bims-aukdir 2026-03-22 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2026–03–22
twenty papers selected by
Mott Given

Evaluation of non-ophthalmologist-led and offline AI-assisted models for diabetic retinopathy screening in India: a pragmatic diagnostic accuracy study.
Machine Learning Driven Insights into Socio-Economic, Demographic, and Health Determinants of Type-II Diabetes: A Tuned Hyperparameter Random Forest Model Using NFHS-5 Data.
Feature reduction using swarm optimization and random forest classifiers for early diabetes risk prediction.
An interpretable machine learning model for detecting vision-threatening diabetic retinopathy among patients with diabetic retinopathy: a web-based cross-sectional study.
Automated diabetic retinopathy classification using vision transformers on optical confocal microscopy images.
Factors Associated with Machine Learning-Based Predictions of Retinal Aging Using Teleretinal Screening Images from Patients with Diabetes.
Development and Validation of a Machine Learning-Based Predictive Model for Peripheral Neuropathy Risk in Elderly Patients with Type 2 Diabetes.
Exploration of a Multimodal Machine Learning Model Integrating Ultrasound and Clinical Indicators for the Diagnosis of Diabetic Peripheral Neuropathy.
Integrated cortical-cognitive signatures identified by machine learning enable early detection of MCI in type 2 diabetes.
Graph-Enhanced Multi-Task Learning for Type 2 Diabetes Comorbidity Risk Prediction.
Interpretable Graph Convolutional Networks for cardiovascular disease risk prediction in patients with Type 2 Diabetes Mellitus.
Detection of Non-diabetic Kidney Disease in Patients with Diabetes Using Machine Learning and Electronic Medical Record Data.
An inflammatory biomarker panel for prediabetes classification using interpretable machine learning.
Fasting Metabolite Panel for OGTT-Free Diagnosis of Gestational Diabetes Mellitus: A Machine Learning Approach Validated in Dual Cohorts.
Estimating prevalence and predictors of glucose-lowering overtreatment among older adults with type 2 diabetes in long-term care and community settings: a machine learning-based cohort study.
Development and external validation of a machine learning model for predicting in-hospital mortality in ICU patients with diabetic kidney disease: a study utilizing the MIMIC database and a Chinese cohort.
Machine learning-based prediction of long-term new-onset diabetes mellitus risk after pancreaticoduodenectomy using radiomics.
Patient Perceptions of Artificial Intelligence in Diabetes Self-Management: Cross-Sectional Survey Study.
Artificial Intelligence in Type 1 Diabetes Management: A Scoping Review of Randomised Controlled Trials.
Personalized Glucose Management With AI: Pilot Study Using a Multiarmed Bandit Approach.

BMJ Open. 2026 Mar 18. 16(3): e106397

Evaluation of non-ophthalmologist-led and offline AI-assisted models for diabetic retinopathy screening in India: a pragmatic diagnostic accuracy study.

Anshul Chauhan, Luke Vale, Ankita Kankaria, Basavaraj Tigari, Sonam Kumar, Mitesh Yadav, Lakshay Kumar, Gagandeep Kaur, Sanjay Kumar Bhadada, Vishali Gupta, Mona Duggal.

   OBJECTIVES: To assess and compare the diagnostic accuracy of non-ophthalmologist-led diabetic retinopathy screening (DRS) at health and wellness centres (HWCs) and offline artificial intelligence (AI)-assisted community-based screening, using specialist grading as the reference standard in India.
DESIGN, SETTINGS AND PARTICIPANTS: Pragmatic diagnostic accuracy study in primary healthcare settings. The settings included HWCs and community-based screening sites in rural Block Boothgarh, Mohali District, Punjab, India. A total of 600 people with diabetes aged ≥30 years were enrolled across three screening models: (1) non-ophthalmologist-led DRS at the HWC, (2) AI-assisted smartphone-based DRS in the community and (3) standard referral-based care. Retinal images were captured using non-mydriatic fundus cameras and independently graded by two masked human graders; a senior retina specialist resolved any disagreements. The AI was assessed for its ability to detect diabetic retinopathy (DR) and referable diabetic retinopathy (RDR). Diagnostic performance metrics were reported.
RESULTS: The non-ophthalmologist-led model demonstrated 86.4% sensitivity (95% CI 65.1% to 97.1%) and 94.3% specificity (95% CI 88.5% to 97.7%) for DR detection, with an ungradability rate of 8%. For RDR, sensitivity reached 95.8% (95% CI 78.9% to 99.9%) and specificity was 93.1% (95% CI 88.0% to 96.5%). The offline AI-assisted model achieved 93.3% sensitivity (95% CI 68.1% to 99.8%) and 85.1% specificity (95% CI 76.9% to 91.2%) for RDR, but with a higher ungradability rate (38%), mainly due to cataracts and poor image quality. Both approaches effectively identified referable cases; however, the non-ophthalmologist-led model demonstrated greater accuracy and operational feasibility.
CONCLUSIONS: This study demonstrates that non-ophthalmologist-led DRS at HWCs can enhance access to primary care. Offline AI-enabled screening demonstrates potential for community use but is currently limited by image quality and binary classification outputs. Integrating both approaches may strengthen DRS coverage in resource-limited settings.
CLINICAL TRIALS REGISTRY OF INDIA: CTRI/2022/10/046283.

Keywords:  Diabetic retinopathy; Epidemiology; PUBLIC HEALTH

DOI:  https://doi.org/10.1136/bmjopen-2025-106397
J Pharm Bioallied Sci. 2026 Feb;18(Suppl 1): S182-S184

Machine Learning Driven Insights into Socio-Economic, Demographic, and Health Determinants of Type-II Diabetes: A Tuned Hyperparameter Random Forest Model Using NFHS-5 Data.

