bims-aukdir 2026-05-31 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2026–05–31
twenty-one papers selected by
Mott Given

A Machine Learning Framework to Quantify Postprandial Glucose Responses in Gestational Diabetes.
Enhancing Early Prediction of Gestational Diabetes Mellitus Through Data Augmentation and Feature Guidance: Model Development and Validation Study.
Non-destructive classification of type 2 diabetic teeth using LIBS-derived elemental biomarkers and machine learning.
A Robust Intelligent CNN Model Enhanced with Gabor-Based Feature Extraction, SMOTE Balancing, and Adam Optimization for Multi-Grade Diabetic Retinopathy Classification.
Retraction: Enhancing diabetic retinopathy classification using deep learning.
Artificial intelligence in diabetic macular edema: Expanding clinical vision.
Using artificial intelligence-assisted retinal imaging devices operated by non-specialist workers in detecting diabetic retinopathy in Sri Lanka.
Plasma protein GDF15 has a good predictive potential for the kidney complications of type 2 diabetes.
Assessing the generalisability of foundation models to ultra-wide field retinal imaging for diabetic retinopathy screening in Denmark and Greenland.
AI driven Hybrid CNN transformer model for early detection and severity assessment of diabetic retinopathy.
Identification and preliminary clinical validation of type 2 diabetes signature genes through machine learning analysis of scRNA-seq data.
Retinal graph neural network for segmentation of retinal vasculature and foveal avascular zone from OCTA images.
Machine Learning Models for Early Prediction of Type 1 Diabetes: A Systematic Review.
DAGs and GRaSP Causal Inference Algorithms Combined and Applied to the Calculation of Insulin Bolus in Patients with Type 1 Diabetes.
A U-Net-based multimodal deep learning model for high-precision blood glucose prediction using non-invasive physiological data.
Leveraging Machine Learning to Assess Post-COVID-19 Glycemic Control in Diabetic Patients.
Comparative Analysis of General-Purpose vs. Domain-Specific Multimodal Models for Diabetic Retinopathy Classification.
AI-driven personalised management platform enhances glycaemic control and remission rates in type 2 diabetes: a quasi-experimental prospective cohort study.
Identification and validation of lactylation-related diagnostic biomarkers for type 2 diabetes by WGCNA.
Unveiling the role of air pollution in diabetic kidney disease: an integrated study combining network toxicology, machine learning, and Mendelian randomization.
Artificial Intelligence Driven Adjusted OCT-Based Classification System for Diabetic Macular Edema (AIDME).

Diabetes Technol Obes Med. 2025 Jan-Dec;1(1):1(1):

A Machine Learning Framework to Quantify Postprandial Glucose Responses in Gestational Diabetes.

Souptik Barua, Tenzin Sangmo, Dhairya Upadhyay, Esra Mucahit, Vanessa Martinez, Lauren Berube, Christina Penfield, Hye Heo, Ling-Jun Li, Shauna Williams, Todd Rosen, Shristi Rawal.

   Objective: To develop a machine learning (ML) framework to identify postprandial glucose responses (PPGR) automatically from continuous glucose monitoring (CGM) data in pregnant adults with gestational diabetes mellitus (GDM).
Methods: Pregnant adults diagnosed with GDM or impaired glucose tolerance (IGT) wore blinded CGMs and logged mealtimes for up to three 14-day time periods after enrollment. A random forest ML algorithm was applied to identify morning PPGRs from daily CGM profiles, and its performance compared against PPGRs derived using self-reported mealtimes.
Results: 21 participants provided analyzable data. Relative to self-reported mealtime, the ML algorithm's predicted mealtimes had an absolute error of a median 30 [IQR: 20,45] minutes. Comparing 1-hour and 2-hour PPGR values from the CGM using self-reported and ML-predicted mealtimes showed a median difference of 8.7 [IQR: 0,22.7] mg/dL and 3.3 [IQR: 0,13.2] mg/dL respectively for the two timepoints.
Conclusion: A random forest ML algorithm accurately identified PPGRs from CGM data in persons with GDM, enabling an automated and convenient approach to monitoring postprandial dysglycemia in this population.

Keywords:  Gestational diabetes mellitus; continuous glucose monitoring; machine learning; postprandial glucose response

DOI:  https://doi.org/10.1089/dtom.2024.0003
JMIR Med Inform. 2026 May 25. 14 e85335

Enhancing Early Prediction of Gestational Diabetes Mellitus Through Data Augmentation and Feature Guidance: Model Development and Validation Study.

Xiekun Chen, Zhifa Jiang, Dong Su, Xiaoping Chen, Aiping Chen, Zhen Zhang, Huabin Wang.

