bims-aukdir Biomed News
on Automated knowledge discovery in diabetes research
Issue of 2025–10–12
seventeen papers selected by
Mott Given



  1. Am J Ophthalmol. 2025 Oct 04. pii: S0002-9394(25)00528-8. [Epub ahead of print]
       BACKGROUND: Diabetic retinopathy (DR) is a leading cause of preventable blindness globally. Although early detection via routine retinal screening significantly reduces vision loss, screening rates remain suboptimal due to workforce shortages and limited accessibility. Autonomous artificial intelligence (AI) systems such as EyeArt offer an FDA-authorized solution for point-of-care DR screening without ophthalmologist oversight METHODS: We conducted a systematic review and meta-analysis following PRISMA-DTA guidelines to assess the diagnostic accuracy of EyeArt in detecting referable diabetic retinopathy (rDR) from color fundus photographs. Searches of PubMed, Embase, and ClinicalTrials.gov through April 2025 identified eligible studies involving adult populations screened with EyeArt. Sensitivity and specificity were pooled using bivariate random-effects models. Subgroup and applicability analyses were conducted to evaluate heterogeneity and clinical relevance.
    RESULTS: Seventeen studies comprising 162,695 examinations were included. EyeArt demonstrated a pooled sensitivity of 95% (95% CI: 92-97%) and specificity of 81% (95% CI: 74-87%). Subgroup analyses indicated consistent accuracy across study designs, economic settings, healthcare contexts, device types, external validation and image gradability. Specificity varied slightly with vendor involvement.
    CONCLUSION: Across 17 real-world studies (162,695 examinations), EyeArt exhibits high diagnostic accuracy for detecting referable diabetic retinopathy (pooled sensitivity 95%, specificity 81%), with high certainty for sensitivity and moderate certainty for specificity. Its consistently strong sensitivity supports autonomous screening in primary care. However, variability in specificity-along with inconsistent reporting/handling of ungradable images-warrants attention and standardized quality-assurance. Successful deployment will depend on workflow/EHR integration, sustainable reimbursement, and targeted implementation in underserved populations to maximize public-health impact.
    Keywords:  Autonomous artificial intelligence; Deep learning; Diabetic retinopathy; Diagnostic accuracy; EyeArt; Fundus photography; Meta-analysis; Point-of-care screening; Systematic review; Tele-ophthalmology
    DOI:  https://doi.org/10.1016/j.ajo.2025.09.045
  2. Medicine (Baltimore). 2025 Oct 03. 104(40): e44871
      Diabetic peripheral vascular disease (DPVD) and diabetic foot (DF) are major complications that lead to disability in diabetic patients, severely impaired their quality of life. Firstly, this study gathered cross-sectional data from 1240 patients with type 2 diabetes and its complications in the the department of vascular surgery and endocrinology of the second affiliated hospital of zhejiang university school of medicine. In the pre-processing part, samples with serious data loss are eliminated, and the data are processed by methods such as MICEforest. After that, random forest (RF), support vector machine (SVM), backpropagation neural network (BPNN), extreme gradient boosting (XGBoost), and SHapley Additive exPlanation (SHAP) were employed to rank the importance of the 27 indicators. The entropy weight method was then applied to comprehensively assign weights to all indexes. Finally, the genetic neural network algorithm (GA-BPNN) was introduced to construct a prediction model for diabetes complications. In addition, the SHAP algorithm was applied to obtain the weight and importance ranking of each risk factor in the prediction model. This study identified the top 17 key indicators through a comprehensive weighting approach. Among the 5 classification models evaluated, the GA-BPNN algorithm exhibited the best performance in both diabetes and DPVD (G1), DPVD and DF (G2), achieving the area under the receiver operating characteristic curve (AUC) values of 0.79 and 0.89, accuracy rates of 0.78 and 0.80, and F1-scores of 0.77 and 0.83, respectively. Furthermore, hypothesis testing results indicate that indicators such as fibrinogen and c-reactive protein show statistically significant differences between groups. SHAP feature importance analysis also highlights the significant influence of these features in identifying diabetic complications. GA-BPNN can be employed as a prediction model for DPVD and DF. In feature selection, the comprehensive weighting method and SHAP analysis identified key features. In summary, this study constructed a comprehensive prediction model based on machine learning and interpretable algorithms, integrating diabetes-specific indicators, traditional cardiovascular risk factors, coagulation function, inflammatory markers, and cardiac structural parameters. It can effectively identify high-risk patients for diabetic complications, uncover potential features, and thereby assist in subsequent efforts to reduce the incidence of these complications.
