bims-aukdir Biomed News
on Automated knowledge discovery in diabetes research
Issue of 2026–10–11
thirteen papers selected by
Mott Given
  1. Real-world validation of an AI-based non-mydriatic fundus camera in the diagnosis of diabetic retinopathy in Liberia.
  2. Integration of ultra-widefield optical coherence tomography angiography images and clinical indicators for identifying early-stage diabetic retinopathy in type 2 diabetes.
  3. Correlation analysis between tongue hyperspectral imaging and tongue coating microbiota and feasibility study for early screening of type 2 diabetes mellitus status.
  4. Chemometrics-assisted FTIR and machine learning for biochemical profiling of human skeletal muscles in relation to type 2 diabetes.
  5. Systemic Diseases and the Eye: A Translational Perspective on Diabetes.
  6. Development and validation of a predictive model for sarcopenia in patients with type 2 diabetes based on the synergistic mechanism of insulin resistance and inflammation.
  7. Development and comparative evaluation of machine learning algorithms and Cox regression for predicting fasting plasma glucose-defined incident prediabetes: a longitudinal cohort study.
  8. Full Spectrum Raman Spectroscopy Combined With Deep Learning for Prediction of Tensile Mechanical Properties of Human Cortical Bone.
  9. Individualized Therapy Optimization for Type 2 Diabetes: Retrospective Study of Model Development, Internal Validation, and Physician Evaluation.
  10. Multimodal Carotid Ultrasound-based Artificial Intelligence Model for Predicting Acute Ischemic Stroke Risk in Patients with Type 2 Diabetes Mellitus.
  11. SEFCA-Net: a semantic-edge-frequency cross-attention network for diabetic retinopathy lesion segmentation.
  12. Retrieval-Augmented Simulation for Interpreting Probabilistic Digital Twins in Type 1 Diabetes.
  13. Quantitative mapping of gut microbiota feature importance to TCM syndrome elements in clinical T2DM stages.
  1. Front Digit Health. 2026 ;8 1938702
    Real-world validation of an AI-based non-mydriatic fundus camera in the diagnosis of diabetic retinopathy in Liberia.
    Vickie Fayiah Sam, Mohammad Rizwan, John Chiemerie Douglas, Amrit Jha, Sirisha Senthil, Padmaja Kumari Rani.
       Introduction: Diabetic retinopathy (DR) is a leading cause of preventable vision loss, yet access to retinal screening remains limited in low-resource settings such as Liberia. This study evaluated the real-world diagnostic performance of an offline artificial intelligence (AI)-enabled smartphone-based non-mydriatic fundus camera for detecting referable diabetic retinopathy in a routine general outpatient screening setting in Liberia.
    Methods: This prospective study was conducted at the General Outpatient Department of John F. Kennedy Medical Center, Monrovia, Liberia. Adults aged 18 years and older were recruited through convenience sampling and underwent non-mydriatic fundus imaging using the Remidio Fundus on Phone (FOP-NM-10) integrated with an offline AI system. Fundus images were independently assessed by two trained optometrists using the International Clinical Diabetic Retinopathy (ICDR) classification, with disagreements adjudicated by a senior retinal ophthalmologist. Referable DR was defined as moderate non-proliferative DR, severe non-proliferative DR, or proliferative DR. Diagnostic performance was evaluated at the participant level among participants with at least one gradable eye.
    Results: A total of 1,612 participants were enrolled, of whom 1,357 were included in the final analytical cohort. Of these, 1,198 (88.3%) had at least one gradable eye and were included in the diagnostic-performance analysis. Eleven participants were classified as having referable DR by the human reference standard. The AI system correctly classified 10 of these 11 participants, yielding a sensitivity of 90.9% (95% CI, 58.7%-100%). Among 1,187 participants without referable DR, 1,183 were correctly classified as negative, resulting in a specificity of 99.7% (95% CI, 99.1%-99.9%). The positive predictive value was 71.4% (95% CI, 41.9%-91.6%), and the negative predictive value was 99.9% (95% CI, 99.5%-99.9%).