   Background: Early prediction of gestational diabetes mellitus (GDM) is critical for improving maternal health outcomes. However, predictive models are often challenged by limited early-pregnancy samples, severe class imbalance in datasets, and complex interrelationships among clinical features.
Objective: This study aimed to develop and evaluate a unified dual-dimensional enhancement framework integrating data augmentation and feature engineering. By addressing data imbalance and leveraging medical prior knowledge, this framework significantly improves early GDM prediction performance.
Methods: We proposed a framework combining Generative Adversarial Network (GAN)-based data augmentation with large language model-inspired feature engineering. GAN sampling was used to generate clinically plausible synthetic minority class samples to mitigate data imbalance. The large language model was guided to organize features into domains (eg, basic demographics, metabolic syndrome, and core liver biomarkers) and generate higher-order composite features, integrating medical prior knowledge. Machine learning models were subsequently developed, and interpretability analyses were performed using Shapley additive explanations to identify key predictors.
Results: This study used a final analytical cohort of 8214 pregnant women, divided into dataset A comprising 966 out of 5251 (18.4%) participants with GDM, and dataset B comprising 598 out of 2963 (20.2%) participants with GDM. The random forest model enhanced by Tabular Variational Autoencoder-based feature augmentation demonstrated the best performance. On the test dataset, it achieved a recall of 0.7559, an accuracy of 0.8444, and an area under the receiver operating characteristic curve (AUROC) of 0.8873. Statistical evaluation confirmed that the Tabular Variational Autoencoder method significantly outperformed the baseline (Cohen d=2.894; P<.001) and the Conditional Tabular Generative Adversarial Network method (Cohen d=1.637; P=.02) in recall enhancement. Shapley additive explanations analysis identified the following 5 features as the most influential predictors: fasting blood glucose, the composite feature (fasting blood glucose+triglycerides)×prepregnancy BMI, activated partial thromboplastin time, leukocyte count, and neutrophil count.
Conclusions: The proposed dual-dimensional enhancement framework effectively alleviates data limitations and captures complex feature interactions in early GDM prediction. This strategy not only improves model performance, particularly in recall, but also provides interpretable biological evidence to support rapid clinical screening, stratified management, and early intervention in pregnancy.

Keywords:  GDM; SHAP analysis; data augmentation; feature enhancement; gestational diabetes mellitus; machine learning

DOI:  https://doi.org/10.2196/85335
Anal Chim Acta. 2026 Aug 08. pii: S0003-2670(26)00565-9. [Epub ahead of print]1410 345615

Non-destructive classification of type 2 diabetic teeth using LIBS-derived elemental biomarkers and machine learning.

Muhammad Mustafa Dastageer, Khurram Siraj, Giorgio Saverio Senesi, Muhammad Aamir Israr, S U Haq.

   BACKGROUND: Type 2 diabetes mellitus (T2DM) is a chronic metabolic illness that severely alters oral health, elevating periodontal infection incidence and dental treatment failures via mechanisms including hyperglycemia. While laser-induced breakdown spectroscopy (LIBS) using machine learning (ML) has demonstrated the presence of diabetes in hair, nails, and urine, mineralized tooth tissues that preserve long-term metabolic signatures remain underexplored as a diagnostic substrate.
RESULTS: A total of 3600 LIBS spectra were collected from four dental tissues (enamel, coronal dentine, radicular dentine, cementum) obtained from 30 individuals (15 healthy and 15 diabetic). LIBS elemental analysis showed higher Fe and Sn in diabetic teeth, along with lower Zn, Si, and K, indicating they are diabetes-related biomarkers. Among the various ML models tested (LR, SVM, ANN) using PCA and correlation-based feature selection (CFS-BFS), the PCA-ANN achieved the greatest performance: mean accuracy 96%, mean sensitivity 94%, and mean specificity 96%. However, PCA-SVM (accuracy 95%) with a narrow confidence interval (94.98-95.02) %, makes it the most stable model in terms of accuracy. No single model was consistently the most stable across all performance evaluation metrics. PCA with 6 Pcs captured 94% of the variance, consistently outperforming CFS-BFS integrated ML algorithms.
SIGNIFICANCE: This work confirms that dental tissues preserve diabetes-induced elemental alterations detectable by LIBS-ML, thus enabling a non-destructive screening method for undiagnosed diabetes in dental practice. By providing pre-treatment risk identification, the LIBS technique can potentially reduce dental treatment failures and improve patient outcomes. Large-scale multicenter validation studies are, however, required to convert these results into therapeutic practice.

Keywords:  Diabetes diagnosis; Human teeth analysis; Laser induced breakdown spectroscopy (LIBS); Machine learning models; Trace elements

DOI:  https://doi.org/10.1016/j.aca.2026.345615
J Imaging. 2026 Apr 27. pii: 188. [Epub ahead of print]12(5):

A Robust Intelligent CNN Model Enhanced with Gabor-Based Feature Extraction, SMOTE Balancing, and Adam Optimization for Multi-Grade Diabetic Retinopathy Classification.

Asri Mulyani, Muljono, Purwanto, Moch Arief Soeleman.

  Diabetic retinopathy (DR) is a leading cause of vision impairment and permanent blindness worldwide, requiring accurate and automated systems for multi-grade severity classification. However, standard Convolutional Neural Networks (CNNs) often struggle to capture fine, high-frequency microvascular patterns critical for diagnosis. This study proposes a Robust Intelligent CNN Model (RICNN) that integrates Gabor-based feature extraction with deep learning to improve DR classification. Specifically, Gabor filters are applied during preprocessing to extract orientation- and frequency-sensitive texture features, which are transformed into feature maps and concatenated with CNN feature representations at the fully connected layer (feature-level fusion). The model also incorporates the Synthetic Minority Oversampling Technique (SMOTE) for data balancing and the Adam optimizer for efficient convergence. This integration enhances sensitivity to microvascular structures such as microaneurysms and hemorrhages. The proposed RICNN was evaluated on the Messidor dataset (1200 images) across four severity levels: Mild, Moderate, Severe, and Proliferative DR. The model achieved an accuracy of 89%, a precision of 88.75%, a recall of 89%, and an F1-score of 89%, with AUCs of 97% for Severe DR and 99% for Proliferative DR. Comparative analysis confirms that the proposed texture-aware Gabor enhancement significantly outperforms LBP and Color Histogram approaches, indicating its potential for reliable clinical decision support.

