bims-aukdir 2025-11-30 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2025–11–30
twenty-two papers selected by
Mott Given

Development of machine learning predictive model for type 2 diabetic retinopathy using the triglyceride-glucose index explained by SHAP method.
Development and comparison of OCT-based prediction models for diabetic retinopathy using LASSO and random forest.
Enhancing AI-based diabetic retinopathy diagnosis through universal cross-camera image adaptation.
Diagnostic Accuracy of Artificial Intelligence in Predicting Anti-VEGF Treatment Response in Diabetic Macular Edema: A Systematic Review and Meta-Analysis.
Automated retinal image analysis systems to triage for grading of diabetic retinopathy: a large-scale, open-label, national screening programme in England.
Machine Learning Predictive Models for Survival in Colorectal Cancer Patients With Diabetes: A Cohort Study of 10 749 Subjects.
Improving Type 2 Diabetes Prediction: Comparative Evaluation of Machine Learning Classifiers Using Balanced Data from the AWI-Gen Cohort.
Double-Layer Simplified Complex Interval Neural Network Stacking for Blood Glucose Prediction of Continuous Glucose Monitoring System.
Can artificial intelligence and wearable technologies transform diabetes foot health?
Predicting cardiovascular disease among diabetic patients in Ethiopia using machine learning models: evidence from Ethiopian public health Institute data (2024/2025).
A Study on the Interpretability of Diabetic Retinopathy Diagnostic Models.
Construction and validation of a hypoglycemia risk prediction model for hospitalized type 2 diabetes patients based on machine learning.
Supervised machine learning identifies impaired mitochondrial quality control in β cells with development of type 2 diabetes.
A dynamic weighted ensemble learning framework for cardiovascular risk prediction in type 2 diabetes: a comparative study with SHAP-based interpretability.
Identifying Risk Groups in 73,000 Patients with Diabetes Receiving Total Hip Replacement: A Machine Learning Clustering Analysis.
AI-Powered Thermography for Diabetic Foot Risk Stratification: Multicenter Cross-Sectional Study.
Artificial intelligence-based diagnosis of diabetic kidney disease using urinary VOC biosensor data.
Artificial Intelligence for Diabetes Complication Prediction: A Systematic Review of Current Applications and Future Directions.
Opportunistic screening of type 2 diabetes with deep metric learning using electronic health records.
Applications of AI in the management of elderly diabetes patients.
Medication-Stratified Analysis of LDL-C Equation Miscalibration in Diabetes: Evidence from the All of Us Research Program and a Medication-Agnostic Machine-Learning Correction.
Neutrophil percentage to albumin ratio and advanced cardiovascular-kidney-metabolic syndrome in diabetes: a machine learning approach.

Front Endocrinol (Lausanne). 2025 ;16 1631647

Development of machine learning predictive model for type 2 diabetic retinopathy using the triglyceride-glucose index explained by SHAP method.

Xiaoqin Liu, Shuying Wu, Yue Yang, Yang Li, Xinting Zhang, Rihui Liu, Ling Qin, Fei Li.

   Introduction: This study aimed to develop a diabetic retinopathy (DR) Prediction model using various machine learning algorithms incorporating the novel predictor Triglyceride-glucose index (TyG). Furthermore, the model was interpreted using the SHapley Additive exPlanations (SHAP) method.
Method: Real-world data were collected from a general hospital in a major city and a county clinic, then divided into the DR Group (1392) and non-DR group (2358). Baseline data were collected, and variables were selected using Recursive Feature Elimination with Cross-Validation (RFECV). The performance of five machine learning algorithms, including Logistic Regression model (LR), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and XGBoost (XGB), was assessed based on accuracy, sensitivity, specificity, and Area Under the Curve (AUC) of the Receiver Operating characteristic Curve (ROC). The optimal model was interpreted using SHAP.
Result: LVM and LR demonstrated superior performance in both the test set and training set (ROC, 0.85 and 0.82, respectively). The top five predictors identified by SHAP analysis included TyG, Insulin therapy, HbA1c, Diabetes Course, HDL. HDL was identified as a protective factor, while the remaining factors were associated with retinopathy.
Conclusion: LR and SVM demonstrated the best performance. To our knowledge, this is the first machine learning-based DR prediction model integrating the triglyceride-glucose index (TyG) as a core predictor, overcoming limitations of insulin resistance (IR) assessment in resource-limited settings. TyG provides a cost-effective alternative to conventional IR biomarkers (e.g., HOMA-IR), enabling practical DR risk stratification in primary care.

Keywords:  SHAP; TyG-index; diabetic retinopathy; machine learning; predictive model

DOI:  https://doi.org/10.3389/fendo.2025.1631647
Medicine (Baltimore). 2025 Nov 21. 104(47): e45901

Development and comparison of OCT-based prediction models for diabetic retinopathy using LASSO and random forest.

Maimaiti Tuerhongjiang, Xuemei Li, Aierken Kalibinuer, Yourong Dong, Aierken Ailiyaer, Feifei Cheng, Wei Gao.

  Diabetic retinopathy (DR) is a leading cause of vision impairment among individuals with type 2 diabetes mellitus (T2DM). This study aimed to construct and evaluate predictive models for DR by integrating clinical data with optical coherence tomography (OCT) parameters, using both Least Absolute Shrinkage and Selection Operator (LASSO) regression and random forest (RF) algorithms. A retrospective analysis was conducted on medical records of T2DM patients admitted between September 2020 and December 2023. After applying inclusion and exclusion criteria, 10,054 cases were selected. Patients were randomly assigned to training (70%) and validation (30%) cohorts. LASSO regression was used for variable selection, followed by logistic modeling. A RF model was also developed using the same features. Model performance was assessed using receiver operating characteristic curves, and differences were analyzed via the DeLong test. Key predictors identified included gender, insulin therapy, duration of diabetes, urinary albumin-to-creatinine ratio, and retinal vessel density. The RF model demonstrated superior performance with an areas under the curve of 0.89, compared to 0.79 for the LASSO model (P < .05). Retinal vessel density was consistently a protective factor, while prolonged diabetes duration and elevated albumin-to-creatinine ratios were associated with increased DR risk. OCT-derived retinal metrics, particularly vessel density, enhance the predictive capability of DR risk models. Among the 2 approaches, the RF model exhibited better classification performance and may serve as a practical tool for early screening and individualized risk assessment in clinical settings.

