bims-aukdir 2026-05-03 papers

bims-aukdir

Biomed News

on Automated knowledge discovery in diabetes research

Issue of 2026–05–03
twenty-six papers selected by
Mott Given

The evolving role of artificial intelligence in optimizing treatment and patient selection in diabetic macular edema.
Correction: Multi-feature integrated machine learning prediction model for early nephropathy in elderly living with type 2 diabetes mellitus.
Diabetic Retinopathy Detection: AI Models and Approaches.
Machine learning-based prediction of NAFLD in patients with type 2 diabetes using routine clinical and biochemical indicators.
Comment on "Exploration of a Multimodal Machine Learning Model Integrating Ultrasound and Clinical Indicators for the Diagnosis of Diabetic Peripheral Neuropathy".
Machine learning in prediction and classification of type 1 diabetes.
Personalised Blood Glucose Time Series Forecasting in Type 1 Diabetes: Deep Collaborative Adversarial Learning.
Integration of Continuous Glucose Monitoring With HbA1c to Improve the Detection of Prediabetes in Asian Individuals: Model Development Study.
Genetic association and machine learning improve the prediction of type 1 diabetes risk.
Identifying Actionable Glucose Patterns in Type 1 and Type 2 Diabetes Using Continuous Glucose Monitoring and Algorithmic Pattern Recognition.
Machine learning-based risk predictive models for depression in patients with diabetes: a systematic review and meta-analysis.
Noninvasive Blood Glucose Estimation via ECG: A Multi-Expert SC-ResNet Model.
Screening of metabolic-related biomarkers linking intervertebral disc degeneration and type 2 diabetes based on comprehensive bioinformatics analysis and machine learning.
Interpretable Machine Learning to Predict Metformin-Induced Vitamin B12 Deficiency: Association with Glycemic Control and Neuropathic Symptoms.
AI-driven retinal image quality monitoring in diabetic retinopathy screening: A retrospective study identifying actionable insights to improve imaging protocols.
Central adiposity outperforms disordered eating behavior in predicting type 2 diabetes risk in young adults: an explainable machine learning approach.
ChatGPT for diabetes education: potential, accuracy, and accessibility in patient support.
Retraction: Artificial Intelligence-Based Neural Network for the Diagnosis of Diabetes: Model Development.
MetaboNet: The Largest Publicly Available Consolidated Data Set for Type 1 Diabetes Management.
Machine Learning Unveils Dietary Antioxidants as Influential Factors for Diabetes-Cancer Comorbidity: Insights From National Health and Nutrition Examination Survey.
Identification of COL5A1 and TGFBI as key basement membrane genes and their molecular mechanisms in gestational diabetes mellitus.
Deep learning-based early prediction of gestational diabetes mellitus through first-trimester placental texture analysis.
Evaluation of large language models for carbohydrate counting in type 1 diabetes mellitus: a comparative study with patient estimations.
Transfer Learning-driven Biofield Image Analysis for Predictive Modeling of Diabetes.
Multi-modal sensor fusion in smart footwear to capture altered gait dynamics in patients with diabetic peripheral neuropathy: System validation and clinical applications.
A Flexible Wearable Glucose Sensor for Noninvasive Diabetes Screening: Functional Equivalence and Model Interpretability.

Indian J Ophthalmol. 2026 May 01. 74(5): 690-697

The evolving role of artificial intelligence in optimizing treatment and patient selection in diabetic macular edema.

Dhanashree Ratra, Ramachandran Rajalakshmi, M Suchetha, Devanjali Relan, J S Deepikasri, Sundaramoorthy Sathishkumar, Aashna Ratra.