Keywords:  Gabor filter; classification; deep learning; diabetic retinopathy; feature extraction; intelligent decision support; medical image

DOI:  https://doi.org/10.3390/jimaging12050188
Digit Health. 2026 Jan-Dec;12:12 20552076261446299

Retraction: Enhancing diabetic retinopathy classification using deep learning.

[This retracts the article DOI: 10.1177/20552076231203676.][This retracts the article DOI: 10.3390/diagnostics13142375.][This retracts the article DOI: 10.1177/20552076231194942.].

DOI: https://doi.org/10.1177/20552076261446299
Indian J Ophthalmol. 2026 Jun 01. 74(6): 929-930

Artificial intelligence in diabetic macular edema: Expanding clinical vision.

Aditya Barigali.

DOI: https://doi.org/10.4103/IJO.IJO_464_26
BMC Glob Public Health. 2026 May 27. pii: 51. [Epub ahead of print]4(1):

Using artificial intelligence-assisted retinal imaging devices operated by non-specialist workers in detecting diabetic retinopathy in Sri Lanka.

Nilmini Wijemunige, Moditha Madhuwanthi Dissanayake, Tanya Vimarshani Gamage, Hasitha Rajakaruna, Pavithra Maniyangamage, Niroshani Wisidagama, Malinda De Silva, Manjula D Nugawela, Mapa Prabhath Piyasena, Sobha Sivaprasad, Sarath Samarage, Shanti Dalpatadu, Ravindra Prasan Rannan-Eliya.

   BACKGROUND: Diabetic retinopathy (DR), a microvascular complication of diabetes, is an important cause of preventable blindness and can cause a significant reduction in the quality of life of working-age adults. Sri Lanka has one of the highest prevalences of diabetes globally, and like many low- and middle-income countries, faces significant barriers to DR screening. Easy-to-use, relatively cheap handheld retinal fundus imaging devices with automatic interpretation using artificial intelligence (AI) presents a new avenue for DR screening by non-specialist healthcare workers. This study piloted the use of one such device by non-specialist healthcare workers in a low-resource setting and aimed to evaluate image gradability and the diagnostic test accuracy (DTA) of the AI interpretation in this context.
METHODS: Adults with and without diabetes were recruited from a research clinic. Non-specialist healthcare workers received brief training to capture macular- and optic-disc-centred retinal images using an AI-assisted handheld retinal imaging camera (Remidio FOP NM-10), before and after mydriasis. The device generated reports identifying referrable DR. Participants also had visual acuity assessed and underwent slit lamp biomicroscopy by an ophthalmologist. We evaluated image gradability of the captured images pre- and post-dilation, as well as the DTA of the AI reports using the ophthalmologist's findings as the gold-standard.
RESULTS: Retinal images were captured on both eyes of 119 participants, including 49 with diabetes. The proportion of participants that had ungradable images of at least one eye prior to mydriasis was 46.2%, and this fell to 15.1% after mydriasis. A higher percentage of older individuals (59% of those aged 55-64 years and 70% of those aged 65-74 years), 85% of those with cataracts and 58% of those with poor visual acuity had at least one eye with ungradable images pre-mydriasis, with these percentages reducing by 40-70% with mydriasis. The sensitivity and specificity for referral of patients compared to the gold-standard was 73.3% and 79.8% respectively, with image gradability a major driver of misclassification.
CONCLUSIONS: The AI-assisted retinal imaging camera used by non-specialist healthcare workers after minimal training, in a low-resource setting shows promise as a feasible tool to identify DR in dilated eyes.

Keywords:  AI; Diabetes; Diabetic retinopathy; Fundoscopy; Screening; Sri Lanka

DOI:  https://doi.org/10.1186/s44263-026-00279-6
Front Endocrinol (Lausanne). 2026 ;17 1758267

Plasma protein GDF15 has a good predictive potential for the kidney complications of type 2 diabetes.

Ming Hao, Houxing Li, Mengyu Xin, Jiatong Li, Rui Sun, Qian Liu, Yujie Zhang, Xinxin Shan, Yuting He, Borui Xu, Qiuyan Guo, Hongyu Kuang, Peng Wang.

   Introduction and aims: Complications of type 2 diabetes are a primary cause of public health challenges in the field of diabetes. The emergence of metabolomics and proteomics provides a direct perspective for revealing the mechanisms of metabolic diseases. Our research aims to explore the relationship between omics components and complications, as well as their clinical predictive performance.
Materials and methods: This prospective study utilized data from the UK Biobank, including over 1,400 proteins and more than 280 metabolites, to analyze outcomes such as type 2 diabetes, microvascular complications, macrovascular complications, neurological complications, kidney complications, retinal complications, cardiovascular complications, peripheral vascular complications, metabolic disorder complications, and all-cause mortality. A total of 50,021 participants without type 2 diabetes were included in the analysis. The baseline time frame spanned from 2006 to 2010, with an average follow-up duration of 12.0 to 12.03 years. Researchers used LASSO Cox and LightGBM to search for new markers of complications, and employed SHAP methods to explain the contributions of these markers within the machine learning models. Subsequently, a comprehensive prediction model was established to reveal the potential of new markers for the early diagnosis of complications under nonlinear patterns, utilizing nine specific machine learning methods (CatBoost, LightGBM, Random Forest, XGBoost, logistic regression, multi-layer perceptron, single-layer neural network, Naive Bayes, and support vector machine).
Results: GDF15 alone is more accurate than blood glucose and HbA1c in reflecting future kidney complications, especially in differentiating those who develop the disease within the next five years (GDF15 AUC=0.94, blood glucose AUC=0.68, HbA1c AUC=0.85). Within the framework of the comprehensive prediction model, the GDF15 model improved the accuracy of early screening for kidney complications compared with models constructed using traditional indicators (5-year Max AUC=0.92, 10-year Max AUC=0.88). In conclusion, both machine learning and statistical methods support the correlation between GDF15 and kidney complications, reflecting its robustness.
Conclusions: The results highlight the association of GDF15 during the early asymptomatic stage of various complications, especially kidney complications, revealing the potential role of GDF15 at the molecular pathological level during disease progression. In distinguishing participants who developed complications after the baseline period, the comprehensive GDF15 model provides a method for the early warning of various complications, particularly kidney complications.