Keywords:  LASSO regression; OCT; diabetic retinopathy; machine learning; random forest; retinal vessel density

DOI:  https://doi.org/10.1097/MD.0000000000045901
BMJ Open Ophthalmol. 2025 Nov 21. pii: e002238. [Epub ahead of print]10(1):

Enhancing AI-based diabetic retinopathy diagnosis through universal cross-camera image adaptation.

Sanil Joseph, Xiaotian Chen, Chi Liu, Zhuoting Zhu, Kim Ramasamy, Thulasiraj D Ravilla, Zongyuan Ge, Mingguang He.

   OBJECTIVE: To evaluate the effectiveness of a deep learning-based style adaptation strategy in improving the diagnostic accuracy and cross-camera generalisability of artificial intelligence (AI) for detecting diabetic retinopathy (DR).
METHODS AND ANALYSIS: This diagnostic study involved prospective recruitment of patients aged 50 years and older attending the outpatient clinic at a tertiary eye hospital in Southern India, between 14 June and 5 August 2022. Paired macula-centred retinal images were captured using two fundus cameras: Optain Resolve (portable, automated) and Topcon NW400 (static, manual). A style adaptation model, the Style-Consistent Retinal Image Transformation Network (SCR-Net), was applied to align image styles across cameras. The AI-based DR detection model, developed using the InceptionNeXt-T architecture, was trained on images from the EyePACS data set and evaluated under three scenarios: (1) training and testing on original images (2) training and testing on SCR-Net-adapted images; and (3) training on a combined (original+adapted) data set and testing on adapted images. Diagnostic accuracy and preservation of image quality were evaluated.
RESULTS: The mixed training/testing approach (scenario 3) achieved the highest diagnostic accuracy for Optain images at 79.2% (95% CI 75.9% to 82.6%) with a Cohen's kappa of 0.893 (95% CI 0.867 to 0.917). Adapted images preserved critical diagnostic features (peak signal-to-noise ratio, 29.35; structural similarity index measure, 0.847). Style adaptation reduced false positives in Optain images while maintaining robust diagnostic performance for Topcon images, effectively addressing cross-camera variability.
CONCLUSION: Style adaptation using SCR-Net enhances the consistency and generalisability of AI-based DR detection systems by reducing false positives and maintaining robust performance across camera systems. This approach has the potential to democratise access to early DR diagnosis in underserved regions. This study was conducted at a single centre using a limited set of fundus cameras, which may affect the generalisability. Nonetheless, further validation across diverse imaging systems and clinical settings is warranted to support broader applicability.

Keywords:  Diagnostic tests/Investigation; Imaging; Public health; Retina; Vision

DOI:  https://doi.org/10.1136/bmjophth-2025-002238
J Clin Med. 2025 Nov 18. pii: 8177. [Epub ahead of print]14(22):

Diagnostic Accuracy of Artificial Intelligence in Predicting Anti-VEGF Treatment Response in Diabetic Macular Edema: A Systematic Review and Meta-Analysis.

Faisal A Al-Harbi, Mohanad A Alkuwaiti, Meshari A Alharbi, Ahmed A Alessa, Ajwan A Alhassan, Elan A Aleidan, Fatimah Y Al-Theyab, Mohammed Alfalah, Sajjad M AlHaddad, Ahmed Y Azzam.

  Background/Objectives: Diabetic macular edema (DME) is a leading cause of vision loss in diabetic patients, with anti-vascular endothelial growth factor (anti-VEGF) therapy being the standard management. However, treatment response varies significantly among patients, necessitating predictive tools. This systematic review and meta-analysis evaluated the diagnostic accuracy of artificial intelligence (AI) models in predicting anti-VEGF treatment response in DME patients. Methods: We conducted a dedicated literature review following PRISMA 2020 guidelines, searching PubMed, Web of Science, Embase, Scopus, and Cochrane Library databases from inception up to 30 September 2025. Studies evaluating AI-based prediction models for anti-VEGF response in DME patients were included. The primary outcomes were sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). A bivariate random-effects meta-analysis was performed using available diagnostic accuracy data. Results: From 3107 participants across 18 studies, six studies with 427 participants provided complete diagnostic accuracy data for meta-analysis. The pooled sensitivity was 86.4% (95% CI: 82.1-90.1%) and the specificity was 77.6% (95% CI: 72.8-82.0%). The summary AUC was 0.89 with a diagnostic odds ratio of 22.0 (95% CI: 12.8-37.9). AI models demonstrated superior performance compared to other methods in 87.5% of comparative studies. Moderate heterogeneity was observed (I2 = 45.2%). Conclusions: AI models demonstrate good diagnostic accuracy for predicting anti-VEGF treatment response in DME patients, with a promising role for improving personalized management strategies and improved outcomes.

Keywords:  deep learning; diabetes; diabetic macular edema; diabetic retinopathy; machine learning

DOI:  https://doi.org/10.3390/jcm14228177
Lancet Digit Health. 2025 Nov 24. pii: S2589-7500(25)00096-2. [Epub ahead of print] 100914

Automated retinal image analysis systems to triage for grading of diabetic retinopathy: a large-scale, open-label, national screening programme in England.