Hitesh Athwani, Kuhu Awasthi, Meghna Athwani, Rajeev Pandey.

   Problem Considered: This investigation attempts to design and deploy a bespoke Random Forest model for early rate prediction of Type-II Diabetes utilizing the dataset available from NFHS-5.
Methods: Hyperparameter optimization was carried out to boost the performance of a Random Forest algorithm. Missing value treatments and class imbalance problems were done utilizing SMOTE. Accuracy, recall, log loss and ROC-AUC were utilized to test the model's performance.
Results: The accuracy, recall, log loss, and ROC-AUC score were 73.48%, 74.40%, 54.17%, and 81.17%, respectively. The study also indicates that demographic, lifestyle, comorbidities, and specific physiological characteristics significantly influence the prediction of diabetes.
Conclusion: Random Forest algorithms are extremely effective with early diabetes detection and could help support the design of focused intervention and policy measures.

Keywords:  Diabetes; NFHS; prediction; random forest; socioeconomic factors

DOI:  https://doi.org/10.4103/jpbs.jpbs_1612_25
Sci Rep. 2026 Mar 21.

Feature reduction using swarm optimization and random forest classifiers for early diabetes risk prediction.

Proshenjit Sarker, Abdullah-Al Nahid, Kwonhue Choi, Md Abdus Samad.

Diabetes is a chronic metabolic disorder caused by excessive blood sugar levels, which leads to severe damage to other organs. Type 2 diabetes is, more often than others, a long-term metabolic disorder in which the body resists insulin or does not produce enough of it. Early diabetes detection with fewer features reduces patient burden, and machine learning makes the process more time-efficient. This study proposes three machine learning approaches that achieve both high performance and effective feature reduction using the Early Stage Diabetes Risk Prediction dataset. Multiple research works have been published. However, they have struggled to achieve efficient feature reduction while maintaining high accuracy and have not provided detailed explanations of the models' nature or misclassifications. This research resolves the issues by showing outstanding performance, including a remarkable feature reduction. Three different swarm-based metaheuristic algorithms: Fox Optimizer, Honey Badger Algorithm, and Tuna Swarm Optimization, have been used, wrapped with a Random Forest Classifier. SHAP, as an explainable AI, is used to present the model's nature and feature importance, including individual predictions. FOX_RF, HBA_RF, and TSO_RF have gained an Accuracy of 99.36%, 99.36%, and 100% without cross-validation. However, TSO_RF has achieved the highest mean 10-fold cross-validation accuracy of 98.14%, F-score of 98.47%, and 98.54% of Precision using only 14 features out of 16. And, FOX_RF has achieved the highest mean Precision of 98,43%. HBA_RF has shown the highest number of feature reduction by selecting 10 features out of 16, maintaining a moderate performance. SHAP has confirmed that Polyuria, Polydipsia, and Gender are the most impacted features for diabetes prediction. SHAP-based individual prediction analysis has revealed that even small changes in these features can influence the model's decisions. This research analyzes the Early Stage Diabetes Risk Prediction dataset, which includes 520 individuals with 16 predictors and one target class, where TSO_RF has outperformed other models by achieving 100% and 98.14% of Accuracy, respectively, without cross-validation and using cross-validation.

DOI: https://doi.org/10.1038/s41598-026-35984-7
Front Endocrinol (Lausanne). 2026 ;17 1776188

An interpretable machine learning model for detecting vision-threatening diabetic retinopathy among patients with diabetic retinopathy: a web-based cross-sectional study.

Mingyang Song, Yimeng Shi.

   Background: Vision-threatening diabetic retinopathy (VTDR) is a severe complication of type 2 diabetes mellitus (T2DM), particularly prevalent in patients with prolonged disease duration, poor glycemic control, and systemic comorbidities. This condition frequently progresses asymptomatically toward irreversible blindness without timely intervention. The early identification of VTDR is challenging due to the lack of validated biomarkers and a reliance on subjective clinical assessments. This study aimed to develop and validate an interpretable machine learning (ML) model to detect VTDR among patients with diabetic retinopathy (DR).
Methods: Retrospective clinical data from T2DM patients with DR were extracted from the electronic medical records at our hospital and categorized into VTDR and non-VTDR (defined as mild-to-moderate non-proliferative diabetic retinopathy) groups. The dataset was partitioned into training and testing sets (7:3 ratio). Eight ML models were trained and evaluated using metrics such as Area Under the Curve (AUC), accuracy, and recall. Model performance was evaluated using a comprehensive scoring system (total score = 64). Shapley Additive Explanations (SHAP) were used to interpret the best-performing model. A web-based application was developed to demonstrate potential clinical utility.
Results: Among 1,124 enrolled patients, the prevalence of VTDR was 36.9%. Key associated factors included diabetic treatment, T2DM duration, glycated hemoglobin levels, albuminuria, and anemia. The Support Vector Machine (SVM) model demonstrated superior performance, with an AUC of 0.879, accuracy of 0.837, precision of 0.833, Brier score of 0.129, and an F1 score of 0.756, outperforming the other ML models. The SVM model achieved the highest total score (57/64) in the testing cohort. Furthermore, decision curve analysis and calibration curves confirmed the robustness and reliability of the models. A simplified calculator derived from the SHAP feature importance rankings maintained strong diagnostic capacity.
Conclusion: The interpretable SVM model effectively detected VTDR among patients with DR using routine clinical data. While requiring external validation, this study serves as a proof-of-concept for a cost-effective screening tool that could assist clinicians in prioritizing high-risk patients and facilitating early intervention to prevent irreversible vision impairment.