Keywords:  complications; machine learning; metabolomics; proteomics; type 2 diabetes

DOI:  https://doi.org/10.3389/fendo.2026.1758267
Int J Med Inform. 2026 May 21. pii: S1386-5056(26)00243-1. [Epub ahead of print]217 106503

Assessing the generalisability of foundation models to ultra-wide field retinal imaging for diabetic retinopathy screening in Denmark and Greenland.

Livie Yumeng Li, Vajira Thambawita, Stine Byberg, Adam Hulman.

   BACKGROUND: Foundation models have shown promising performance in ophthalmology image analysis, but their ability to generalize to unseen imaging types and populations remains unknown. We evaluated the generalizability of ophthalmology foundation models to ultra-wide field (UWF) retinal images for diabetic retinopathy (DR) screening in a Danish and a Greenlandic population.
METHODS: Three ophthalmology foundation models (RETFound DINOv2, VisionFM, and EyeCLIP) were fine-tuned and evaluated using 6,374 UWF retinal images from 1,760 participants in Denmark and 6,558 images from 1,146 participants in Greenland. Binary DR classification (normal vs. any retinopathy) was performed under four experimental settings: fine-tuning on the Danish dataset, fine-tuning on the Greenlandic dataset, external validation of Danish-fine-tuned models on the Greenlandic dataset, and sequential fine-tuning from the Danish to the Greenlandic dataset. Model discrimination and calibration were assessed.
RESULTS: DR prevalence differs between the Danish and Greenlandic datasets, with 45% and 14% of all images having DR, respectively. When fine-tuned and evaluated within the same population, discrimination was similar in Denmark and Greenland, with RETFound DINOv2 achieving the highest AUROC (0.76 [95% CI: 0.73, 0.78] and 0.76 [0.73, 0.80], respectively). External validation on the Greenlandic dataset showed worse performance across models (AUROC 0.59-0.62). Sequential fine-tuning improved discrimination (AUROC 0.70-0.78). However, calibration remained poor across all settings, with calibration intercepts ranging from -1.69 to 0.37 and slopes from 0.25 to 0.78.
CONCLUSION: Foundation models showed limited generalizability when applied to unseen imaging contexts and populations, with disparities in model performance observed across Danish and Greenlandic populations. Local fine-tuning improved discrimination, but did not resolve calibration issues, underscoring the importance of careful calibration evaluation to ensure clinical relevance.

Keywords:  Clinical prediction; Deep learning; Diabetic retinopathy; Foundation models; Ophthalmology; Ultra-wide field retina image

DOI:  https://doi.org/10.1016/j.ijmedinf.2026.106503
Sci Rep. 2026 May 28.

AI driven Hybrid CNN transformer model for early detection and severity assessment of diabetic retinopathy.

Priyadharshini Sekar, Kanaga Suba Raja S.

  As the major cause of sight impairment in working age adults is diabetic retinopathy (DR) which requires timely and proper diagnosis to avoid the irreversible vision loss. Traditional retinal fundus image based manual diagnosis is time consuming, subjective, and requires highly skilled ophthalmologists, and therefore is not scalable in resource constrained environments. Automated DR grading systems are essential to overcome this issue especially in underserved areas. Although Convolutional Neural Networks (CNNs) are suitable in capturing the fine grained local features namely Microaneurysms and exudates, they are incapable of considering the global retinal context, which is required to accurately grade the severity. Transformer models, on the other hand, are more effective in capturing long range dependencies, at the cost of low spatial resolution. To address these shortcomings, this research suggests Fusion Net a hybrid DL model that combines a CNN branch that extracts local features in the form of GoogLeNet with a Vision Transformer (ViT) branch that captures global context. This is a dual stream architecture that combines complementary embeddings to categorize DR into five ordinal levels of severity. The model has been trained on a mixed dataset of APTOS 2019 and Messidor-2 and tested on three datasets that are not used during training. Fusion Net had overall accuracy percent of 98.85, AUC-ROC of 0.981 and weighted F1-score percent of 97.62. It was also highly sensitive in detecting early stage DR and balanced in all the severity levels. The thorough ablation experiments, ordinal assessment scales, and feature map confirms the strength and clarity of the suggested framework. The findings indicate that Fusion Net is an effective, explainable, and computationally efficient system to use in grading automated DR with a high chance of implementation into clinical practice and tele-ophthalmology systems.

Keywords:  Convolutional Neural Network; Deep Learning; Diabetic Retinopathy; Explainable AI; Fundus Imaging; Hybrid Model; Severity Grading; Vision Transformer

DOI:  https://doi.org/10.1038/s41598-026-50608-w
Front Med (Lausanne). 2026 ;13 1801737

Identification and preliminary clinical validation of type 2 diabetes signature genes through machine learning analysis of scRNA-seq data.

Fang Tang, Xin Zuo, Weiyan Wang, Rui Sun, Heng Cheng, Weihua Wu.