ARIAS Research Group

BACKGROUND: The global prevalence of diabetes is rising, alongside costs and workload associated with screening for diabetic eye disease (diabetic retinopathy). Automated retinal image analysis systems (ARIAS) could replace primary human grading of images for diabetic retinopathy. We evaluated multiple ARIAS in a real-life screening programme.
METHODS: Eight of 25 invited and potentially eligible CE-marked systems for diabetic retinopathy detection from retinal images agreed to participate. From 202 886 screening encounters at the North East London Diabetic Eye Screening Programme (between Jan 1, 2021, and Dec 31, 2022) we curated a database of 1·2 million images and sociodemographic and grading data. Images were manually graded by up to three graders according to a standard national protocol. ARIAS performance overall and by subgroups of age, sex, ethnicity, and index of multiple deprivation (IMD) were assessed against the reference standard, defined as the final human grade in the worst eye for referable diabetic retinopathy (primary outcome). Vendor algorithms did not have access to human grading data.
FINDINGS: Sensitivity across vendors ranged from 83·7% to 98·7% for referable diabetic retinopathy, from 96·7% to 99·8% for moderate-to-severe non-proliferative diabetic retinopathy, and from 95·8% to 99·5% for proliferative diabetic retinopathy. Sensitivity was largely consistent for moderate-to-severe non-proliferative and proliferative diabetic retinopathy by subgroups of age, sex, ethnicity, and IMD for all ARIAS. For mild-to-moderate non-proliferative diabetic retinopathy with referable maculopathy, sensitivity across vendors ranged from 79·5% to 98·3%, with greater variability across population subgroups. False positive rates for no observable diabetic retinopathy ranged from 4·3% to 61·4% and within vendors varied by 0·5 to 44 percentage points across population subgroups.
INTERPRETATION: ARIAS showed high sensitivity for medium-risk and high-risk diabetic retinopathy in a real-world screening service, with equitable performance across population subgroups. ARIAS could provide a cost-effective solution to deal with the rising burden of screening for diabetic retinopathy by safely triaging for human grading, substantially increasing grading capacity and rapid diabetic retinopathy detection.
FUNDING: NHS Transformation Directorate, The Health Foundation, and The Wellcome Trust.

DOI: https://doi.org/10.1016/j.landig.2025.100914
J Dig Dis. 2025 Nov 27.

Machine Learning Predictive Models for Survival in Colorectal Cancer Patients With Diabetes: A Cohort Study of 10 749 Subjects.

Junjie Huang, Claire Chenwen Zhong, Zhaojun Li, Yu Jiang, Zehuan Yang, Jinqiu Yuan, Jonathan Poon, Qi Dou, Martin C S Wong.

   OBJECTIVES: To identify the risk factors for the survival of colorectal cancer (CRC) patients with type 2 diabetes mellitus (T2DM), compare the predictive performance of models based on different algorithms, and develop a risk score system to predict the survival risk of the target population.
METHODS: We analyzed data from the Hong Kong Hospital Authority Data Collaboration Laboratory (HADCL), including 10 749 CRC patients with T2DM from 2000 to 2020. We employed traditional statistical methods and machine learning algorithms to compare their performance using the area under the receiver operating characteristic curve (AUC). The SHapley Additive exPlanations (SHAP) analysis was conducted to identify risk factors and attribute model outputs. A risk score system was developed using the AutoScore-Survival package for risk stratification.
RESULTS: Key predictors of CRC survival among T2DM patients included age at cancer diagnosis, sex, T2DM duration, alcohol consumption, central obesity, hypertension, levels of low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol, and serum potassium, and anti-lipid drug usage. Among the models assessed, the random survival forest showed the best performance. The risk score system was calibrated as follows: age at diagnosis, T2DM duration, LDL-C, glycated hemoglobin, creatinine, and body mass index. The AUCs for 1, 3, and 5 years of the tuned risk score system were 0.746, 0.718, and 0.677, respectively.
CONCLUSIONS: The random survival forest model provides superior survival prediction compared to other models evaluated. A validated risk score system has been established, facilitating risk stratification for clinicians to manage these patients.

Keywords:  colorectal neoplasms; diabetes mellitus; machine learning; random survival analysis; risk score

DOI:  https://doi.org/10.1111/1751-2980.70019
Res Sq. 2025 Nov 04. pii: rs.3.rs-8019155. [Epub ahead of print]

Improving Type 2 Diabetes Prediction: Comparative Evaluation of Machine Learning Classifiers Using Balanced Data from the AWI-Gen Cohort.

Richmond Balinia Adda.

Background: Type 2 diabetes mellitus (T2DM) is an escalating public health concern across Africa, but regionally tailored predictive models are scarce. Advances in machine learning (ML) offer potential for early identification, though previous research has been constrained by methodological issues such as data leakage, class imbalance, and overfitting, limiting clinical deployment, especially in digital health contexts. Methods: This study analysed data from 2,010 participants in the H3Africa AWI-Gen cohort in northern Ghana to develop and evaluate ML-based prediction models tailored to African settings. Rigorous preprocessing steps, including handling class imbalance with SMOTE and excluding diagnostic biomarkers prone to target leakage, were applied. Eight ML classifiers underwent robust Bayesian hyperparameter optimisation. Model performance was assessed via stratified 5-fold cross-validation and confirmed through extensive sensitivity and calibration analyses. Results: The optimised XGBoost model yielded an AUC of 0.845 (95% CI: 0.812-0.878) and a sensitivity of 78.2% on unseen data. Including glucose as a predictor increased performance by 11.5%, underscoring the necessity of its exclusion to avoid biased evaluation. Models using only anthropometric and lifestyle variables (AUC = 0.783) demonstrated robust predictive capacity, with waist circumference, physical activity, and BMI standing out as the most stable predictors across analyses. Conclusion: Our findings demonstrate that ML models constructed from routinely collected clinical and lifestyle data can attain clinically meaningful diabetes prediction suitable for digital health applications in low-resource African contexts. This study addresses prior methodological gaps and offers a data-driven framework that is both robust and clinically plausible for early T2DM detection, with potential implications for public health policy and digital screening programmes in similar populations.

