bims-aukdir 2026-11-01 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2026–11–01
28 papers selected by
Mott Given

Development of an explainable machine learning framework for early diabetes detection, multiclass classification, and risk stratification across age groups.
Diffusion-augmented contrastive learning for diabetic retinopathy severity classification.
Performance Assessment of an Artificial Intelligence System Used for Case Detection in a Large Community-based Diabetic Retinopathy Screening Program: A Retrospective Diagnostic Accuracy Validation Study.
Artificial Intelligence Algorithms for Early Detection and Risk Stratification of Type 2 Diabetes Mellitus, Multimodal Bioinformatic Phenotyping, and Its Complications: A Technical Review of Classical Machine Learning, Deep Learning, and Post-Hoc Feature Attribution.
Machine learning predictors of cardiovascular autonomic neuropathy in type 2 diabetes: insights from a prospectively collected single-center registry.
A personalized hybrid federated learning framework for glucose prediction in type 1 diabetes.
Assessing Diabetes-Associated Bone Changes Using Chest Computed Tomography Radiomics and Extreme Gradient Boosting: A Retrospective Study.
AI-Assisted Triage for Diabetic Retinopathy Integrated into Routine Diabetes Care: A Real-World Diagnostic Accuracy Study of the DAIRET® Algorithm.
[Construction of predictive models for type 2 diabetic retinopathy based on the principle of differentiation of syndrome elements].
Diagnostic model based on machine learning with multi-feature analysis for diabetic peripheral neuropathy in type 2 diabetes: A single-center retrospective study.
An attention-guided dense convolutional network for automated diabetic foot ulcer segmentation.
National Use of Artificial Intelligence for Eye Screening in Singapore.
Deep Learning for Multi-Disease OCT Classification in Clinical Retina Practice Using Attention Mechanisms and Contrastive Learning.
PhysioXGluNet: a physiology-informed network architecture with attributable branches for explainable blood glucose prediction.
Identification of ageing-related genes in prostate cancer with type 2 Diabetes using machine learning: a multi-omics study focusing on SOX4, PRNP, and FOXO4.
AIDCare: A Technical Proof-of-Concept for Integrating Personalized Nutrition, Physical Activity, and AI-Assisted Care in Diabetes Lifestyle Management.
GlucoSense: early diabetes risk assessment via fusion of human activity recognition and glucose prediction.
PLXNB2 Regulates Monocyte-Mediated Inflammation and Intercellular Communication in the Pathogenesis of Diabetic Retinopathy.
Development of a risk identification model for sPTB in patients with GDM based on ensemble learning algorithms.
A quantum-inspired cross-attention fusion network for multi-class eye disease classification from retinal fundus images.
Clinical Effectiveness, Safety, and Implementation of Artificial Intelligence-Assisted Insulin Management Systems in Type 1 and Type 2 Diabetes: A Systematic Review of Randomized Controlled Trials.
Short-Term Continuous Glucose Forecasting with Large-Language-Model-Derived Nutrient Estimates from Real-World Chinese Dietary Records.
Generative Forecasting of Longitudinal 3D-OCT Under Anti-VEGF Therapy for Diabetic Macular Edema.
Non-Invasive Blood Glucose Monitoring Using Machine Learning and Optical Approaches: A Systematic Literature Review and a Proposed Framework for Reproducible and Clinically Reliable Research.
Comparative efficacy of large language models (Google Gemini and ChatGPT 4.0) in screening patients for diabetic retinopathy.
Forecasting accuracy is not enough: safety-oriented evaluation of neural digital twins prior to in-silico insulin and carbohydrate intervention in type 1 diabetes.
Individualized statin associated type 2 diabetes risk estimation with a deep causal model.
Concurrent Diabetes Mellitus and Periodontitis in U.S. Adults: Mortality Associations and Machine Learning Risk Prediction.

Front Digit Health. 2026 ;8 1952669

Development of an explainable machine learning framework for early diabetes detection, multiclass classification, and risk stratification across age groups.

Nosizo Lukhele, Fahad Mostafa.

   Objective: To develop and evaluate a novel machine learning (ML) framework tailored to a clinical diabetes dataset and to assess whether demographic stratification enhances model performance and interpretability for multiclass diabetes classification.
Methods: A clinical dataset of 264 patients' records was used to classify individuals including non-diabetic (36.4%), prediabetic (15.2%), and diabetic (48.5%) individuals. Several supervised learning models were trained using 80:20 train-test split and optimized using RandomizedSearchCV and 10-fold cross validation. Model performance was evaluated using the metrics accuracy, precision, recall and the F1-score. Area under the receiver operating characteristic curve (AUC) was calculated for the best generalizing model. A structured ML framework was developed for this dataset, incorporating preprocessing, model optimization, age stratification analysis age (<35 vs ≥35 years) and gender. SHAP was developed for model interpretability.
Results: Ensemble methods demonstrated superior performance in comparison to linear or single-tree approaches, with Gradient Boosting showing the most stable generalization with a test accuracy of 0.981 and stable cross validation accuracy of 0.972. AUC-ROC analysis using Gradient Boosting yielded good discriminative ability across the three diabetes classes: 0.991 (non-diabetic), 0.986 (prediabetic) and 0.972 (diabetic). Stratified analysis showed improved reliability in individuals aged ≥35 years (accuracy = 0.94, F1-score = 0.92), while performance in younger individuals was unstable due to small sample size. SHAP analysis identified HbA1c, BMI, and age as dominant predictors.
Conclusion: This study presents a ML framework integrating age stratified modelling with explainable ML frameworks to improve interpretability. The findings offer clinically relevant results that can support clinical decision-making systems, individualized risk assessment, and potential applications for targeted intervention in diabetes progression.

Keywords:  diabetes; explainable AI; machine learning; multiclass classification; risk stratification

DOI:  https://doi.org/10.3389/fdgth.2026.1952669
Front Artif Intell. 2026 ;9 1932259

Diffusion-augmented contrastive learning for diabetic retinopathy severity classification.

Ali Abidalkareem, Ali K Ibrahim, Hanqi Zhuang.

  Automated diabetic retinopathy (DR) severity classification faces persistent challenges from limited annotated datasets, severe class imbalance across the five-level International Clinical Diabetic Retinopathy (ICDR) scale, and domain shift between imaging environments. This paper presents an integrated deep learning framework that extends our earlier Fused Ensemble Convolutional Inception Network (93.2% accuracy) through three complementary innovations. First, a class-conditional latent denoising diffusion probabilistic model (DDPM) generates anatomically realistic synthetic fundus images for underrepresented DR severity levels, addressing data scarcity and class imbalance simultaneously. Second, a SimCLR contrastive self-supervised pre-training protocol, applied to the combined real-and-synthetic corpus, learns discriminative retinal representations prior to supervised fine-tuning. Third, a five-model ensemble (EfficientNet-B3, ResNet-50, VGG16, VGG19, and DenseNet-121) fused via a learned Pattern-Net achieves improved generalization through architectural diversity. Evaluated on the APTOS 2019 dataset under 5-fold stratified cross-validation, the integrated framework achieves 95.1% accuracy, 0.928 quadratic weighted kappa (QWK), 0.91 F1-score, and 0.97 AUC. Class-wise improvements over the EfficientNet-B3 baseline reach +23.8 percentage points for Severe DR (61.8 → 85.6%) and +17.7 for Mild DR (71.2 → 88.9%). Independent zero-shot validation on EyePACS yields 78.4% accuracy, 0.809 QWK, 0.738 macro F1, and 0.887 macro AUC, quantifying the expected performance reduction under domain shift. Fold-level variability and paired out-of-fold statistical analyses are additionally reported, and Grad-CAM visualization is included for qualitative model interpretability. These results demonstrate the synergistic benefit of generative augmentation with self-supervised representation learning for imbalanced medical image classification, with implications for scalable automated DR screening in resource-limited settings.

Keywords:  class imbalance; contrastive learning; convolutional neural networks; denoising diffusion probabilistic models; diabetic retinopathy; domain shift; ensemble learning; external validation

DOI:  https://doi.org/10.3389/frai.2026.1932259
Acta Med Philipp. 2026 ;60(17): 55-67

Performance Assessment of an Artificial Intelligence System Used for Case Detection in a Large Community-based Diabetic Retinopathy Screening Program: A Retrospective Diagnostic Accuracy Validation Study.

Maria Kristina Dg Eduardo, Junn R Pajarillo.