    Keywords:  SHapley Additive exPlanation; diabetes; genetic neural network algorithm; machine learning
    DOI:  https://doi.org/10.1097/MD.0000000000044871
  3. Sci Rep. 2025 Oct 08. 15(1): 35076
      Diabetic Retinopathy (DR) continues to be the leading cause of preventable blindness worldwide, and there is an urgent need for accurate and interpretable framework. A Multi View Cross Attention Vision Transformer (MVCAViT) framework is proposed in this research paper for utilizing the information-complementarity between the dually available macula and optic disc center views of two images from the DRTiD dataset. A novel cross attention-based model is proposed to integrate the multi-view spatial and contextual features to achieve robust fusion of features for comprehensive DR classification. A Vision Transformer and Convolutional neural network hybrid architecture learns global and local features, and a multitask learning approach notes diseases presence, severity grading and lesions localisation in a single pipeline. Results show that the proposed framework achieves high classification accuracy and lesion localization performance, supported by comprehensive evaluations on the DRTiD dataset. Attention-based visualizations further enhance interpretability, indicating the framework's potential for clinical use. This framework establishes a criterion for improving state-of-the-art retinal image analysis for DR diagnosis which may result in better patient results and final clinical decision.
    Keywords:  Cross-attention mechanism; Diabetic retinopathy classification; Domain adaptation in retinal analysis; Multi-view retinal image fusion; Vision transformer (ViT)
    DOI:  https://doi.org/10.1038/s41598-025-18742-z
  4. Eur J Med Res. 2025 Oct 10. 30(1): 961
       BACKGROUND: Cardiovascular disease remains the predominant cause of morbidity and mortality in individuals with type 2 diabetes mellitus (T2DM). Traditional risk models are limited in predictive accuracy. Pericoronary adipose tissue (PCAT), a novel imaging biomarker of vascular inflammation, may offer additional prognostic value. Therefore, this study aimed to develop and validate a machine learning model that integrates PCAT parameters with clinical risk factors to improve the accuracy of cardiovascular risk prediction in individuals with T2DM.
    METHODS: This study retrospectively enrolled 686 hospitalized T2DM patients from four branches of Guangdong Provincial Hospital of Chinese Medicine between January 2017 and December 2021. PCAT-FAI and volume index were measured using coronary CTA. Major adverse cardiovascular events (MACE) were recorded during follow-up. Eight machine learning algorithms were applied, and multiple evaluation metrics were used to compare the predictive performance of the models. Feature contributions in the best-performing model were interpreted using both feature importance ranking and SHapley Additive exPlanations (SHAP) values.
    RESULTS: A total of 183 patients experienced MACE during the mean 38.4 months of follow-up. Among the eight machine learning models evaluated, the XGBoost model performed the best in predicting MACE in patients with T2DM. In the internal validation of the training set, the AUC was 0.818 (95% CI 0.777-0.858), and in the external test set, the AUC was 0.809 (95% CI 0.700-0.918). Additionally, the XGBoost model outperforms other models in all evaluation metrics (accuracy = 0.824, specificity = 0.882, F1 score = 0.654, Brier score = 0.248). In the feature importance analysis of the prediction model, RCA-FAI in the PCAT parameters consistently ranked among the top three in eight ML models. Further SHAP analysis indicated that RCA-FAI, body mass index (BMI), and the monocyte/high-density lipoprotein cholesterol ratio (MHR) were the most influential factors for MACE in patients with T2DM.
    CONCLUSION: This study demonstrates the independent predictive value of PCAT parameters for long-term cardiovascular risk in patients with T2DM. The XGBoost model showed promise as a potential clinical decision support tool. Integrating PCAT parameters with conventional risk factors may improve the identification of high-risk individuals and enhance the ability to predict MACE in this population. Clinical trial registration ChiCTR2400079869.