    Conclusions: In this prospective real-world study, the offline AI-based non-mydriatic fundus camera showed high sensitivity and specificity for detecting referable diabetic retinopathy among participants with at least one gradable eye in a general outpatient setting in Liberia. These findings support its feasibility for AI-assisted detection of referable DR in resource-limited settings.
    Keywords:  Liberia (West Africa); artificial intelligence - AI; diabetic retinopathy (DR); fundus photography; low-resource setting; non-mydriatic imaging; retinal screening
    DOI:  https://doi.org/10.3389/fdgth.2026.1938702
  2. Front Endocrinol (Lausanne). 2026 ;17 1941613
    Integration of ultra-widefield optical coherence tomography angiography images and clinical indicators for identifying early-stage diabetic retinopathy in type 2 diabetes.
    Cong Zhang, Li Wang, Shimeng Huang, Ruoxuan Zheng, Jingjing Li, Xiaotong Sun, Shuang Wu, Tianpeng Zhang, Xinyu Zhang, Qian Xu, Yangwen Chen, Chengye Xu, Hongyu Kuang.
       Background: Diabetic retinopathy (DR) is a common microvascular complication of diabetes and a leading cause of irreversible vision loss among working-age adults, making early detection essential for preventing permanent visual impairment. This study aimed to identify early-stage DR in patients with type 2 diabetes mellitus (T2DM) using a multimodal deep learning (DL) approach integrating ultra-widefield optical coherence tomography angiography (UW-OCTA) images with clinical factors.
    Methods: This cross-sectional study enrolled patients with T2DM from an urban tertiary hospital and a community primary care setting between July 2024 and April 2026. Baseline demographic and laboratory data and 26 × 21 mm UW-OCTA images were obtained. Automatically segmented UW-OCTA slabs were compared to identify the most discriminative layer. Clinical variables were selected using between-group comparisons, XGBoost-based Shapley additive explanations (SHAP), and multivariable logistic regression. Selected clinical variables and UW-OCTA image features were integrated into a multimodal model. Performance was evaluated in internal test and external validation sets using receiver operating characteristic analysis. Gradient-weighted class activation mapping (Grad-CAM) was used for visualization.
    Results: Overall, 1,676 patients were included: 1,047 and 453 tertiary-care patients formed the training and internal test sets, and 176 community-based patients formed the external validation set. Diabetes duration, glycated hemoglobin, low-density lipoprotein cholesterol and systolic blood pressure were independently associated with early-stage DR. The retinal slab showed the best image-only performance with AUCs of 0.906 (95%CI: 0.870-0.942) and 0.828 (95%CI: 0.725-0.931) in the internal and external sets. The multimodal model achieved AUCs of 0.921 (95%CI: 0.887-0.954) and 0.849 (95%CI: 0.760-0.953), outperforming image-only and clinical-only models. Grad-CAM highlighted macular and peripheral retinal regions.
    Conclusion: Integrating UW-OCTA with clinical factors improved early-stage DR identification in T2DM and may support non-invasive risk assessment.
    Keywords:  clinical factors; diabetic retinopathy; multimodal deep learning; type 2 diabetes mellitus; ultra-widefield optical coherence tomography angiography
    DOI:  https://doi.org/10.3389/fendo.2026.1941613
  3. Front Endocrinol (Lausanne). 2026 ;17 1936566
    Correlation analysis between tongue hyperspectral imaging and tongue coating microbiota and feasibility study for early screening of type 2 diabetes mellitus status.
    Zhao-Wei Liu, Yan Zhang, Lei Fu, Jing-Ya Xie, Xin-Yu Zhang, Ming Liu, Jing Zhao.
       Objective: To investigate the feasibility of applying hyperspectral imaging technology combined with machine learning algorithms for early non-invasive screening of type 2 diabetes mellitus (T2DM).
    Methods: In this study, tongue images of 153 subjects were acquired using a hyperspectral imaging system. Concurrently, 53 tongue coating samples were collected for 16S rRNA sequencing. Spearman correlation analysis was performed between tongue hyperspectral data and tongue coating microbiota. Finally, machine learning was integrated to construct a T2DM disease status identification model based on tongue hyperspectral data.
