bims-aukdir 2026-08-02 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2026–08–02
eighteen papers selected by
Mott Given

Transfer Learning and Optimized Machine Learning Techniques for Multiclass Diabetic Retinopathy Classification Using Retinal Images.
Artificial intelligence for early detection of diabetic retinopathy: A vision transformer-based approach.
A Roadmap for National Diabetic Retinopathy Screening in Croatia: Integrating European Evidence, Telemedicine, and AI.
Screening for diabetic retinopathy with artificial intelligence in a primary care setting: a comparative cost analysis.
Joint Multi-Task Deep Learning with Cross-Task Attention for Simultaneous Lesion Segmentation, Detection, and Severity Grading of Diabetic Retinopathy on Fundus Images: An External Validation Study.
Toward Reliable Diabetic Retinopathy Screening.
Diagnostic accuracy and clinical performance of deep learning models for grading diabetic retinopathy: a systematic review and meta-analysis.
VesselMetaKAN: vessel-guided meta-learned interpretable classification for diabetic retinopathy grading.
Content-based retrieval of fundus images and diabetic retinopathy detection using variants of local texture features.
Artificial intelligence-based analysis of retinal vascular changes in the preclinical and early stages of diabetic retinopathy using ultra-widefield fundus imaging: an observational cross-sectional study.
A large sensor foundation model pretrained on continuous glucose monitor data for diabetes management.
Intelligent Staging Performance of Diabetic Retinopathy Based on Fundus Fluorescein Angiography Images with Different Angiographic Phases.
Identification of COX7B, GLRX5, and OR52H1 as Vascular Endothelial Injury-associated Biomarkers in Gestational Diabetes Mellitus.
INSIGHT: Ingest and Standardize Glucose Harmonization Tool.
Single-cell transcriptomic integrated with machine learning reveals human retinal cell-specific biomarkers in diabetic retinopathy.
Active Learning Under Expert-Budget Constraints: A Human-in-the-Loop Pipeline for Diabetic Retinopathy Lesion Detection.
Machine learning algorithms for predicting glycemic control and weight loss outcomes in GLP-1 receptor agonist users.
Human-in-the-loop AI predictive digital twin to extend virtual precision diabetes care between visits.

Diagnostics (Basel). 2026 Jul 14. pii: 2189. [Epub ahead of print]16(14):

Transfer Learning and Optimized Machine Learning Techniques for Multiclass Diabetic Retinopathy Classification Using Retinal Images.

Mohammad Reza Yousefi, Ali Bakrani, Elias Ebrahimzadeh, Amin Dehghani.

  Background/Objectives: Diabetic Retinopathy (DR) is a prevalent and severe complication of diabetes, caused by prolonged hyperglycemia that damages retinal microvasculature and may ultimately lead to vision loss or blindness. While convolutional neural networks (CNNs) have shown promise in automating DR detection via retinal imaging, traditional approaches often suffer from limited diagnostic accuracy, long training times, and reliance on small or imbalanced datasets. Objective: This study evaluates an integrated transfer-learning using adaptive training strategies for multiclass retinal image classification. Methods: The proposed framework integrates transfer learning, feature-space dimensionality reduction, and adaptive training strategies based on an ImageNet pretrained ResNet50 backbone to improve training stability, computational efficiency, and multiclass retinal image classification performance. Results: The proposed Transfer Learning (TL)-based model was trained and evaluated on a large, publicly available dataset of retinal images, achieving an overall accuracy of 84%, maximum class-specific accuracy of 89%, sensitivity of up to 97%, and an F1-score of 92%. These results demonstrate reasonable overall classification performance under constrained data conditions. Conclusions: The proposed framework demonstrates the feasibility of integrating transfer learning and adaptive training strategies for multiclass retinal image classification under constrained benchmark conditions. However, the study is limited by the use of heavily downsampled retinal images, and further validation on high-resolution clinical datasets is required before practical deployment. Future methodological refinement and validation on high-resolution clinical datasets may support development of computer-assisted retinal image analysis systems.

Keywords:  convolutional neural networks (CNNs); deep learning; diabetic retinopathy (DR); machine learning (ML); medical image analysis; retinal imaging; transfer learning (TL)

DOI:  https://doi.org/10.3390/diagnostics16142189
PLoS One. 2026 ;21(7): e0350854

Artificial intelligence for early detection of diabetic retinopathy: A vision transformer-based approach.

Asma ElAdel, Imen Filali, Mourad Zaied.

BACKGROUND: Early identification of diabetic retinopathy (DR), which is a primary cause of vision impairment globally, is a crucial phasis for effective intervention and treatment. Traditional screening workflows rely on manual diagnosis by ophthalmologists, which remains the gold standard but can be time-consuming and subject to variability due to human factors. To support and enhance the screening process, artificial intelligence (AI)-based tools have shown promise in automating DR detection, particularly with recent advances in deep learning. However, medical images with long-range dependencies and spatial linkages can be challenging for CNN-based algorithms to handle.
METHODS: This paper proposes a Vision Transformer (ViT)-based model, specifically using a Compact Convolutional Transformer (CCT), for early automated detection of DR. The model uses self-attention techniques to improve feature extraction and classification performance; combining three main stages: the CCT tokenizer, transformer encoder, and sequence pooling. The proposed approach was trained on public datasets (EyePACS and APTOS 2019) and evaluated against state-of-the-art deep learning architectures.
RESULTS: Our experimental findings demonstrate that ViT performs among the best in the current state of the art with an overall accuracy of 97% and F1-scores above 0.95 across all DR severity levels. Our system is primarily designed for the pre-screening stage of diabetic retinopathy workflows, enabling rapid and reliable identification of potential DR cases for further clinical evaluation.
CONCLUSION: These results highlight the potential of transformer-based designs in medical picture analysis, as well as the implications for telemedicine and e-health solutions in real-time, especially in cases of low-resource settings.