Keywords:  detection model; machine learning; non-vision-threatening retinopathy; shap; type 2 diabetes; vision-threatening diabetic retinopathy

DOI:  https://doi.org/10.3389/fendo.2026.1776188
Appl Opt. 2025 Oct 01. 64(28): 8151-8160

Automated diabetic retinopathy classification using vision transformers on optical confocal microscopy images.

Mohamed R Shoaib, Hamada Esmaiel, Abdulrahman Al Ayidh, Hany S Hussein, Fathi E Abd El-Samie.

In this paper, we employed the Vision Transformer architecture to effectively classify diabetic retinopathy (DR) in retinal images. Our dataset consisted of validated images, categorized into two classes: "normal" and "abnormal." The model demonstrated robust performance, achieving a training accuracy of 98.91% and a validation accuracy of 98.79%, alongside a commendable test accuracy of 100%. The training process spanned 30 epochs, during which the model exhibited consistent improvements in accuracy and reductions in loss. We further evaluated the model efficacy through a classification report, which revealed precision, recall, and F1-scores of 1.00 for both classes, indicating perfect classification performance on the test set. The results highlight the model potential for clinical applications in DR detection, ensuring its capability to assist healthcare professionals in diagnosing DR with high accuracy. Future work will focus on enhancing the model generalizability across diverse datasets and exploring the integration of additional clinical features.

DOI: https://doi.org/10.1364/AO.562201
Ophthalmol Sci. 2026 Apr;6(4): 101112

Factors Associated with Machine Learning-Based Predictions of Retinal Aging Using Teleretinal Screening Images from Patients with Diabetes.

Tuyet Thao Nguyen, Maria Jessica Cruz, Tanvi Chokshi, Valencia Lenora Potter, Micah Joy Patag San Agustin, Angel Rose Joshy, Neha Sarabu, Xin Liu, Glenn Yiu.

   Purpose: To identify factors associated with accelerated retinal aging based on machine learning predictions of age using fundus images from teleretinal screening of patients with diabetes.
Design: Cross-sectional study of retinal images.
Subjects: Ten thousand, five hundred thirty eye images from 2939 patients with diabetes who underwent teleretinal screening at the University of California clinics.
Methods: We trained a vision transformer (ViT) model to predict chronological age from retinal fundus photographs of 2939 patients with diabetes who underwent teleretinal screening as part of the Collaborative University of California Teleophthalmology Initiative (CUTI), and validated it using images from the Artificial Intelligence Ready and Exploratory Atlas for Diabetes Insights data set. We collected demographic, lifestyle, and systemic health factors, and analyzed their association with prediction errors, known as the retinal age gap.
Main Outcome Measures: Association between demographic, lifestyle, and systemic factors with retinal age gap.
Results: Our model accurately predicted chronological age from teleretinal images (mean absolute error 4.43 years; R2 = 0.84). Saliency maps showed model predictions primarily informed by the optic disc and proximal retinal vasculature. The retinal age gap was associated with predicted 10-year risk for cardiovascular diseases including heart failure and stroke (all P < 0.05). Retinal aging appears lower in Black patients (-1.36 years, P = 0.009) and increased in active smokers (+1.24 years, P = 0.044), as well as patients with severe obesity (+0.88 years, P = 0.033), hypertension (+0.86 years, P = 0.012), hyperlipidemia (+1.01 years, P = 0.002), and diabetic neuropathy (+1.80 years, P < 0.001). Key limitations include the cross-sectional study design and potential biases in medical record data.
Conclusions: Machine learning predictions of retinal aging using teleretinal images from patients with diabetes may predict cardiovascular risk and are accelerated by systemic comorbidities.
Financial Disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Keywords:  Cardiovascular risk; Diabetes; Retinal age gap; Teleophthalmology; Teleretinal screening

DOI:  https://doi.org/10.1016/j.xops.2026.101112
Risk Manag Healthc Policy. 2026 ;19 573535

Development and Validation of a Machine Learning-Based Predictive Model for Peripheral Neuropathy Risk in Elderly Patients with Type 2 Diabetes.

Jinling Peng, Dandan Xue, Juanjuan Li, Lihua Wei, Yanmei Wang.

   Background: Diabetic peripheral neuropathy (DPN) is highly prevalent among elderly patients with type 2 diabetes; however, existing models exhibit suboptimal performance and lack specificity. This study aims to develop and validate a machine learning-based model for early identification of DPN risk.
Methods: We retrospectively collected the data of 1450 elderly patients with type 2 diabetes using the electronic medical record system of the National Metabolic Management Center (MMC) at a tertiary hospital in Shanghai's Pudong New Area from March 2022 to March 2025. The dataset included general information, disease-related indicators, and laboratory results. We randomly divided the dataset into training and testing sets in a 7:3 ratio. After feature preprocessing and selection, four machine learning algorithms-logistic regression, naïve Bayes, random forest, and extreme gradient boosting (XGBoost)-were used to construct prediction models. Hyperparameter tuning was executed through grid search combined with 5-fold cross-validation, and model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, precision, recall, F1-score, calibration curves, and Decision Curve Analysis (DCA). The SHapley Additive exPlanations (SHAP) analysis was applied for model interpretation.
Results: The prevalence of DPN was 42.9% (623/1450). Nine variables were identified as independent predictors: diabetes duration, HbA1c, sleep quality, Charlson Comorbidity Index, sugar-sweetened beverage intake, peripheral arterial disease, sedentary behavior, smoking, and hypertension. Among the models, XGBoost performed best with an AUC of 0.951, accuracy of 0.878, precision of 0.876, recall of 0.834, F1-score of 0.855, and Brier score of 0.087. SHAP analysis confirmed the dominant contribution of diabetes duration and HbA1c to model predictions.
Conclusion: The XGBoost-based risk prediction model exhibited robust predictive performance and clinical utility for DPN in elderly patients with type 2 diabetes, offering potential for early identification of high-risk individuals and guiding targeted clinical interventions.