   Background: In the Philippines, Artificial Intelligence (AI) systems used for diabetic retinopathy (DR) screening are still an emerging technology. There are several AI and Deep Learning Systems (DLS) developed and licensed abroad that are currently undergoing FDA approval in the Philippines. The challenge is to validate these systems on Filipino diabetic cohorts within the uncontrolled environment of a community-based setting before it could be adapted and integrated into routine DR screening programs locally.
Objective: This study aimed to determine the performance of an AI system as a population-based mass screening tool to detect referable cases of DR.
Methods: This study was designed as a retrospective diagnostic accuracy validation study using fundus images obtained from a community-based screening program by the Ilocos Training & Regional Medical Center Eye Mobile Team in the province of La Union. The diagnostic accuracy of the iCare Illume System (Thirona RetCAD + iCare DRS plus) in detecting any stage DR (ADR) and referable DR (RDR) was validated against a reference standard of two-field fundus photographs read by expert human graders (retina specialists). The diagnostic performance was evaluated using various metrics, including Accuracy, Sensitivity (SN), Specificity (SP), Negative Predictive Value (NPV), Positive Predictive Value (PPV), Negative Likelihood Ratio (LR-), and Positive Likelihood Ratios (LR+).
Results: A total of 284 eyes (142 diabetic patients, 717 images) were included in the study. Human expert graders classified 187 eyes (65.85%) as no apparent DR while 77 eyes (27.11%) as ADR. 63 eyes (22.18%) were considered RDR. The iCare Illume AI system achieved an accuracy of 83.21 % and 92.37% for ADR and RDR, respectively. The SN, SP, NPV, PPV, LR-, and LR+ for ADR were 92.11%, 79.57%, 96.1%, 64.81%, 0.10, and 4.50, respectively. The SN, SP, NPV, PPV, LR-, and LR+ for RDR were 79.03%, 96.50%, 93.69%, 87.50%, 0.22, and 22.58, respectively. Technical failure rate was 0.70% (99.3% Imageability).
Conclusion: The iCare Illume system exhibits strong diagnostic accuracy in screening for referable DR in our cohort of Filipino Diabetics. Although its sensitivity is marginally below the UK NICE guideline, the system demonstrates high specificity, negative predictive value, and positive predictive value. This combination highlights its potential as a valuable tool for diabetic retinopathy screening in the Philippines. Future research should focus on refining local AI validation protocols, optimizing data sets of commercially available AI systems, assessing the impact on patient outcomes and disease burden, and evaluating cost-effectiveness.

Keywords:  Icare Illume system; Mulat Mata DR Screening Program; artificial intelligence; diabetic retinopathy

DOI:  https://doi.org/10.47895/amp.v60i17.13381
Bioengineering (Basel). 2026 Sep 28. pii: 1132. [Epub ahead of print]13(10):

Artificial Intelligence Algorithms for Early Detection and Risk Stratification of Type 2 Diabetes Mellitus, Multimodal Bioinformatic Phenotyping, and Its Complications: A Technical Review of Classical Machine Learning, Deep Learning, and Post-Hoc Feature Attribution.

Kattia Orozco-Romero, Jonas Pariona-Torres, Jose Cornejo.

  Type 2 diabetes mellitus (T2DM) affects hundreds of millions of people worldwide, and nearly half of all cases remain undiagnosed. This technical review, conducted under conducted under the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines and structured around the Population, Intervention, Comparison, Outcome and Context (PICOC) framework, identifies and synthesises the evidence on artificial intelligence (AI) algorithms for the early detection of T2DM and the risk stratification of its complications, evaluating their performance against conventional diagnostic methods and characterising their clinical implementation contexts. A systematic search in Scopus with a four-stage screening process, synthesised across five research dimensions, yielded 255 records, of which 47 studies met the inclusion criteria (2025-2026). Ensemble methods based on gradient boosting, particularly particularly extreme gradient boosting (XGBoost), emerged as the dominant primary modelling approach, while logistic regression was the most frequently reported algorithm, predominantly as a baseline comparator, with area under the receiver operating characteristic curve (AUC-ROC) values ranging from 0.63 to 0.97. AI models consistently outperformed glycated haemoglobin (HbA1c), whose reported sensitivity for prediabetes was only 49%, reaching accuracies of up to 96.7%, although that upper figure comes from a narrowly defined cohort with a high baseline prevalence and does not represent general screening performance. Multimodal approaches achieved the highest discrimination, combining wearables, continuous glucose monitoring, and genomics (AUC 0.96 internal, 0.90 external, for the separation of self-reported glycaemic states), and a real-world clinical deployment integrated into electronic health records improved glycaemic control outcomes among high-risk patients. An assessment of the eight anchor studies with thePrediction model Risk Of Bias ASsessment Tool updated for AI (PROBAST + AI) found high concern regarding quality in six of them, driven mainly by cross-sectional designs and self-reported outcomes. Overall, AI stands out as a viable and scalable complement to conventional screening, although gaps remain in multicentre prospective validation, reporting standardisation, and algorithmic equity across populations.

Keywords:  artificial intelligence; deep learning; diabetes prediction; diabetic complications; early detection; feature selection; machine learning; predictive models; risk stratification

DOI:  https://doi.org/10.3390/bioengineering13101132
Front Endocrinol (Lausanne). 2026 ;17 1965013

Machine learning predictors of cardiovascular autonomic neuropathy in type 2 diabetes: insights from a prospectively collected single-center registry.

Katalin Farkas, Viktor Vedelek, Máté Vámos, Éva Csajbók, István Baczkó, János Zádori, Andrea Szentesi, Péter Hegyi, Tamás Várkonyi, Csaba Lengyel, Anna Vágvölgyi.

   Background: Cardiac autonomic neuropathy (CAN) is a severe complication of type 2 diabetes (T2DM) associated with increased cardiovascular morbidity and mortality. Machine learning (ML) can integrate multidimensional clinical data and identify features associated with CAN. We aimed to characterize predictors of CAN categories using ML.
Methods: Data were obtained from a prospective single-center observational registry of patients with T2DM undergoing standardized cardiovascular autonomic reflex tests (CARTs) based on Ewing's protocol. Demographic, anthropometric, clinical, laboratory, and medication data were analyzed. Patients were classified as CAN0 (no or borderline CAN), CAN1 (early CAN), or CAN2 (definite or severe CAN). Three extreme gradient boosting models were evaluated: a CART-inclusive model, a CART-excluded model restricted to patients with T2DM, and a CART-excluded model supplemented with non-diabetic controls without abnormal CART results. Model performance was assessed using area under the receiver operating characteristic curve (AUC), and feature contributions were interpreted using SHapley Additive exPlanations values.
Results: Among 693 patients with T2DM, 19 (2.7%) had completely normal autonomic function, whereas 674 (97.3%) exhibited at least one borderline or abnormal CART result. Greater CAN severity was associated with older age, longer diabetes duration, higher systolic blood pressure, and poorer renal function. The CART-inclusive model achieved perfect discrimination in the best iteration (weighted AUC, 1.00; accuracy, 100%), with all five CARTs ranking among the leading predictors. After CART exclusion, performance declined markedly (weighted AUC, 0.64), with class-specific AUCs of 0.59 for CAN0, 0.59 for CAN1, and 0.68 for CAN2. Inclusion of non-diabetic controls improved the weighted AUC to 0.78, with the greatest improvement for CAN0 (AUC, 0.85), followed by CAN2 (0.78) and CAN1 (0.70). Age, diabetes duration, and renal function markers were recurrent non-CART predictors, while anthropometric, hematological, metabolic, and treatment-related features showed category- and model-specific contributions. Uniform Manifold Approximation and Projection demonstrated substantial overlap among CAN categories, particularly for CAN1.
Conclusions: CARTs remained essential for accurate CAN classification and staging. Routinely available clinical variables provided complementary phenotypic information but were insufficient to robustly distinguish CAN categories without CART data, particularly CAN0 and CAN1. ML may support multidimensional phenotyping and hypothesis generation but cannot replace standardized autonomic testing.

Keywords:  T2DM; autonomic score; cardiovascular autonomic neuropathy; diabetes duration; machine learning

DOI:  https://doi.org/10.3389/fendo.2026.1965013
Artif Intell Med. 2026 Oct 26. pii: S0933-3657(26)00213-7. [Epub ahead of print]183 103561

A personalized hybrid federated learning framework for glucose prediction in type 1 diabetes.

Carlos Gallardo-García, Clara Fuertes, M Elena Hernando, David Subías, Mercedes Rigla, Gema García-Sáez.