DOI: https://doi.org/10.1371/journal.pone.0350854
Medicina (Kaunas). 2026 Jun 29. pii: 1251. [Epub ahead of print]62(7):

A Roadmap for National Diabetic Retinopathy Screening in Croatia: Integrating European Evidence, Telemedicine, and AI.

Toma Babić, Martina Tomić, Nenad Vukojević, Ivo Dumić-Čule, Sonja Jandroković, Tea Čaljkušić Mance.

  Diabetic retinopathy (DR) is a leading cause of preventable blindness among working-age adults. Systematic screening programmes in the United Kingdom, Ireland, and the Nordic countries have reduced diabetes-related visual loss, yet many European countries, including Croatia, lack organised screening. This narrative review examines European DR screening programmes, evaluates telemedicine and artificial intelligence (AI) as enabling technologies, and proposes a phased roadmap for a national programme in Croatia. Croatia has nearly 400,000 registered persons with diabetes, a national diabetes registry (CroDiab) linked to the Central Health Information System (CEZIH), and pilot screening data showing 40% DR prevalence among screened patients with type 2 diabetes. The proposed programme combines decentralised telemedicine-based imaging at primary care sites with centralised grading, a staged rollout across three phases targeting 400,000 annual screenings, and stepwise AI integration for triage. Successful European programmes share standardised digital imaging, trained grading workforces, embedded quality assurance, and registry linkage. With dedicated funding and sustained political commitment, Croatia could adopt a hybrid telemedicine-AI model and, under favourable implementation conditions, reduce the burden of advanced DR over the coming decade.

Keywords:  artificial intelligence; diabetic retinopathy; screening; telemedicine

DOI:  https://doi.org/10.3390/medicina62071251
BMJ Open Ophthalmol. 2026 Jul 30. pii: e002904. [Epub ahead of print]11(3):

Screening for diabetic retinopathy with artificial intelligence in a primary care setting: a comparative cost analysis.

Livio Di Matteo, Noelle Whitestone, Vishaal Bhambhwani.

   OBJECTIVE: Cost analysis of autonomous artificial intelligence (AI)-based screening of diabetic retinopathy (DR) for adults with diabetes at a primary care clinic.
METHODS AND ANALYSIS: This study provides a comparative cost analysis of actual results using AI-based DR screening with counterfactual results based on all patients going through the physician-based referral system. A cost analysis is conducted using cost data from published sources, provincial billing codes, statistical sources, and patient characteristics from a clinical study to compare autonomous AI-based screening for DR versus physician-based screening. Costs considered include direct costs of operating the AI system, physician fees, and indirect costs to patient time. Along with total cost comparisons, a cost per DR case detected is estimated and a sensitivity analysis based on variations in AI costs is provided.
RESULTS: Over the study period, 202 participants were screened for DR using autonomous AI. The majority (93.6%, n=189) of AI-based DR screening exams were completed successfully. The AI-based scenario results in total direct costs of $C7919.04 and indirect costs of $C5728.80, resulting in total costs of $C13 647.84 per 100 patients. The traditional physician-based approach results in total direct costs of $C8240 and indirect costs of $C19 998.09, resulting in total costs of $C28 238.09 for 100 patients. When costs are converted to costs per unit outcome, the total cost per diagnosed DR case is $C620.36 for the AI-based approach and $C1283.55 for the physician-based approach; the AI-based cost per diagnosed case was 52% lower.
CONCLUSION: Given the lower cost per diagnosed case of the AI-based approach, there are advantages to the implementation of AI-based screening for DR.

Keywords:  Imaging; Retina

DOI:  https://doi.org/10.1136/bmjophth-2026-002904
Photodiagnosis Photodyn Ther. 2026 Jul 31. pii: S1572-1000(26)00267-X. [Epub ahead of print] 105600

Joint Multi-Task Deep Learning with Cross-Task Attention for Simultaneous Lesion Segmentation, Detection, and Severity Grading of Diabetic Retinopathy on Fundus Images: An External Validation Study.

Shabnam Jafarpoor Nesheli, Saleh Rouhi, Alireza Motamedi, Maryam Rashidi, Fatemeh Salamat, Sara Mardani, Salar Bijari, Masoumeh Kheirkhahzadeh, Seyed Masoud Rezaeijo.