DOI: https://doi.org/10.21203/rs.3.rs-8019155/v1
Biosensors (Basel). 2025 Oct 22. pii: 707. [Epub ahead of print]15(11):

Double-Layer Simplified Complex Interval Neural Network Stacking for Blood Glucose Prediction of Continuous Glucose Monitoring System.

Shaowei Kong, Yusheng Fu, Jingshan Duan, Jian Yan.

  Diabetes is a metabolic disorder characterized by persistent hyperglycemia, with its incidence steadily rising worldwide. Blood glucose monitoring is a core measure in diabetes management, and continuous glucose monitoring provides more comprehensive and accurate glucose data compared to traditional fingerstick testing. To collect continuous glucose data from patients, precise glucose prediction algorithms can help them better control their blood glucose fluctuations. Therefore, by addressing the issues of low prediction accuracy, complex input features, and poor generalization performance in existing glucose prediction methods, this paper proposes a glucose prediction model based on a double-layer SCINet stack using time-series analysis methods. SCINet effectively captures multi-scale dynamic features in time-series data through recursive down-sampling and convolution operations, making it suitable for glucose prediction tasks. Experimental data were sourced from real-world continuous glucose monitoring records of patients at Yixing People's Hospital. Model input features were optimized through variable selection and data preprocessing, with predictive performance validated on a test dataset. The results demonstrate that the proposed model outperforms existing time-series prediction models across varying prediction horizons and patient datasets, exhibiting high predictive accuracy and stability.

Keywords:  SCINet; blood glucose prediction; continuous glucose monitoring; diabetes; time series analysis

DOI:  https://doi.org/10.3390/bios15110707
Curr Opin Endocrinol Diabetes Obes. 2025 Nov 07.

Can artificial intelligence and wearable technologies transform diabetes foot health?

Wing May Kong, Farah Saif.

   PURPOSE OF REVIEW: Diabetes foot ulcers (DFUs) affect millions globally, and are a global health challenge, contributing significantly to morbidity, mortality, and healthcare costs. This review article critically evaluates advances in artificial intelligence and new technologies and their potential to transform diabetes foot complications.
RECENT FINDINGS: Artificial intelligence-based thermal and clinical image analysis offer the potential for early detection, remote diagnosis and monitoring and to mitigate disparities in specialist access. Incorporating novel pressure and temperature sensing, wearable technologies could enhance foot monitoring and enable personalized care and intervention. However, ethical challenges with artificial intelligence, including accountability, limited explainability, data security and equitable access will have to be addressed.
SUMMARY: Artificial intelligence and wearable technologies could herald a paradigm shift in diabetes foot health and research. However, these exciting tools are not yet ready for adoption in clinical practice. Larger, well funded clinical intervention studies and greater collaboration between clinicians, artificial intelligence scientists and product engineers, working in partnership with people with diabetes, is needed if these approaches are to fulfil their potential.

Keywords:  artificial intelligence; diabetes; foot ulcer; health inequalities; wearable technology

DOI:  https://doi.org/10.1097/MED.0000000000000942
BMC Public Health. 2025 Nov 24. 25(1): 4114

Predicting cardiovascular disease among diabetic patients in Ethiopia using machine learning models: evidence from Ethiopian public health Institute data (2024/2025).

Ashagrie Anteneh, Muluken Belachew, Sefefe Birhanu, Maru Meseret.

   INTRODUCTION: Cardiovascular disease (CVD) is the leading cause of death among individuals with diabetes, accounting for nearly 50% of diabetes-related mortality. In Ethiopia, the burden of diabetes is increasing, yet there is a lack of predictive tools for identifying those at highest risk of developing CVD. In Ethiopia recent studies report a CVD prevalence of 37.26% among diabetic patients. This study employed machine earning to predict CVD among Ethiopia diabetic patients using Ethiopian public Health Institute (EPHI) datasets, with a focus on identifying the most influential risk factors for public health decision-making.
OBJECTIVE: The main objective of this study is to predict CVD among diabetic patients in Ethiopia using machine learning techniques.
METHOD: The dataset comprised of 9030 instances with 22 features sourced from Ethiopian Public Health Institute. This prediction of cardiovascular disease (CVD) incorporated socio-demographic, behavioral, and clinical measurement data. Logistic regression, decision tree, Support Vector Machine, Random forest, Gradient boosting machine and artificial neural network were employed. Those models were trained on 80% of the data and tested on the remaining 20%. The analysis was conducted with python using 3.10.
RESULTS: According to the results analyzed, Gradient Boosting Model (GBM) demonstrated the highest overall performance, achieving an accuracy of 93%, followed closely by Logistic Regression (LR) with 90% accuracy. In terms of precision, GBM and LR performed comparably, while the LR achieved the highest recall at 88%. Regarding the F1 score, GBM attained 82%, indicating a strong balance between precision and recall. Additionally, the receiver operating characteristics (ROC) analysis showed that GBM had the largest area under the curve (AUC) of 0.96, reflecting superior descriptive ability 0.96.
CONCLUSION: The gradient boosting machine (GBM) model demonstrated the highest performance compared to the other models, achieving an accuracy of 93%. The most significant factors influencing the GBM model were total cholesterol, hypertension, and fasting blood glucose levels. The gradient boosting model shows potential for future integration into clinical decision-support systems, pending external validation and early prediction of cardiovascular disease in individuals with diabetes.

Keywords:  Cardiovascular disease; Diabetes; Gradient boosting model; Machine learning

DOI:  https://doi.org/10.1186/s12889-025-24850-2
Bioengineering (Basel). 2025 Nov 10. pii: 1231. [Epub ahead of print]12(11):

A Study on the Interpretability of Diabetic Retinopathy Diagnostic Models.

Zerui Zhang, Hongbo Zhao, Li Dong, Lin Luo, Hao Wang.