  Deep learning methods for glucose prediction in type 1 diabetes (T1D) are often constrained by data availability, restricted access to data, and substantial heterogeneity across treatment settings and institutions. Federated Learning offers a privacy-preserving alternative to centralized training, but its performance under clinically realistic, non-independent and non-identically distributed (non-IID) data remains challenging. This work proposes a hybrid federated-personalized framework that combines cross-silo training with patient-level personalization for multi-horizon glucose prediction. Clinically defined silos based on treatment modality and insulin type are used to explicitly model real-world heterogeneity. A Temporal Fusion Transformer is trained using federated optimization (FedProx), incorporating clinician-prescribed basal insulin profiles as known future inputs. The resulting global model is subsequently adapted at the patient level, enabling personalization. The proposed approach is evaluated on the T1-DEXI dataset and externally validated on the OhioT1DM cohort. Centralized and federated personalization are compared to quantify performance differences between in-distribution and external evaluation and to assess patient-specific adaptation. Results show that the federated global models reduce prediction errors at longer horizons compared to centralized training on unseen data distributions, whereas centralized training achieves lower in-distribution errors. After patient-level personalization, centralized models show slightly lower absolute errors than federated models. These findings are consistent with a balance between population-level behavior and individual adaptation under the evaluated configurations. They support the feasibility of combining cross-silo training with patient-level personalization for privacy-preserving glucose prediction in heterogeneous clinical environments. Furthermore, incorporating scheduled basal insulin profiles enhances in-distribution prediction accuracy, underscoring the importance of clinically meaningful inputs in model design.

Keywords:  Temporal fusion transformer; Type 1 diabetes; attention models; deep learning; federated learning

DOI:  https://doi.org/10.1016/j.artmed.2026.103561
J Vis Exp. 2026 Oct 23.

Assessing Diabetes-Associated Bone Changes Using Chest Computed Tomography Radiomics and Extreme Gradient Boosting: A Retrospective Study.

Ting Li, Nadeer M Gharaibeh, Yunfeng Han, Gang Wu.

Chest computed tomography (CT) provides quantitative information on bone that may reveal diabetes-associated skeletal changes not readily identified by routine image assessment. This study aimed to establish a prediction model for diabetes using chest CT radiomics features and identify bone features with the highest contributions to diabetes prediction. One hundred patients with diabetes and 100 healthy controls underwent chest CT. The thoracic vertebrae, scapulae, and sternum were automatically segmented using a convolutional neural network (CNN), and 93 radiomics features were extracted from each structure. Max-Relevance and Min-Redundancy (mRMR), least absolute shrinkage and selection operator (LASSO), and extreme gradient boosting (XGBoost) were successively used for feature selection and model development. The area under the receiver operating characteristic curve (AUC) was evaluated using internal-validation data and an external-test cohort of 50 participants, including 25 patients with diabetes. Shapley Additive Explanations (SHAP) were used to interpret the XGBoost model. After mRMR, and LASSO, 17 features were selected. The model achieved AUCs of 0.95 and 0.90 on the internal-validation and external-test data, respectively. gldm_SmallDependenceLowGrayLevelEmphasis_T11, glcm_Idmn_T8, and gldm_LargeDependenceHighGrayLevelEmphasis_T4 showed the highest contributions to diabetes prediction in the SHAP analysis. Glycated hemoglobin (HbA1c) levels correlated with several bone CT features. These findings demonstrate that chest CT radiomics combined with interpretable machine learning can characterize diabetes-associated bone changes and distinguish patients with diabetes from healthy controls.

DOI: https://doi.org/10.3791/74212
Diagnostics (Basel). 2026 Oct 21. pii: 3422. [Epub ahead of print]16(20):

AI-Assisted Triage for Diabetic Retinopathy Integrated into Routine Diabetes Care: A Real-World Diagnostic Accuracy Study of the DAIRET® Algorithm.

Alberto Vassallo, Alessio Delfino, Greta Ghizzardi, Andrea Carbone, Gloria Groppelli, Gisella Boselli, Noemi Busca, Matteo Cerutti, Sara Lodigiani, Caterina Golinelli, Guido Grignaffini, Valeria Guazzoni.

  Background: Screening for diabetic retinopathy (DR) is resource-intensive. We evaluated the real-world diagnostic accuracy of the DAIRET® Artificial Intelligence (AI) algorithm, integrated into the electronic medical record, as a triage tool in routine diabetes care. Methods: Retrospective, STARD-AI-compliant study of consecutive adults with diabetes screened with non-mydriatic fundus photography between April and December 2025. The DAIRET® output ("suspected"/"not suspected") was compared with grading by an ophthalmologist masked to the AI result. Sensitivity, specificity and predictive values with exact 95% confidence intervals were calculated for any DR and referable (moderate or worse) DR. Results: Among 1051 patients, 185 (17.6%) had DR: 160 mild and 25 moderate non-proliferative DR; none had severe DR. DAIRET® classified 271 patients (25.8%) as suspected. For any DR, sensitivity was 59.5% (52.0-66.6) and specificity 81.4% (78.7-83.9). For referable DR, sensitivity was 100.0% (86.3-100.0), specificity 76.0% (73.3-78.6) and negative predictive value 100.0% (99.5-100.0). Specificity was lower in patients aged <65 years and with type 1 diabetes. Conclusions: No referable case was missed, although sensitivity was based on 25 cases (lower limit 86.3%) and 15 AI-negative patients lacked complete verification. The 74% classified as "not suspected" is a potential, not measured, reduction in referrals. Performance for severe or proliferative DR could not be assessed, and multi-center confirmation is needed.

Keywords:  STARD-AI; artificial intelligence; diabetes; diabetic retinopathy; diagnostic accuracy; screening; triage

DOI:  https://doi.org/10.3390/diagnostics16203422
Nan Fang Yi Ke Da Xue Xue Bao. 2026 Oct 20. pii: 1673-4254(2026)10-2456-15. [Epub ahead of print]46(10): 2456-2470

[Construction of predictive models for type 2 diabetic retinopathy based on the principle of differentiation of syndrome elements].

Shijie Qiao, Jinshun You, Chang Wang, Hongping Xiong, Shunan Li.

   OBJECTIVES: To construct a prediction model for occurrence of type 2 diabetic retinopathy (T2DR) by combining the principle of differentiation of syndrome elements with machine learning (ML).
METHODS: Cross-sectional data were collected from 273 patients with T2DR and 376 T2DM patients for comparison of routine test results and parameters of syndrome elements in traditional Chinese medicine (TCM) identify the candidate TCM indicators for predicting T2DR occurrence. LASSO regression analysis was used to select the syndrome features with the highest predictive value. Six ML algorithms (LR, ANN, XGBoost, RF, SVM and KNN) were used to construct the prediction models, followed by internal validation of the model and evaluation of the importance of each predictive feature.
RESULTS: LASSO analysis identified the liver, lung, and yang excess syndrome elements, among others, as the most predictive features for T2DR occurrence. Among the 6 ML models, LR and KNN models showed good discrimination power. The LR model yielded highly reliable and robust probability output in global calibration, and the KNN model had good performance for probability predictions. Both the LR and KNN models provided positive net clinical benefits over a wide range of threshold probabilities, but the net benefit curve of the LR model was above that of the KNN model in the whole range, especially in the validation set.
CONCLUSIONS: The ML-based prediction models (especially the ML model) combining the features of syndrome elements effectively use the advantage of TCM for predicting disease and allow comprehensive prediction of T2DR in T2DM patients to assist clinical decision-making for early interventions.

Keywords:  Traditional Chinese medicine; machine learning; prediction model; syndrome elements; type 2 diabetic retinopathy

DOI:  https://doi.org/10.12122/j.issn.1673-4254.2026.10.15
J Diabetes Investig. 2026 Oct 30.

Diagnostic model based on machine learning with multi-feature analysis for diabetic peripheral neuropathy in type 2 diabetes: A single-center retrospective study.

Xi Zheng, Keke Yang, Ruyi Qu, Xiaomin Peng, Ling Yang.