   Objective: Type 2 diabetes (T2DM) is a highly prevalent metabolic disorder with substantial molecular heterogeneity, and traditional bulk transcriptomic approaches often fail to capture cell-specific changes critical to disease pathogenesis. This study aims to identify and validate key signature genes for T2DM by integrating single-cell RNA sequencing (scRNA-seq) with machine learning, providing new insights into disease mechanisms and potential biomarkers.
Methods: We analyzed scRNA-seq data to characterize cellular heterogeneity across 10 distinct cell types. Differential expression analysis identified 455 candidate genes, which were refined using LASSO regression. The diagnostic potential of identified genes was evaluated using ROC curve analysis on an independent dataset. Functional enrichment and cell communication analyses were performed to elucidate biological processes and intercellular signaling networks. Finally, expression changes of the candidate genes were validated in peripheral blood from a separate clinical cohort (15 T2DM patients, 20 controls) using qRT-PCR.
Results: Four core genes (PNLIP, BUB1, CTSB, NAMPT) were identified as candidate signature genes. ROC analysis showed AUC values of 0.819, 0.931, 0.882, and 0.694, respectively, suggesting promising but variable diagnostic accuracy. Enrichment analyses indicated these genes participate in processes including extracellular matrix remodeling, digestion/absorption, and signal transduction. Cell communication analysis suggested a potential central role of Alpha and Beta cells in diabetic signaling networks, with the MK and SPP1 pathways showing complementary expression patterns. In addition, qRT-PCR confirmed significantly up-regulated expression of PNLIP, BUB1, and CTSB along with down-regulated NAMPT in T2DM patients, supporting their potential as circulating candidate biomarkers.
Conclusion: This study integrates machine learning with scRNA-seq to identify PNLIP, BUB1, CTSB, and NAMPT as potential T2DM signature genes. These findings offer candidate diagnostic biomarkers and provide preliminary mechanistic insights into disease-associated pathways.

Keywords:  Signature genes; clinical validation; machine learning; single-cell RNA sequencing; type 2 diabetes

DOI:  https://doi.org/10.3389/fmed.2026.1801737
Biomed Tech (Berl). 2026 May 27.

Retinal graph neural network for segmentation of retinal vasculature and foveal avascular zone from OCTA images.

Nisan Pranavah Raja, Anju Thomas, Bibin Francis, Varun P Gopi, Palanisamy Ponnusamy.

   OBJECTIVES: Diabetic Retinopathy (DR) causes major vision loss, requiring precise segmentation of retinal vessels and the Foveal Avascular Zone (FAZ). Accurate structural masks enable quantitative biomarkers that support early diagnosis and long-term monitoring.
METHODS: We propose a Retinal Graph Neural Network (RGNNNet) for OCTA segmentation. It combines multi-scale feature extraction with a graph representation, where node relations derive from an affinity matrix of feature maps. A symmetric normalization strategy stabilizes graph propagation and integrates local-global vascular context. A hybrid Dice-Focal loss refines fine-structure segmentation.
RESULTS: On OCTA-500, RGNNNet achieved superior Dice and IoU to existing methods. For FAZ, it attained Dice values of 96.78 % (6 mm) and 98.02 % (3 mm), and maintained 0.915 on ROSE-0 without retraining. It outperformed baselines by 1-3 % Dice for other classes and remained lightweight (0.83 M params, 11.25 ms per 400 × 400 image).
CONCLUSIONS: By coupling residual feature learning with graph-based relational reasoning, RGNNNet provides accurate structure-specific masks that can serve as a foundation for downstream biomarker extraction. Its compact design and stable generalization highlight its potential for large-scale ophthalmic screening and integration into clinical workflows.

Keywords:  biomedical image processing; deep learning; diabetic retinopathy; graph neural networks; image segmentation

DOI:  https://doi.org/10.1515/bmt-2025-0312
Cureus. 2026 Apr;18(4): e107547

Machine Learning Models for Early Prediction of Type 1 Diabetes: A Systematic Review.

Rayan Saad Aldeen Mohammed Saad Aldeen, Nazar Mohamed, Shima Abdoelmonem A Mohamad, Ali Hadi M Alhajri, Wigdan Adam Obeid Hagar, Hussam Altaj Abdalkarim Khalifa, Samah Abdelhadi Eisa Mohamed, Mohamed Hosny Saadeldien Mohamed Nagla.

  Type 1 diabetes (T1D) is a chronic autoimmune condition with a rising global incidence. Early prediction of disease onset and detection of preclinical progression are critical for timely intervention. Machine learning (ML) offers the ability to analyze complex, high-dimensional data and may improve risk prediction across different stages of T1D development. This systematic review evaluates the application and performance of ML models for predicting T1D onset and early disease-related outcomes. Following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, a structured search was conducted in PubMed, British Medical Journals, Scopus, IEEE Xplore, and Web of Science for studies published between 2021 and 2025. Eligible studies included those that developed or validated ML models for T1D prediction or early detection. Study selection, data extraction, and risk of bias assessment (using Prediction model Risk of Bias Assessment Tool (PROBAST)) were performed, and findings were synthesized narratively due to heterogeneity in study design, populations, prediction targets, and outcome measures. Fourteen studies were included, with sample sizes ranging from 32 to over 800,000 participants. ML approaches included logistic regression, random forests, support vector machines, and gradient boosting methods. Reported performance varied (area under the receiver operating characteristic curve (AUROC) 0.73-0.92), with prediction horizons spanning short-term outcomes (minutes to hours) to long-term disease onset (up to 10 years). However, study heterogeneity was substantial, and only three studies performed external validation. While most studies were rated as low risk of bias, several high-performing models were based on small samples or limited validation, raising concerns about overfitting and generalizability. ML models demonstrate potential for improving prediction of T1D onset and early disease-related outcomes, but current evidence is limited by variability in methods, inconsistent validation, and uncertain clinical applicability. Future research should prioritize large, prospective, and externally validated studies, with greater emphasis on model transparency, generalizability, and real-world implementation.