  This study focuses on the interpretability of diabetic retinopathy classification models. Seven widely used interpretability methods-Gradient, SmoothGrad, Integrated Gradients, SHAP, DeepLIFT, Grad-CAM++, and ScoreCAM-are applied to assess the interpretability of four representative deep learning architectures, VGG, ResNet, DenseNet, and EfficientNet, on fundus images. Through saliency map visualization, perturbation curve analysis, and trend correlation analysis, combined with four quantitative metrics-saliency map entropy, AOPC score, Recall, and Dice coefficient-the interpretability performance of the models is comprehensively assessed from both qualitative and quantitative perspectives. The results show that model architecture greatly influences interpretability quality: models with simpler structures and clearer feature extraction paths (such as VGG) perform better in terms of interpretability, while deeper or lightweight architectures exhibit certain limitations.

Keywords:  diabetic retinopathy classification models; fundus images; interpretability evaluation; model interpretability

DOI:  https://doi.org/10.3390/bioengineering12111231
BMC Endocr Disord. 2025 Nov 22.

Construction and validation of a hypoglycemia risk prediction model for hospitalized type 2 diabetes patients based on machine learning.

Caixia Liu, Zhaoxu Huang, Tuonan Liu, Yangyuan Ge, Jie Yuan, Yue Lin, Chunli Wang, Jinjuan Zhang, Xiaoli Wang, Yan Hua, Rui Qi.



Keywords:  Hospitalized type 2 diabetes; Hypoglycemia; Machine learning; Prediction model; XGBoost algorithm

DOI:  https://doi.org/10.1186/s12902-025-02104-x
bioRxiv. 2025 Oct 23. pii: 2025.10.22.683335. [Epub ahead of print]

Supervised machine learning identifies impaired mitochondrial quality control in β cells with development of type 2 diabetes.

Mirza Muhammad Fahd Qadir, Charles Dana, Paul Mauvais-Jarvis, Nazmul Haque, Theodore Dos Santos, Patrick E MacDonald, Franck Mauvais-Jarvis.

In type 2 diabetes (T2D), molecular pathways driving β cell failure are difficult to resolve with standard single cell analysis. Here we developed an interpretable, supervised machine learning framework that couples sparse rule-based classification (SnakeClassifier), pathway constrained modelling (BlackSwanClassifier), and β cell mitochondrial fitness stratification (Kolmogorov-Arnold Neural Networks KANN), linking and integrating them into disease mechanisms in single cell RNA sequencing (scRNA-seq) from 52 human donors. SnakeClassifier trained on 50 genes accurately predicted T2D at single cell resolution, outperforming classical ensemble machine learning classifier models, and yielded donor level diabetes scores that correlated with chronic hyperglycemia. The clustering of β cell populations (β1-4) revealed a resilient non-diabetic (ND) β1 subtype characterized by preserved β cell identity genes and lower disease risk, whereas T2D β2-4 subtypes exhibited upregulation of genes involved in cellular and mitochondrial stress and suppression of genes promoting oxidative phosphorylation and insulin secretion. Mitophagy emerged as the dominant program linked to T2D and a mitophagy focused BlackSwanClassifier nominated PINK1, BNIP3 , and FUNDC1 as key regulators. PINK1 was enriched in ND β1, decreased with T2D disease score and connected sex stratified mitophagy. We generated a KANN derived mitochondrial fitness index (MFI) integrating mitophagy, mitochondrial proteostasis, biogenesis and oxidative phosphorylation into a single interpretable score (R 2 = 0.934 vs module-based mitochondria quality index), which identified mitophagy PINK1, SQSTM1, PRKN and BNIP3 as top contributors to T2D progression. These transparent models unify prediction with T2D disease mechanism and identify the mitophagy receptor PINK1 as a central determinant of β cell metabolic fitness.

DOI: https://doi.org/10.1101/2025.10.22.683335
Sci Rep. 2025 Nov 22.

A dynamic weighted ensemble learning framework for cardiovascular risk prediction in type 2 diabetes: a comparative study with SHAP-based interpretability.

ChunHong Yuan, Ziyang Liu, Xiangyu Li, Xiang Zhou, Dongjun Wang, Yadong Fan, Xuan Sun, Zhikui Tian.

  Diabetes mellitus is one of the major issues in global public health and its cardiovascular complications are the primary cause of death in patients. Traditional risk assessment models, such as the Framingham Risk Score and the UKPDS Risk Engine, are built primarily on linear hypotheses, so they fail to capture complex non-linear relationships and show poor generalization across different populations. In this study, a multi-index dynamic weighted ensemble model was built by innovatively integrating TCM tongue diagnosis indexes with modern medical biomarkers. Specifically, adopting a cross-sectional design, this study initially enrolled 3,111 Type 2 diabetes patients, of which 2,895 were included in the final analysis after excluding 216 participants with excessive missing data ([Formula: see text]), built base models using Random Forest (RF), Gradient Boosting Decision Tree (GBDT), K-nearest Neighbors (KNN), and eXtreme Gradient Boosting (XGBoost) algorithms, and developed a dynamic weighted ensemble framework for optimizing model integration. According to the study results, the ensemble model achieved an accuracy of 95.68%, a sensitivity of 94.92%, and a specificity of 96.21%, remarkably outperforming the existing models. The SHAP analysis revealed that indexes such as chest tightness, ESR, and tongue purple had a significant non-linear influence on diabetes risk. Moreover, the ablation test results further proved the superiority of the ensemble framework over single-algorithm frameworks. This study effectively integrates TCM diagnosis indexes with Western medical indexes, provides a reliable decision-making tool for early screening and personalized intervention in diabetic complications, and shows great values for clinical application and transformations.

Keywords:  Cardiovascular complications; Ensemble learning; Machine learning; SHAP value; TCM tongue features; Type 2 diabetes

DOI:  https://doi.org/10.1038/s41598-025-28786-w
J Pers Med. 2025 Nov 05. pii: 537. [Epub ahead of print]15(11):

Identifying Risk Groups in 73,000 Patients with Diabetes Receiving Total Hip Replacement: A Machine Learning Clustering Analysis.