    Results: The experimental results revealed that 10 genera exhibited significant differences among the three groups (P < 0.05). Among these, the relative abundance of unclassified_p_Bacillota, Pseudoleptotrichia, Xylanibacter, Gemella, norank_f_Mitochondria, and Enhydrobacter showed significant correlations with hyperspectral tongue images (P < 0.05). The T2DM disease status identification models constructed based on tongue hyperspectral data achieved favorable modeling results, with the optimal model demonstrating ACC = 0.8966 and AUC = 0.8029.
    Conclusion: The results of this study indicate that certain correlations exist between tongue hyperspectral features and tongue coating microbial genera in T2DM patients at different disease stages. The T2DM disease status identification model constructed based on tongue hyperspectral imaging and machine learning algorithms can be used for the identification of HC, IGR, and T2DM, and this method may provide a noninvasive and rapid screening approach for early identification of T2DM.
    Keywords:  disease risk prediction; hyperspectral tongue imaging; machine learning algorithms; tongue coating microbiota; type 2 diabetes mellitus
    DOI:  https://doi.org/10.3389/fendo.2026.1936566
  4. Spectrochim Acta A Mol Biomol Spectrosc. 2026 Sep 28. pii: S1386-1425(26)01464-2. [Epub ahead of print]366 128893
    Chemometrics-assisted FTIR and machine learning for biochemical profiling of human skeletal muscles in relation to type 2 diabetes.
    Mohamed Elwy Abdelhamed Abdelmonaem, Chiedozie Kenneth Ugwoke, Armin Alibegović, Erika Cvetko, Nejc Umek, Jože Grdadolnik, Barbara Zupančič.
      Skeletal muscle is the major site of insulin-stimulated glucose uptake, making it highly susceptible to metabolic disorders. This study investigates Type 2 diabetes mellitus-associated alterations in the macromolecular composition of functionally distinct human skeletal muscles. These modifications are crucial to understanding the molecular basis of diabetic myopathies but remain underexplored. Five skeletal muscles from 80 individuals (40 controls and 40 individuals with Type 2 diabetes) were analysed. To overcome the limitations of traditional histochemical assays, we employed infrared spectroscopy combined with a chemometric decomposition approach to resolve spectra into five spectral components, each dominated by a specific macromolecular group. The corresponding component weights quantified and compared relative macromolecular proportions between the diabetic and control groups for different muscle types. Compared with controls, muscles from individuals with diabetes showed a consistent pattern of higher lipid levels and lower glycogen levels, with a statistically significant increase in lipids within the diaphragm. This biochemical profiling was enriched by building machine learning classifiers, which achieved low-to-modest but statistically significant AUC performance of up to 72% (p < 0.05) in discriminating diabetic from healthy muscle. Our work demonstrates an efficient methodology for the simultaneous screening of various macromolecular constituents using spectral data obtained in a single experiment. It provides medically relevant insights into the biological patterns of diabetes-associated skeletal muscle remodelling to guide future research. However, an independent validation should be performed to overcome this study limitation and draw definitive conclusions.
    Keywords:  Chemometrics; Infrared spectroscopy; Machine learning; Macromolecular composition; Skeletal muscles; Type 2 diabetes mellitus
    DOI:  https://doi.org/10.1016/j.saa.2026.128893
  5. Chronic Dis Transl Med. 2026 Oct 08.
    Systemic Diseases and the Eye: A Translational Perspective on Diabetes.
    Matteo Capobianco, Marco Zeppieri, Simonetta Gaia Nicolosi, Niccolò Castellino, Francesco Cappellani, Irene Gattazzo, Caterina Gagliano.