   OBJECTIVE: Diabetic peripheral neuropathy (DPN) is a common and serious complication of diabetes, which imposes a significant economic burden on both society and individuals. Therefore, early identification and diagnosis are crucial for the prevention and treatment of DPN. This study aims to develop machine learning (ML) models based on a variety of different characteristics to predict the risk of DPN in patients with type 2 diabetes mellitus.
METHODS: We retrospectively reviewed the data of 770 patients screened for DPN with type 2 diabetes mellitus in the Department of General Practice, Jiangwan Town Subdistrict Community Health Service Center between May 2024 and July 2024. The data included 91 variables for baseline description analysis. The dataset was divided into a training and test set for the model evaluation in a 7:3 ratio, and feature selection was performed with the Boruta algorithm. Eleven ML algorithms were used to develop diagnostic models for DPN. The area under the receiver operating characteristic curve, calibration curves, and decision curves analysis were applied to evaluate the performance of the models, and the Shapley Additive Explanations (SHAP) method was adopted to interpret the performance of the models and the importance of features.
RESULTS: The study included 770 type 2 diabetes mellitus inpatients, with a mean age of 70.88 ± 5.88 years, and 26.62% (205/770) had DPN. Integrated with the results of evaluation, we selected logistic regression (LR), extreme gradient boosting (XGBoost), and light gradient boosting machine (LGBM) models including neurological examinations as the top-performing algorithms. Based on variable importance ranking from the three models with overlap analysis, we identified 12 key predictors. Among the three models, the tuning fork test, ankle reflex, and temperature sensation test were all identified as the top three important predictors. Meanwhile, the low-density lipoprotein cholesterol (LDL-C) in the XGBoost model and the total cholesterol (Tch) in the LGBM were the important features for predicting DPN. In addition, excluding neurological examination findings, both LDL-C and Tch were important indicators with high contribution degree.
CONCLUSIONS: The ML approach helped establish diagnostic models with stable performance. Meanwhile, the use of interpretable SHAP methods can enhance the clinical applicability of the models.

Keywords:  Diabetic peripheral neuropathy; Risk factors; machine learning

DOI:  https://doi.org/10.1111/jdi.70473
Front Endocrinol (Lausanne). 2026 ;17 1940557

An attention-guided dense convolutional network for automated diabetic foot ulcer segmentation.

Rampriya R S, Sabari Nathan, Arumuga Arun R.

   Introduction: A large number of diabetic patients have diabetic foot ulcers (DFU), and it is essential to segment the regions of the ulcer in a timely and accurate manner to treat them. The segmentation task is very difficult due to the highly irregular shape of the ulcers and very small areas of lesions, up to 0.04% of the DFU image.
Methods: We aim to develop an attention-guided dense convolutional network, called UlcerNet, which is able to extract fine-grained ulcer features at different scales. It introduces dense convolutions, multi-scale residual attention blocks (MRADB), and modified receptive-field modules (MRFB) to enhance feature propagation and context modeling. We use the public DFUC 2022 dataset that contains 4000 close-up foot images to train and evaluate the UlcerNet model.
Results: Our model is able to perform a Dice of 73.50% and mIoU of 76.04%, which gives better performance than the standard segmentation baselines (FCN8 49.93%, SegNet 47.68%, DeepLabV3 + 62.77%) and recent DFU segmentation baselines (TranUNet 42.61%, HarDNet-DFUS 59.81%, UFOS-Net 73.09%).
Discussion: An ablation study further shows that each component, such as dense blocks, channel attention, spatial attention, MRADB, and MRFB, is significant for the accuracy. These findings show that UlcerNet offers good and precise DFU segmentation, which can be useful for clinical settings in the real-world.

Keywords:  attention mechanism; deep learning; diabetic foot ulcer; medical image segmentation; multi-scale feature learning

DOI:  https://doi.org/10.3389/fendo.2026.1940557
NEJM AI. 2024 Dec;1(12): AIcs2400404

National Use of Artificial Intelligence for Eye Screening in Singapore.

Dinesh Visva Gunasekeran, Steven Miller, Wynne Hsu, Mong Li Lee, Hon Tym Wong, Mun Tuck Lee, Ecosse Lamoureau, Daniel Shu Wei Ting, Gavin Siew Wei Tan, Tien-Yin Wong.

AbstractDiabetes is a major health care challenge, affecting 10% of the global population. One third of patients with diabetes have an ocular complication known as diabetic retinopathy (DR). DR progression to manifestations such as vision-threatening diabetic retinopathy (VTDR) remains the leading cause of blindness in working-aged adults. Yearly DR screening is a universally recommended practice in primary care settings for patients with diabetes, but it is often difficult to implement due to a lack of staffing and screening capacity in primary care. This case study highlights our experience with developing a medical artificial intelligence (AI) software-as-a-medical-device (SaMD) solution for DR screening and implementing it at a national level to provide the capacity needed for DR screening in Singapore. Our approach involved two broad phases. First, we established a national telemedicine screening program, Singapore Integrated Diabetic Retinopathy Program (SiDRP), for population screening of DR in primary care run by trained, nonclinician human graders. Second, we deployed a deep learning-based AI solution, Singapore Eye Lesion Analyzer (SELENA+), into the SiDRP to scale-up the DR screening process by the human graders. We demonstrated the cost-effectiveness of this solution, and obtained medical device regulatory approval for clinical use in health care settings. We report the prospective evaluation of SELENA+ in SiDRP using real-world pilot data from the first 1712 patients consecutively recruited. Sensitivity and specificity of SELENA+ in detection of referable DR cases were 94.7% (95% confidence interval [CI] 88.0% to 98.3%) and 82.2% (95% CI 80.8% to 83.5%), respectively. In comparison, sensitivity and specificity of human graders were 98.9% (95% CI 94.0% to 99.9%) and 97.2% (95% CI 96.6-97.8%), respectively. For patients with VTDR, SELENA+ demonstrated a substantial advantage of higher sensitivity compared with human performance, reflecting the benefit of the fine-tuning of SELENA+ that we performed to enhance the AI solution's ability to detect VTDR. We outline the clinical, technical, operational, regulatory, and governance challenges encountered as well as the lessons learnt in this AI algorithm implementation journey. We also present a conceptual framework with considerations and strategies for the broader adoption of medical AI SaMD solutions in the field of ophthalmology and beyond.

DOI: https://doi.org/10.1056/AIcs2400404
Healthc Technol Lett. 2026 Jan-Dec;13(1):13(1): e70106

Deep Learning for Multi-Disease OCT Classification in Clinical Retina Practice Using Attention Mechanisms and Contrastive Learning.

Khosro Rezaee, Safoura Farsi, Hossein Ghayoumi Zadeh.

  Artificial intelligence (AI) is increasingly used in ophthalmic imaging to support early detection of vision-threatening retinal diseases such as age-related macular degeneration, diabetic retinopathy (DR) and choroidal neovascularisation (CNV). However, optical coherence tomography (OCT)-based deep learning models may remain sensitive to image degradation, struggle to resolve subtle inter-class morphological differences and show reduced reliability beyond controlled benchmark settings. We propose AESNet-SEC, a lightweight attention-enhanced SqueezeNet framework integrating multi-scale dilated convolutions, spatial attention, squeeze-and-excitation recalibration and contrastive representation learning. The framework was evaluated on the Retinal OCT-C8 dataset (24,000 images; eight classes) and a prospectively collected clinical dataset from Zeyaei Hospital, Iran (9732 OCT B-scans; four classes). AESNet-SEC achieved 98.56% accuracy on OCT-C8 and 96.33%-96.98% across four repeated clinical evaluations. Ablation and ROC analyses showed the strongest overall discrimination for the complete configuration, while controlled Gaussian-noise testing revealed greater degradation for DR and drusen. Grad-CAM visualisations highlighted anatomically plausible retinal regions, with independent retinal-specialist review providing qualitative support for the observed activation patterns. Overall, AESNet-SEC combines strong multi-class discrimination, computational efficiency, controlled robustness and visual interpretability, while broader patient-disjoint, multi-centre validation remains necessary.

Keywords:  artificial intelligence; deep learning; diabetic retinopathy; optical coherence tomography; retinal diseases

DOI:  https://doi.org/10.1049/htl2.70106
Med Biol Eng Comput. 2026 Oct 29.

PhysioXGluNet: a physiology-informed network architecture with attributable branches for explainable blood glucose prediction.

Shuang Wen, Hongru Li, Xiaoyu Sun.

  For people with type 1 diabetes (T1D), poor management of blood glucose (BG) can lead to hyperglycemia and hypoglycemia, which are the main culprits of health risks. Blood glucose prediction plays a crucial role in BG management. Patients can understand future BG change trends and take actions in advance to maintain euglycemia. Multivariate prediction models based on deep learning have achieved outstanding performance, but insufficient consideration of physiological consistency and physiological explainability has led to a lack of clinical confidence. To address this, we propose a novel physiology-informed network architecture with attributable branches, named PhysioXGluNet, which structurally integrates domain knowledge and individual data. To ensure physiological consistency, we determine the branch composition according to the physiological compartment model and also design the physiology-informed recurrent neural network and physiology-informed composite loss. To enhance physiological explainability, PhysioXGluNet not only outputs future BG levels but also provides the net BG changes caused by each input variable via corresponding attributable branch, allowing a clear traceability of the dominant factors at each time step driving future BG changes. PhysioXGluNet outperforms baseline models in both numerical and clinical metrics when evaluated on the real-world dataset. PhysioXGluNet has the potential to assist clinical decision-making in T1D BG management for this studied cohort.