    Keywords:  Machine learning; Major adverse cardiovascular events; Pericoronary adipose tissue; Predictive model; Type 2 diabetes mellitus
    DOI:  https://doi.org/10.1186/s40001-025-03237-4
  5. J Eval Clin Pract. 2025 Oct;31(7): e70284
       OBJECTIVE: Diabetes mellitus is a chronic disease that presents significant health challenges worldwide. Accurate diabetes prediction facilitates early intervention and personalized healthcare strategies, thereby improving patient care and reducing healthcare processing costs. Ensemble-based machine learning (ML) methods enhance predictive performance.
    METHOD: This study explores various ML classifiers, both individually and in ensemble configurations, including decision trees, random forests, k-nearest neighbors, Naive Bayes, AdaBoost (AB), XGBoost (XB), and multilayer perceptron (MLP) for prediction. The performance of each method is evaluated through rigorous experimentation and comparative analysis across multiple aspects.
    RESULTS: The performance of the best ML model, MLP, is compared with that of the proposed CatBoost classifier and the ensemble model to identify the most effective approach for diabetes prediction in minimal duration. The proposed CatBoost classifier's execution time of 4.27 s, which is approximately 98.64% faster than the ensemble model's 314.96 s. This demonstrates CatBoost's significant advantage in computational efficiency over ensemble-based classifiers.
    CONCLUSION: By leveraging the diverse and complementary strengths of ML classifiers, this study contributes to the advancement of precision medicine and personalized healthcare for individuals at risk of diabetes.
    Keywords:  AdaBoost; XGBoost; diabetes mellitus; ensemble‐based machine learning; multilayer perceptron; performance metrics; personalized healthcare
    DOI:  https://doi.org/10.1111/jep.70284
  6. Biomed Phys Eng Express. 2025 Oct 07.
      Cardiovascular disease (CVD) is a major cause of morbidity and mortality in diabetic populations. Early detection of cardiovascular risk in diabetes is crucial to reduce complications, particularly in resource-limited settings. This study aimed to develop and evaluate a hybrid machine learning framework that integrates Long Short-Term Memory (LSTM) networks with traditional algorithms to improve cardiovascular risk prediction in diabetic patients. The hybrid model, which included structured data and time-series health data, was tested on a sample of 1,000 diabetes patients. Using 10-fold cross-validation, the model achieved impressive predictive performance (accuracy 98.7%, AUC 0.99). There are three main conclusions from this study. Initially, the hybrid model demonstrated a significant increase in CVD prediction accuracy when compared to independent machine-learning techniques. Second, the model provided reasonable predictions across different demographic groupings, ensuring equitable outcomes. Finally, the model's high performance supports its potential for future use in clinical decision-support systems aimed at improving outcomes and optimizing resource allocation. Increased CVD screening rates in diabetic patients, better access to care for communities with limited resources, and the advancement of health equity are all possible outcomes of incorporating machine learning and deep learning techniques. The proposed hybrid model also demonstrates strong potential for clinical deployment in cardiovascular risk prediction among diabetic populations, supporting earlier interventions and improved patient outcomes.
    Keywords:  Cardiovascular disease in diabetes; Diagnostic accuracy; Hybrid model; predicting cardiovascular disease
    DOI:  https://doi.org/10.1088/2057-1976/ae103a
  7. PLoS One. 2025 ;20(10): e0333388
       OBJECTIVE: As an emerging insulin resistance marker, the relationship between estimated glucose disposal rate (eGDR) and frailty needs further exploration. This study examines the eGDR-frailty link, develops a machine learning predictive model to address this gap, and explores diabetes mellitus (DM) as a mediator, providing new insights for clinical intervention.
    METHODS: Using National Health and Nutrition Examination Survey (NHANES) 2005-2010 data, we analyzed glucose disposal and frailty associations. Feature selection used LASSO, and class imbalance was handled by SMOTEN. The resampled data were split 7:3 into a training set (n = 29,309) and a test set (n = 12,561).Ten machine learning models were built, with discrimination, calibration, and clinical utility evaluated to identify the optimal model. Confusion matrices visualized performance. Mediation analysis assessed DM's role in the eGDR-frailty relationship.
