wearable and Internet of Things devices, emphasizing the need for ethical and equitable implementation in future diabetes management strategies.

Keywords:  Artificial intelligence; Clinical decision support; Diabetes management; Mobile health; Personalized self-management; Predictive analytics

DOI:  https://doi.org/10.5662/wjm.v16.i1.107488
Diabetes Obes Metab. 2026 Mar 08.

Artificial Intelligence in Screening and Grading Diabetic Eye Diseases: A Systematic Review From Algorithms to Clinic.

Jiamin Zhou, Chen Hu, Jiaqi Chen, Jinshen Liu, Xiang Li.

   SIGNIFICANCE: This systematic review comprehensively synthesises the progress of artificial intelligence in the grading diagnosis of diabetes-related ocular diseases, with a specific focus on the translational gaps from algorithm development to clinical implementation.
PURPOSE: This study aimed to systematically review and summarise the technological evolution, advantages of clinical application, and limitations of artificial intelligence in the grading diagnosis of diabetes-related ocular diseases (such as diabetic retinopathy and diabetic macular edema), clarify its clinical translation pathways and propose future research directions.
METHODS: A systematic literature review was conducted according to the PRISMA guidelines. Relevant English-language articles published between 2021 and 2025 were searched in databases such as the Web of Science using Boolean operators. A total of 74 core publications were included, including AI algorithm types, performance evaluations, clinical validations and translational research.
RESULTS: AI has formed a technical system primarily based on supervised learning with integrated algorithms for grading the diagnosis of diabetes-related eye diseases, demonstrating significant advantages over traditional manual diagnosis in screening efficiency, diagnostic consistency and healthcare accessibility. However, limitations remain in the identification of early-stage lesions, diagnosis in multidisease comorbidity scenarios and cross-device generalisation. Optimisation strategies include data augmentation using generative adversarial networks, multimodal fusion and the enhancement of interpretability. In clinical translation, screening models based on portable devices have emerged; however, challenges persist regarding data security and standardised validation.
CONCLUSIONS: Artificial intelligence holds significant clinical value and application potential in the graded diagnosis of diabetes-related ocular diseases, improving screening efficiency and consistency, particularly in resource-limited settings. Future efforts should focus on enhancing algorithmic adaptability in complex scenarios, promoting deeper integration of technology with clinical workflows, and establishing robust data security and validation standards to facilitate large-scale and high-quality clinical implementation.

Keywords:  artificial intelligence; clinical translation; diabetes‐related ocular diseases; grading diagnosis

DOI:  https://doi.org/10.1111/dom.70621
Diabetes Obes Metab. 2026 Mar 09.

Identifying Predictors of Early Treatment Intensification in Individuals With Type 2 Diabetes Treated With GLP-1 Receptor Agonists: A Machine Learning-Based Analysis of a Large Real World Cohort.

Pierpaolo Falcetta, Rita Zilich, Fabio Baccetti, Walter Baronti, Lelio Morviducci, Nicoletta Musacchio, Marco Muselli, Alessandro Ozzello, Federico Pisani, Antonio Rossi, Enrica Salomone, Damiano Verda, Riccardo Candido, Paola Ponzani.

   AIMS: Despite the proven efficacy of GLP-1 receptor agonists (GLP-1 RAs), many patients with type 2 diabetes (T2DM) are not able to achieve glycaemic targets with these agents and they require additional therapies. Timely identification of individuals at higher risk of early intensification may improve outcomes and reduce therapeutic inertia.
MATERIALS AND METHODS: In this retrospective cohort study, we analysed data from 69 194 individuals with T2DM initiating GLP-1 RA. We applied logic learning machine (LLM), an explainable machine-learning algorithm, to identify-at the time of GLP-1 RA prescription-predictors of early intensification (≤ 12 months from GLP-1 RA initiation). For this purpose, we developed two distinct models: Model 1, which compared characteristics of individuals intensified early (≤ 12 months) versus those not intensified in the first year, and Model 2, comparing individuals intensified early vs. those never intensified (> 5 years of GLP-1 RA treatment).
RESULTS: Both models identified the same clinical phenotype, characterised by longer diabetes duration, unstable glycaemic control, prior insulin use, and cardio-renal complications. Model 1 showed limited discriminative ability (AUC 0.65) with many false positives, suggesting therapeutic inertia in real-world practice. Model 2, while selecting same predictors of Model 1, achieved better performance (AUC 0.78), suggesting clearer differentiation between patient cohorts.
CONCLUSIONS: Explainable AI identified a reproducible phenotype of patients associated with early intensification after GLP-1 RA initiation. These models may support earlier, more personalised treatment decisions in routine diabetes care.