Keywords:  early prediction; machine learning; predictive modeling; systematic review; type 1 diabetes

DOI:  https://doi.org/10.7759/cureus.107547
Entropy (Basel). 2026 May 01. pii: 506. [Epub ahead of print]28(5):

DAGs and GRaSP Causal Inference Algorithms Combined and Applied to the Calculation of Insulin Bolus in Patients with Type 1 Diabetes.

Rocío Contreras-Jiménez, Juan Carlos Olivares-Rojas, Adriana Del Carmen Téllez-Anguiano, Jesús Eduardo Alcaráz-Chávez, José Antonio Gutiérrez-Gnecchi, Enrique Reyes-Archundia.

  Type 1 diabetes mellitus (T1DM) is a chronic, non-preventable, and incurable disease that requires lifelong insulin administration. The principal challenge is calculating the prandial insulin bolus to avoid hypoglycemia and hyperglycemia. Traditional bolus calculators are based on limited number of variables, but there are many variables that define the complex interactions among glucose levels, like carbohydrate intake, physical activity, mood, and contextual factors. While recent artificial intelligence (AI) approaches have shown promise in glucose prediction, most remain correlational and offer limited interpretability for clinical decision support. This study evaluates a causal inference-based framework for insulin bolus calculation using Directed Acyclic Graphs (DAGs) and the Greedy Relaxation of the Sparsest Permutation (GRaSP). Historical data from individuals with T1DM were analyzed, incorporating domain knowledge constraints to guide structure learning. A bootstrap-based stability analysis was conducted to evaluate the robustness of inferred relationships. Results show that integrating prior medical knowledge reduces graph complexity and improves interpretability. However, bootstrap stability reflects robustness of the learning procedure rather than causal validity. The findings suggest that the proposed framework is useful for generating plausible causal hypotheses, but not for confirming causal relationships. Further validation using conditional independence testing, equivalence class analysis, and temporal causal methods is required. However, the proposed framework focuses on generating plausible causal hypotheses rather than establishing causal validity, which requires further refutation-based validation.

Keywords:  DAGs; GRaSP; artificial intelligence; causal inference; health technology; insulin bolus calculation; type 1 diabetes

DOI:  https://doi.org/10.3390/e28050506
Comput Methods Biomech Biomed Engin. 2026 May 25. 1-11

A U-Net-based multimodal deep learning model for high-precision blood glucose prediction using non-invasive physiological data.

Ruting Wang, Li-Ang Gao, Yuhao Xu, Qunfeng Tang, Cheng Fang.

  This study proposes a novel U-Net-based convolutional neural network for blood glucose prediction by integrating historical continuous glucose monitoring data with non-invasive physiological parameters. The model achieved robust performance using the OhioT1DM dataset-demonstrating superiority with MAEs of 8.4761 mg/dL (30 min) and 14.0170 mg/dL (60 min), RMSEs of 13.1315 and 20.6470 mg/dL, and R2 values of 0.9397 and 0.8549. This framework offers a practical, data-driven strategy for short-term glucose forecasting and metabolic health management in insulin-dependent T1DM patients, while the quantitative validation of its non-invasiveness provides a robust basis for wearable sensor-based glucose monitoring.

Keywords:  Blood glucose prediction; U-Net architecture; deep learning; multimodal physiological signals

DOI:  https://doi.org/10.1080/10255842.2026.2676676
Int J Environ Res Public Health. 2026 May 12. pii: 644. [Epub ahead of print]23(5):

Leveraging Machine Learning to Assess Post-COVID-19 Glycemic Control in Diabetic Patients.

N3C Consortium

  Hemoglobin A1c is a central biomarker for long-term glycemic control and a key predictor of diabetes-related complications. The COVID-19 pandemic disrupted routine healthcare delivery and introduced potential metabolic effects of SARS-CoV-2 infection, yet the long-term impact of COVID-19 on glycemic trajectories in individuals with diabetes remains unclear. In this retrospective study, we leveraged harmonized electronic health record data from the National Clinical Cohort Collaborative to evaluate changes in HbA1c before and after documented SARS-CoV-2 infection in adults with diabetes (n = 93,320). Patients were required to have repeated HbA1c measurements pre- and post-infection and stable exposure to key antihyperglycemic medications. A paired statistical analysis was used to identify individuals with statistically significant post-infection changes in HbA1c. We then developed and evaluated multiple supervised machine learning classifiers using an 80/20 train-test split and cross-validation to assess demographic, clinical, and structural factors associated with significant glycemic change. Most patients (71%) did not experience a statistically significant change in average HbA1c following COVID-19 infection, and among those who did, decreases were more common than increases. A random forest classifier achieved the best overall performance, and feature importance and SHAP analyses highlighted body mass index, insulin use, age, and socioeconomic proxies as key contributors. These findings suggest that while COVID-19 infection does not substantially alter long-term glycemic control for most patients with diabetes, individual-level clinical and structural factors influence post-infection glycemic variability.