      Diabetes mellitus is a global chronic disease that affects the whole body and not just glucose metabolism. In diabetes, non-invasive ocular changes can help understand microvascular, neurodegenerative, and inflammatory processes. This analysis looks at the eye as a biomarker platform for diabetes-related disease, with retinal imaging informing systemic pathophysiology and the emergence of AI-assisted approaches. With improvements in optical coherence tomography (OCT) and OCT angiography, it is possible to observe neurovascular changes that can occur before the clinical onset of diabetic retinopathy. Issues in the retina have also been connected to problems with the heart, brain, kidneys, and cognition. However, much of the evidence for problems involving the heart and cognition comes from studies on people without diabetes. Also, most of the available data are observational, rather than coming from controlled studies. As a result, retinal measurements should be considered candidate biomarkers rather than clinically validated predictors or surrogate endpoints for systemic outcomes. Artificial Intelligence (AI) has reached a phase of clinical maturity such that it can be used in the automated screening of diabetic retinopathy. However, it remains investigational for systemic risk prediction, such as for chronic kidney disease, cardiovascular outcomes, and cognitive decline. Equally, the therapeutic implications of retinal neurodegeneration and other imaging biomarkers are under-validated for systemic decisions. Before retinal biomarkers can be used in precision diabetes care, we need more diabetes-specific prospective studies to establish standardized measurements, the clinical meaning of thresholds, incremental predictive value above and beyond established risk factors, reproducibility, cost-effectiveness, and evidence of improved outcomes.
    Keywords:  artificial intelligence; biomarkers; diabetes mellitus; diabetic retinopathy; optical coherence tomography; retinal vessels
    DOI:  https://doi.org/10.1002/cdt3.70071
  6. Front Physiol. 2026 ;17 1922889
    Development and validation of a predictive model for sarcopenia in patients with type 2 diabetes based on the synergistic mechanism of insulin resistance and inflammation.
    Ajinisha Akeaji, Baihetinisha Akeaji, Yihe Han, Ning Wang.
       Background: Type 2 diabetes mellitus (T2DM) complicated by sarcopenia represents a considerable public health challenge, yet early detection remains challenging due to a lack of practical screening tools. The aim of this study was to develop and internally validate a clinically feasible, mechanism-based machine learning model for predicting sarcopenia risk in patients with T2DM.
    Methods: A total of 904 patients with T2DM treated at the First Affiliated Hospital of Xinjiang Medical University from May 2024 to May 2026 were retrospectively enrolled. Sarcopenia was diagnosed according to the Asian Working Group for Sarcopenia (AWGS) 2025 consensus. The dataset was randomly partitioned into training (70%) and internal validation (30%) sets. Key predictive features were identified by intersecting the variables selected via least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm. Eight machine learning algorithms were subsequently developed and evaluated.
    Results: The prevalence of sarcopenia among the patients with T2DM was 29.20%. Eight predictors were selected: age, sex, body mass index, serum albumin, alanine aminotransferase, hemoglobin, metabolic score for insulin resistance, and monocyte-to-high-density lipoprotein cholesterol ratio. LightGBM demonstrated the best overall predictive performance in the validation set, yielding an area under the receiver operating characteristic curve (AUC) of 0.928, an accuracy of 0.882, a sensitivity of 0.897, a specificity of 0.876, a precision of 0.745, and an F1 score of 0.814. Calibration and decision curve analysis indicated that the model yielded clinical net benefit. A web-based calculator developed using LightGBM enabled three-tier risk stratification: low (<15%), moderate (15%-40%), and high (>40%).
    Conclusion: A machine learning model for predicting sarcopenia risk in patients with T2DM was developed and internally validated, resulting in the implementation of an online risk calculator. This mechanism-based model demonstrated promising internal predictive performance and may provide a practical and cost-effective screening approach for diabetes management.
    Keywords:  inflammatory marker; insulin resistance; machine learning; sarcopenia; type 2 diabetes mellitus
    DOI:  https://doi.org/10.3389/fphys.2026.1922889
  7. Front Endocrinol (Lausanne). 2026 ;17 1862258
    Development and comparative evaluation of machine learning algorithms and Cox regression for predicting fasting plasma glucose-defined incident prediabetes: a longitudinal cohort study.
    Su Hu, Changshun Yan, Xinyuan Tian, Menghan Xu, Bo Cheng, Tong Xu, Guiqiu Cao.
       Objective: To develop, compare, and externally evaluate machine learning (ML) and Cox regression models for predicting fasting plasma glucose (FPG)-defined incident prediabetes.