Keywords:  diabetic peripheral neuropathies; elderly; machine learning; predictive model; type 2 diabetes

DOI:  https://doi.org/10.2147/RMHP.S573535
J Ultrasound Med. 2026 Mar 20.

Exploration of a Multimodal Machine Learning Model Integrating Ultrasound and Clinical Indicators for the Diagnosis of Diabetic Peripheral Neuropathy.

Bo-Yu She, Meng-Lu Song, Wen-Bin Chen, Kun-Bin Wu, Zhen-Han Lai.

   OBJECTIVES: Based on ultrasound technology and clinical indicators, this study intends to develop multiple risk prediction models for diabetic peripheral neuropathy (DPN), conduct comparative analyses of these models, and further evaluate and validate the diagnostic efficacy of the optimal model for DPN as well as its potential in clinical application.
METHODS: The study included 235 patients grouped according to criteria for diagnosis of DPN. Ultrasound and clinical data were collected concurrently. The dataset was randomly partitioned into training and testing sets at a 7:3 ratio. Four machine learning models were developed: logistic regression (LR), extreme gradient boosting (XGBoost), random forest (RF), and support vector machine (SVM). These models were evaluated using a 10-fold cross-validation, comparing accuracy and areas under the curve (AUC) to determine the most optimal model. Based on SHAP (Shapley additive explanation) value visualization, decision curve analysis (DCA) and clinical impact curve (CIC) were employed to assess clinical utility.
RESULTS: Comparisons of different models indicate that the RF model performed best overall across all models, reaching an AUC of 0.852 in the test set while also producing the highest recall rates and F1 scores. SHAP analysis revealed key risk factors identified in the RF model, in order of importance: body mass index (BMI), 2-hour C-peptide (2hC-P), diabetes duration, triglycerides (TG), CSA-CPN1, and age. In addition, DCA and CIC demonstrate the model's clinical utility.
CONCLUSIONS: The RF model demonstrated optimal performance for DPN prediction and shows significant potential for clinical application in DPN risk assessment.

Keywords:  diabetes mellitus; diabetic peripheral neuropathy; machine learning model; ultrasonography

DOI:  https://doi.org/10.1002/jum.70237
Sci Rep. 2026 Mar 14.

Integrated cortical-cognitive signatures identified by machine learning enable early detection of MCI in type 2 diabetes.

Komal Verma Saluja, Prasad Balachandran, Drishya Pillai, Deelip Meena, Sangeeta Saxena, Harshwardhan Khokhar, Girish Khandelwal, S Manoj, Saurabh Chittora, Pratiti Bhadra.

  Type 2 Diabetes Mellitus (T2DM) confers a significant risk for Mild Cognitive Impairment (MCI), yet robust biomarkers for early detection remain limited. In this study, 150 age-matched participants (50 Healthy Controls, 50 T2DM, 50 T2DM with MCI) were assessed using high-resolution structural MRI, neuropsychological testing, and serological profiling to identify sensitive neuroanatomical and cognitive markers. Cortical thinning was observed, most prominently in the Left Pars Opercularis (LPO), which exhibited stepwise unidirectional atrophy across the diagnostic continuum, highlighting its potential as a structural marker for cognitive deterioration in T2DM. Significant deficits in episodic memory, processing speed, executive function, and verbal memory were also observed, reflecting disruptions in medial-temporal and frontoparietal networks. A Random Forest classifier integrating multimodal features achieved high discriminatory performance (AUC-ROC = 0.95) for distinguishing T2DM with MCI from T2DM patients. SHAP, an Explainable AI method, identified cortical thickness at LPO, and executive function assessed by TMTB as the most influential predictors. These findings establish the LPO as a key neuroanatomical substrate of T2DM-related cognitive impairment and demonstrate that combining targeted neuroimaging with domain-specific cognitive assessments provides a clinically viable framework for early identification of at-risk T2DM patients, offering critical opportunities for preventive intervention.

Keywords:  Biostatistics; Cortical Thickness; Left Pars Opercularis; MCI; Machine Learning; T2DM; Trail Making Test

DOI:  https://doi.org/10.1038/s41598-026-44608-z
IEEE J Biomed Health Inform. 2026 Mar 19. PP

Graph-Enhanced Multi-Task Learning for Type 2 Diabetes Comorbidity Risk Prediction.

Liyun Tang, Jiaxin Lu, Daohua Pan, Zhongfu Zuo, Xiqiao He, Haoqiang Zhang, Bing Song.

Early prediction of Type 2 diabetes mellitus (T2DM) complications holds significant clinical importance for improving patient outcomes and reducing healthcare burden, yet existing prediction methods exhibit notable limitations. This paper proposes a Graph-Enhanced Multi-Task Learning (GEMTL) framework for simultaneously predicting the occurrence risk of multiple diabetic complications. The framework constructs disease relation graphs through a hybrid strategy that linearly combines a data-driven statistical graph derived from disease co-occurrence patterns with a knowledge-driven prior graph encoding clinically established association strengths, employs graph neural networks to capture higher-order dependencies among diseases, designs cross-attention mechanisms to achieve heterogeneous information fusion between patient features and disease graph embeddings, and utilizes multi-gating expert network architecture for task-specific modeling. Large-scale experimental validation was conducted on a dataset constructed from MIMIC-IV. Results demonstrate that the GEMTL framework achieves macro-averaged F1 score of 0.723, micro-averaged F1 score of 0.856, and mean Average Precision of 0.759, significantly outperforming baseline methods across all evaluation metrics, including traditional machine learning methods, deep multi-task learning methods, graph neural network methods, and multi-expert architectures. This study provides an effective technical framework for complex medical multi-task prediction problems, with broad application prospects in diabetes precision management and clinical decision support.