   BACKGROUND: Diabetic retinopathy (DR) severity is graded from the type and distribution of retinal lesions, yet most automated methods address grading or lesion analysis in isolation. This study develops one framework that performs lesion segmentation, lesion detection, and DR grading together, evaluated within and across datasets.
METHODS: A hierarchical multi-task network with a shared ResNet-50 encoder and a feature pyramid feeds three heads, in which a cross-task attention module conditions the grading representation on the predicted lesion features. The three losses are combined by homoscedastic uncertainty weighting under a progressive three-stage schedule that fits the partial-label structure of the data. The model was developed on DDR (13,673 images, 757 with lesion annotation) using stratified five-fold cross-validation over three seeds, and evaluated without tuning on the independent IDRiD dataset. Results are reported for the training, validation, and test splits, with 95 percent bootstrap confidence intervals, paired significance tests, calibration, and an ablation.
RESULTS: On the DDR test split, the model reached a mean lesion-segmentation Dice of 0.535, a detection means average precision at intersection over union (IoU) 0.5 of 0.289, a grading accuracy of 0.823, and a quadratic weighted kappa of 0.858, with a referable-DR sensitivity of 0.928 and a vision-threatening-DR sensitivity of 0.871. On the external IDRiD set, segmentation Dice was 0.489 and grading accuracy and kappa were 0.781 and 0.821, quantifying the reduction under domain shift; the external segmentation estimates in particular rest on only 81 images and are indicative rather than definitive. In the ablation, removing the cross-task attention produced the largest single-component reduction in grading performance, while collapsing the framework to single-task training lowered it further.
CONCLUSIONS: Jointly learning lesion localization and grading, with external validation and calibration, produces a transparent framework whose grading is accompanied by lesion-level evidence. However, external performance declines moderately (Dice: 0.535 → 0.489), and the small external segmentation sample (n = 81) limits generalizability claims. Prospective validation on larger multi-ethnic cohorts is needed before clinical use.

Keywords:  Diabetic retinopathy; external validation; fundus imaging; image grading; lesion segmentation; multi-task learning

DOI:  https://doi.org/10.1016/j.pdpdt.2026.105600
Sensors (Basel). 2026 Jul 16. pii: 4515. [Epub ahead of print]26(14):

Toward Reliable Diabetic Retinopathy Screening.

Hendrio Bragança, Ítalo P Caliari, Wington L Vital, Antonio Fontenele, Sergio Cavalcante, Glaucio Messias.

  Diabetic retinopathy (DR) grading requires reliable five-grade severity assessment under substantial acquisition variability and cross-dataset distribution shift. We propose PRISM-DR, a multi-objective five-grade DR grading framework trained under a gradient-partitioned strategy. The architecture is organized as a feedforward pipeline: a data-driven preprocessing stage followed by a ConvNeXtV2-Base backbone, a Recurrent BiFPN neck for multi-scale feature fusion, a Frequency-Aware Fusion module, a lightweight multi-scale reasoning transformer, dual classification heads with gradient-isolated pathways (categorical and ordinal), and a prototype memory module for embedding regularization. The CORAL ordinal head operates through a dedicated projection layer and is gradient-isolated from the backbone; the backbone is shaped by the cross-entropy, prototype contrastive, and view-consistency objectives, which carry indirect ordinal signal through severity-weighted class penalties and grade-indexed cluster regularization. The model is trained in a multi-crop setting with a phased loss curriculum designed for severely imbalanced DR datasets. Evaluated across six datasets under Fixed-Source, Multi-Target (FSMT) protocols, PRISM-DR trained on EyePACS + DDR achieves QWK of 0.835 on IDRiD, 0.865 on APTOS2019, and 0.720 on Messidor-2, with in-domain QWK = 0.920 and AUC-PR = 0.941 on EyePACS, outperforming RETFound, RETFound-Green, and MedGemma-4B in AUC-PR across all evaluated datasets. Quantitative interpretability evaluation against 755 expert-annotated lesion images yields 8.0× Energy Ratio Enrichment and a FAF gate retention ratio of 4.4× inside lesion regions, confirming that anatomically plausible spatial priors emerge from grade-level supervision alone, without pixel-level annotation. PRISM-DR establishes a superior accuracy-robustness-capacity trade-off for scalable, automated DR screening.

Keywords:  diabetic retinopathy grading; explainable artificial intelligence; frequency-aware fusion; multi-scale feature fusion; ordinal learning

DOI:  https://doi.org/10.3390/s26144515
Front Endocrinol (Lausanne). 2026 ;17 1853785

Diagnostic accuracy and clinical performance of deep learning models for grading diabetic retinopathy: a systematic review and meta-analysis.

Xin Yan, Shiqi Lei, Lifen Hu, Mu Qin, Na Wu.