    RESULTS: Among 26,282 participants, eGDR negatively correlated with frailty. Higher eGDR significantly reduced frailty risk in subgroups: women, age ≤ 60, normal/high BMI, never/current smokers, and alcohol users. LASSO selected 12 predictors. Across 10 models, CatBoost performed best on the test set (AUC = 0.970, accuracy = 0.920, F1 = 0.918), with robust calibration and decision-curve net benefit. SHAP interpretation ranked eGDR among the most influential predictors: SHAP summary and dependence plots indicated that higher eGDR decreased the model's predicted probability of frailty. Confusion matrices validated classification accuracy. Mediation analysis showed DM partially mediated the eGDR-frailty relationship: indirect effect β=-0.003 (95% CI -0.003 to -0.002; P < 0.001), mediation proportion = 8.71%.
    CONCLUSION: This first NHANES-based study demonstrates a significant negative correlation between eGDR and frailty, confirming DM's partial mediating role. The developed machine learning models effectively support early frailty risk assessment and intervention.
    DOI:  https://doi.org/10.1371/journal.pone.0333388
  8. Medicine (Baltimore). 2025 Oct 03. 104(40): e44732
      The global prevalence of type 2 diabetes mellitus (T2D) has been increasing dramatically as well as diabetic kidney disease (DKD). We aimed to compare the accuracies of 4 machine learning (Mach-L) methods with multiple linear regression (MLR) in predicting future estimated glomerular filtration rate (eGFR) in T2D patients and to rank the importance of DKD risk factors. The study was conducted from 2013 to 2019. Nine hundred and seven T2D patients were followed up for 4 years. Data of potential DKD risk factors were collected and calculated. We used 4 different Mach-L methods to predict the eGFR, including classification and regression tree, random forest, artificial neural network, and eXtreme Gradient Boosting. Simple correlation was applied to overview the relationships between baseline risk factors and eGFR at the end of follow-up (eGFRend). Besides, traditional MLR was used as a benchmark to evaluate if Mach-L methods could outperform MLR. For model interpretability, Shapley additive explanation was applied to explain the contribution of each feature and directions of impacts in the prediction model. In 4 different Mach-L methods, random forest, classification and regression tree, and eXtreme Gradient Boosting were more superior than MLR in the prediction of the eGFRend. The first 6 important risk factors in predicting diabetic eGFRend were body mass index (BMI), baseline high-density lipoprotein cholesterol (HDL-C), baseline urine microalbumin creatinine ratio (MCR), baseline low-density lipoprotein cholesterol (LDL-C), duration of diabetes, and age. By applying Shapley additive explanation, it appeared that age, duration of diabetes, HDL, and LDL were positively related to eGFRend and BMI and MCR were negatively related to eGFRend. Mach-L methods were proved to be more accurate in predicting eGFRend than traditional MLR. BMI presented the most influential factor for eGFRend, followed by HDL-C, baseline urine MCR, LDL-C, duration of diabetes, and age. These findings highlight the potential of Mach-L to enhance early risk stratification for DKD, enabling timely interventions to preserve renal function in T2D patients.
    Keywords:  estimated glomerular filtration rate; learning; machine; type 2 diabetes
    DOI:  https://doi.org/10.1097/MD.0000000000044732
  9. Sci Rep. 2025 Oct 10. 15(1): 35430
      Diabetes is a lifelong condition that occurs when the pancreas loses its ability to secrete insulin or experiences a significant reduction in insulin production. Early identification of high-risk patients is crucial for timely interventions and improved outcomes. Traditional clinical risk prediction models rely on regression analysis using clinical, sociodemographic, and anthropometric data; however, they have limitations in terms of accuracy and generalizability. This research proposes a diagnostic strategy leveraging machine learning (ML) techniques, specifically the XGBoost algorithm optimised with Optuna, to enhance high-risk prediction based on laboratory parameters. The study utilises an open-access diabetes dataset incorporating patient demographics, laboratory test results, and clinical outcomes. Data preprocessing, including cleaning, normalisation, and feature extraction, is performed using an Adaptive Tree-Structured Parzen Estimator (ATPE) and XGBoost model. The proposed model outperforms conventional classification models, achieving 83% accuracy, 80% precision, 78% recall, and a 78% F1 score. A comprehensive correlation and confusion matrix evaluation highlights the model's effectiveness in distinguishing high-risk patients. Findings indicate that integrating machine learning (ML)-based risk classification frameworks with laboratory test-based diagnostic strategies improves predictive accuracy and patient stratification. However, data quality, population diversity, and real-time applicability remain challenges. Future research should explore the integration of real-time data from wearable devices and expand model deployment to other chronic and rare diseases, enhancing adaptability and clinical decision-making.