  Anti-vascular endothelial growth factor (anti-VEGF) therapy is the mainstay of management for diabetic macular edema (DME), but marked variability in response, high injection frequency, and cumulative treatment burden highlight the need for tools that can individualize treatment beyond protocol-driven regimens. Artificial intelligence (AI) offers a pathway toward more individualized risk stratification and prognostic support primarily by capturing statistical associations rather than biological mechanisms. Deep learning systems have achieved great accuracy in detecting diabetic retinopathy (DR) and DME. Several autonomous DR/DME screening solutions are in clinical use. Recent advances have applied supervised machine learning, convolutional neural networks, generative adversarial networks, and ensemble methods to multimodal data from fundus images, baseline and follow-up optical coherence tomography (OCT), along with clinical and biochemical data, to classify likely responders and non-responders. These models automatically quantify and track imaging biomarkers to accurately predict central subfield thickness and vision outcomes after loading doses, and estimate future injection burden. AI-driven decision-support tools analyze vast amounts of patient data, treatment histories, and integrate multimodal data, including fundus images, OCT images, and systemic data to provide recommendations for optimal treatment and follow-up, tailored to each individual profile. The AI systems can potentially generate individualized risk and response profiles that can support decisions on initiating therapy, choosing between agents, tailoring treat-and-extend intervals, and timing switches to steroids or combination strategies. However, issues of generalizability, transparency, workflow integration, and ethical deployment need to be systematically addressed. AI-enabled decision support for patient selection and treatment response prediction is poised to become an integral component of anti-VEGF therapy.

Keywords:  Anti-VEGF therapy; artificial intelligence; convolutional neural network; deep learning; diabetic macular edema

DOI:  https://doi.org/10.4103/IJO.IJO_3152_25
Front Endocrinol (Lausanne). 2026 ;17 1807912

Correction: Multi-feature integrated machine learning prediction model for early nephropathy in elderly living with type 2 diabetes mellitus.

Tingting Fang, Yuanyuan Yang, Feng Zhuo, Xinran Xie, Jialun Song, Linghua Kong.

  [This corrects the article DOI: 10.3389/fendo.2025.1660903.].

Keywords:  early nephropathy; elderly; machine learning; prediction model; type 2 diabetes mellitus

DOI:  https://doi.org/10.3389/fendo.2026.1807912
J Ophthalmol. 2026 ;2026 8857887

Diabetic Retinopathy Detection: AI Models and Approaches.

Meftah Mohamed Mohamed Madi, Peter Clarke- Farr, Dirk Bester.

Background: Diabetic retinopathy (DR), a major cause of vision loss worldwide, results from chronic diabetes damage to retinal blood vessels. Vision loss can be prevented if DR is detected early, but traditional retinal screening by eye care takes time and expertise. Recent advances in AI technology, including classical machine learning and deep learning, can be more accurate in DR detection. This article provides a comprehensive review of current AI models and approaches of DR screening.
Methods: We searched PubMed, Web of Science, Scopus, ScienceDirect, and EBSCOhost using the keywords: diabetes, retinopathy, screening, and early detection. The search was limited to English language and studies published between 2020 and 2025.
Results: The findings suggest that AI models have become crucial for early DR diagnosis. While traditional machine learning previously lacked effectiveness, deep learning has now significantly improved diagnostic performance. The models, such as the URNet system, the vision transformer (ViT) model, the ResNet-50 and EfficientNetB0 models, the DenseNet model, and the ResNet-18 model, have achieved high-performance metrics using publicly available datasets. DR screening devices, like ADX-DR, have shown commendable performance. The EyeArt modality demonstrated exceptional sensitivity across diverse populations, detecting around 98.5% of vision-threatening DR, while Google AI matched specialist performance in specificity and surpassed it in sensitivity.
Conclusion: AI methods using deep learning frameworks such as CNNs have attained expert-level accuracy in DR classification, in addition to real-world validation. Semiautonomous systems like the IDx-DR and EyeArt have robust clinical performance and scalability, especially in countries with few ophthalmologists. Although research has been mainly conducted in Asia, there is a lack of research from Africa and low-income countries. Future techniques, including ensemble models and federated learning, will enhance accuracy and reliability further, aiding early diagnosis and prevention of vision loss globally.