   Background: Diabetic retinopathy (DR) is a leading cause of preventable visual impairment worldwide, and its precise severity grading is critical for optimizing clinical management. Conventional frameworks, notably the International Clinical Diabetic Retinopathy (ICDR) scale, are often hindered by substantial inter-observer variability and high dependency on specialist expertise. While deep learning (DL) has recently emerged as a transformative approach for automated stratification, a comprehensive synthesis of evidence regarding its diagnostic performance and clinical application remains lacking.
Objective: This systematic review and meta-analysis aimed to comprehensively assess the diagnostic accuracy of fundus image-based deep learning models in the grading of diabetic retinopathy.
Methods: PubMed, Embase, Web of Science, and the Cochrane Library were systematically searched for relevant studies published up to October 28, 2025. Diagnostic accuracy studies utilizing DL algorithms alongside ICDR criteria for diabetic retinopathy grading were included. Literature screening and data extraction were performed independently by two researchers, and the risk of bias was assessed using the QUADAS-2 tool.
Results: A total of 41 studies were included, encompassing various DL architectures and multiple public and private fundus image datasets. In the five-class classification task based on ICDR criteria, the pooled sensitivities of DL-based models varied significantly across severity levels: 95.19% (95% CI: 93.00%-97.00%) for no DR (stage 0), 72.06% (95% CI: 62.06%-81.09%) for mild NPDR (stage 1), 84.33% (95% CI: 78.90%-89.10%) for moderate NPDR (stage 2), 75.84% (95% CI: 68.42%-82.57%) for severe NPDR (stage 3), and 78.82% (95% CI: 71.76%-85.13%) for PDR (stage 4). In the simplified four-class classification task, sensitivities markedly improved across all grades: 96.85% (95% CI: 90.18%-99.93%) for stage 0, 92.94% (95% CI: 79.50%-99.72%) for stage 1, 92.75% (95% CI: 79.31%-99.61%) for stage 2, and 88.19% (95% CI: 68.99%-98.93%) for stage 3.
Conclusion: DL exhibits high sensitivity and substantial potential for DR grading, particularly in screening for no DR and vision-threatening DR. Nevertheless, precisely differentiating between adjacent non-proliferative stages remains a clinical challenge. The observed heterogeneity underscores the imperative for methodological standardization, rigorous external validation, and multimodal data integration. Future research should prioritize enhancing clinical utility and generalizability to facilitate their translation into real-world clinical practice.
Systematic Review Registration: https://www.crd.york.ac.uk/PROSPERO/, identifier CRD420261338867.

Keywords:  artificial intelligence; automated grading; deep learning; diabetic retinopathy; meta-analysis; systematic review

DOI:  https://doi.org/10.3389/fendo.2026.1853785
Front Med (Lausanne). 2026 ;13 1846788

VesselMetaKAN: vessel-guided meta-learned interpretable classification for diabetic retinopathy grading.

TianQi Yang, GuoYong Chen, Lu Liu.

   Introduction: Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide, affecting approximately 93 million people globally. Accurate automated grading is critical for large-scale public health screening, yet existing deep learning methods face challenges related to structure-coupled sparse evidence and domain shift across imaging conditions.
Methods: We propose VesselMetaKAN, a two-stage framework integrating explicit vessel guidance with meta-learning. Stage 1 employs GMF-SwinUnet for topology-aware vessel segmentation using Frangi-guided attention fusion, reliability gating, and clDice loss. Stage 2 employs KAN-MAML, combining Kolmogorov-Arnold networks with radial basis function expansions as an interpretable decision head and Reptile-style meta-learning over augmentation-defined tasks. Vessel probability maps from Stage 1 guide Stage 2 classification through vessel-gated feature fusion.
Results: On APTOS 2019, VesselMetaKAN achieved 74.9 ± 0.4% accuracy, 58.7 ± 0.6% macro-F1, and 0.838 ± 0.005 QWK over five runs. The best single run reached 75.44% accuracy, 59.44% macro-F1, and 0.843 QWK, outperforming EfficientNet-B4 on the principal grading metrics with paired bootstrap significance (p < 0.05). Ablation studies confirmed the synergistic contributions of vessel guidance, KAN-based interpretability, and meta-learning.
Discussion: VesselMetaKAN provides a principled, structure-aware, and interpretable solution for robust DR grading in public health screening programs.

Keywords:  Kolmogorov-Arnold networks; deep learning; diabetic retinopathy; fundus imaging; interpretable classification; meta-learning; public health screening; vessel segmentation

DOI:  https://doi.org/10.3389/fmed.2026.1846788
Radiol Phys Technol. 2026 Jul 29.

Content-based retrieval of fundus images and diabetic retinopathy detection using variants of local texture features.

Arpita Santra, Imtiyaz Ahmad, Vibhav Prakash Singh.

  Content-based retinal image analysis plays a crucial role in the early diagnosis of ocular diseases. In this study, we proposed a novel approach for efficient content-based retinal image retrieval and Diabetic Retinopathy (DR) detection using variants of local texture features derived from Local Binary Pattern (LBP), Local Ternary Pattern (LTP), and Gray Level Co-occurrence Matrix (GLCM). The methodology begins with meticulous image preprocessing to enhance feature extraction, followed by the extraction of LBP, LTP, and GLCM features, which capture intricate texture patterns and enrich the feature space for robust analysis. Subsequently, we trained machine learning models, including Support Vector Machine (SVM), Decision Tree, and Random Forest, on the extracted features to effectively retrieve retinal images and detect DR. A comparative analysis between preprocessed and raw images highlights the impact of preprocessing techniques on performance. A key innovation of this study lies in the fusion of multiple texture-based features, creating a comprehensive representation that integrates high-level semantic information with fine-grained local patterns. This hybrid approach enhances the system's capability to handle diverse retinal image variations, leading to improved retrieval accuracy and robustness. Further, a metaheuristic approach for feature selection and optimization is employed, comparing Differential Evolution, Genetic Algorithm, and Particle Swarm Optimization to identify the most effective features for retrieval. Differential Evolution achieved the highest precision of 90.67 % for retrieving the top 10 relevant images. The proposed hybrid approach demonstrates the effectiveness of integrating classical image analysis methods with machine learning for DR detection and content-based image retrieval. This research contributes to precision medicine and healthcare innovation by advancing ML-driven retinal image analysis.