    Keywords:  Diabetes; Diagnostic strategies; Laboratory parameters; Machine learning; Optuna; Risk prediction; XGBoost
    DOI:  https://doi.org/10.1038/s41598-025-19295-x
  10. Saf Health Work. 2025 Sep;16(3): 355-360
       Background: Diabetes contributes significantly to death in the U.S., with many working-age individuals affected. This research determined the independent and joint associations of long working hours and night work with diabetes risk in U.S. workers, and their contribution to risk prediction.
    Methods: This prospective study included 1,454 workers from the Midlife in the United States (MIDUS) study with 9-year follow-up. Long working hours included those working 55 or more hours per week. Night work involved those working 16 or more nights per year. Diabetes was determined by self-reported diagnosis or treatment. Multivariable Poisson regression analysis was applied to examine the prospective association of these work-related factors at baseline with incident diabetes. A gradient boosting machine learning model was used to investigate the contributions of both factors in predicting incident diabetes.
    Results: Long working hours (RR and 95% CI = 1.60 [1.04, 2.46], p < 0.05) and night work (RR and 95% CI = 1.66 [1.05, 2.62], p < 0.05) were independently associated with the risk for diabetes, while controlling for baseline covariates. Gradient boosting analysis suggested long working hours and night work facilitated diabetes incidence. Exposure to both long working hours and night work increased the risk for diabetes (RR and 95% CI = 3.02 [1.64, 5.58], p < 0.001), suggesting additive interaction.
    Conclusion: Organizations may consider reducing hours on duty and improving shift systems for primary prevention of diabetes.
    Keywords:  diabetes; long working hours; machine learning; night work; prospective cohort
    DOI:  https://doi.org/10.1016/j.shaw.2025.05.005
  11. Sci Rep. 2025 Oct 09. 15(1): 35381
      Type 2 diabetes mellitus (T2DM) is a chronic metabolic disease that increases the risk of cardiovascular complications. The atherogenic index of plasma (AIP) is a risk marker for T2DM and cardiovascular disease on the basis of lipid profiles. T2DM and CVD risk are also associated with nonlipid biomarkers, including oxidative stress, inflammation, and mitochondrial dysfunction, and are linked to diabetes progression. This study applies hierarchical random forest (HRF) machine learning to identify stage-specific predictors of AIP in normoglycemic, prediabetic, and diabetic individuals. Participants were divided into normal (< 5.7%), prediabetic (5.7-6.4%), and diabetic (≥ 6.5%) groups based on their HbA1c values. Clinical, oxidative, inflammatory, and mitochondrial biomarkers were included in the study. Lipid measures directly contributing to the AIP calculation were excluded to minimize collinearity. Predictive models were developed via random forest (RF) and hierarchical random forest (HRF) approaches. HRF incorporates repeated threefold cross-validation to improve stability and feature importance across subgroups. Model performance was evaluated via the coefficient of determination (R²) and mean squared error (MSE). HRF models revealed distinct biomarker profiles associated with AIP and diabetes progression associated with inflammation, oxidative stress, and mitochondrial function variables. Waist-to-height ratio was the main contributing variable in the stratified dataset. For the stratified data, mitochondrial redox markers (p66Shc, humanin) were among the top predictors in the normoglycemia group. In individuals with prediabetes, the importance of these cytokines decreased, whereas oxidative stress-associated biomarkers (GSH, 8-OHdG) provided more accurate classifications. In the diabetes group, 8-OHdG remained moderately predictive, whereas the mitochondrial peptide MOTSc and inflammatory markers (IL-1β) were key features. These results indicate that the progression from mitochondrial-associated changes in the early stages of diabetes to immunometabolic dysfunction in individuals with established diabetes is correlated with AIP. Hierarchical random forest machine learning combined with glycemic stratification reveals evolving biomarker associations with the atherogenic index of plasma linked with diabetes progression. Mitochondrial and immune markers contribute differently across disease stages, supporting their potential use in stage-specific risk stratification and targeted intervention in T2DM management.