    Methods: We performed a secondary analysis of a publicly available Chinese health-examination cohort and conducted an external comparative evaluation in a separate hospital-based cohort. The development cohort of adults with normal baseline fasting glucose was divided into training and internal validation sets. Candidate predictors were selected using least absolute shrinkage and selection operator (LASSO) regression. Six ML models and an initial proportional-hazards Cox model were compared using discrimination and threshold-based metrics. After identifying FPG nonproportionality, four Cox specifications were compared. Model selection considered fit, performance in the internal validation set and external cohort, complexity, and parsimony. The final model was evaluated at the primary 3-, 4-, and 5-year horizons using time-dependent area under the receiver operating characteristic curve (AUC), Brier scores, calibration, and decision curve analysis. SHapley Additive exPlanations (SHAP), sensitivity analyses, and an online calculator supported interpretation and implementation.
    Results: Among 10, 592 participants, 1, 287 (12.2%) developed FPG-defined incident prediabetes over a mean follow-up of 3.05 years. LASSO retained 11 of 17 predictors. Random forest showed the highest apparent training performance, but its performance declined in the internal validation set and external cohort, whereas the initial Cox model demonstrated more consistent discrimination. FPG violated the proportional-hazards assumption. The extended Cox model incorporating FPG × log(t/3) (M2) had the lowest Akaike information criterion, accommodated the time-varying FPG effect, and performed comparably to more complex alternatives; it was therefore selected. M2 retained age, body mass index, diastolic blood pressure, FPG, and family history of diabetes. Its 3- and 4-year AUCs were 0.800 and 0.796 in training, 0.778 and 0.755 in internal validation, and 0.750 and 0.764 in the external cohort. Five-year performance was exploratory and less stable externally. SHAP identified FPG as the dominant contributor, and sensitivity analyses generally supported robustness.
    Conclusion: The final extended Cox model (M2) and online risk calculator may provide a practical framework for interpretable, individualized risk assessment of FPG-defined incident prediabetes, supporting early risk stratification and preventive risk-factor management using routinely available clinical indicators.
    Keywords:  SHAPley additive explanations; diabetes mellitus; machine learning; online risk calculator; prediabetes
    DOI:  https://doi.org/10.3389/fendo.2026.1862258
  8. J Biophotonics. 2026 Oct;19(10): e70371
    Full Spectrum Raman Spectroscopy Combined With Deep Learning for Prediction of Tensile Mechanical Properties of Human Cortical Bone.
    Muhammet Varli, Sasidhar Uppuganti, Jeffry S Nyman, Ahmet Kayabasi, Mustafa Unal.
      Type 2 diabetes mellitus (T2DM) alters bone quality at the matrix level in ways that current methods do not fully capture. Raman spectroscopy provides molecular information on bone composition and is well suited to probing such alterations. Full-spectrum Raman data were acquired from cortical bone of fresh-frozen cadaveric femurs from 60 non-diabetic and 60 T2DM donors (equal numbers of men and women, aged 50-97 years), followed by uniaxial tensile testing. Machine learning (ML) models (random forest [RF], gradient boosting machine [GBM], adaptive boosting [AdaBoost], stacking) and a one-dimensional convolutional neural network (1D CNN) were trained on the full spectrum to predict tensile properties. A combined 1D CNN and multilayer perceptron (MLP) model that added clinical variables (age, sex, group) reached the highest performance, with R2 values from 0.83 to 0.89, while spectral-only models also performed well. These findings show that full-spectrum Raman analysis with deep learning (DL) predicts tensile mechanical behavior of human cortical bone. SHapley Additive exPlanations (SHAP) analysis further showed that collagen associated spectral regions, rather than mineral-related regions, were the primary contributors to the model predictions.
    Keywords:  Raman spectroscopy; Type 2 diabetes mellitus; artificial intelligence; bone mechanical properties; bone quality; convolutional neural network; deep learning; machine learning; tensile properties
    DOI:  https://doi.org/10.1002/jbio.70371
  9. JMIR Form Res. 2026 Oct 09. 10 e92877
    Individualized Therapy Optimization for Type 2 Diabetes: Retrospective Study of Model Development, Internal Validation, and Physician Evaluation.
    Arina Lozhkina, Camillo Dario Piazza, Zeina Gabr, Maurice Rupp, David Herzig, Lia Bally, Andre Jaun.