DOI: https://doi.org/10.1155/joph/8857887
Front Endocrinol (Lausanne). 2026 ;17 1761846

Machine learning-based prediction of NAFLD in patients with type 2 diabetes using routine clinical and biochemical indicators.

Yongjun Zeng, Jinlin Wang, Kun Lin, Lu Ao, Chenggong Zhu, Qiang Yi, Ya Fu, Yongbin Zeng.

   Objectives: This study aimed to develop and evaluate machine learning (ML) models for predicting non-alcoholic fatty liver disease (NAFLD) in patients with type 2 diabetes mellitus (T2DM) using readily accessible clinical and biochemical indicators.
Methods: A total of 2,459 patients with T2DM were enrolled in this cross-sectional study. Eight ML algorithms, logistic regression (LG), k-nearest neighbors (k-NN), support vector machine (SVM), decision tree (DT), random forest (RF), gradient boosting machine (GBM), extreme gradient boosting (XGBoost), and naïve Bayes (NB), were developed to construct predictive models. Feature selection was performed using Boruta, recursive feature elimination, and LASSO regression. Model performance was assessed using several metrics, including the area under the receiver operating characteristic curve (AUC), accuracy, recall, F1 score, and decision curve analysis.
Results: Among the study population, 1,309 individuals (53.23%) were diagnosed with NAFLD. Sixteen variables, including BMI, waist circumference, systolic blood pressure, triglycerides, HDL-C, ALT, GGT, bilirubin fractions, albumin, BUN, GFR, fasting insulin, RBC, and hemoglobin, were selected as key predictors. The SVM model demonstrated the best overall performance, achieving an AUC of 0.920, accuracy of 0.839, and specificity of 0.898 in the training set, and an AUC of 0.833 and accuracy of 0.733 in the validation set. Decision curve analysis confirmed superior clinical utility of the SVM model compared with other algorithms.
Conclusions: ML-based models, particularly the SVM algorithm, effectively predicted NAFLD among patients with T2DM using easily accessible clinical and biochemical indicators. These findings highlight the potential utility of ML-assisted screening tools for improving early identification and risk stratification of NAFLD in diabetic populations.

Keywords:  clinical indicators; machine learning; non-alcoholic fatty liver disease; risk prediction; type 2 diabetes mellitus

DOI:  https://doi.org/10.3389/fendo.2026.1761846
J Ultrasound Med. 2026 Apr 25.

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

Madhuri Tribhuvan, Shashikant N Phatke.



Keywords:  diabetic peripheral neuropathy; machine learning; random forest model; risk prediction; ultrasound

DOI:  https://doi.org/10.1002/jum.70275
Nat Genet. 2026 Apr 30.

Machine learning in prediction and classification of type 1 diabetes.

Yangxi Li, Constantin Polychronakos.

DOI: https://doi.org/10.1038/s41588-026-02571-5
J Pers Med. 2026 Apr 08. pii: 210. [Epub ahead of print]16(4):

Personalised Blood Glucose Time Series Forecasting in Type 1 Diabetes: Deep Collaborative Adversarial Learning.

Heydar Khadem, Hoda Nemat, Jackie Elliott, Mohammed Benaissa.