    Keywords:  Atherogenic index of plasma; Cardiometabolic risk prediction; Machine learning models; Mitochondrial biomarkers; Type 2 diabetes mellitus
    DOI:  https://doi.org/10.1038/s41598-025-19289-9
  12. Cureus. 2025 Sep;17(9): e91520
      Diabetes mellitus (DM) is a long-term metabolic condition involving persistent hyperglycemia, which causes morbidity, mortality, and economic stress. This article examines the role of artificial intelligence (AI)-initiated precision medicine in optimizing type 2 diabetes mellitus management in the Indian population. An exhaustive review of AI-based platforms used for diabetic treatment was performed, highlighting the combination of multidimensional data sets involving genetic, epigenetic, phenotypic, and environmental variables. The envisioned AI platform aims to offer personalized glycemic forecasts, tailored therapeutic interventions, complication monitoring, and stage-by-stage disease progression predictions. Precision medicine enabled by AI has shown promising outcomes in improving diabetes management through the administration of patient-specific treatment regimens, early glycemic change detection, and real-time monitoring of diabetes-related complications. The use of AI applications enables patients to follow evidence-based self-management behaviors, such as diet modifications, physical activity changes, insulin management, and continuous glucose monitoring. This patient-centered strategy enhances clinical efficacy, prevents long-term complications, and lowers healthcare costs. Additional longitudinal and multicentric trials are needed to confirm outcomes among heterogeneous cohorts and to fine-tune AI algorithms for increased clinical relevance and translational use.
    Keywords:  ai tools; diabetes mellitus; glycemic fluctuations; indian phenotype; personalized medicine (pm)
    DOI:  https://doi.org/10.7759/cureus.91520
  13. JMIR Med Inform. 2025 Oct 10. 13 e71994
       BACKGROUND: Machine learning (ML) has shown great potential in recognizing complex disease patterns and supporting clinical decision-making. Diabetic foot ulcers (DFUs) represent a significant multifactorial medical problem with high incidence and severe outcomes, providing an ideal example for a comprehensive framework that encompasses all essential steps for implementing ML in a clinically relevant fashion.
    OBJECTIVE: This paper aims to provide a framework for the proper use of ML algorithms to predict clinical outcomes of multifactorial diseases and their treatments.
    METHODS: The comparison of ML models was performed on a DFU dataset. The selection of patient characteristics associated with wound healing was based on outcomes of statistical tests, that is, ANOVA and chi-square test, and validated on expert recommendations. Imputation and balancing of patient records were performed with MIDAS (Multiple Imputation with Denoising Autoencoders) Touch and adaptive synthetic sampling, respectively. Logistic regression, support vector machine (SVM), k-nearest neighbors, random forest (RF), extreme gradient boosting (XGBoost), Bayesian additive regression trees, and artificial neural network were trained, cross-validated, and optimized using random sampling on the patient dataset. To evaluate model calibration and clinical utility, calibration curves, Brier scores, and decision curve analysis (DCA) were performed.
    RESULTS: The exploratory dataset consisted of 700 patient records with 199 variables. After dataset cleaning, the variables used for model training included age, smoking status, toe systolic pressure, blood pressure, oxygen saturation, hemoglobin, hemoglobin A1c, estimated glomerular filtration rate, wound location, diabetes type, Texas wound classification, neuropathy, and wound area measurement. The SVM obtained a stable accuracy of 0.853 (95% CI 0.810-0.896) with an area under the receiver operating characteristic curve of 0.922 (95% CI 0.889-0.955). The RF and XGBoost acquired an accuracy of 0.838 (95% CI 0.793-0.883) and 0.815 (95% CI 0.768-0.862), respectively, with areas under the receiver operating characteristic curve of 0.917 (95% CI 0.883-0.951) for RF and 0.889 (95% CI 0.849-0.929) for XGBoost. SVM, RF, and XGBoost were well-calibrated, with average Brier scores around 0.127 (SD 0.013). DCA showed that the SVM provided the highest net clinical benefit across relevant risk thresholds.