       Background: Type 2 diabetes is a widespread chronic condition in which blood glucose and body weight management constitute essential therapeutic targets. Emerging technologies have the potential to aid complex therapeutic pharmacotherapy choices that are optimally tailored to individual needs.
    Objective: In this study, we developed and evaluated an AI model combining guidelines with clinical features and continuous glucose monitoring (CGM) to optimize therapeutic decision-making.
    Methods: Therapeutic guidelines were first encoded using a rule-based model and trained on a feed-forward neural network to predict the probability of therapeutic success for individual treatment recommendations. This approach relied on real-world evidence from a specialist diabetes outpatient clinic, using historical clinical data generated between 2009 and 2023. We used data from 533 patients with a diagnosis of type 2 diabetes and complete baseline data for weight and hemoglobin A1c within relevant therapy windows, resulting in a total of 853 treatment regimens. Transfer learning was used to optimize for glucose-lowering therapies that led to successful treatment outcomes, defined as an absolute 0.3% reduction in hemoglobin A1c (when it is over 6.5%) without weight gain in patients with a BMI over 28 kg/m2. Recommendations that deviated from the guidelines were described using Shapley values and tested in digital twins for statistical significance. Four CGM-derived glucose-insulin response dynamic factors served as additional biomarkers.
    Results: Dual glycemic and weight targets were achieved in actual clinical practice in 51.2% (131/256) of cases, increasing to 54% (20/37) when clinical guidelines were followed. After selecting outcomes in the test set that followed individualized recommendations, this increased further to 58% (21/36) when using only phenotypic markers and to 65% (22/34) when adding CGM-derived dynamic factors.
    Conclusions: Tested on the limited number of patients available, our findings show that our AI model was associated with improved retrospective outcomes compared to the guidelines in complex type 2 diabetes cases by integrating multiple data sources, drawing on experiential clinical insights, and selecting treatments most likely to meet each patient's clinical targets for glucose and weight control. Future research is needed with a larger dataset.
    Keywords:  AI; artificial intelligence; continuous glucose monitoring; glucose-lowering pharmacotherapy; precision medicine; type 2 diabetes
    DOI:  https://doi.org/10.2196/92877
  10. Acad Radiol. 2026 Oct 09. pii: S1076-6332(26)00771-3. [Epub ahead of print]
    Multimodal Carotid Ultrasound-based Artificial Intelligence Model for Predicting Acute Ischemic Stroke Risk in Patients with Type 2 Diabetes Mellitus.
    Yaning Sun, Bojian Feng, Yanming Zhang, Di Ou, Dangin Pan, Saisai Wang, Fengqin Shen, Xinchen Li, Dong Xu.
       RATIONALE AND OBJECTIVES: Patients with type 2 diabetes mellitus (T2DM) are at high risk of acute ischemic stroke (AIS), yet accurate risk stratification remains challenging. This study aimed to develop and validate a multimodal model based on carotid ultrasound for improved AIS risk prediction.
    MATERIALS AND METHODS: In this multicenter retrospective study, a total of 480 patients with T2DM who underwent carotid ultrasound were recruited from two centers, with one center (n=394) used for model development and the other (n = 86) for independent external testing. A dual-scale deep learning framework based on a Swin Transformer was developed to capture both local plaque features and surrounding vascular context. Deep learning features were integrated with radiomic and clinical features to construct a multimodal model. Model performance was evaluated using the area under the curve (AUC), along with calibration and decision curve analysis.
    RESULTS: The multimodal model achieved AUCs of 0.952 (95% CI: 0.910-0.984) in the internal cohort and 0.939 (95% CI: 0.838-0.996) in the external cohort, showing improved performance compared to single-modality models. In the reader study, AI-assisted interpretation improved diagnostic performance across all readers, with AUC increases ranging from 0.760 to 0.857 and from 0.868 to 0.929 for senior readers, and from 0.450 to 0.700 and from 0.548 to 0.659 for junior readers. Generalized estimating equation analyses showed significant improvements in diagnostic correctness among both senior and junior readers, with P values of 0.00235 and 0.000343, respectively.
    CONCLUSION: A multimodal framework integrating carotid ultrasound with clinical, radiomic, and deep learning features enables improved AIS risk stratification in patients with T2DM.