Keywords:  CBMIR; Diabetic retinopathy; Gray-level co-occurrence matrix (GLCM); Local binary pattern (LBP); Local ternary pattern (LTP); Metaheuristic approach

DOI:  https://doi.org/10.1007/s12194-026-01106-1
Front Med (Lausanne). 2026 ;13 1906162

Artificial intelligence-based analysis of retinal vascular changes in the preclinical and early stages of diabetic retinopathy using ultra-widefield fundus imaging: an observational cross-sectional study.

Tingting Sun, Yingjun Min, Caihong Wang, Nurbiya Abduriyim, Chenting Wen, Qinghua Qiu.

   Background: Subclinical retinal microvascular remodeling may occur before clinically detectable diabetic retinopathy (DR). This study investigated artificial intelligence (AI)-derived ultra-widefield retinal vascular metrics in patients with type 2 diabetes mellitus (T2DM) with and without non-proliferative DR (NPDR), and evaluated their potential for early vascular phenotyping and diagnostic discrimination.
Methods: In this observational cross-sectional single-center study, 237 participants were included: 63 healthy controls (103 eyes), 101 patients with T2DM without DR (No-DR; 201 eyes), and 73 patients with NPDR (132 eyes). Non-mydriatic 200-degree ultra-widefield fundus images were analyzed using an AI-based vascular segmentation and quantification system. AI-exported values coded as -1 were treated as missing, and sparse parameters were excluded from primary inference. Intergroup comparisons of retained vascular parameters were performed using age- and sex-adjusted mixed-effects models with participant as a random intercept, followed by Benjamini-Hochberg false discovery rate (FDR) correction. Multiparameter logistic models were evaluated using 5-fold subject-level cross-validation.
Results: After quality control, covariate adjustment, and FDR correction, 38 vascular-parameter rows remained significant. Whole-field vessel density was highest in the No-DR group, intermediate in NPDR, and lowest in controls [control, 0.017 (0.008-0.023); No-DR, 0.026 (0.020-0.032); NPDR, 0.021 (0.015-0.026); FDR P < 0.001]. Fractal-dimension metrics showed similar early alterations, with arterial fractal dimension increased in No-DR and intermediate in NPDR [control, 1.287 (1.164-1.359); No-DR, 1.380 (1.328-1.420); NPDR, 1.329 (1.271-1.374); FDR P < 0.001]. Whole-field mean vessel diameter was lower in No-DR than in controls, whereas total vessel length decreased across controls, No-DR, and NPDR (FDR P < 0.001 for both). Regional heatmaps showed that significant density and fractal-dimension signals clustered mainly in the superotemporal, superonasal, and inferotemporal regions. Cross-validated models showed good discrimination for No-DR versus control (AUC, 0.853; 95% CI, 0.807-0.893), NPDR versus control (AUC, 0.785; 95% CI, 0.720-0.841), and No-DR/NPDR versus control (AUC, 0.830; 95% CI, 0.788-0.873), but more modest discrimination between NPDR and No-DR (AUC, 0.658; 95% CI, 0.596-0.720).
Conclusion: AI-derived ultra-widefield retinal vascular metrics demonstrate early, spatially heterogeneous microvascular remodeling in T2DM before clinically apparent DR. Vessel density, fractal dimension, vessel diameter, and vessel length provide complementary information, and multiparameter vascular modeling may support early detection and risk stratification. External validation and longitudinal studies are required before clinical implementation.

Keywords:  artificial intelligence; diabetic retinopathy; retinal vascular parameters; type 2 diabetes mellitus; ultra-widefield fundus photography

DOI:  https://doi.org/10.3389/fmed.2026.1906162
Npj Health Syst. 2025 Sep 16. pii: 35. [Epub ahead of print]2(1):

A large sensor foundation model pretrained on continuous glucose monitor data for diabetes management.

Junjie Luo, Abhimanyu Kumbara, Mansur Shomali, Rui Han, Anand Iyer, Grazia Aleppo, Ritu Agarwal, Gordon Gao.

Continuous glucose monitoring (CGM) combined with AI offers new opportunities for proactive diabetes management through real-time glucose forecasting. However, most existing models are task-specific and lack generalization across patient populations. Inspired by the autoregressive paradigm of large language models, we introduce CGM-LSM, a Transformer decoder-based Large Sensor Model (LSM) pretrained on 1.6 million CGM records from patients with different diabetes types, ages, and genders. We model patients as sequences of glucose time steps to learn latent knowledge embedded in CGM data and apply it to the prediction of glucose readings for a 2-h horizon. Compared with prior methods, CGM-LSM significantly improves prediction accuracy and robustness: a 48.51% reduction in root mean square error in 1-h horizon forecasting and consistent zero-shot prediction performance across held-out patient groups. We analyze model performance variations across patient subgroups and prediction scenarios and outline key opportunities and challenges for advancing CGM foundation models.

DOI: https://doi.org/10.1038/s44401-025-00039-y
Bioengineering (Basel). 2026 Jul 10. pii: 791. [Epub ahead of print]13(7):

Intelligent Staging Performance of Diabetic Retinopathy Based on Fundus Fluorescein Angiography Images with Different Angiographic Phases.