Alishah Ahmadi, Anthony J Kaywood, Alejandra Chavarria, Oserekpamen Favour Omobhude, Adam Kiss, Mateusz Faltyn, Jason S Hoellwarth.

  Background/Objective: Diabetes mellitus (DM) is a highly prevalent condition that contributes to adverse outcomes in patients undergoing total hip arthroplasty (THA). This study applied machine learning clustering algorithms to identify comorbidity profiles among diabetic THA patients and evaluate their association with postoperative outcomes. Methods: The 2015-2021 National Inpatient Sample was queried using ICD-10 CM/PCS codes to identify DM patients undergoing THA. Forty-nine comorbidities, complications, and clinical covariates were incorporated into clustering analysis. The Davies-Bouldin and Calinski-Harabasz indices determined the optimal number of clusters. Multivariate logistic regression assessed risk of non-routine discharge (NRD), and Kruskal-Wallis H testing evaluated length-of-stay (LOS) differences. Results: A total of 73,606 patients were included. Six clusters were identified, ranging from 107 to 61,505 patients. Cluster 6, enriched for urinary tract infection and sepsis, had the highest risk of NRD (OR 7.83, p < 0.001) and the longest median LOS (9.0 days). Clusters 1-4 had shorter recoveries with median LOS of 2.0 days and narrow variability, while Cluster 5 showed intermediate outcomes. Kruskal-Wallis and post hoc testing confirmed significant differences across clusters (p < 0.001). Conclusions: Machine learning clustering of diabetic THA patients revealed six distinct groups with varied comorbidity profiles. Infection-driven clusters carried the highest risk for non-routine discharge and prolonged hospitalization. This approach provides a novel framework for risk stratification and may inform targeted perioperative management strategies.

Keywords:  clustering; diabetes mellitus; machine learning; non-routine discharge; total hip arthroplasty

DOI:  https://doi.org/10.3390/jpm15110537
JMIR Form Res. 2025 Nov 27. 9 e81289

AI-Powered Thermography for Diabetic Foot Risk Stratification: Multicenter Cross-Sectional Study.

Meshari F Alwashmi, Mustafa Alghali, Waseem Abu-Ashour, Abdullah Mohammed Arabe, Abdullah Al Soheimi, Nasibh Abdulrahman Alharbi, Hani Mohammed Badahdah.

   BACKGROUND: Diabetic foot complications are among the most severe and costly outcomes associated with diabetes, with high prevalence particularly in the Middle East and North Africa region. Current screening tools are often limited by subjectivity, invasiveness, or scalability challenges, underscoring the need for innovative approaches.
OBJECTIVE: This multicenter study aimed to evaluate the performance of an artificial intelligence (AI)-powered thermographic system, Thermal Foot Scan (TFScan), in identifying patients at elevated risk of diabetic foot complications through noninvasive temperature profiling.
METHODS: A multicenter cross-sectional analysis of deidentified routine screening data across 4 regions in Saudi Arabia was conducted enrolling 1120 individuals with diabetes. Participants underwent thermal imaging using a smartphone-compatible infrared camera with AI algorithms analyzing angiosomal temperature patterns and asymmetries. Risk was stratified into 4 categories (very low, low, moderate, and high). Associations between TFScan classifications and clinical risk factors, symptoms of neuropathy, and thermal abnormalities were assessed.
RESULTS: While 90.7% (1016/1120) of the participants were classified as very low or low risk, 9.3% (104/1120) were identified as moderate or high risk. This higher-risk group exhibited significantly greater prevalence of key diabetic complications (P<.001). Peripheral artery disease was present in 20.2% (21/104) of moderate- and high-risk participants compared to just 0.8% (8/1016) of lower-risk individuals. Cardiovascular disease (60/104, 57.7% vs 313/1016, 30.8%), neuropathy (12/104, 11.5% vs 37/1016, 3.6%), foot deformities (15/104, 14.4% vs 6/1016, 0.6%), and symptoms of loss of protective sensation (53/104, 51% vs 354/1016, 34.8%) were all significantly more frequent in the high-risk subgroup than in the low-risk group, respectively. Thermal imaging further revealed pronounced abnormalities: temperature asymmetries of ≥2.2 °C were observed in 7.1% (79/1120) of the patients overall, with the highest asymmetry and thermal change index scores concentrated in the moderate- and high-risk groups. These individuals also exhibited greater deviations in angiosomal temperature differences-exceeding 2.2 °C in key vascular territories such as the medial plantar and lateral plantar arteries-suggesting both early inflammatory states and critical perfusion deficits.
CONCLUSIONS: The TFScan system effectively stratified patients with diabetes into clinically meaningful risk categories, with moderate- and high-risk groups exhibiting a significantly higher burden of vascular, neuropathic, and thermal abnormalities. However, the cross-sectional design, partial reliance on self-report, and low prevalence of advanced complications may limit causal inference. These findings highlight the potential of AI-enhanced thermography to serve as a scalable, objective screening tool for proactive diabetic foot management. Further longitudinal studies are warranted to validate its predictive power and support widespread clinical adoption.

Keywords:  AI; artificial intelligence; diabetic foot ulcer; digital health; risk stratification; screening; thermography

DOI:  https://doi.org/10.2196/81289
BMC Nephrol. 2025 Nov 26. 26(1): 669

Artificial intelligence-based diagnosis of diabetic kidney disease using urinary VOC biosensor data.

Chatchai Kreepala, Watcharapong Anakkamatee, Anawin Pechbooranin.