DOI: https://doi.org/10.1109/JBHI.2026.3675904
J Biomed Inform. 2026 Mar 13. pii: S1532-0464(26)00039-0. [Epub ahead of print]178 105015

Interpretable Graph Convolutional Networks for cardiovascular disease risk prediction in patients with Type 2 Diabetes Mellitus.

Ioannis Siachos, Maria Athanasiou, Konstantia Zarkogianni, Anastasia C Thanopoulou, Konstantina S Nikita.

   BACKGROUND: Cardiovascular disease (CVD) is the most prevalent complication of Type 2 Diabetes Mellitus (T2DM) and a leading cause of mortality in this population. Early and accurate CVD risk prediction is essential for timely intervention, yet traditional clinical risk calculators may overlook complex, non-linear relationships between risk factors, and have exhibited varying performance. Graph neural networks (GNNs) are able to capture these complex relationships but often lack interpretability.
OBJECTIVE: The present study aims to develop and evaluate the first interpretable Graph Neural Network (GNN)-based framework for Cardiovascular Disease (CVD) risk prediction in patients with Type 2 Diabetes Mellitus (T2DM). We introduce a novel approach that integrates a GNN classifier with a rule-based surrogate model to generate clinically meaningful explanations for the model's predictions, addressing the critical need for both high accuracy and transparency in clinical AI.
METHODS: A population graph of 560 T2DM patients was constructed using demographic, lifestyle, laboratory, and treatment data. A GNN-based classifier was trained by leveraging a loss function originally designed for graph-based anomaly detection to address class imbalance. Post-hoc interpretability was achieved through the deployment of a RuleFit surrogate model, combining decision tree ensembles and a sparse linear model to extract global, rule-based explanations.
RESULTS: The proposed model achieved an AUC of 0.786±0.076, exceeding all benchmark methods and outperforming prior best-reported results on this dataset by over 7% in terms of the AUC, and produced well-calibrated probabilities (Brier score: 0.053±0.021). RuleFit explanations aligned with established CVD risk factors, while revealing intermediate-risk patterns, such as residual dyslipidemia despite treatment, that may warrant earlier intervention.
CONCLUSION: The proposed interpretable GNN framework demonstrated its ability to provide reliable CVD risk estimates while offering transparent, clinically relevant explanations. These findings support its potential integration into CVD risk screening tools for patients with T2DM, paving the way for real-world clinical implementation.

Keywords:  Anomaly detection; Cardiovascular disease; Explainable artificial intelligence; Global surrogate model; Graph Neural Networks; Machine learning; Type 2 Diabetes Mellitus

DOI:  https://doi.org/10.1016/j.jbi.2026.105015
Kidney Blood Press Res. 2026 Mar 18. 1-21

Detection of Non-diabetic Kidney Disease in Patients with Diabetes Using Machine Learning and Electronic Medical Record Data.

Zhongyu Liu, Chunyang Li, Jia Yao, Zhiye Ying, Ling Li, Ping Fu, Xiaoxi Zeng.

INTRODUCTION: The identification of non-diabetic kidney disease (NDKD) in diabetic patients is critically important. Unlike diabetic nephropathy, NDKD often requires additional therapeutic interventions beyond standard diabetes care. There is a need to develop computational methods using electronic medical record data to identify NDKD in diabetic patients for whom kidney biopsy is not an option.
METHODS: The study included 1136 diabetic patients who underwent kidney biopsy at a tertiary teaching hospital. We collected 103 parameters from electronic medical records, including demographic characteristics, physical examination results, laboratory tests, and the status of diabetic retinopathy. We developed seven models to detect NDKD, including k-nearest neighbors, random forest, extreme gradient boosting (XGB), lasso Logistic regression, support vector machine, naïve bayes, and multilayer perceptron (MLP), in the training set (n=908), and compared their performances in the testing set (n=228). The SHapley Additive exPlanations (SHAP) approach was used to analyze the importance of features.
RESULTS: Biopsy-confirmed NDKD was present in 53% of the 1136 participants. In the testing set, the area under the receiver operating characteristic curve (AUC) for NDKD detection using XGB, Lasso regression, and MLP reached 0.8, with performances that were stable regardless of whether variable normalization was performed. Among them, XGB revealed the highest AUC (0.833; 95% CI: 0.800 to 0.864) without feature normalization, which was statistically superior to the other models according to DeLong's tests. After feature normalization, SVM achieved the highest AUC of 0.841 (95% CI: 0.817to 0.861) among all models. In addition to established predictive factors for NDKD (e.g., hematuria and absence of diabetic retinopathy), SHAP analysis identified several features, such as low IgG levels, that contributed significantly to the differentiation models.
CONCLUSION: Despite performance variations in different modeling techniques, machine learning models may have the potential to facilitate the detection of NDKD for patients with contraindications for kidney biopsy. Further efforts are warranted to improve accuracy and facilitate their translation into clinical practice.

DOI: https://doi.org/10.1159/000551589
PLoS One. 2026 ;21(3): e0341195

An inflammatory biomarker panel for prediabetes classification using interpretable machine learning.

Maher Maalouf, Maram Tammam, Sana Kurungadan, Asmaa Alsereidi, Muhammad Afzal, Herbert F Jelinek.