Keywords:  COVID-19; HbA1c; diabetes; disease diagnosis; electronic health records; epidemic prediction; machine learning

DOI:  https://doi.org/10.3390/ijerph23050644
Diagnostics (Basel). 2026 May 15. pii: 1504. [Epub ahead of print]16(10):

Comparative Analysis of General-Purpose vs. Domain-Specific Multimodal Models for Diabetic Retinopathy Classification.

Mohammad Iqbal Nouyed, Mohammad Al-Mamun, Donald A Adjeroh, Gangqing Hu.

  Background/Objectives: General-purpose and domain-specific multimodal foundation models show considerable promise in medical image analysis. In this study, we evaluated the classification accuracy of diabetic retinopathy vs. normal fundus images using general-purpose conversational models (Gemini 3 Flash, GPT-5.2, and Pixtral-Large), a medical conversational model (MedGemma-1.5), and its image-encoder (MedSigLIP), as well as ophthalmology-specific models (RETFound and EyeCLIP). Methods: We applied zero-/few-shot to general-purpose conversational models, linear probing, and fine-tuning approaches to domain-specific models for evaluation purposes. Results: We found that the zero-shot accuracies for Pixtral-Large (70.7%) and fine-tuned RETFound (77.1%) were comparable but lower than those of GPT-5.2 (77.9%), MedGemma-1.5 (88.2%), and Gemini 3 (88.5%) as well as the fine-tuned EyeCLIP (85.8%) and MedSigLIP (94.8%). The accuracy gains from few-shot prompting were substantial for Pixtral-Large (+7.4%) but were limited for GPT-5.2 (+3.6%), Gemini 3 (-3.4%), and MedGemma-1.5 (-1.1%). Embedding-based linear probing further improved accuracy over fine-tuning for RETFound (+9.7%) and yielded only marginal gains for EyeCLIP (+2.3%) but did not benefit MedSigLIP (-0.8%). Overall, with minimal prompting enhancement, general-purpose conversational models such as Gemini 3 and GPT-5.2 achieved performance comparable to ophthalmology-specific models that were either fine-tuned or enhanced via embedding-based linear probing, but remained inferior to MedSigLIP and its conversational counterpart, MedGemma-1.5. Conclusions: The findings highlight a trade-off between specialization and flexibility, where domain-specific models provide higher accuracy and stability, while general-purpose multimodal models offer greater accessibility, adaptability, and interactive reasoning, serving as complementary tools for retinal disease screening and clinical decision support.

Keywords:  diabetic retinopathy; fundus images; image classification; large multimodal models

DOI:  https://doi.org/10.3390/diagnostics16101504
Front Digit Health. 2026 ;8 1778918

AI-driven personalised management platform enhances glycaemic control and remission rates in type 2 diabetes: a quasi-experimental prospective cohort study.

Reziwanguli Amuti, Areziguli Kadier, Mengting Jiang, Abudureheman Maihemuti, Bingquan Yang.

   Background and aim: Although artificial intelligence (AI) technologies show promise in diabetes management, real-world evidence supporting AI-driven personalized care remains limited. We developed an AI-powered platform that generates dynamic self-management plans (DSMPs) and enables continuous, whole-process patient monitoring. This study aimed to evaluate its effectiveness in patients with type 2 diabetes mellitus (T2DM).
Design: This was a 12-month, multicenter, prospective, open-label, quasi-experimental cohort study. A total of 1,452 eligible patients with T2DM were enrolled between April 2022 and April 2023 and assigned by an AI-driven platform, based on enrollment sequence, to either the intervention group (n = 726) or the control group (n = 726). All participants received integrated diabetes management via the AI-driven platform. The intervention group was provided with DSMPs, whereas the control group received routine care.
Findings: Among the 1,343 patients who completed the study (655 in the intervention group and 688 in the control group), the intervention group demonstrated significantly better glycemic control at 3 months. Reductions in glycated hemoglobin (HbA1c) (1.92% vs. 1.31%) and fasting plasma glucose (FPG) (3.23 vs. 2.31 mmol/L) were significantly greater in the intervention group, and significantly more patients achieved treatment targets for both HbA1c (57% vs. 26%) and FPG (68% vs. 39%). Furthermore, the intervention led to significant improvements (all p < 0.05) in Self-Monitoring of Blood Glucose (SMBG) frequency, hypoglycemia incidence, Body Mass Index (BMI), Urine Albumin-to-Creatinine Ratio (UACR), and most blood lipids compared to the control group and baseline, though Total Cholesterol (TC) remained unchanged (p = 0.53). Notably, the co-primary outcome was achieved by 18.2% (119/655) of patients in the intervention group, which was significantly higher than the 5.5% (38/688) observed in the control group (p < 0.001).
Conclusions: AI-driven personalized management was feasible and effective for the management of T2DM and was associated with an increased rate of diabetes remission. However, further evidence is required to justify its integration into routine clinical practice, and improvements in cholesterol management remain necessary.

Keywords:  AI-driven management platform; DSME; diabetes management; personalized management care; type 2 diabetes

DOI:  https://doi.org/10.3389/fdgth.2026.1778918
J Clin Biochem Nutr. 2026 May 01. 78(3): 238-250

Identification and validation of lactylation-related diagnostic biomarkers for type 2 diabetes by WGCNA.

Najun Zhu, Suping Gu, Ying Shen, Lvxia Zhou, Weiping Tu.