    Keywords:  Acute ischemic stroke; Carotid ultrasound; Deep learning; Risk stratification; Type 2 diabetes mellitus
    DOI:  https://doi.org/10.1016/j.acra.2026.09.036
  11. J Biomed Opt. 2026 Oct;31(10): 106002
    SEFCA-Net: a semantic-edge-frequency cross-attention network for diabetic retinopathy lesion segmentation.
    Jin Xie, Jingtao Wang, Duo Xu, Yuzhong Peng, Kunlong Wu, Fangning Gu, Xue Zhang, Zhizhong Zhang, Xinjian Chen, Jianping Liao.
       Significance: Diabetic retinopathy (DR) lesion segmentation provides pixel-level lesion localization in fundus images. However, such lesions are often tiny, sparsely distributed, low-contrast, morphologically diverse, and visually similar to vessels or surrounding retinal structures, making accurate localization difficult. Many existing methods emphasize multi-scale spatial representation learning and handle boundary ambiguity and subtle texture differences mainly through encoder-decoder fusion. We therefore examine boundary and frequency representations as explicit, complementary inputs to semantic decoding.
    Aim: We aim to develop SEFCA-Net, a semantic-edge-frequency cross-attention network for diabetic retinopathy lesion segmentation. The main objective is to examine whether semantic, edge, and frequency-domain representations can be fused under semantic guidance to improve pixel-level lesion localization without compromising the semantic decoding stream. The evaluation focuses on four lesion categories, including hard exudates, hemorrhages, soft exudates, and microaneurysms, using public benchmarks and a DDR-to-IDRiD cross-dataset setting.
    Approach: SEFCA-Net uses HRNet-W48 as the semantic backbone to extract multi-scale semantic features and explicitly constructs edge and frequency-domain representations from input fundus images. The learnable multi-scale edge branch extracts boundary-sensitive features, while the all-channel multi-scale discrete cosine transform (DCT) frequency-domain branch generates texture-sensitive responses to low-contrast lesions. The semantic-edge-frequency cross-attention decoder uses the semantic representation as the primary stream and injects edge and frequency-domain features through semantic-guided window-based cross-attention. Performance is evaluated using the area under the precision-recall curve (AUPR), Dice, and Intersection-over-Union (IoU).
    Results: On DDR, SEFCA-Net achieves 47.85% mean AUPR (mAUPR), 46.76% mean Dice, and 31.22% mean IoU. On IDRiD, it reaches 69.74% mAUPR; when trained on DDR and tested on IDRiD, it obtains 59.83% mAUPR. Ablation studies show that either auxiliary branch improves the semantic-only baseline and that the complete configuration gives the highest mAUPR. Multi-scale Sobel modeling and DCT outperform the tested single-scale and wavelet alternatives, respectively.
    Conclusions: SEFCA-Net shows that semantic features can remain the main decoding stream while boundary-sensitive features and texture-sensitive responses are integrated in a controlled manner. Under the evaluated protocols, the model shows competitive in-domain performance and records the highest DDR-to-IDRiD mAUPR among the reproduced methods.
    Keywords:  complementary feature fusion; diabetic retinopathy; lesion segmentation; semantic-guided window-based cross-attention
    DOI:  https://doi.org/10.1117/1.JBO.31.10.106002
  12. J Diabetes Sci Technol. 2026 Oct 07. 19322968261484150
    Retrieval-Augmented Simulation for Interpreting Probabilistic Digital Twins in Type 1 Diabetes.
    Omer Mujahid, Ivan Contreras, Aleix Beneyto, Josep Vehi.
       BACKGROUND: Personalized digital twins for type 1 diabetes (T1D) can simulate future glucose responses under different therapeutic conditions, but their probabilistic outputs require structured interpretation before supporting clinician-facing reasoning. We introduce retrieval-augmented simulation (RAS), a modular framework connecting probabilistic digital-twin simulation with guideline-grounded report generation.
    METHODS: Retrieval-augmented simulation summarizes simulated glucose trajectories, converts them into structured clinical descriptors, retrieves relevant guideline evidence, and generates uncertainty-aware reports for clinician review. We evaluated RAS using data from 35 individuals with T1D in a controlled ablation study involving 2 locally hosted large language models (LLMs) (LLaMA 3.1:8b and Qwen 3:8b). Two independent LLM-based evaluators, GPT-5.5 Thinking and Kimi 2.6 Thinking, assessed report quality using a standardized rubric.