OBJECTIVE: Prediabetes is a silent condition that often goes undetected. However, timely interventions could prevent its progression to type 2 diabetes. Traditional glycemic markers, such as hemoglobin A1c (HbA1c), have limitations, creating a need for new diagnostic biomarkers. In this study, our objective was to develop an interpretable machine learning model using biomarkers related to oxidative stress, inflammation, and lipid metabolism to classify prediabetes independently of traditional glycemic markers, such as HbA1c. We also compared multiple biomarker panels to determine which biomarkers offer the highest predictive accuracy.
METHODS: We developed and validated interpretable machine learning models using clinical and biomarker data from 545 participants (405 healthy controls and 140 with prediabetes). To ensure robust and generalizable findings, we employed a nested cross-validation technique, managed feature collinearity using the variance inflation factor (VIF), and interpreted the final model with Shapley Additive exPlanations (SHAP) [Kapoor S, Narayanan A. Patterns. 4(9):100804 (2023); Vabalas A, et al. PLoS One. 14(11):e0224365 (2019); Lundberg SM, Lee SI. Adv Neural Inf Process Syst. 30:4768-77 (2017)].
RESULTS: Our approach identified a distinct panel of inflammatory biomarkers (IL-10, IGF-1, and CRP) capable of classifying prediabetes independently of traditional glycemic markers. This non-glycemic model achieved a promising Area Under the Curve (AUC) of 0.711 on holdout validation, establishing inflammation as a key and measurable indicator of early metabolic dysfunction.
CONCLUSION: Our findings introduce a novel panel of inflammatory biomarkers that show promise in the identification of prediabetes independently of traditional glucose-based measures. By highlighting inflammation as an early indicator of metabolic dysfunction, this approach may enhance precision in the detection of prediabetes. Longitudinal studies with larger and more diverse populations are essential to clinically validate these biomarkers and confirm their value in improving the early diagnosis and management of metabolic health.

DOI: https://doi.org/10.1371/journal.pone.0341195
Diabetes Metab Syndr Obes. 2026 ;19 572688

Fasting Metabolite Panel for OGTT-Free Diagnosis of Gestational Diabetes Mellitus: A Machine Learning Approach Validated in Dual Cohorts.

Binbin Yin, Yiyun Shen, Qianwen Zhang, Rongchang Chen, Xue Zhang, Ziqing Kong, Yuning Zhu.

   Objective: This study aimed to develop a fasting serum metabolite-based method for screening and risk assessment of gestational diabetes mellitus (GDM), potentially reducing dependence on the oral glucose tolerance test (OGTT).
Methods: Using a retrospective discovery cohort (n = 435; April-May 2021) with prospective validation (n = 473; November 2018-May 2021) design, 1,053 pregnant women completing standard 75g OGTT were initially enrolled. Fasting serum samples underwent targeted metabolomic profiling. A diagnostic model was constructed using machine learning (random forest) in combination with univariate analysis and rigorous validation protocols. Model performance was evaluated using the area under the receiver operating characteristic curve (ROC).
Results: Eight metabolites demonstrated significant differential expression between GDM and non-GDM groups (FDR <0.05). Based on the feature importance rankings, we developed a multivariate logistic regression model incorporating seven metabolites: 2-hydroxybutyric acid, 1,5-anhydroglucitol, glycine, 3-methyl-2-oxobutyric acid, 3-methyl-2-oxovaleric acid, tyrosine, and oleic acid. The composite model (fasting glucose + risk factors + metabolites) demonstrated significantly higher discriminative performance in the discovery cohort (AUC = 0.78) compared to fasting glucose alone (AUC = 0.62), with sustained performance in external validation (AUC = 0.71).
Conclusion: This fasting metabolite detection protocol demonstrates promising potential for GDM screening and risk stratification, offering the prospect of reducing reliance on OGTT in specific clinical settings.

Keywords:  diagnosis model; fasting glucose; gestational diabetes mellitus; machine learning; metabolite biomarker

DOI:  https://doi.org/10.2147/DMSO.S572688
BMJ Open. 2026 Mar 13. 16(3): e106707

Estimating prevalence and predictors of glucose-lowering overtreatment among older adults with type 2 diabetes in long-term care and community settings: a machine learning-based cohort study.

Greg Carney, Sean Burnett, Anshula Ambasta, Wade Thompson, Linda Lapp, Colin Dormuth.

   OBJECTIVE: To estimate the prevalence of potential overtreatment of type 2 diabetes mellitus (T2DM) among older adults and to develop and compare predictive models to identify patient and physician characteristics associated with overtreatment.
DESIGN: Population-based retrospective cohort study with predictive modelling.
SETTING: A province-wide, publicly funded healthcare system in British Columbia, Canada, using linked administrative health claims data from 2016 to 2023.
PARTICIPANTS: Residents of long-term care facilities over age 65, and community-dwelling individuals over age 75, with a diagnosis of T2DM and a glycated haemoglobin (A1C) laboratory value ≤7.0%. Participants were required to have ≥365 days of continuous provincial health insurance coverage prior to their index A1C test. Patients receiving palliative care and those with missing physician information were excluded.
PRIMARY AND SECONDARY OUTCOME MEASURES: Potential overtreatment of T2DM, defined a priori as overlapping prescriptions for ≥2 glucose-lowering medications or ≥1 insulin or sulfonylurea dispensing within 90 days after the index A1C test.Model performance outcomes included discrimination (area under the curve (AUC), sensitivity, specificity, positive predictive value and negative predictive value). Performance metrics were calculated with 95% CIs using a 25% temporally distinct test dataset (2021-2023). No changes were made to outcome definitions after protocol development.
RESULTS: Among 133 773 patients with an A1C≤7.0%, 38 074 (28.5%) were classified as overtreated. These patients had a mean age of 79.6 years, were 47% female, and had a median A1C of 6.4%. The gradient boost model was the best performing model overall, using a combination of expert-selected variables and data-driven variables, achieving an AUC of 0.87, sensitivity of 0.81 and negative predictive value of 0.89. The top predictors of overtreatment included use of blood glucose test strips, A1C test volume, polypharmacy, specialist involvement and measures of diabetes severity.
CONCLUSIONS: Overtreatment of T2DM was prevalent among older adults in our cohort. Machine learning algorithms that integrate clinical expertise with data-driven variable selection performed the best in predicting T2DM overtreatment. We identified several patient and physician characteristics as key contributors that may inform future clinical practice and quality improvement initiatives, although external validation is required before clinical implementation.