    CONCLUSIONS: Handling missing values, feature selection, and addressing class imbalance are critical components of the key steps in developing ML applications for clinical research. Seven models were selected for comparing their predictive power regarding complete wound healing, and each model representing a different branch in ML. In this initial DFU dataset used as an example, the SVM achieved the best performance in predicting clinical outcomes, followed by RF and XGBoost. The model's calibration and clinical utility were determined through calibration curves, Brier scores, and DCA, demonstrating its potential relevance in clinical decision-making.
    Keywords:  Bayesian additive regression trees; artificial neural network; complete wound healing; diabetic foot ulcer; extreme gradient boosting; k-nearest neighbor; logistic regression; machine learning; random forest; support vector machine
    DOI:  https://doi.org/10.2196/71994
  14. Anal Chem. 2025 Oct 10.
      Aldehyde compounds are significantly associated with diabetes mellitus. The metabolic profile of aldehydes can enhance understanding of the mechanisms underlying development of diabetes. This study employed a pair of stable isotope labeling (SIL) reagents, N-((1-phenyl-1H-1,2,3-triazol-4-yl)methyl)hydroxylamine (PTMH) and N-((1-(phenyl-d5)-1H-1,2,3-triazol-4-yl)methyl)hydroxylamine (PTMH-d5), for aldehyde profiling, address challenges related to selectivity, isomer formation, and transamination that occur with conventional labels, such as hydrazide or amine reagents. The metabolic profiling of 28 aldehydes on the serum samples of patients with type 2 diabetes mellitus (T2DM, n = 39) and gestational diabetes mellitus (GDM, n = 37) was carried out using PTMH/PTMH-d5. Furthermore, comparative metabolomic analyses of T2DM and GDM against healthy controls were performed. Moreover, advanced informatics approaches, including PCA, ROC, and PLS-DA, were employed for statistical evaluation. A machine learning classification model was also developed. The results revealed that 4-hydroxyhexenal, methylglyoxal, and trans-2-pentenal may serve as potential biomarkers for T2DM, whereas 4-hydroxyhexenal, methylglyoxal, heptanal, 5-hydroxymethylfurfural, and trans-2-octenal can be employed as potential biomarkers for GDM. The established model demonstrated significant potential as a prototype for early and accurate diagnosis of T2DM and GDM and may be translated into routine clinical diagnostics.
    DOI:  https://doi.org/10.1021/acs.analchem.5c04501
  15. J Wound Care. 2025 Oct 01. 34(Sup10): S18-S29
       OBJECTIVE: Diabetes frequently results in diabetic foot ulcers (DFUs), which can lead to lower limb amputation if left untreated. Current DFU management consists of a multidisciplinary team approach, including physicians, podiatrists, wound care specialists, nursing staff and patients. Traditional diagnoses of DFUs can be expensive, lengthy, and generally reliant on local and private clinical evaluation. There is a need for an automated, remote, diagnostic option for patients with suspected DFUs.
    METHOD: This paper introduces MGWONET, a new model for deep learning (DL) on classification problems, using the Modified Grey Wolf Optimisation (MGWO) algorithm to automatically find the optimal configuration of hyperparameters. The automation settings are intended to govern the model's operation, and include the number of layers, learning rate and filter sizes. To improve the accuracy and efficiency of DL models, it is essential to carefully select the right hyperparameters. However, choosing the best combination involves searching through a large number of possible settings, which is computationally challenging and complex. Unlike manual tuning, which is time-consuming and inefficient, the MGWO algorithm efficiently explores the large and complex hyperparameter search space to improve the model's accuracy and robustness. MGWONET is intended for classifying skin patches as healthy or ulcerated, and was trained with an augmented dataset of 2200 DFU images. Generally recognised metrics, such as accuracy, recall, precision, specificity, balanced classification rate and the receiver operating characteristic curve were used to evaluate optimal model performance.
    RESULTS: The optimal model had a classification accuracy equating to 98.64% and was superior to a selection of well-known DL architectures: AlexNet; VGG16; GoogLeNet; and a monochrome baseline Grey Wolf optimised model. This study is centred on binary image-based classification and does not constitute a clinical grading system; however, the method presented here has the potential to be a valuable supportive addition to clinical practice.