  Background/Objectives: Blood glucose prediction (BGP) for individuals with type 1 diabetes (T1D) is a clinically essential yet highly challenging task in time series forecasting (TSF) and an important problem in personalised medicine. Accurate bespoke BGP is crucial for individualised T1D management, reducing complications, and supporting patient-specific glycaemic risk mitigation. However, the pronounced volatility of glycaemic fluctuations in T1D, combined with the need for mathematical rigor and clinical relevance, hampers reliable prediction. This complexity underscores the demand to explore and enhance more advanced techniques. While adversarial learning is adept at modelling intricate data variability, its potential for BGP remains largely untapped. Methods: This work presents a novel approach for BGP by addressing a key limitation in conventional adversarial learning when applied to this task. Typically, these methods optimise prediction accuracy within a set horizon by minimising adversarial loss. This focus overlooks how predictions align with longer-term patterns, which are critical for clinical relevance in BGP, thereby yielding suboptimal results. To overcome this limitation, we introduce collaborative augmented adversarial learning, designed to improve the model's temporal awareness. Incorporating collaborative interaction optimisation, this approach enables the model to reflect extended time dependencies beyond the immediate horizon, thereby improving both the clinical reliability of predictions and overall predictive performance. We develop and evaluate four learning systems for BGP: independent learning, adversarial learning, collaborative learning, and adversarial collaborative learning. The proposed systems were evaluated for two clinically relevant prediction horizons, namely 30 min and 60 min ahead. Results: The interdependent collaboratively augmented learning frameworks, validated using the well-established Ohio T1D datasets, demonstrate statistically significant superior performance in both clinical and mathematical evaluations. Conclusions: Beyond advancing BGP accuracy and clinical reliability, the proposed approach supports personalised medicine by improving subject-specific glucose forecasting from CGM data, with potential relevance for more individualised diabetes monitoring and decision support. The proposed approach also opens new avenues for advancements in other complex TSF domains, as outlined in our future work.

Keywords:  artificial intelligence; blood glucose prediction; deep learning; personalised diabetes management; time series forecasting

DOI:  https://doi.org/10.3390/jpm16040210
JMIR Diabetes. 2026 Apr 27. 11 e81520

Integration of Continuous Glucose Monitoring With HbA1c to Improve the Detection of Prediabetes in Asian Individuals: Model Development Study.

Michelle H Lee, Shihui Jin, Eveline Febriana, Maybritte Lim, Sonia Baig, Shahmir H Ali, Ian Yi Han Ang, Tze Ping Loh, Ashna Nastar, Kee Seng Chia, Alice Pik-Shan Kong, Faidon Magkos, Alex R Cook, Sue-Anne Toh.

   Background: Glycated hemoglobin (HbA1c) is a convenient tool to evaluate glycemic status but its ability to detect individuals at risk for type 2 diabetes is limited.
Objective: Exploiting the glycemic variability captured in continuous glucose monitoring (CGM), we used a well-characterized Asian cohort study from Singapore to assess whether utilizing CGM features in a machine learning model can improve the detection of prediabetes as compared to using HbA1c alone.
Methods: In this study, 406 nondiabetic Asian participants underwent an oral glucose tolerance test and had their fasting and 2-hour plasma glucose concentrations measured, together with HbA1c, to classify them as with normoglycemia or prediabetes. They also wore a CGM sensor for 14 days. CGM profile features were extracted and prediction models were constructed with random subsampling validation to evaluate predictive efficacy. The use of CGM and HbA1c data alone or in combination was assessed for the ability to correctly distinguish prediabetes from normoglycemia.
Results: In this cohort (N=406), 189 (46.6%) individuals had prediabetes. The majority of the cohort were women (n=236, 58.1%) and of Chinese ethnicity (n=267, 65.8%). Those with prediabetes were slightly older, heavier, and had higher glucose levels with more variability than the normoglycemia group. A 2-step approach was used where those with HbA1c ≥5.7% were automatically categorized as having prediabetes; the model then focused on the prediction capability of the CGM features among individuals with HbA1c <5.7%. The prediction models with CGM outperformed the benchmark for comparison defined by HbA1c ≥5.7%, where they yielded an area under the receiver operating characteristic curve of 0.866-0.876, with a lower specificity of 78%-80% but a vastly improved sensitivity of 76%-78%.
Conclusions: Adding CGM to HbA1c in a 2-step approach greatly improved the sensitivity of detecting prediabetes in an Asian population. Given the benefits to optimizing lifestyle behaviors and its growing acceptability among the nondiabetic population, CGM is a promising alternative for type 2 diabetes mellitus risk screening.

Keywords:  Asian ; HbA1c; continuous glucose monitoring ; glycated hemoglobin; machine learning ; prediabetes; screening

DOI:  https://doi.org/10.2196/81520
Nat Genet. 2026 Apr 30.