Keywords:  Diabetes Mellitus, Type 2; Machine Learning; PUBLIC HEALTH; Polypharmacy; Primary Care; Quality Improvement

DOI:  https://doi.org/10.1136/bmjopen-2025-106707
Front Endocrinol (Lausanne). 2026 ;17 1699647

Development and external validation of a machine learning model for predicting in-hospital mortality in ICU patients with diabetic kidney disease: a study utilizing the MIMIC database and a Chinese cohort.

YuNan Han, RuMeng Mao, ChengYue Xiong, YongXiang Wang, Lin Li, HongLian Wang.

   Background: Patients with diabetic kidney disease (DKD) admitted to the intensive care unit (ICU) face an exceptionally high risk of in-hospital mortality. Currently, effective tools for their early risk stratification are critically lacking. Therefore, this study aimed to develop and externally validate an interpretable machine learning (ML) model for predicting in-hospital mortality in this high-risk ICU-DKD patient population.
Methods: This retrospective cohort study involved developing and evaluating eight ML algorithms. Model performance was rigorously assessed using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) provided model interpretability. Data from DKD patients with ≥24-hour ICU stays were extracted from the MIMIC-IV database (n=3,403) for model development. An independent external validation cohort (n=260) was collected from the First Affiliated Hospital of Yangtze University (YTU-ICU). The primary outcome was in-hospital mortality. Lasso regression identified key predictors. Model evaluation focused on the area under the ROC curve (AUROC), calibration, and net clinical benefit.
Results: Ten features were selected for model development. Among the tested algorithms, XGBoost demonstrated superior predictive performance, achieving an AUROC of 0.738 (internal validation) and 0.746 (external validation), with corresponding accuracies of 72.18% and 72.69%. SHAP analysis highlighted respiratory failure, lymphocyte count, SOFA score, RDW, age, and SAPS II as the six most important predictors.
Conclusions: The developed XGBoost model demonstrates good predictive performance for in-hospital mortality in ICU-DKD patients, exhibiting satisfactory generalizability and interpretability. This tool supports early risk stratification and facilitates personalized treatment strategies in critical care settings.

Keywords:  MIMIC-IV; SHAP; all-cause mortality; diabetes nephropathy; machine learning

DOI:  https://doi.org/10.3389/fendo.2026.1699647
Digit Health. 2026 Jan-Dec;12:12 20552076261415943

Machine learning-based prediction of long-term new-onset diabetes mellitus risk after pancreaticoduodenectomy using radiomics.

Jihyun Yoon, Seon Min Lee, Byoungduck Han, Yang-Hyun Kim, Young Jae Kim, Jaehun Yang, Yeon Ho Park, Doojin Kim, Doo-Ho Lee, Kwang Gi Kim.

   Background: Pancreaticoduodenectomy carries substantial metabolic consequences, with 20-32% of patients developing new-onset diabetes mellitus (NODM) within three years, leading to increased morbidity and healthcare burden. Current predictive models relying primarily on clinical variables demonstrate limited accuracy, underutilizing tissue-level information available in routine CT imaging. This study aimed to develop and validate a multimodal machine learning framework integrating clinical data with CT-derived radiomics features for long-term NODM risk prediction.
Methods: This retrospective cohort study analyzed 126 patients who underwent pancreaticoduodenectomy at Gachon University Gil Medical Center (2005-2023). Using PyRadiomics, 186 radiomic features were extracted from preoperative and postoperative CT scans (93 features per timepoint). Combined with 10 clinical variables (196 total features), Recursive Feature Elimination identified 10 key predictors. Logistic Regression, Support Vector Machine, Random Forest, and Extreme Gradient Boosting were evaluated using 5-fold cross-validation. SHAP analysis ensured model interpretability.
Results: Long-term NODM developed in 47 patients (37.3%). The Logistic Regression model demonstrated optimal performance with AUC 0.77 (95% CI: 0.68-0.84), sensitivity 70% (95% CI: 0.57-0.83), and specificity 72% (95% CI: 0.62-0.82). Key predictors included pancreatic volume changes, preoperative hypertension, and texture features (Strength, GrayLevelNonUniformity) from both imaging timepoints. The multimodal approach significantly outperformed clinical-only models (P < .05). Subgroup analyses confirmed consistent model performance across gender (P = .72) and age groups (P = .83).
Conclusions: The proposed approach that integrates CT radiomics with clinical data quantitatively improved the prediction performance for NODM after pancreatectomy. This multimodal strategy offers a more robust alternative to single-modality models and may facilitate personalized risk stratification and targeted postoperative surveillance.

Keywords:  Machine learning; artificial intelligence; new-onset diabetes mellitus; pancreaticoduodenectomy; predictive modeling; radiomics

DOI:  https://doi.org/10.1177/20552076261415943
JMIR Form Res. 2026 Mar 16. 10 e83030

Patient Perceptions of Artificial Intelligence in Diabetes Self-Management: Cross-Sectional Survey Study.

Nataly Martini, Navneet Kaur Dhaliwal, Elena Alipour, Shane Scahill, Laszlo Sajtos.