   BACKGROUND: Diabetic kidney disease (DKD) remains a leading cause of chronic kidney disease worldwide. However, current diagnostic methods rely on indirect biomarkers or invasive renal biopsy. This study aimed to evaluate the feasibility of urinary volatile organic compound (VOC) profiling, combined with machine learning, for non-invasive classification of DKD.
METHODS: Urine samples were collected from 127 participants divided into four diagnostic groups: healthy controls, patients with type 2 diabetes without nephropathy, biopsy-confirmed DKD, and patients with primary nephrotic syndromes. Samples were analyzed using a chemiresistive VOC biosensor. A total of 15,240 signal-derived features were extracted based on sensor response dynamics. Synthetic Minority Over-sampling Technique (SMOTE) was applied to balance class sizes. Four machine learning classifiers-Random Forest, Support Vector Machine, k-Nearest Neighbors, and Naïve Bayes-were trained and validated using stratified data. Performance was assessed using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC).
RESULTS: Random Forest achieved the best test performance, with 86% accuracy, 0.91 precision, 0.86 recall, F1-score of 0.86, and an AUC of 0.95. K-fold cross-validation confirmed the model's robustness and generalizability. Random Forest consistently outperformed other models in distinguishing DKD from both diabetic and non-diabetic glomerular diseases, demonstrating its strong discriminative capability.
CONCLUSIONS: Urinary VOC-based machine learning models provide proof-of-concept evidence for non-invasive DKD diagnosis. Random Forest, in particular, shows potential as a triage tool to differentiate DKD from other glomerular conditions, which may in the future help reduce reliance on biopsy and support earlier identification in nephrology practice.

Keywords:  Artificial intelligence; Biosensor; Diabetic kidney disease; Urine analysis; VOCs

DOI:  https://doi.org/10.1186/s12882-025-04608-z
J Diabetes Sci Technol. 2025 Nov 29. 19322968251384314

Artificial Intelligence for Diabetes Complication Prediction: A Systematic Review of Current Applications and Future Directions.

Francesca Pescol, Pietro Bosoni, Stefania Ghilotti, Pasquale De Cata, Lucia Sacchi, Riccardo Bellazzi.

  Importance and aims:Diabetes can lead to microvascular and macrovascular complications. Modeling the complex relationships between risk factors has motivated the use of Artificial Intelligence (AI) to develop predictive models. Recent advancements, including foundation models and generative AI, have significantly changed how this technology is applied across various contexts. In this review, we summarize the current state of research on AI for predictive diabetes complications, investigating the present and future implications of these innovations.
METHODS: We conducted the literature search on PubMed, Scopus, Ovid MEDLINE, CINAHL, and IEEE databases. Our analysis focused on predicted complications, population characteristics, use of AI-based approaches, models' performance, predictor variables, and feature importance evaluation results.
RESULTS: The 49 studies selected in our analysis considered different conditions as prediction outcomes. Eye-related complications were included in 29 studies (59%), emerging as the most frequent predicted diseases. Among the 48 studies employing AI algorithms specifically for the prediction task, 26 (54%) developed only Machine Learning models, 4 (8%) only Deep Learning models, and 18 (38%) applied both approaches. Foundation models and recent AI innovations included in the query were not used by any study. Moreover, only five studies (10%) dealt with unstructured data (signals and images). In the feature importance evaluation, age and glycated hemoglobin consistently emerged as important predictors.
CONCLUSIONS: Despite the extensive existing literature on AI for predicting diabetes complications, several emerging challenges persist. These include the effective utilization of unstructured data and the integration of recent advancements introduced by foundation models and generative AI.

Keywords:  artificial intelligence; diabetes complications; foundation models; risk prediction

DOI:  https://doi.org/10.1177/19322968251384314
Sci Rep. 2025 Nov 25. 15(1): 41892

Opportunistic screening of type 2 diabetes with deep metric learning using electronic health records.

Qixuan Jin, Haoran Zhang, Lukasz Szczerbinski, Jiacheng Zhu, Walter Gerych, Xuhai Xu, Kai Wang, Sarah Hsu, Ravi Mandla, Aaron J Deutsch, Alisa Manning, Josep M Mercader, Thomas Hartvigsen, Miriam S Udler, Marzyeh Ghassemi.

Deep learning models leveraging electronic health records (EHR) for opportunistic screening of type 2 diabetes (T2D) can improve current practices by identifying individuals who may need further glycemic testing. Accurate onset prediction and subtyping are crucial for targeted interventions, but existing methods treat the tasks separately, thus limiting clinical utility. In this paper, we introduce a novel deep metric learning (DML) model that unifies both tasks by learning a latent space based on sample similarity. In onset prediction, the DML model predicts the onset of T2D 7 years later with an AUC of 0.754, outperforming logistic regression (AUC 0.706), clinical risk factors (AUC 0.693), and glycemic measures (AUC 0.632). For subtyping, we identify three subtypes with varying prevalences of obesity-related, cardiovascular, and mental health conditions. Additionally, the subtype with fewer comorbidities shows earlier metformin initiation and a greater reduction in HbA1c. We validated these findings using data from 300 U.S. hospitals in the All of Us program (T2D, n = 7567) and the Massachusetts General Brigham Biobank (T2D, n = 3298), demonstrating the transferability of our model and subtypes across cohorts.

DOI: https://doi.org/10.1038/s41598-025-25759-x
Ageing Res Rev. 2025 Nov 25. pii: S1568-1637(25)00306-X. [Epub ahead of print] 102960

Applications of AI in the management of elderly diabetes patients.

Tianci Xie, Ziyu Wang, Jieyu He, Jingjing Zhang.

  The prevalence of diabetes among the elderly population has been rising rapidly. Elderly individuals with diabetes frequently exhibit atypical symptoms, and face an increased risk of severe complications, comorbidities and geriatric syndromes, including hypoglycemia, sarcopenia, cognitive impairment, depression, and falls. These issues not only diminish quality of life, but also contribute to elevated disease burden. Conventional diabetes management strategies for this population often fall short due to their invasive nature, limited capacity for real-time monitoring, and poor patient adherence. The emergence of artificial intelligence (AI) in medicine offers a transformative solution to these challenges. Fueled by large-scale datasets, advanced machine learning algorithms, and state-of-the-art computational techniques, AI enables non-invasive, real-time monitoring systems and delivers personalized diagnostic and treatment solutions for this population. This review comprehensively explores the role of AI in managing diabetes and its comorbid conditions among older adults, highlighting its applications in screening, diagnostic, monitoring, and therapeutic strategies. It also addresses the practical challenges and ethical considerations of integrating AI into clinical practice.