  Lactylation, a novel post-translational histone modification, has emerged as a critical regulatory mechanism in various metabolic disorders. However, its role in the pathogenesis of type 2 diabetes (T2D) remains poorly understood. This study aims to investigate the potential of lactylation-related genes as diagnostic biomarkers for T2D. Differential analysis and weighted gene co-expression network analysis (WGCNA) were performed on the GSE164416 dataset. Genes obtained from these analyses were intersected with the lactylation-related genes to screen candidate genes. The LASSO, SVM-RFE and random forest algorithms were applied to screen the characteristic genes, and their diagnostic efficacy was verified in the independent cohort. The functions and immune associations were analyzed by GSVA, ssGSEA, and TF-miRNA regulatory network analysis, and qRT-PCR, Western blot and CCK-8 experiments were conducted in the T2D cell model for verification. Lactylation-related IKZF1, S100A4, and VIM were identified as potential diagnostic markers for T2D. These three genes were significantly upregulated in T2D samples and exhibited excellent diagnostic performance (AUC >0.80) in both the training set and validation set. The GSVA analysis revealed that these three genes were involved in key biological processes such as immune regulation, transcriptional modification, metabolic homeostasis and cytoskeleton remodeling. Cell experiments demonstrated that the three genes were upregulated in T2D cell models and knockdown of their expression could promote cell viability. This study identified and validated three potential diagnostic markers related to lactylation for T2D, providing new molecular evidence for the early diagnosis and mechanism research of this disease.

Keywords:  WGCNA; biomarkers; lactylation; machine learning; type 2 diabetes

DOI:  https://doi.org/10.3164/jcbn.25-245
Ren Fail. 2026 Dec;48(1): 2668270

Unveiling the role of air pollution in diabetic kidney disease: an integrated study combining network toxicology, machine learning, and Mendelian randomization.

Yi Kang, Qian Jin, Size Li, Yaping Lv, Xuezhe Wang, Xinyan Jin, Mengqi Zhou, Danwen Li, Jie Lv, Huijuan Zheng, Yaoxian Wang.

  Diabetic kidney disease (DKD) is a major and severe complication associated with diabetes. Air pollution is not only an independent risk factor for metabolic disorders but also an "accelerator" of DKD progression. This study seeks to investigate the molecular pathways connecting air pollution to DKD. Multiple databases were integrated to obtain potential target genes of 10 common air pollutants. The gene expression omnibus (GEO) database was employed to acquire DKD datasets. Differential expression analysis and weighted correlation network analysis (WGCNA) were performed to identify DKD-related genes. 12 machine learning algorithms were utilized to generate 113 unique predictive models, which were employed to select hub genes. Subsequently, MR, single-cell, gene set enrichment analysis (GSEA), and immune infiltration analyses were performed, followed by molecular docking of hub genes with air pollutants. A total of 714 targets were identified from 10 air pollutants, and 80 potential targets were identified from DKD transcriptomic data. Machine learning methods identified 5 hub genes that are closely associated with DKD. Mendelian randomization (MR) analysis indicated that, among the five hub genes, only NOS3 demonstrated a statistically significant causal association with DKD. Immune infiltration analysis found that hub genes were closely related to immune cells. Molecular docking validation indicated that certain air pollutants can stably bind with hub genes such as NOS3 and PTGS2. Air pollutants may be linked to alterations in various biological processes, potentially involving key genes such as ADH5, CASP3, NOS3, PTGS2, and SDHB, including potential metabolic reprogramming, inflammatory processes, and immune microenvironment changes.

Keywords:  Air pollution; Mendelian randomization; diabetic kidney disease; machine learning; molecular docking; network toxicology

DOI:  https://doi.org/10.1080/0886022X.2026.2668270
Clin Ophthalmol. 2026 ;20 604743

Artificial Intelligence Driven Adjusted OCT-Based Classification System for Diabetic Macular Edema (AIDME).

Rodrigo Abreu-González, Gonzalo Quezada-Peralta, Patricia Udaondo, María Antonia Gil-Hernández, Juan Donate-López.

   Objective: To develop and validate an adjusted scoring-based classification system for Diabetic Macular Edema (DME) using quantitative and qualitative parameters. The system aims to improve clinical decision making and correlate disease severity with visual outcomes.
Methods: A cohort of patients with DME was analyzed using optical coherence tomography (OCT) data. Key quantitative variables included retinal thickness (RT), intraretinal fluid (IRF), and subretinal fluid (SRF), which were scored according to severity thresholds (0-2 points per parameter). Qualitative biomarkers, including disorganization of the retinal inner layers (DRIL) and epiretinal membrane (ERM), were incorporated as severity modifiers (+2 for DRIL and +1 for ERM). The patients were classified based on the total score as follows: Mild (0-3 points), Moderate (4-6 points), and severe (>7 points). Visual acuity (logMAR) was analyzed to validate the system and to assess its correlation with disease severity. Statistical analyses included descriptive summaries and ANOVA for intergroup comparisons.
Results: A total of 71 patients were classified into three groups: mild (48 patients, mean logMAR 0.45), moderate (18 patients, mean logMAR 0.56), and severe (5 patients, mean logMAR 0.92). The system showed significant differences in logMAR values (p = 0.014). ROC curve analysis showed an area under the curve (AUC) of 0.89. This adjusted scoring method aligned well with the clinical expectations, emphasizing the impact of fluid accumulation and structural biomarkers on vision impairment.
Conclusion: The adjusted scoring-based classification system for DME demonstrated a robust correlation between the disease severity and visual acuity. Integrating quantitative measures and qualitative markers offers a practical and clinically relevant approach for stratifying DME severity. This method has the potential to enhance decision making in routine practice and research applications.

Keywords:  artificial intelligence; diabetic macular edema; diagnostic/tests investigation; imaging; optical coherence tomography; vision

DOI:  https://doi.org/10.2147/OPTH.S604743