Wei Wang, Zhenpeng Chen, Mingming Li, Shuang Li, Kang Wang, Haiyun Li.

  Fundus fluorescein angiography (FFA) is the gold standard for diabetic retinopathy (DR) staging, yet whether angiographic phase affects deep learning performance remains unknown. Using 7508 FFA images from 863 eyes, stratified into venous, recirculation, and late phases, we developed Swin Transformer- and ConvNeXt-based DR staging models under the International five-grade and Chinese six-grade classification systems. Multiclass and binary (NPDR vs. PDR) classification tasks were evaluated. This study provides the first systematic quantification of angiographic phase effects on FFA-based DR staging. Phase-related performance was analyzed using generalized linear mixed-effects models with beta regression. ConvNeXt generally outperformed Swin Transformer, particularly in multiclass classification. Across all experimental settings, performance showed a consistent numerical decline from the venous to the late phase. Under the International five-grade system, ConvNeXt accuracy declined from 86.67% to 81.87%; however, Bonferroni-adjusted comparisons revealed no statistically significant phase-related differences (all adjusted p > 0.05), with small effect sizes (|SMD| < 0.2). Binary classification remained highly stable across phases, with accuracies exceeding 91%. Grad-CAM visualizations demonstrated progressively diffuse model attention in later phases. These findings support phase-flexible FFA acquisition for binary DR screening, whereas venous- or recirculation-phase images remain preferable for high-precision multiclass staging, guiding phase-aware AI development.

Keywords:  ConvNeXt; Swin Transformer; angiographic phase; diabetic retinopathy; fluorescein fundus angiography; grading model

DOI:  https://doi.org/10.3390/bioengineering13070791
Curr Med Chem. 2026 Jul 22.

Identification of COX7B, GLRX5, and OR52H1 as Vascular Endothelial Injury-associated Biomarkers in Gestational Diabetes Mellitus.

Yueping Wang, Weiwei Feng.

   INTRODUCTION: This study aimed to screen and validate such biomarkers to clarify the role of vascular endothelial injury (VEI) in the pathogenesis of gestational diabetes mellitus (GDM) and provide clues for future diagnostic and mechanistic studies.
METHODS: VEI scores were calculated using a published VEI-related gene set. VEI-correlated gene modules were classified by weighted gene co-expression network analysis (WGCNA), with module genes subjected to GO/KEGG enrichment analysis and intersected with differentially expressed genes (DEGs) from the GSE70493 dataset. Core genes were filtered using three machine learning algorithms, including random forest, LASSO regression, and SVM-RFE, followed by analyses of diagnostic performance, hallmark pathway correlations, and immune infiltration associations.
RESULTS: DEGs were enriched in endoplasmic reticulum protein localization, oxidative stress, and mitochondrial pathways and overlapped with the yellow gene module. Machine learning identified COX7B, GLRX5, and OR52H1 as the candidate biomarkers for GDM. These genes showed favorable diagnostic performance and distinct correlations with hallmark pathways and immune infiltration features, suggesting their potential involvement in VEI-related molecular changes in GDM.
DISCUSSION: This study identified VEI-related hub genes in GDM but is limited by its reliance on a single dataset, a small validation sample, and the lack of experimental validation, warranting further confirmation.
CONCLUSION: COX7B, GLRX5, and OR52H1 were identified as VEI-associated diagnostic biomarkers in GDM. These findings support the involvement of VEI-related molecular alterations in GDM but should be regarded as hypothesis-generating, requiring further validation in larger clinical cohorts and experimental models.

Keywords:  Gestational diabetes mellitus; cytochrome C oxidase subunit 7B; glutaredoxin 5; machine learning; olfactory receptor family 52 subfamily H member 1; vascular endothelial injury

DOI:  https://doi.org/10.2174/0109298673497893260708194320
J Diabetes Sci Technol. 2026 Jul 27. 19322968261467778

INSIGHT: Ingest and Standardize Glucose Harmonization Tool.

Ben Ehlert, Krish Shah, Dev Katarya, Yue Wu, Tao Wang, Michael P Snyder.

   BACKGROUND: Publicly available continuous glucose monitoring (CGM) datasets are fragmented across schemas, limiting reuse, pooling, and machine learning development. Although large biobanks provide harmonized resources, many smaller CGM studies remain difficult to access because ingestion and standardization require substantial manual effort.
METHODS: We developed INSIGHT (Ingest and Standardize Glucose Harmonization Tool), an open-source large language model (LLM)-assisted harmonization pipeline that standardizes public CGM datasets into a minimal common schema: Subject_ID, Timestamp, and Glucose. INSIGHT combines deterministic rules with structured LLM assistance to identify CGM files, classify file roles, generate schema-specific read plans, normalize timestamps and glucose units, resolve subject identifiers, detect overlapping exports, merge outputs, and retain auditable JSON artifacts. We evaluated INSIGHT on 14 public CGM datasets with curated reference outputs using pre-specified training and held-out testing partitions. Performance was assessed using file-level precision/recall, subject recovery, temporal alignment, glucose fidelity, scalar CGM feature accuracy, and composite INSIGHT scores.
RESULTS: INSIGHT is open-source at https://github.com/benehlert/CGMHarmonization. Across 90 model-dataset evaluations spanning 9 model variants and 10 datasets, top configurations achieved near-perfect held-out harmonization fidelity. GPT-5.4 and Gemini 3.1 Pro Preview achieved mean INSIGHT scores of 0.990, with glucose mean absolute error (MAE) below 0.03 mg/dL and within 5 mg/dL agreement of 0.9996. Errors primarily reflected complex subject-timestamp recovery rather than glucose conversion, with failures concentrated in deeply nested datasets containing overlapping or co-located non-CGM exports.
CONCLUSION: INSIGHT demonstrates that LLM-assisted code generation can generalize CGM harmonization across heterogeneous datasets while producing auditable loaders. By reducing reuse costs, INSIGHT complements curated repositories and supports large-scale pooled analyses and foundation model development for glycemic physiology.