Keywords:  Blood glucose prediction; Deep learning; Knowledge and data hybrid-driven model; Physiological explainability; Type 1 diabetes

DOI:  https://doi.org/10.1007/s11517-026-03693-y
Artif Cells Nanomed Biotechnol. 2026 Dec;54(1): 890-907

Identification of ageing-related genes in prostate cancer with type 2 Diabetes using machine learning: a multi-omics study focusing on SOX4, PRNP, and FOXO4.

Wei Zhang, Xueqian Zhang, Chenghao Zhang, Xinyang Shi, Yiheng Deng, Yan Gao, Kaipeng Bi, Zheng Wang, Guanqiu Chen, Zhoujie Ye, Ziyan An, Weijun Fu.

  Prostate cancer (PCa) demonstrates complex associations with type 2 diabetes mellitus (T2DM) and ageing, though their interconnected molecular mechanisms remain poorly defined. This study aimed to identify and validate diagnostic biomarkers at the intersection of these pathways to enable early detection and risk stratification in high-risk populations. Bioinformatic analysis was performed on PCa and T2DM transcriptomic data downloaded from the GEO database, followed by differential expression analysis, weighted gene co-expression network analysis (WGCNA), and functional enrichment. Eight machine learning algorithms were applied to identify hub ageing-related genes (ARGs), with SHAP analysis interpreting feature importance. A nomogram model was constructed and validated across multiple cohorts. Immune infiltration, drug sensitivity, single-cell RNA sequencing, and spatial transcriptomics were used for mechanistic exploration. Three key ARGs (SOX4, PRNP, and FOXO4) were identified and used to develop a diagnostic model with high accuracy (AUCs: 0.726-0.905). SOX4 promoted immune evasion, PRNP exhibited context-dependent duality, and FOXO4 functioned as a tumor suppressor. Spatial and single-cell analyses revealed cell-type-specific expression and fibroblast-epithelial communication via MIF, LRP1, and FZD1. Drug sensitivity profiling identified abiraterone and docetaxel as potential therapeutics for high- and low-risk patients, respectively. In conclusion, SOX4, PRNP, and FOXO4 are key biomarkers linking PCa, T2DM, and ageing, and the diagnostic model shows robust performance and clinical utility, providing insights into microenvironment-driven mechanisms and personalized treatment strategies for PCa patients with T2DM.

Keywords:  Prostate cancer; Type 2 diabetes mellitus; ageing-related genes; diagnostic model; machine learning; tumor microenvironment

DOI:  https://doi.org/10.1080/21691401.2026.2737622
Nutrients. 2026 Oct 13. pii: 3353. [Epub ahead of print]18(20):

AIDCare: A Technical Proof-of-Concept for Integrating Personalized Nutrition, Physical Activity, and AI-Assisted Care in Diabetes Lifestyle Management.

Adnan Kavak, Muhammad Jamil, Halil Bahadır Arın, Alpaslan Burak İnner, Hossein Fotouhi, Emre Gezer, Özlem Alkan, Zeynep Cantürk, Alev Selek, Gautam Srivastava.

  Background/Objectives: Medical nutrition therapy and structured lifestyle self-management are fundamental to diabetes care, yet conventional mobile health tools typically track diet, physical activity, and glucose measurements in isolation, with limited connection to clinical oversight. This study presents AIDCare, an integrated mobile health platform that connects camera-based nutritional assessment, wearable physical activity tracking, physiological monitoring, and clinician-supervised conversational artificial intelligence (AI) to support personalized dietary and lifestyle self-management in diabetes. Methods: AIDCare links a patient mobile application with wearable sensors, backend analytical services, and an interactive clinician portal. Dietary assessment uses smartphone camera imaging to identify foods and estimate meal contents, specifically carbohydrates, protein, and dietary fiber, against nutritional reference data. Physical activity and energy expenditure are captured through smartphone and smartwatch motion sensors. To support contextual lifestyle guidance, an open-access medical conversational language model was adapted using 24,612 structured digital health and diabetes educational articles. A clinician-in-the-loop mechanism allows healthcare providers to inspect AI-generated recommendations and adjust personalization parameters. System feasibility was evaluated through benchtop and system-level testing without patient recruitment, using synthetic patient profiles across 400 data-acquisition pipeline trials and an intent-classification benchmark spanning nutrition, activity, glucose monitoring, and physician consultation. Results: Across the 400 software-pipeline trials, the platform achieved data acquisition, transmission, and processing completion rates of 97.5%, 96.5%, and 95.5%, respectively, with backend response times under 1.8 s. In the recommendation-intent benchmark, the medical language model achieved 81.0% accuracy, 83.5% precision, 79.0% recall, and an F1-score of 81.2%. The platform successfully linked image-based meal nutrient estimates (carbohydrate, protein, and dietary fiber) with physical activity metrics and physiological tracking within a unified interface under clinician supervision. Conclusions: AIDCare establishes the technical feasibility of a unified digital platform that integrates dietary nutrient estimation, physical activity tracking, and clinician-guided AI decision support for diabetes self-management. By bridging daily nutritional intake with metabolic tracking and professional oversight, the architecture provides a functional framework for personalized lifestyle care. Prospective clinical trials in patient cohorts are required to establish nutritional estimation accuracy against weighed food records, long-term user adherence, and clinical efficacy on glycemic control.

Keywords:  AIDCare; Internet of Medical Things (IoMT); Medical Large Language Models; clinician in the loop; diabetes lifestyle management; mobile health (mHealth); personalized nutrition; physical activity monitoring

DOI:  https://doi.org/10.3390/nu18203353
Front Digit Health. 2026 ;8 1932366

GlucoSense: early diabetes risk assessment via fusion of human activity recognition and glucose prediction.

Md Fokrul Islam Khan, Md Zia Uddin, Mohammed Shamsul Alam, Adel Baselizadeh, Abdul Kadar Muhammad Masum.

  Non-Insulin-Dependent Diabetes Mellitus (NIDDM) is a chronic and progressive metabolic disease characterized by chronically high blood glucose levels. It is a disease with an often asymptomatic onset, which means that a high percentage of affected individuals remain unidentified until quite far in the disease process, thereby underlining the importance of a good, reliable screening instrument for proactive detection. In this paper, GlucoSense is introduced as a single framework for early NIDDM risk stratification by combining Human Activity Recognition (HAR) with continuous blood glucose forecasting using two Graph Attention Network version 2 (GATv2)-based architectures. The first is a HAR module trained with smartphone triaxial accelerometer data that covers 23 activity classes, while the second is a glucose prediction module trained with the OhioT1DM dataset to predict glucose levels for a 30 min horizon. Unlike sequential models, GlucoSense encodes complex dependencies between movement patterns, cardiac activity, nutrition, and glycemic behaviour by modeling relationships in a graph structure. The study was longitudinal, spanning 90 days, with six participants. Three important clinical parameters, mean glucose level, time above hyperglycemic level, and glycemic coefficient of variance were calculated. The system was able to correctly identify metabolically healthy individuals and those with prediabetic patterns. The framework was cross-validated with HbA1c laboratory tests to verify the clinical consistency of the framework. Both modules were benchmarked against sequential, convolutional, attention-based, and graph-based baselines under an identical protocol with five independent runs. The glucose module lowered the root mean squared error from 31.82 mg/dL for an LSTM baseline to 25.60 mg/dL, and the activity module reached 98.50% accuracy, with every difference confirmed by paired significance testing and reported with 95% confidence intervals. Understanding glucose dynamics based on smartphone sensor data and a limited number of physiological markers, GlucoSense offers a non-invasive, scalable and cost-effective alternative to traditional clinical glucose screening. This paper demonstrates the feasibility of smartphone-based longitudinal diabetes risk assessment and motivates future large-scale population studies.

Keywords:  NIDDM; continuous glucose monitoring; early diabetes diagnosis; graph attention network; human activity recognition; type 2 diabetes

DOI:  https://doi.org/10.3389/fdgth.2026.1932366
Curr Med Chem. 2026 Oct 01.

PLXNB2 Regulates Monocyte-Mediated Inflammation and Intercellular Communication in the Pathogenesis of Diabetic Retinopathy.

Qin Wang, Peiran Zhang, Fengtao Ji, Yongrong Li, Peirong Lu.