    RESULTS: Structured clinical descriptors were essential for guideline selection: their removal substantially altered the retrieved evidence, demonstrating that qualitative clinical framing influenced evidence selection more strongly than numerical values alone. For LLaMA, the full RAS pipeline achieved the strongest overall performance, maintaining high numerical faithfulness (NumCov: 0.98±0.03) and consistent guideline coupling. In contrast, Qwen exhibited substantial model-specific safety risks and frequently distorted simulation-derived numerical values (NumCov: [Formula: see text]). Directional bias analysis showed that Qwen systematically under-reported time in range by a mean of -10.6 percentage points and over-reported time below range, thereby inflating hypoglycemia risk. These safety findings were consistent across both independent evaluators, with a mean score difference of >2.1 points.
    CONCLUSION: Retrieval-augmented simulation provides a reproducible framework for the expert-supervised interpretation of digital-twin simulations. However, aggregate framework scores alone are insufficient for clinical validation. Downstream language models must be rigorously evaluated for numerical faithfulness and directional bias to prevent unsafe, hallucinated clinical narratives.
    Keywords:  artificial intelligence safety; clinical decision support; digital twin; glycemic control; large language models; retrieval-augmented generation; type 1 diabetes; uncertainty quantification
    DOI:  https://doi.org/10.1177/19322968261484150
  13. Front Cell Infect Microbiol. 2026 ;16 1910757
    Quantitative mapping of gut microbiota feature importance to TCM syndrome elements in clinical T2DM stages.
    Ruifang Liu, Yingrong Liu, Jiawei Jiang, Zhaoyang Yang.
       Introduction: The clinical phenotype of modern type 2 diabetes mellitus (T2DM) has diverged from the classical traditional Chinese medicine (TCM) syndrome of Xiao Ke (wasting-thirst), with obesity replacing emaciation as the dominant presentation and challenging traditional symptom-based syndrome differentiation. The TCM syndrome element system, a quantifiable framework inherently compatible with molecular-level parameters, positions the gut microbiota - causally linked to T2DM, stage-specific in composition, and amenable to targeted intervention - as an ideal molecular anchor for refining syndrome differentiation. However, the quantitative feature importance of specific gut microbes for individual syndrome elements remains undetermined.
    Methods: We enrolled 154 participants across three T2DM stages (51 pre-diabetes, 53 T2DM, 50 T2DM with complications) for 16S rDNA sequencing, syndrome element assessment, and machine learning.
    Results: Mendelian randomization identified 21 gut microbiota taxa and functional pathways causally implicated in T2DM in East Asian populations, providing biological priors for subsequent analyses. Support vector machine (SVM) models with SHAP values quantified microbial feature importance for individual syndrome elements, revealing progressive evolution from dampness, qi deficiency, and spleen (pre-diabetes) to dampness, phlegm, heat, and kidney (complications). Bacteroides and Faecalibacterium showed the highest feature importance for spleen (4.9% each), Veillonella showed the highest feature importance for dampness (3.4%), and Bifidobacterium was the top feature for phlegm (1.0%). Fecal microbiota transplantation (FMT) in 34 patients provided interventional evidence consistent with the model: 26.7% of post-FMT differentially abundant genera (4 of 15) and 7.5% of differentially abundant species (8 of 106) overlapped with top-ranking SVM features.
    Discussion: This proof-of-concept study provides the first quantitative mapping of gut microbiota feature importance to TCM syndrome elements, establishing an integrated framework combining MR-based causal prioritization, ML-based feature importance mapping, and FMT-based intervention validation for micro-syndrome differentiation.
    Keywords:  Mendelian randomization; TCM syndrome elements; fecal microbiota transplantation; gut microbiota; machine learning; type 2 diabetes mellitus
    DOI:  https://doi.org/10.3389/fcimb.2026.1910757
This page is being maintained by Thomas Krichel. It was last updated on 2026‒10‒11 at 0:09.