    CONCLUSION: The MGWONET framework has proved to be highly reliable and provides robust discriminative power, making it a strong candidate for automated DFU diagnosis. It has a potential role in supporting clinicians, reducing diagnostic burdens, and accessing early urgent interventions through smart health services.
    Keywords:  DFU classification; MGWONET; deep learning; diabetic foot ulcer; hyperparameter optimisation; modified Grey Wolf optimisation; wound; wound care
    DOI:  https://doi.org/10.12968/jowc.2024.0297
  16. Int Emerg Nurs. 2025 Oct 04. pii: S1755-599X(25)00131-4. [Epub ahead of print]83 101700
       AIMS: This study aims to develop and validate an artificial intelligence -driven survival prediction model using the Random Forest algorithm to support clinical decision-making in diabetic emergency cases. The model is designed to assist emergency nurses in triage prioritization and resource allocation to improve patient outcomes.
    METHODS: A retrospective cross-sectional study was conducted using medical records of 1,047 diabetic emergency patients treated at regional hospital in Indonesia, from 2019 to 2024. Key clinical variables, including age, gender, blood glucose levels, Glasgow Coma Scale, triage classification, and insulin use, were analyzed. Logistic regression identified significant survival predictors, and random forest model was developed for survival prediction. Model performance was evaluated using accuracy, sensitivity, specificity, and the area under the receiver operating characteristic AUC (Area Under Curve).
    RESULTS: The random forest model identified GCS and triage classification as the most significant predictors of survival. Patients with higher GCS scores and immediate triage classification (P1) had a greater likelihood of survival. The model demonstrated high predictive performance, achieving an accuracy of 94.9 %, sensitivity of 95.6 %, specificity of 93.7 %, and an AUC of 0.96.
    CONCLUSION: The AI-based random forest model demonstrated excellent predictive accuracy, supporting its integration into emergency nursing workflows. Implementing AI-driven decision-support systems in emergency departments may enhance triage accuracy, to improve survival outcomes in diabetic emergencies, future studies should focus on external validation and the integration of additional clinical parameters to further refine model performance.
    Keywords:  Artificial intelligence; Diabetic emergencies; Glasgow Coma Scale; Random forest; Survival prediction; Triage
    DOI:  https://doi.org/10.1016/j.ienj.2025.101700
  17. J Eval Clin Pract. 2025 Oct;31(7): e70295
       AIMS: This synthetic simulation, using no real patient data, study aimed to evaluate and compare the performance of three prominent large language models (LLMs)-ChatGPT-4.1, Grok-3 and DeepSeek-in generating medical nutrition therapy aligned dietary plans for adults with type 2 diabetes mellitus (T2DM).
    METHODS: A simulation-based design was employed using 24 standardized virtual patient profiles differentiated by gender and body mass index (BMI) category. Each LLM was prompted in Turkish to generate 3-day meal plans. Outputs were assessed for energy and macro-/micronutrient accuracy, adherence to national and international T2DM guidelines and alignment with the nutrition care process (NCP).
    RESULTS: ChatGPT-4.1 showed the highest alignment with energy requirements (70.9%) but overestimated fat intake. Grok-3 demonstrated superior energy accuracy (83.1%) but failed to meet several micronutrient targets. DeepSeek adjusted protein intake according to BMI but underdelivered carbohydrates. None of the models demonstrated full concordance with the NCP framework, particularly in the diagnosis and monitoring components. Frequent hallucinations and lack of clinical contextualization were noted. Integration of retrieval-augmented generation (RAG) was identified as a potential improvement strategy.
    CONCLUSION: While LLMs showed promise in generating baseline dietary guidance in a simulated context, these results reflected concordance with guideline documents only and concordance with guideline documents only and should not be interpreted as evidence of equivalence to dietitian-led care. These findings reflected model behaviour in synthetic scenarios only and highlighted the need for RAG integration and expert supervision before any clinical application.
    Keywords:  artificial intelligence; large language models; medical nutrition therapy; retrieval‐augmented generation; type 2 diabetes mellitus
    DOI:  https://doi.org/10.1111/jep.70295