Keywords:  diabetes mellitus type 2; glucagon‐like peptide‐1 receptor agonists; insulin; machine learning; therapeutic inertia; therapy intensification

DOI:  https://doi.org/10.1111/dom.70648
Front Digit Health. 2026 ;8 1685842

From raw data to actionable insights: preprocessing real-world data for machine learning in diabetes care.

Marco Montagna, Aleksandar Svilenov Rabadzhiev, Alberto Traverso, Emanuela Setola, Edoardo Draetta, Alessio Dimonte, Simone Barbieri, Bruno Fabiani, Lorenzo Piemonti, Antonio Esposito, Carlo Tacchetti, Patrizia Rovere Querini.

   Background: Machine Learning (ML) applied to healthcare Real World Data (RWD) may improve patient management. RWD, however, requires extensive preprocessing to make it ML-ready. Our aim was to explore the impact of preprocessing on ML models applied to RWD from 20 years of type 2 diabetes patients visits.
Methods: Our cohort consisted of patients with at least two glycated hemoglobin (HbA1c) measurements three years apart. We set up three different experimental settings consisting of different data preprocessing pipelines. Logistic Regression (LR), XGBoost and a Decision Tree Classifier (DTC) were then applied and tuned to optimize precision.
Results: The final dataset comprised 12 variables from 1,651 patients treated between 2003 and 2023. 921 (56%) patients had a HbA1c decrease at three years. This group had a higher baseline HbA1c, higher BMI and shorter first visit gap from the date of diagnosis (p < 0.0001). Precision scores for LR, XGBoost did not vary across different experimental conditions while DTC benefitted from missing data imputation. Shapley Additive Explanations confirmed the Exploratory Data Analysis findings, with worse baseline values being predictors of HbA1c decrease at three years.
Conclusions: ML models' performance and their explanation did not vary substantially across experimental conditions, with worse baseline values being predictors of HbA1c decrease at three years. Insights such as this, extracted by ML application to RWD, enable clinical discussion and may foster improvements in patient management.

Keywords:  explainability; glycated hemoglobin; machine learning; predictive modeling; preprocessing pipeline; real world data; type 2 diabetes mellitus

DOI:  https://doi.org/10.3389/fdgth.2026.1685842
J Diabetes Sci Technol. 2026 Mar 10. 19322968261422628

Evaluating Large Language Models-Generated Health Education Materials for Discharged Patients with Diabetes: A Comparative Analysis.

Rongping Zha, Jingbang Liu, Li Wang, Shan Li, Lei Mei, Xiaoyan Gong, Yijing Weng, Xiaofang Jiang, Xuehua He.

   BACKGROUND: Diabetes is a chronic condition requiring long-term management, and continuous health education is vital for improving disease awareness and self-management. Large language models (LLMs), advanced artificial intelligence systems trained on large text data sets, have shown promise in generating diabetes-related educational materials. While LLMs can generate accurate and readable content, most studies focus on general education based on guidelines, rather than tailoring content to individual patients' clinical profiles. This study addresses these gaps by comparing the performance of three major LLMs (ChatGPT-4o, Doubao 1.5, and DeepSeek R1) in generating health education materials for discharged patients with diabetes.
METHODS: Ten de-identified medical records of discharged patients with diabetes were uploaded to the LLMs. Each model generated health education materials based on these records. Experienced diabetes nursing experts evaluated the quality of the generated materials.
RESULTS: The comprehensibility scores pass rates for all models were above 70%, with DeepSeek R1 performing the best (P < .01). The actionability scores pass rates were below 70% for all models, with no significant differences (P > .01). Accuracy scores for all models were ≥98%, and there were no significant differences in accuracy (P > .01). Similarly, no significant differences were observed in personalization and effectiveness scores (P > .01). DeepSeek R1 achieved the highest safety score, while Doubao 1.5 had the lowest safety score (P < .01).
CONCLUSION: While ChatGPT-4o, Doubao 1.5, and DeepSeek R1 generate accurate and comprehensible materials, concerns remain regarding their actionability and safety. These findings suggest that LLMs should be used as auxiliary tools in diabetes education, requiring further refinement for personalized and actionable content.