Genetic association and machine learning improve the prediction of type 1 diabetes risk.

Carolyn McGrail, Timothy J Sears, Emily N Griffin, Alexandra L Ghaben, Patrick Smadbeck, Jason Flannick, Parul Kudtarkar, Hannah Carter, Kyle Gaulton.

Type 1 diabetes (T1D) has a large genetic component, and expanded genetic studies of T1D can enhance biological and therapeutic discovery and improve risk prediction. Here we performed genome-wide genetic association and fine-mapping analyses in 20,355 T1D and 797,363 nondiabetic individuals of European ancestry and in 10,107 T1D and 19,639 nondiabetic individuals at the MHC locus, which identified 160 risk signals. We trained a machine learning model, T1GRS, to predict T1D using genetic risk, which improved classification in Europeans and performed similarly in African Americans, compared to previous scores. T1GRS particularly improved prediction in T1D, with fewer high-risk HLA haplotypes and more complex risk profiles, and revealed 154 nonlinear interactions between MHC and non-MHC loci. Finally, we identified four genetic subclusters based on T1GRS features with significant differences in age of onset and diabetic complications. Overall, improved genetic discovery and prediction will have wide clinical, therapeutic and research applications for T1D.

DOI: https://doi.org/10.1038/s41588-026-02578-y
J Diabetes Sci Technol. 2026 Apr 26. 19322968261438523

Identifying Actionable Glucose Patterns in Type 1 and Type 2 Diabetes Using Continuous Glucose Monitoring and Algorithmic Pattern Recognition.

Timor Glatzer, Ajandek Peak, Eemeli Leppäaho, Patrick Lustenberger, Pau Herrero, Magí Andorrà, Ellen van Maren.

   BACKGROUND: Continuous glucose monitoring (CGM) provides real-time glucose data, aiding diabetes management. Identifying glucose patterns is difficult for patients due to data overload, hindering self-management. This study aimed to systematically identify glucose patterns using Accu-Chek SmartGuide and quantify their impact on glucose management.
METHODS: This retrospective, observational analysis included real-world CGM data from 3379 individuals with type 1 diabetes (T1D; N = 2198) or type 2 diabetes (T2D; N = 1181), encompassing 23 486 valid user-weeks. An algorithm identified 29 predefined glucose patterns weekly. Pattern prevalence, demographic influence, persistence, their attribution to time above range/time below range (TAR/TBR) as well as their potential impact on time in range (TIR) in case of pattern resolution were analyzed.
RESULTS: Resolving glucose patterns, defined as repeatedly occurring glucose events, showed varying potential for glycemic improvement. Cumulatively, actionable patterns contributed significantly to total TAR (T1D: 66.2 ± 14.7%, T2D: 58.0 ± 14.3%) and TBR (T1D: 56.3 ± 2.6%, T2D: 42.2 ± 1.4%). For instance, resolving the day-time hyperglycemia pattern could improve TIR by up to +10.72% (4.26, 16.9) in T1D and +5.16% (0.0, 12.92) in T2D, addressing an average of 9.33 (8.0, 10.75) events per week in T1D and 9.29 (8.0, 10.67) in T2D.
CONCLUSION: The majority of glucose excursions in T1D and T2D can be explained by recurring glucose patterns. Detecting these actionable patterns provides an opportunity to improve TIR. Targeting therapy and behavior change toward resolving these patterns is a critical step toward more personalized diabetes management.

Keywords:  artificial intelligence; continuous glucose monitoring (CGM); glucose patterns; glucose variability; hyperglycemia; hypoglycemia; pattern recognition

DOI:  https://doi.org/10.1177/19322968261438523
Front Endocrinol (Lausanne). 2026 ;17 1816661

Machine learning-based risk predictive models for depression in patients with diabetes: a systematic review and meta-analysis.

Xingxin Cai, Guiying Guo, Jun Zhou, Mengqi Han, Yuanyuan Cui, Zhenglin Chen.