Keywords:  Artificial intelligence; Diabetes mellitus; Geriatric syndromes; Machine learning; Older adults

DOI:  https://doi.org/10.1016/j.arr.2025.102960
medRxiv. 2025 Nov 17. pii: 2025.10.23.25338682. [Epub ahead of print]

Medication-Stratified Analysis of LDL-C Equation Miscalibration in Diabetes: Evidence from the All of Us Research Program and a Medication-Agnostic Machine-Learning Correction.

Ronald Doku, Nana Yaw Osafo, John Kwagyan, William M Southerland.

Objective: Standard LDL-C equations were derived in cohorts largely untreated with modern combination diabetes therapies. With medication-treated patients comprising 84% on statins, 53% on insulin, and 25% on GLP-1 receptor agonists-often in combination-we quantified medication-specific miscalibration in LDL-C equations and evaluated a machine learning correction that operates without requiring medication data.
Research Design and Methods: Using All of Us Research Program data (n=3,477; test =696), we compared Friedewald, Martin-Hopkins, and Sampson (NIH) Equation 2 against direct LDL-C measurements. We developed a stacked ensemble model (elastic net, random forest, XGBoost, neural network) trained solely on routine laboratory values. Accuracy was assessed within medication groups allowing for combination therapy: insulin users, GLP-1 users, and statin users. Primary endpoints: mean absolute error (MAE) with 95% bootstrap confidence intervals and calibration (ordinary least squares regression of true on predicted LDL-C). Secondary endpoint: Net Reclassification Index at 100 mg/dL.
Results: Among 696 test participants, 587 (84%) used statins, 366 (53%) insulin, and 175 (25%) GLP-1 agonists. Patients on triple therapy (insulin+GLP-1+statin) showed the most severe miscalibration: Friedewald slope 0.29, representing 71% compression of the prediction range. In all GLP-1 users (77% also on insulin), standard equations severely underestimated LDL-C with calibration slopes of 0.42-0.48 versus ideal 1.0. Specifically, Friedewald showed slope 0.42 (95% CI 0.27-0.56) with intercept +62 mg/dL; Sampson (NIH) Equation 2 slope 0.48 (0.32-0.64) with intercept +55 mg/dL; Martin-Hopkins slope 0.47 (0.31-0.63) with intercept +55 mg/dL. The machine learning model maintained better calibration (slope 0.83 [0.56-1.09]; intercept -2.2 mg/dL) and reduced MAE by 17% versus Friedewald. Insulin users showed similar improvement: Friedewald slope 0.55 (0.45-0.65) versus the machine learning (ML) model 0.95 (0.78-1.12), with 16% lower error. The medication-by-triglyceride interaction was significant (p=0.002). In patients with insulin exposure and triglycerides ≥200 mg/dL, Net Reclassification Index was 0.240 versus 0.022 overall, indicating greater misclassification risk in hypertriglyceridemia.
Conclusions: Standard LDL-C equations systematically underestimate true levels in medication-treated diabetes patients, with errors greatest in combination therapy. A machine learning model trained on routine laboratories-without medication data-achieved near-ideal calibration (slopes 0.83-1.03) and reduced errors by 8-20% across medication groups. These observational findings suggest direct LDL-C measurement or ML-assisted correction should be considered when equation estimates approach treatment thresholds, particularly for patients on combination therapy.

DOI: https://doi.org/10.1101/2025.10.23.25338682
Eur J Med Res. 2025 Nov 27.

Neutrophil percentage to albumin ratio and advanced cardiovascular-kidney-metabolic syndrome in diabetes: a machine learning approach.

Zifan Zhuang, Xinyao Liu, Hongrui Lu, Qian Wu, Yupeng Chen, Qing Ni.

   BACKGROUND: Diabetes mellitus (DM) has become an increasingly significant global health challenge, with rising complications and mortality rates. Patients with DM are at a higher risk for advanced Cardiovascular-kidney-metabolic (CKM) syndrome, underscoring the importance of early detection and precise prevention strategies. The neutrophil percentage to albumin ratio (NPAR), a composite biomarker, may be indicative of inflammatory dysregulation and nutritional status in patients with advanced CKM and DM. This study aims to explore the association between NPAR and advanced CKM in patients with diabetes.
METHODS: Data were derived from six National Health and Nutrition Examination Survey (NHANES) cycles (1999-2020), including 9375 adults. Multivariable logistic regression analyses were conducted to assess the association of NPAR with advanced CKM in diabetes, and restricted cubic spline (RCS) regression was further applied to explore potential nonlinear relationships. Subgroup analyses were performed to explore differences across various population factors. Feature selection was carried out using the Boruta algorithm, and predictive performance was evaluated through receiver operating characteristic (ROC) analysis, calibration curves, and decision curve analysis (DCA).
RESULTS: This study found a positive association between NPAR and risk of advanced CKM in diabetes. The RCS analysis revealed that this positive correlation was linear. Subgroup analysis showed no significant interactions across groups. Feature selection identified 22 relevant variables, and the machine learning model demonstrated excellent predictive accuracy (Area under curve: 0.873 and 0.872 for training and validation sets). The calibration curves and DCA affirmed the model's clinical relevance.
CONCLUSIONS: This study suggests that an elevated NPAR may serve as a potential marker for advanced CKM in patients with diabetes. It holds potential as an adjunct tool for detection and management of advanced CKM in patients with DM, offering valuable insights for clinical practice.

Keywords:  Cardiovascular-kidney-metabolic syndrome; Diabetes mellitus; Machine learning; NHANES; Neutrophil percentage to albumin ratio

DOI:  https://doi.org/10.1186/s40001-025-03560-w