Keywords:  ChatGPT; DeepSeek; artificial intelligence; health education material; large language models; patients with diabetes

DOI:  https://doi.org/10.1177/19322968261422628
Diabetes Technol Ther. 2026 Mar 12. 15209156261432144

Glucose Predictions Improve Glycemic Control: A Digital Twin Evaluation.

Pau Herrero, Magí Andorrà, Marc D Breton, Ajandek Peak, Matthias Koehler, Yannick Klopfenstein, Eemeli Leppäaho, Mattia Zanon, Christian Ringemann, Patrick Lustenberger, Timor Glatzer.

   BACKGROUND: Glucose predictions aim to empower continuous glucose monitoring (CGM) users by enabling preventive actions to reduce adverse glycemic events. The Accu-Chek® SmartGuide Predict app offers several AI-enabled predictive features, driven by machine learning algorithms. These include notifications for a low glucose predict within 30 min (LGP) and for nighttime low glucose risk, as well as a 2-h continuous glucose forecast.
AIMS: This study aimed to quantify the potential glycemic benefits of using the Predict app's predictive features in an adult population with type 1 diabetes (T1D).
METHODS: A comparative in silico study was conducted using the clinically backed University of Virginia Replay digital twin simulator. A control arm, simulating standard hypoglycemia and hyperglycemia mitigation strategies in line with international guidelines, was compared against intervention arms that incorporated probabilistic user behavior models responding to the app's predictive features. The evaluation was performed on 204 digital twins, representing 29,929 days of data, generated from the REPLACE-BG clinical trial dataset.
RESULTS: Results demonstrated that using the app's predictive features has the potential to improve glycemic control in adults with T1D. The simulated intervention led to an average 2.9 percentage point reduction in time below range (<70 mg/dL), and a clinically significant increase of more than 3.6 percentage points in time in range (70-180 mg/dL). Furthermore, the daily number of CGM hypoglycemia alarms (<70 mg/dL) was reduced by 67%. The findings also suggest that consuming 10 g of fast-acting carbohydrates in response to LGP notifications provides an optimal balance, effectively preventing hypoglycemia while limiting rebound hyperglycemia.
CONCLUSIONS: This in silico evaluation provides strong evidence supporting the potential clinical utility of the Accu-Chek SmartGuide Predict app for improving glycemic management in adults with T1D.

Keywords:  artificial intelligence; continuous glucose monitoring; digital twin; glucose prediction; mHealth; machine learning; type 1 diabetes

DOI:  https://doi.org/10.1177/15209156261432144
JMIR AI. 2026 Mar 09. 5 e86960

AI-Enabled Personalization of Semaglutide Therapy in Type 2 Diabetes: Systematic Review With an Integration Framework.

Ghinwa Barakat, Samer El Hajj Hassan, Hanane Akhdar, Nghia Duong-Trung, Wiam Ramadan.