   Background: Currently, numerous studies have employed machine learning (ML) methods to develop predictive models for depression risk in patients with diabetes mellitus (DM); however, the findings remain inconsistent. Therefore, this study aims to clarify the current state of research and emerging trends in this field by systematically evaluating the performance, strengths, and limitations of existing prediction models.
Objective: This systematic review evaluates the performance and clinical applicability of ML-based depression risk prediction models for patients with DM, providing reliable evidence to assist healthcare professionals in selecting and optimizing more appropriate prediction models.
Methods: We conducted a systematic search of clinical studies employing ML approaches to predict depression risk in patients with DM across the PubMed, Embase, Cochrane Library, and Web of Science databases, from their inception to January 2026. The primary performance metric for the models was the area under the receiver operating characteristic curve (AUC) along with its 95% confidence interval (95% CI). Two independent researchers screened the literature, extracted data, and used PROBAST-AI to assess the risk of bias and clinical applicability of the included studies. Pooled AUC was estimated using the Der Simonian and Laird random-effects model.
Results: A total of 14 studies comprising 64 distinct ML models were included. All included studies were assessed as high risk of bias and high clinical applicability. A pooled analysis of the best-performing ML prediction models reported in each study showed a pooled AUC of 0.822 (95% CI, 0.789-0.858), indicating relatively good overall predictive performance. However, there was substantial heterogeneity among the studies (I² = 97.4%; P < 0.001). Subgroup analysis based on ML model types revealed the following pooled AUC values: 0.765 (95% CI 0.706-0.829) for traditional regression models, 0.789 (95% CI 0.747-0.834) for general machine learning models, and 0.802 (95% CI 0.769-0.836) for deep learning models. Notably, logistic regression (LR) (n = 10) was the most frequently employed ML method for developing depression risk prediction models in patients with DM. To evaluate model generalizability and avoid overfitting, the included studies adopted three validation strategies: 5-fold cross-validation yielded a pooled AUC of 0.913 (95% CI 0.781-1.067), 10-fold cross-validation yielded 0.819 (95% CI 0.781-0.858), and random split validation yielded 0.747 (95% CI 0.648-0.862). The most commonly used predictors in the included models were age, sex, and body mass index (BMI), which are readily available in clinical settings and strongly associated with depression risk.
Conclusions: ML-based depression risk prediction models for patients with DM demonstrate overall satisfactory predictive performance. However, most existing studies had relatively small sample sizes and lacked external validation. Future research should prioritize refining study design and optimizing clinical data processing to improve the generalizability and stability of these models in clinical practice.
Systematic Review Registration: https://www.crd.york.ac.uk/PROSPERO/view/CRD420251243343, identifier CRD420251243343.

Keywords:  depression; diabetes mellitus; machine learning; meta-analysis; predictive model; systematic review

DOI:  https://doi.org/10.3389/fendo.2026.1816661
IEEE J Biomed Health Inform. 2026 Apr 28. PP

Noninvasive Blood Glucose Estimation via ECG: A Multi-Expert SC-ResNet Model.

Jialin Zhang, Junsong Wang, Weitong Feng, Syeda Shamaila Zareen, Jianfang Wu, Chao Luo.