   INTRODUCTION: Diabetic retinopathy (DR) is a severe microvascular complication of diabetes, with a complex pathogenesis in which inflammatory responses and abnormal intercellular communication play pivotal roles. As crucial immune cells, monocytes play a central role in the inflammatory processes of DR by secreting various signaling molecules and interacting with other retinal cells, thereby jointly driving disease progression. However, the key molecules that regulate monocyte function and intercellular communication in DR remain incompletely elucidated. This study aims to investigate the role of PLXNB2 in the pathogenesis of DR, particularly its regulatory effects on monocyte-mediated inflammation and intercellular communication. The novelty of this study lies in the integration of WGCNA, differential expression analysis, machine learning, SHAP interpretation, single-cell transcriptomics, and cell-cell communication analysis to systematically identify PLXNB2 as a potential regulator of monocyte-associated inflammation in DR.
METHODS: This study employed bioinformatics approaches to systematically identify candidate genes associated with DR. WGCNA was leveraged to construct gene co-expression networks using DR and NDR samples, and differential gene expression analysis was performed to identify gene modules and key genes associated with DR. Subsequently, functional enrichment analysis was conducted on the relevant genes. Machine learning, in conjunction with SHAP value analysis, was employed to accurately screen for key genes driving the DR phenotype. To further investigate cellular mechanisms, single-cell transcriptomic and cell-cell communication analyses were performed, focusing on monocyte characteristics and intercellular signaling in DR patient samples. Ultimately, this study identified and focused on PLXNB2 as a candidate molecule potentially involved in monocyte-mediated inflammation and pathological intercellular communication in DR.
RESULTS: After comprehensive gene screening using WGCNA, machine learning, and single-cell transcriptomics, PLXNB2 was found to be the most important gene. PLXNB2 showed strong diagnostic relevance for distinguishing DR from NDR samples and was identified by SHAP analysis as a major contributor to the model prediction. It was discovered that its expression level was closely related to the improvement of Diabetic Retinopathy (DR). Single-cell analysis showed that PLXNB2 significantly altered the molecular landscape of monocytes, and there were clear enhancements in the expression patterns of their genes. Single-cell transcription may have an association with Diabetic Retinopathy (DR), but it does not demonstrate that PLXNB2 causes transcriptional remodeling.
DISCUSSION: Our findings indicate that PLXNB2 can serve as a diagnostic biomarker and a potential therapeutic target for DR. Mechanistically, PLXNB2 may regulate monocyte communication via the ANXA1-FPR1 ligand-receptor pair, mediating anti-inflammatory pathways to inhibit inflammation and cell migration, and participating in apoptotic clearance as well as immune cell chemotaxis and activation, thereby providing mechanistic evidence for PLXNB2 as a therapeutic target for DR.
CONCLUSION: This integrative multi-omics analysis identifies PLXNB2 as a promising monocyte- associated biomarker linked to inflammatory regulation and altered intercellular communication in diabetic retinopathy.

Keywords:  Diabetic retinopathy; PLXNB2; machine learning; monocyte; single cell RNA sequencing

DOI:  https://doi.org/10.2174/0109298673487078260917082010
BMC Pregnancy Childbirth. 2026 Aug 04. pii: 1246. [Epub ahead of print]26(1):

Development of a risk identification model for sPTB in patients with GDM based on ensemble learning algorithms.

Jie Hao, Ying Rong, Jingru Xue, Lifang Cheng, Hongwei Ling.

   BACKGROUND: The global prevalence of gestational diabetes mellitus (GDM) continues to rise, and GDM significantly increases the risk of spontaneous preterm birth (sPTB), posing serious threats to maternal and neonatal health. Existing models for assessing sPTB risk in patients with GDM remain limited in accuracy, nonlinear relationship capture, and the handling of complex feature interactions. This study aimed to systematically evaluate the performance differences between eight individual machine learning algorithms and a Stacking ensemble learning architecture, and to achieve risk identification and stratification through optimized model-building strategies.
METHODS: Clinical data were collected from 661 patients with GDM who were hospitalized for delivery at the Affiliated Hospital of Xuzhou Medical University from March 2021 to March 2025. The patients were randomly divided into a training set (n = 463) and a testing set (n = 198) at a 7:3 ratio. Clinical characteristics were initially screened using univariate analysis. LASSO regression and the Boruta algorithm were then jointly used to determine the full-set feature indicators, and Wald tests were further applied to identify independent risk factors and construct the reduced-set feature space. Based on the two feature sets, eight individual machine learning models were developed, including logistic regression (LR), k-nearest neighbors (KNN), support vector machine (SVM), decision tree (DT), random forest (RF), naïve bayes (NB), artificial neural network (ANN), and gradient boosting machine (GBM). To address the limitations of individual models in calibration and generalization performance, a Stacking ensemble learning framework was introduced for model fusion optimization, and a visual online risk calculator was developed.
RESULTS: LASSO regression and the Boruta algorithm selected nine variables for the full-set feature group: fasting plasma glucose on OGTT, 1-hour post-load glucose, white blood cell count (WBC), neutrophil-to-lymphocyte ratio (NLR), hemoglobin (Hb), fibrinogen (FIB), alanine aminotransferase (ALT), alkaline phosphatase (ALP), and uric acid-to-creatinine ratio (UA/Cr). Wald tests confirmed that ALP, ALT, fasting plasma glucose (FPG), UA/Cr, and Hb were independent influencing factors and constituted the reduced-set core indicators. Comparison of individual models showed that the full-set GBM model performed well in risk recognition, whereas the LR model had advantages in calibration and clinical applicability. The Stacking ensemble model integrating GBM and LR showed the best overall performance, with an AUC of 0.879 and accuracy, sensitivity, and specificity of 0.768, 0.767, and 0.768, respectively, in the testing set. Two patients were selected to apply the online sPTB risk calculator for patients with GDM based on the ensemble model. The estimated risk was 3.26% for a non-sPTB patient and 93.96% for a patient with sPTB.
CONCLUSION: The Stacking ensemble learning model showed good assessment performance in the present dataset and may serve as an exploratory model for near-delivery sPTB risk stratification in patients with GDM. The online risk calculator developed from this model may support preliminary near-delivery risk stratification in patients with GDM; external validation is required before clinical implementation.

Keywords:  Clinical risk identification model; Ensemble learning algorithms; Gestational diabetes mellitus; Machine learning algorithms; Spontaneous preterm birth

DOI:  https://doi.org/10.1186/s12884-026-09690-1
Front Ophthalmol (Lausanne). 2026 ;6 1975827

A quantum-inspired cross-attention fusion network for multi-class eye disease classification from retinal fundus images.

Abinash P, Krishnamoorthy N.

  Automated screening of blinding eye diseases from colour fundus photographs is an active area of medical image analysis, yet many reported systems are validated on a single dataset and offer limited insight into their internal decision process. This study introduces QICAF-Net, a dual-branch deep learning framework that couples a Swin Transformer global-context branch with an EfficientNet local-lesion branch and fuses their representations through a novel Quantum-Inspired Cross-Attention Fusion (QICAF) module. The QICAF module reformulates cross-attention using the mathematical formalism of quantum states: token features are amplitude-encoded as complex unit vectors, transformed by a learnable norm-preserving unitary gate, compared through a fidelity-inspired similarity, normalised by a Born-rule measurement rule, and combined through an entanglement-inspired bilinear interaction. All operations are executed as classical linear algebra on conventional graphics hardware; no physical quantum computer and no quantum speed-up are claimed. The model was trained and evaluated on a balanced four-class retinal dataset of 4,217 fundus images spanning cataract, diabetic retinopathy (DR), glaucoma and normal categories. Under stratified five-fold cross-validation with six-view test-time augmentation (TTA), QICAF-Net attained a macro-F1 of 0.9358 ± 0.0113, an accuracy of 0.9367 ± 0.0112 and a macro one-versus-rest area under the receiver operating characteristic curve (AUC) of 0.9914 ± 0.0023; a three-seed evaluation of the full model gave a macro-F1 of 0.9420 ± 0.0071. An 18-run multi-seed component ablation (six fusion configurations × three seeds) confirmed that the quantum-inspired components achieve accuracy statistically indistinguishable from strong classical fusion baselines, and a paired three-seed significance test between QICAF-Net and a matched dual-branch concatenation model returned no statistically significant accuracy difference (paired t-test p = 0.887; Wilcoxon p = 1.000). Against five single-backbone baselines under an identical preprocessing and TTA protocol, QICAF-Net was competitive but did not exceed the strongest backbones (Swin-Tiny 0.9583, ViT-Small 0.9487). The principal contribution of the module is therefore a novel and interpretable formulation rather than a raw accuracy gain. External validation on a heterogeneous independent cohort (N = 3,907) showed heterogeneous cross-dataset transfer - better preserved for cataract and normal and substantially degraded for glaucoma; under an observed external-domain shift the diabetic-retinopathy sensitivity fell to 39.61%, and under colour-constancy correction with six-view TTA the external accuracy was 0.8741 (95% CI 0.8633-0.8841) and the macro-F1 0.7188 (95% CI 0.6969-0.7398). Two complementary explainability methods together with visualisation of the QICAF attention indicated that the model attends to clinically plausible regions, most strongly the optic nerve head for glaucoma. Calibration, quantitative explanation-faithfulness (deletion/insertion) and profiler-based computational analyses were additionally performed. QICAF-Net is thus presented as an interpretable and transparently evaluated framework whose accuracy is on par with established fusion strategies while contributing a new attention formulation to the retinal imaging literature.