Keywords:  continuous glucose monitoring; data harmonization; diabetes; health sensors; large language model

DOI:  https://doi.org/10.1177/19322968261467778
Hum Mol Genet. 2026 Jul 28. pii: ddag069. [Epub ahead of print]35(16):

Single-cell transcriptomic integrated with machine learning reveals human retinal cell-specific biomarkers in diabetic retinopathy.

Sen Lin, Luning Yang, Yiwen Tao, Qi Pan, Tengda Cai, Yunyan Ye, Jianhui Liu, Yang Zhou, Yongqing Shao, Quanyong Yi, Zen H Lu, Lie Chen, Gareth McKay, Richard Rankin, Fan Li, Weihua Meng.

   BACKGROUND: Diabetic retinopathy (DR) is a leading cause of vision impairment worldwide, yet the cell-type-specific molecular alterations associated with disease progression remain incompletely understood. This study aimed to characterize transcriptional changes across retinal cell types in diabetes and DR and identify candidate disease-associated biomarkers using single-cell transcriptomics and machine-learning approaches.
METHODS: We generated a single-cell RNA sequencing (scRNA-seq) atlas comprising 297 121 high-quality retinal cells from 20 eyes of 13 Chinese donors, including non-diabetic controls (NON), diabetes without retinopathy (DM), and DR samples. Following quality control, batch correction, clustering, and cell-type annotation, differential expression analyses were performed across disease states within each retinal cell type. Candidate biomarkers were further prioritized using a machine-learning framework incorporating L1-regularized logistic regression, recursive feature elimination with cross-validation, and stability selection.
RESULTS: We identified 10 major retinal cell populations and characterized extensive cell-type-specific transcriptional alterations associated with diabetes and DR. Pathway enrichment analyses consistently highlighted immune activation, oxidative stress, neurodegeneration, and synaptic dysfunction across multiple retinal cell types. A total of 707 cell-type-specific candidate marker genes were identified, providing a comprehensive resource for investigating disease-associated molecular mechanisms and potential therapeutic targets.
CONCLUSIONS: This study establishes a single-cell transcriptomic atlas of the Chinese diabetic retina and reveals cell-type-specific molecular signatures associated with DR progression. These findings provide biological insights into retinal disease mechanisms and nominate candidate biomarkers for future functional and translational studies.

Keywords:  Chinese human retina; cell atlas; cell-type-specific biomarkers; diabetic retinopathy; differential expression analysis; feature selection; machine learning classification; neurodegeneration; oxidative stress; single-cell RNA sequencing; transcriptomic atlas

DOI:  https://doi.org/10.1093/hmg/ddag069
Bioengineering (Basel). 2026 Jun 29. pii: 762. [Epub ahead of print]13(7):

Active Learning Under Expert-Budget Constraints: A Human-in-the-Loop Pipeline for Diabetic Retinopathy Lesion Detection.

Hyeok Kim, Seok-Min Chang, Bo-Young Lim, Soo Young Lee, Ho-Gil Jung.

  Early diagnosis of Diabetic Retinopathy (DR) is critical for preventing irreversible vision loss, but precise lesion annotation by ophthalmologists is the dominant cost in building any clinical-grade DR detection model. The structural problem in real hospital settings is not labeling cost per se, but expert availability: ophthalmologists' time is bounded by clinical duties, so the active-learning (AL) cycle can iterate only a handful of times in practice. We frame this constraint explicitly and ask which AL designs work best under a tight expert budget. We propose Virtuous Cycle, a Human-in-the-Loop (HITL) pipeline that integrates (i) a YOLOv8x-based object detector for microaneurysms, hemorrhages, and exudates, (ii) four AL sampling strategies (Average Confidence, Random, Hybrid-Diversity, Monte Carlo Dropout), and (iii) an in-hospital annotation platform (Diavision Studio) in which clinicians refine AI pre-labels rather than draw from scratch. We evaluate Virtuous Cycle on a real-world fundus dataset from the National Medical Center (NMC) across eight AL rounds, expanding the labeled pool from 81 images (R0) to 481 images (R8) within the actual expert-time budget of two ophthalmologists. Across three independent random seeds, random sampling dominates at cold start (mean mAP@50 0.14→0.25 over R0-R1), whereas Hybrid-Diversity converges to the highest mAP@50, Precision, and Recall by R7 (431 images; mAP@50 0.40, Precision 0.55, Recall 0.41), with MC Dropout close behind; by R8, the labeled pool is exhausted and all strategies converge to the same final model. A clinician crossover analysis of 36 paired clinical images, controlling for per-clinician speed bias and per-image difficulty bias, shows no statistically significant difference in overall per-image labeling time between AI-assisted and manual annotation (p=0.52), but a statistically significant increase in confirmed lesion detections under AI assistance (p=0.0058), driven predominantly (84-100% of the net increase) by microaneurysms, the lesion type most prone to being missed unaided. The results indicate that, under expert-budget constraints, AL strategy choice should be staged: random sampling for cold start, uncertainty-and-diversity sampling once the model has matured, and that AI assistance trades a modest, lesion-burden-dependent time cost for a measurable gain in the sensitivity of microaneurysm detection.