   BACKGROUND: Type 2 diabetes mellitus (T2D) is a rapidly growing global health concern requiring innovative treatment methods. Ozempic (semaglutide), a glucagon-like peptide-1 receptor agonist, has proven consistent effectiveness in lowering blood glucose levels, supporting weight loss, and minimizing cardiovascular complications. In parallel, artificial intelligence (AI) elevates diabetes care yet complements these efforts by converting raw data from wearable devices, electronic health records, and medical imaging into practical insights for efficient, tailored, and customized treatment plans.
OBJECTIVE: The objective of this systematic review is to examine current evidence of AI-driven methods to optimize Ozempic-based T2D therapy.
METHODS: A total of 18 peer-reviewed articles were identified, revealing four dominant thematic clusters: (1) patient stratification and risk prediction, (2) AI-enhanced imaging for body composition changes, (3) cardiovascular and metabolic risk assessment, and (4) personalized AI-driven dosage.
RESULTS: Across multiple metrics, such as glycated hemoglobin reduction, weight loss, cardiovascular benefits, and adverse event mitigation, AI-based approaches outperformed standard fixed-dose regimens. A theoretical framework is proposed for AI-Ozempic integration, with continuous data collection, AI processing, clinical decision support, real-time support, and real-time feedback and modeling iteration refinement cycles.
CONCLUSIONS: Significant gaps remain a persistent challenge, including the need for large-scale randomized controlled trials, longer follow-up periods, explainable AI models, regulatory validation, and practical strategies for routine clinical implementation. The findings emphasize the AI's potential to transform semaglutide therapy while delineating important paths for future research.

Keywords:  Ozempic; artificial intelligence; clinical decision support; personalized medicine; semaglutide; type 2 diabetes mellitus

DOI:  https://doi.org/10.2196/86960
Front Digit Health. 2026 ;8 1660815

Machine learning and engagement insights for personalized blood glucose management.

Inbar Breuer Asher, David L Horwitz, Omar Manejwala, Yifat Fundoiano-Hershcovitz.

   Introduction: Diabetes is a chronic metabolic disorder characterized by elevated blood glucose (BG) levels, with poor control linked to serious long-term complications. Managing BG effectively requires personalized strategies, given the influence of demographic, lifestyle, and clinical factors. Machine learning (ML) offers a powerful framework for analyzing complex, real-world data to uncover individual patterns of glycemic control. Coupled with digital health platforms that enable real-time monitoring and behavioral engagement, these tools can transform diabetes care. This study leverages data from the Dario digital health platform to examine BG trends moderated by clinical and engagement variables, aiming to inform personalized digital interventions.
Objective: To apply ML techniques to digital health data to identify moderating factors that influence individual BG trajectories, supporting data-driven and personalized diabetes management.
Methods: A retrospective cohort study was conducted using real-world data from users with type 2 diabetes and baseline BG ≥180 mg/dL who measured BG over at least two separate months between 2020 and 2024. A piecewise linear mixed-effects model characterized BG changes over time. Generalized Linear Mixed Effects Tree models identified subgroups with distinct BG trajectories based on demographic (age, gender, BMI, ethnicity), clinical (insulin use, comorbidities, diagnosis year), and monitoring factors. An additional model tested whether lifestyle engagement (e.g., meal and activity logging) moderated BG improvement across age groups.
Results: Data from 22,414 users (49.9% male; mean age 57.5; BMI 34.5) showed significant reductions in monthly average BG over 12 months (B = -6.8 in months 1-4; B = -0.3 in months 4-12; both p < .001). Age strongly moderated outcomes; users >60 showed the largest sustained improvements. Clinical factors such as insulin use and diagnosis duration further stratified responses, with insulin users diagnosed within five years showing the greatest reduction within the first 4 months (B = -14.3, p < .001). Higher frequency of BG monitoring (>12/month) was associated with greater and sustained improvements.
Conclusion: Machine learning can reveal distinct glycemic trajectories moderated by demographic, clinical, and engagement factors. These findings underscore the potential of digital health platforms to personalize diabetes care and improve blood glucose management through adaptive, data-driven strategies.

Keywords:  chronic condition; diabetes; digital health; engagement; machine learning (ML)

DOI:  https://doi.org/10.3389/fdgth.2026.1660815
Diagnostics (Basel). 2026 Mar 04. pii: 773. [Epub ahead of print]16(5):

Identification of Retinal Diseases Using Light Convolutional Neural Networks and Intrinsic Mode Function Technique.