Keywords:  EfficientNet; Grad-CAM++; cross-attention fusion; diabetic retinopathy; explainable artificial intelligence; external validation; fundus image classification; glaucoma

DOI:  https://doi.org/10.3389/fopht.2026.1975827
Diseases. 2026 Oct 13. pii: 378. [Epub ahead of print]14(10):

Clinical Effectiveness, Safety, and Implementation of Artificial Intelligence-Assisted Insulin Management Systems in Type 1 and Type 2 Diabetes: A Systematic Review of Randomized Controlled Trials.

Victória Dogani Rodrigues, Virgínia Maria Cavallari Strozze Catharin, Mauro Audi, Eliana de Souza Bastos Mazuqueli Pereira, Jesselina Francisco Dos Santos Haber, Débora Audi, Vitor Cavallari Strozze Catharin, Maria Angélica Miglino, Lucas Fornari Laurindo.

   BACKGROUND: Artificial intelligence (AI) is increasingly being integrated into insulin management through decision-support systems, adaptive bolus calculators, conversational platforms, and automated insulin delivery (AID) technologies. However, the clinical effectiveness and safety of AI-assisted insulin management remain uncertain.
OBJECTIVE: To systematically evaluate evidence from randomized controlled trials (RCTs) investigating AI-assisted insulin management systems in individuals with type 1 or type 2 diabetes.
METHODS: A systematic literature review was conducted following PRISMA guidelines. PubMed/MEDLINE, Embase, and Cochrane CENTRAL were searched for RCTs evaluating AI-enabled insulin management interventions in insulin-treated individuals with diabetes. Eligible studies compared AI-based systems with standard care, clinician-guided insulin titration, conventional bolus calculators, or other AID algorithms. Outcomes included glycemic control, insulin titration, safety, and implementation-related measures. Owing to substantial heterogeneity, findings were synthesized narratively.
RESULTS: We included seven RCTs published between 2020 and 2025. Interventions comprised clinical decision-support systems, adaptive bolus calculators, conversational AI platforms, neural-network-based artificial pancreas algorithms, and exercise-aware AID systems. AI-assisted interventions consistently demonstrated favorable safety profiles, with no increase in severe hypoglycemia, diabetic ketoacidosis, or serious adverse events. Some systems, particularly fully automated closed-loop platforms and conversational AI-guided insulin titration, showed improvements in time in range, glucose control, adherence, and insulin optimization. User acceptance and adherence emerged as important determinants of effectiveness.
CONCLUSIONS: Current randomized evidence suggests that AI-assisted insulin management systems are safe, feasible, and capable of achieving glycemic outcomes comparable to conventional care. Although certain technologies demonstrate promising clinical benefits, the evidence base remains limited by small sample sizes, short follow-up durations, and heterogeneity across interventions. Larger, long-term randomized trials are needed to establish comparative effectiveness, optimize human-AI interaction, and support broader clinical implementation.

Keywords:  artificial intelligence; automated insulin delivery; closed-loop systems; diabetes mellitus; insulin therapy; insulin titration; machine learning; randomized controlled trials; systematic review

DOI:  https://doi.org/10.3390/diseases14100378
Health Data Sci. 2026 ;6 0471

Short-Term Continuous Glucose Forecasting with Large-Language-Model-Derived Nutrient Estimates from Real-World Chinese Dietary Records.

Leyao Ma, Jie Hao, Lin Yang, Xiaowei Xu, Jiao Li.

Background: Accurate short-term glucose forecasting is important for diabetes self-management, yet dietary information remains difficult to quantify and integrate into forecasting models in free-living settings, especially for complex Chinese meals. Methods: We investigated whether large language models (LLMs) can convert free-text Chinese dietary records into structured nutritional features for short-term continuous glucose forecasting. Five Chinese-capable LLMs were evaluated as nutrient estimators, and the resulting nutritional features were integrated with continuous glucose monitoring histories in a seasonal-patch transformer for glucose forecasting. Performance was assessed in 2 real-world cohorts (ShanghaiT1DM and ShanghaiT2DM) for both full-period and postprandial glucose forecasting. Results: LLM-derived nutrient estimates significantly outperformed semantic embeddings as dietary representations for glucose forecasting (P < 0.001). In the primary comparison setting, the best-performing LLM-based model achieved a postprandial 3-h root mean square error (RMSE) of 34.668 mg/dl, reducing the error by 1.967 mg/dl compared to non-LLM settings. In the full-period setting, the optimal LLM-based model achieved an RMSE of 27.035 mg/dl, reducing the error by 0.42 mg/dl relative to non-LLM settings. Across both forecasting tasks, direct nutrient estimation by DeepSeek-v3.1 achieved the best average RMSE. Forecasting gains were positively associated with nutrient-estimation quality, particularly for carbohydrate, dietary fiber, and fat. Architecture ablations further showed that patch-based encoding was the most influential component, with additional contributions from temporal decomposition and side branch design. Conclusion: These findings show that LLM-based nutrient estimation provides a practical and physiologically interpretable way to convert free-text dietary records into structured meal-related inputs for glucose forecasting.

DOI: https://doi.org/10.34133/hds.0471
Diagnostics (Basel). 2026 Oct 17. pii: 3382. [Epub ahead of print]16(20):

Generative Forecasting of Longitudinal 3D-OCT Under Anti-VEGF Therapy for Diabetic Macular Edema.

Yuxiang Feng, Chaofeng Liang, Shujun Wang.

  Background/Objectives: Forecasting the individual anatomical response to anti-vascular endothelial growth factor (anti-VEGF) therapy from optical coherence tomography (OCT) could support the management of diabetic retinal disease. Direct volumetric generation is difficult because treatment-associated change is local while most of the retina is stationary, so generative models regress to the mean and blur retinal layers. Methods: We propose a two-stage framework that decouples baseline anatomy from longitudinal change in 3D-OCT. A Retinal Response Tokenizer compresses the latent residual between baseline and follow-up volumes into discrete action tokens via Finite Scalar Quantization (FSQ), and a conditional decoder renders the follow-up volume from baseline features and these tokens. An Action Transformer predicts the tokens from the baseline scan, the follow-up interval, and baseline covariates. Results: On the OLIVES trial cohort (diabetic retinopathy and macular edema; 738 test pairs from five held-out eyes), our method achieved the highest structural similarity in every test eye (SSIM 0.657 vs. 0.543-0.623) and tracked the central subfield thickness (CST) change measured by a frozen regressor (Δ-r=0.88). In a five-fold cross-validation over 52 eyes, it improved PSNR over an identity copy by 1.27 dB (95% CI 0.93-1.57) and reduced the CST change error by 10.5 μm (95% CI 3.1-19.4). Conclusions: Prospective validation is required before clinical use.

Keywords:  anti-VEGF therapy; diabetic macular edema; diabetic retinopathy; discrete latent dynamics; generative forecasting; optical coherence tomography

DOI:  https://doi.org/10.3390/diagnostics16203382
Sensors (Basel). 2026 Oct 16. pii: 6572. [Epub ahead of print]26(20):

Non-Invasive Blood Glucose Monitoring Using Machine Learning and Optical Approaches: A Systematic Literature Review and a Proposed Framework for Reproducible and Clinically Reliable Research.

Syed Saim Ahmed, Norbert Noury.