Keywords:  Monte Carlo dropout; YOLOv8; active learning; diabetic retinopathy; human-in-the-loop; medical image annotation; object detection; uncertainty quantification

DOI:  https://doi.org/10.3390/bioengineering13070762
Front Artif Intell. 2026 ;9 1861563

Machine learning algorithms for predicting glycemic control and weight loss outcomes in GLP-1 receptor agonist users.

Tadesse M Abegaz, Gabriel Frietze.

   Introduction: Glucagon-like peptide-1 receptor agonists (GLP-1 RAs) are widely used for the management of type 2 diabetes mellitus and obesity; however, substantial inter-individual variability in glycemic and weight loss outcomes remains. This study aimed to develop and validate machine learning (ML) models to predict glycemic control and weight loss outcomes following GLP-1 RA initiation using real-world data and to identify key features associated with treatment response.
Methods: We conducted a retrospective cohort study using data from the All of Us Research Program. Adult participants initiating GLP-1 RA therapy with available baseline and follow-up measurements were included. Two cohorts were constructed: a glycemic control cohort (n = 3,975) and a weight loss outcome cohort (n = 11,420). Glycemic control was defined as hemoglobin A1c (HbA1c) <7% at follow-up, and weight improvement was defined as achieving a body mass index (BMI) <30 kg/m². Multiple ML models, including logistic regression, random forest (RF), extreme gradient boosting (XGBoost), support vector machine, neural network, LightGBM, and CatBoost, were developed using 10-fold cross-validation. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and precision-recall curves. SHapley Additive exPlanations (SHAP) were used to improve model interpretability.
Results: For weight loss outcome prediction, ensemble models demonstrated superior performance, with RF and XGBoost achieving the highest discrimination (AUC ≈ 0.94) and accuracy (0.89-0.90). For glycemic control prediction, RF and XGBoost achieved modest performance (accuracy ≈ 0.73; AUC ≈ 0.79). SHAP analysis identified baseline BMI and body weight as the most influential features of weight improvement, while duration of diabetes, baseline HbA1c, and use of sulfonylureas or insulin were among the most important features of glycemic control.
Discussion: Machine learning models, particularly tree-based ensemble methods, demonstrated strong potential for predicting treatment response to GLP-1 RA therapy. Integration of explainable ML approaches with real-world data may support personalized treatment strategies, and facilitate identification of patients most likely to benefit from GLP-1 RA therapy.

Keywords:  All of US research program; GLP-1 receptor agonists; explainable artificial intelligence; glycemic control; machine learning; weight loss

DOI:  https://doi.org/10.3389/frai.2026.1861563
Npj Health Syst. 2026 Jul 08. pii: 59. [Epub ahead of print]3(1):

Human-in-the-loop AI predictive digital twin to extend virtual precision diabetes care between visits.

Jing Wang, Syed Hasib Akhter Faruqui, Adel Alaeddini, Yan Du, Shiyu Li, Yijiong Yang, Kumar Sharma.

In a 6-month randomized trial, we evaluated a digitally enabled "human-in-the-loop" care support model using a predictive artificial intelligence (AI) digital twin to provide personalized daily short message service (SMS) feedback for adults with type 2 diabetes (T2D). The parent study enrolled 40 adults aged ≥18 years with T2D who completed 3 months of baseline observation followed by a 3-month intervention period, generating 6467 longitudinal data points across weight, dietary intake, physical activity, and glucose monitoring (mean follow-up: 174 days). For this ancillary AI intervention, a subset of 19 participants was randomized to receive either AI-generated individualized daily feedback (AI group, n = 10) or no daily feedback (control group, n = 9). The online human-in-the-loop predictive control model incorporated a transfer-learning artificial neural network predictive digital twin trained on participant self-monitoring data, including weight, food logs, physical activity, and glucose values. A particle swarm optimization controller identified personalized behavioral recommendations aligned with glucose and weight goals, and the digital twin was retrained weekly using newly accrued data. The model achieved ≥80% prediction accuracy across all diet-condition subgroups. During the intervention period, participants receiving AI-generated feedback demonstrated trends toward increased daily step counts and improved adherence to caloric and carbohydrate intake targets. The AI intervention group achieved significantly greater weight loss than controls (mean loss 5.87 lbs vs 3.57 lbs; p < 0.012) while maintaining stable glucose levels throughout the study period (p = 0.661). These findings suggest that AI-enabled predictive digital twin models may offer a scalable approach for extending precision diabetes self-management support beyond clinic visits.

DOI: https://doi.org/10.1038/s44401-026-00118-8