Preethi Kulkarni, Konda Srinivasa Reddy.

  Background/Objectives: Fundus imaging provides a detailed view of the interior surface of the eye and plays a crucial role in the early diagnosis of retinal diseases. However, automated interpretation of fundus images remains challenging due to variations in illumination, noise, and structural complexity. Methods: A novel hybrid model that integrates the Intrinsic Mode Function (IMF) filter, derived from Empirical Mode Decomposition (EMD), with a Light Convolutional Neural Network (LightCNN) for enhanced fundus image classification was proposed. The IMF filter effectively decomposes the input signal into intrinsic components, isolating high-frequency noise and preserving critical retinal patterns. These refined components are subsequently processed by the LightCNN architecture, which offers lightweight yet highly discriminative feature extraction and classification capabilities. Results: Experimental results on DIARETDB fundus datasets demonstrate that the proposed IMF + LightCNN model achieves 99.4% accuracy, 99.1% precision, 98.87% recall, and a 98.31 F1-score, significantly outperforming conventional CNN and ResNet-based models. Conclusions: Integrating advanced signal processing with lightweight deep learning improves both diagnostic accuracy and computational efficiency. This hybrid framework establishes a promising pathway for reliable and real-time clinical screening of retinal diseases.

Keywords:  Diabetic Retinopathy; Intrinsic Mode Function; Light Convolutional Neural Network; fundus images

DOI:  https://doi.org/10.3390/diagnostics16050773
Front Endocrinol (Lausanne). 2026 ;17 1667159

Comparative assessment of large language models in diabetic foot infection management: alignment with IWGDF/IDSA guidelines.

Hongxia Wu, Jiayi Deng, Xu Qiu, Li Xu, Lumeng Lu, Mingna Fan, Danni Yu, Chuanbo Liu, Zhaohuan Chen, Kai Wang, Yuyan Wang, Haifang Zhou, Liyang Chang, Hanbin Wang.

   Objective: To assess the clinical utility of artificial intelligence (AI) models (ChatGPT-4o, DeepSeek-R1, Grok-3 and Claude-3.7) in aligning with international guidelines for diabetic foot infection (DFI) management.
Background: AI systems have demonstrated their potential application value in numerous fields. However, the specific effects of these technologies in the medical and health sector still require in-depth exploration. DFI is a relatively common and serious complication among diabetic patients, and the accurate transmission of relevant information is of great significance. Therefore, it is particularly important to evaluate whether artificial intelligence can serve as an effective clinical auxiliary tool.
Methods: Responses from ChatGPT-4o, DeepSeek-R1, Grok-3 and Claude-3.7 were evaluated against DFI guidelines using four clinical dimensions (Accuracy, Overconclusiveness, Supplementary Value, and Completeness) using a 5-point Likert scale, and assessed for readability using Flesch Reading Ease (FRE) and Flesch-Kincaid Grade Level (FKGL). Statistical analyses included ANOVA and post hoc comparisons.
Results: No significant differences were found across models for Accuracy and Overconclusiveness (p > 0.05). However, Supplementary Value differed significantly (p < 0.001), the performance of Grok-3 is superior to that of ChatGPT-4o (p < 0.0001), DeepSeek-R1 (p=0.003), and Claude-3.7 (p < 0.0001). Meanwhile, there are significant differences in terms of Completeness (p=0.005), Grok-3 outperforms ChatGPT-4o (p=0.016)and Claude-3.7 (p=0.010) significantly.Readability also varied: DeepSeek-R1 responses were more complex than ChatGPT-4o (p = 0.046).
Conclusion: All models perform comparably in terms of accuracy and in avoiding over-conclusions. Grok-3 outperformed the other models in the dimensions of complementarity and completeness. DeepSeek-R1 generated the most complex text. These findings validate the feasibility of AI in the standardized management of DFI, but the models still need to be further verified through clinical trials to determine their value in the real-world decision-making process.

Keywords:  adherence; artificial intelligence; diabetic foot infection; guideline; large language models

DOI:  https://doi.org/10.3389/fendo.2026.1667159