   Background: Artificial intelligence (AI) is increasingly applied in chronic disease management, including diabetes, where it has the potential to support real-time data interpretation, improve clinical decision-making, and enhance patient engagement. Although AI tools are often developed to increase efficiency and personalization, there is limited evidence on how patients perceive the role of AI in managing their condition, particularly in relation to shared decision-making (SDM) and the patient-provider relationship.
Objective: This study explored how people with diabetes perceive the usefulness of AI across key self-management tasks and examined their preferences for AI versus health care provider (HCP) involvement. It also assessed predictors of AI preference and proposed a conceptual foundation for integrating AI into a triadic SDM model involving patients, HCPs, and AI.
Methods: We conducted a cross-sectional online survey of adults with diabetes in New Zealand. Participants were asked to rate 7 diabetes self-management tasks in terms of (1) current HCP involvement, (2) perceived usefulness of AI, (3) comfort with HCPs using AI, and (4) preference for AI, HCP, or both in completing each task. Tasks included data collection, data interpretation, medication adherence, treatment decision-making, lifestyle management, personal reflection, and evaluation of treatment options. Both ordinary least squares regression and ordinal logistic regression (proportional odds models) were used to identify predictors of AI preference.
Results: A total of 48 participants completed the survey. Of these participants, 38 (79%) were female, 27 (56%) were aged 26 to 45 years, and 26 (54%) had higher education. Mean HCP involvement across tasks was 2.82 (SD 1.23; range 1-5). AI was viewed as moderately useful overall (mean 3.67, SD 1.20), with highest usefulness for tracking (mean 4.23, SD 1.06) and interpreting information (mean 4.40, SD 0.87). Actual AI use was reported by 15/48 (31%) participants. Participants preferred HCP involvement for tasks involving treatment decision-making (17/48, 35% vs 9/48, 19%) and personal reflection (23/48, 48% vs 9/48, 19%). Across regression models, perceived usefulness of AI was a significant predictor of preference for AI in 4 tasks: data collection (P=.02), data interpretation (P=.005), treatment decision-making (P=.04), and lifestyle management (P=.046). The patient-HCP relationship significantly predicted lower preference for AI in treatment decision-making (P=.03) and medication adherence (ordinary least squares P=.005). Comfort with HCPs using AI was generally nonsignificant. Effects were modest (adjusted R²=0.08-0.21).
Conclusions: Patients demonstrated task-specific openness to AI involvement in diabetes management, particularly for structured, data-intensive activities. These findings provide a foundation for future development and evaluation of AI-integrated SDM models. Broader exploration of technology types, relationship dynamics, and collaborative decision-making will be essential as AI becomes increasingly embedded in chronic care management.

Keywords:  artificial intelligence; chronic disease management; diabetes self-management; digital health; patient preferences; shared decision-making

DOI:  https://doi.org/10.2196/83030
Diabetes Obes Metab. 2026 Mar 19.

Artificial Intelligence in Type 1 Diabetes Management: A Scoping Review of Randomised Controlled Trials.

Yucen Wu, Jun Pang, Shaoyong Xu, Lalantha Leelarathna, Aaron M Lett.

   BACKGROUND: Artificial intelligence is emerging in healthcare systems. In type 1 diabetes, AI-enabled tools are increasingly used to support nutrition assessment and insulin decision-making, yet their clinical utility and safety remain unclear.
METHODS: The study aims to identify and map the evidence on the clinical utility of AI-based diabetes management tools in people with type 1 diabetes. We conducted a scoping review following PRISMA-ScR guidelines, searching PubMed, CINAHL and Web of Science up to January 2026 for eligible randomised controlled trials.
RESULTS: Our findings indicate that the evidence base is small and concentrated in high-income settings, with most trials assessing clinical utility using CGM outcomes and showing mixed improvements across interventions. No serious safety events were reported, but small sample sizes, short follow-up and inconsistent safety reporting limit confidence.
CONCLUSIONS: Future research should prioritise larger, longer-term real-world evaluations that use standardised safety endpoints and patient-centred outcomes, including in low- and middle-income countries to support equitable implementation.

Keywords:  artificial intelligence; diabetes management; randomised controlled trials (RCTs); safety; type 1 diabetes

DOI:  https://doi.org/10.1111/dom.70671
JMIR Form Res. 2026 Mar 19. 10 e70826

Personalized Glucose Management With AI: Pilot Study Using a Multiarmed Bandit Approach.

Shinji Hotta, Mikko Kytö, Saila Koivusalo, Seppo Heinonen, Pekka Marttinen.

   Background: Personalized behavioral recommendations through mobile apps have proven effective in preventing serious chronic diseases such as diabetes. Recent studies have primarily focused on optimizing personalized recommendations using reinforcement learning. However, the main problem with these approaches is that they focus on behavioral changes and overlook clinical outcomes.
Objective: This study aimed to propose a method for online planning of dietary and exercise recommendations to optimize postprandial glucose levels through behavioral changes directly.
Methods: The proposed method is a multiarmed bandit based on a two-stage reward prediction model, where an action is a combination of the total carbohydrate intake and postprandial walking duration, and the reward is the reduction in postprandial glucose levels. We implemented the prediction of the reward for each action based on the predicted behavioral responses to an action, and subsequently, the postprandial glycemic response.
Results: In a simulation experiment, we demonstrated that the proposed online algorithm can significantly improve postprandial glucose levels with personalized recommendations, compared to the randomized policy. Furthermore, we conducted a small real-world experiment with a simplified proposed method involving a single update of the recommendation policy into a personalized one. For 6 participants, compared to the randomized policy, we observed a 23% improvement, on average, in actual glucose responses along with the behavioral adherence to the recommendations concerning carbohydrate intake and postprandial walking.
Conclusions: The preliminary effectiveness of the proposed method was demonstrated from both the simulation experiment and the small real-world experiment. However, further longitudinal real-world experiments in patients with diabetes are needed to validate and generalize the findings.

Keywords:  diabetes; dietary and exercise recommendation; glucose management; mobile intervention; multiarmed bandit; personalization

DOI:  https://doi.org/10.2196/70826