  Diabetes Mellitus is the leading healthcare challenge. For effective management of this disease, patients need to check their blood glucose levels regularly; however, the standard method of taking a drop of blood and placing it on a test strip does not allow for continuous monitoring and is dreaded by most people, who find it painful and uncomfortable. Researchers worldwide are trying to develop non-invasive blood glucose monitoring (NIBGM) devices; most of them have reported promising accuracy using a combination of optical techniques and machine learning, but the CE/FDA marked solution still cannot reach the market. The main objective of this systematic review is to critically analyse recent quantitative research works in optical NIBGM, highlight their strengths, and identify the main factors that hinder regulatory approvals and the commercialisation of these devices. Web of Science, IEEE Xplore, PubMed, ScienceDirect, SPIE Digital Library, and Optica Publishing Group databases were searched between 1st September 2015 and 1st August 2026 using specific keywords. From the initial 1143 records, 245 studies were considered for full-text quality assessment, and 208 were excluded based on a post hoc quality score threshold. The sensing modality, wavelength, sample size, reference device, glucose range, validation strategy, accuracy metrics, and use of demographic variables as model inputs for the 37 included studies are reported. By applying an author-defined three-tier rubric, the risk of shortcut learning in each study was highlighted. Based on these findings, a common framework of fifteen concrete, evidence-grounded recommendations for future NIBGM studies was proposed.

Keywords:  blood glucose; common framework; diabetes mellitus; machine learning; non-invasive monitoring; optical techniques; shortcut learning

DOI:  https://doi.org/10.3390/s26206572
Indian J Ophthalmol. 2026 Nov 01. 74(11): 1721-1724

Comparative efficacy of large language models (Google Gemini and ChatGPT 4.0) in screening patients for diabetic retinopathy.

Ahana Sen, Rohan Basak, Sarang Gokhale, Bhagyashree Verma, Rupak Roy.

DOI: https://doi.org/10.4103/IJO.IJO_1266_26
Front Endocrinol (Lausanne). 2026 ;17 1958918

Forecasting accuracy is not enough: safety-oriented evaluation of neural digital twins prior to in-silico insulin and carbohydrate intervention in type 1 diabetes.

Ahmed Ibrahim, Ameen Banjar, Simon Harper, Ala Alarood, Eesa Alsolami.

  Neural networks trained on a single patient's continuous glucose monitoring (CGM) record are increasingly proposed as "digital twins" that could let clinicians rehearse insulin and carbohydrate decisions in silico before applying them to the patient. Before any such use, a candidate twin should respond to treatment in a direction consistent with established pharmacology: a model that merely extrapolates the recent glucose trend can report an excellent forecasting error while remaining effectively blind to insulin, and this failure is invisible to the root-mean-square error (RMSE) that dominates the glucose-forecasting literature. We examined this gap in two independent cohorts: a longitudinal, multimodal University of Manchester cohort (17 participants enrolled; 13 formed the Core analytical cohort after data-quality screening) and, as an independent replication of the protocol, the OhioT1DM dataset (12 participants). Patient-specific two-layer long short-term memory (LSTM) models were trained under a strict chronological 60/20/20 split and evaluated across five random seeds. A physiology-guided variant added insulin-on-board and carbohydrate-on-board states with a double-monotonicity training penalty encouraging insulin to lower and carbohydrate to raise predicted glucose. Two findings replicated across both cohorts and all seeds. First, on the accuracy axis, a physiology-free Vanilla LSTM had the lowest one-step RMSE (Manchester 7.00, Ohio 5.02 mg/dL; Friedman χ 2(5) = 43.4 and 46.7, both p < 10 - 7), and under two-hour rollout a non-learning persistence baseline beat every neural model at 60 and 120 minutes-even after rollout-aware training-so ranking twins by accuracy alone would select a model that cannot represent treatment at all. Second, on a graded insulin dose-response audit, the physiology-guided model produced the steepest correctly-directed (glucose-lowering) response among the primary configurations in both cohorts (slope significantly better than the unconstrained model: Holm-corrected p = 0.002 and 0.003), whereas removing the directional constraint left a wrong-signed response; a supplementary ablation without masking was steeper still, pointing to the directional penalty, not masking, as the driver. Its advantage over the simpler Vanilla model was significant on for this cohort dependence. Accuracy and treatment-response plausibility are separable properties; we provide a reproducible protocol auditing both before a neural twin is used for intervention simulation.

Keywords:  LSTM; continuous glucose monitoring; digital twin; dose-response audit; glucose forecasting; in-silico intervention simulation; treatment-response plausibility; type 1 diabetes

DOI:  https://doi.org/10.3389/fendo.2026.1958918
Mol Med. 2026 Aug 03. pii: 198. [Epub ahead of print]32(1):

Individualized statin associated type 2 diabetes risk estimation with a deep causal model.

Hao Zhou, Jorge Passamani Zubelli, Haralampos Hatzikirou, Andreas Henschel, Laurent Alain Najman, Daniel E Platt, Antonello Maruotti, Siobhan O'Sullivan, Lithe Basbous, Cynthia Al Hageh, Mariam Al Harbi, Antoine Abchee, Pierre Zalloua.

   BACKGROUND: Statin therapy reduces cardiovascular risk, but recorded statin exposure is associated with subsequent type 2 diabetes (T2D) in some populations. Individualized risk estimation and separation of low-density lipoprotein cholesterol (LDL-C)-mediated from non-LDL components remain challenging in observational data.
METHODS: We analyzed adults free of diagnosed diabetes and without substantial baseline hyperglycemia in UK Biobank, the All of Us Research Program, and the Abu Dhabi Public Health Center cohort. We used weighted survival and regression analyses, UK Biobank exposure-definition and propensity-score sensitivity analyses, two-stage residual inclusion Mendelian randomization mediation, and CausalT2DNet, a balanced deep counterfactual prediction model.
RESULTS: Recorded statin exposure was associated with higher incident T2D risk over cohort-specific horizons. UK Biobank matched-method 10-year sensitivity analyses yielded adjusted risk ratios of 2.18-2.25, and a broader baseline-only IPTW sensitivity model yielded a weighted risk ratio of 1.95. In a separate CausalT2DNet 12-year absolute-risk scenario, predicted T2D risk increased from 2.24% under no statin exposure to 4.05% under statin exposure. Mediation and counterfactual analyses suggested that much of the excess risk was not explained by LDL-C lowering, under their respective assumptions.
CONCLUSIONS: Recorded statin exposure was consistently associated with higher incident T2D risk across three longitudinal cohorts. The framework supports assumption-explicit research risk estimation and hypothesis generation and requires external recalibration and prospective validation before clinical use.

Keywords:  Cardiovascular; Causal inference; Diabetes; LDL; Machine learning; Statin

DOI:  https://doi.org/10.1186/s10020-026-01589-z
J Clin Med. 2026 Oct 20. pii: 8138. [Epub ahead of print]15(20):

Concurrent Diabetes Mellitus and Periodontitis in U.S. Adults: Mortality Associations and Machine Learning Risk Prediction.

Jincheng Li, Tian Zhu, Haruhiko Kashiwazaki.

  Background/Objectives: We examined mortality associations by diabetes mellitus (DM)-periodontitis status and the incremental predictive value of composite inflammatory indices among adults with concurrent disease. Methods: We analyzed 10,632 adults from the National Health and Nutrition Examination Survey 2009-2014 linked to mortality data through 2019. Groups included neither DM nor periodontitis (NDM + NP), periodontitis alone (P), DM alone (DM), or concurrent DM and periodontitis (DM + P). Survey-weighted Cox and Fine-Gray models assessed all-cause and cardiovascular mortality. Among 978 participants with DM + P, nine survival models underwent five repeats of fivefold cross-validation to assess 5-year all-cause mortality prediction and the incremental value of eight composite inflammatory indices. Results: DM + P was associated with higher all-cause mortality than NDM + NP (HR, 1.51; 95% CI, 1.17-1.94) and P (HR, 1.31; 95% CI, 1.09-1.57). For cardiovascular mortality, DM + P was associated with higher mortality than NDM + NP (subdistribution hazard ratio [sHR], 1.51; 95% CI, 1.003-2.276), although the confidence interval only narrowly excluded 1.0. DM + P also had higher cardiovascular mortality than P (sHR, 1.55; 95% CI, 1.14-2.10). Neither outcome differed significantly between DM + P and DM. The Combined extreme gradient boosting Cox model had a concordance index and 5-year AUC of 0.754. Adding inflammatory indices did not significantly improve discrimination, prediction error, calibration, or net benefit. Observed 5-year mortality across predicted-risk tertiles was 3.7%, 7.4%, and 21.2%. Conclusions: Concurrent diabetes and periodontitis were associated with higher all-cause and cardiovascular mortality than periodontitis alone, but not with significantly higher mortality than diabetes alone. Composite inflammatory indices did not significantly improve prediction beyond clinical predictors.

Keywords:  cardiovascular mortality; diabetes mellitus; inflammatory indices; machine learning; mortality; periodontitis

DOI:  https://doi.org/10.3390/jcm15208138