Diabetes mellitus, a chronic metabolic disorder, has seen a steep rise in global incidence, posing a significant health challenge. Traditional blood glucose estimation methods are invasive and limited, failing to reflect blood glucose fluctuations comprehensively. However, these fluctuations elicit responses from the autonomic nervous system, resulting in electrocardiographic (ECG) alterations, offering a novel avenue for noninvasive Blood Glucose (BG) estimation. In this work, we propose a Multi-Expert SC-ResNet model to enhance the accuracy and reliability of glucose prediction utilizing ECG signals. Firstly, tailored to the characteristics such as dynamic changes and periodicity of ECG signals, the SC-ResNet is designed, which integrates the Spatial and Channel reconstruction Convolution (SCConv) module into ResNet, effectively reducing spatial and channel redundancy and facilitating the learning of representative features. Secondly, we combine an innovative Multi-Expert method with SC-ResNet model to comprehensively capture the features of ECG signals at different blood glucose levels. ECG datasets are categorized based on their corresponding blood glucose into three different categories: Hypoglycaemic (L), Normoglycaemic (N), and Hyperglycaemic (H). For each category, we use the SC-ResNet model to extract deep features and then perform feature fusion. These fused features are finally fed into the Random Forest model for blood glucose prediction. Experimental validation on the D1NAMO dataset demonstrates the superior performance of the proposed method in terms of Root Mean Square Error (RMSE), Mean Absolute Relative Difference (MARD), and Clarke Error Grid Analysis (CEGA). The results demonstrate the effectiveness of the method for ECG-based blood glucose estimation and indicate its potential for practical diabetes management.

DOI: https://doi.org/10.1109/JBHI.2026.3687523
Biochem Biophys Rep. 2026 Jun;46 102593

Screening of metabolic-related biomarkers linking intervertebral disc degeneration and type 2 diabetes based on comprehensive bioinformatics analysis and machine learning.

Ba Qiu, Tao Chen, Qi Hong, Xiaoxing Xie, Mintao Xue, Ning Xu, Changgui Shi, Guohua Xu, Weiheng Wang, Muhammad Sulaiman, Yanhai Xi.

   Background: Intervertebral disc degeneration (IVDD) is a prominent etiology of lower back pain. Type 2 diabetes (T2D), the most prevalent metabolic disorder, may expedite IVDD progression through mechanisms involving hyperglycemia, advanced glycation end products, and microvascular complications. We aim to identify the biomarkers for T2D and IVDD.
Methods: We retrieved two datasets pertaining to IVDD and T2D respectively from the Gene Expression Omnibus (GEO) database. Initially, differential expression analysis was conducted to identify differentially expressed genes (DEGs) in each group. Subsequently, machine learning algorithms including Boruta algorithm, Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost) were utilized to explore co-biomarkers for IVDD and T2D by analyzing the intersection of DEGs sets between the two groups. The clinical diagnostic value of these biomarkers was evaluated using ROC curve. Additionally, we employed the CYBERSORT and Single-cell sequencing to investigate the infiltration of immune cells in two diseases. Finally, the expression of biomarkers and the correlation between BCAA and IVDD was experimentally validated.
Results: AMBP, a shared biomarker of IVDD and T2D, were identified using differential analysis and machine learning algorithms. In addition, CYBERSORT analysis and single-cell sequencing results revealed significant correlations between AMBP and cellular immunity. Finally, experimental validation on clinical samples and HNPC cells confirmed upregulation of AMBP, and the correlation between BCAA metabolism and AMBP.
Conclusion: AMBP, as the shared biomarker between IVDD and T2D, has great clinical diagnostic value, which therefore may be a potential regulatory factor for two diseases.

Keywords:  AMBP; BCAA; Intervertebral disc degeneration; Machine learning; Type 2 diabetes

DOI:  https://doi.org/10.1016/j.bbrep.2026.102593
Metabolites. 2026 Mar 30. pii: 227. [Epub ahead of print]16(4):

Interpretable Machine Learning to Predict Metformin-Induced Vitamin B12 Deficiency: Association with Glycemic Control and Neuropathic Symptoms.

Yasmine Salhi, Meriem Yazidi, Amine Dhraief, Elyes Kamoun, Melika Chihaoui, Tamim Alsuliman, Layth Sliman.

  Background/Objectives: Vitamin B12 deficiency is a common but often underdiagnosed complication in patients with type 2 diabetes (T2D) undergoing long-term metformin therapy. Accurate early prediction could enable targeted screening and timely intervention. This study aimed to develop and interpret a machine learning model for predicting vitamin B12 deficiency in metformin-treated patients with T2D, using eXtreme Gradient Boosting (XGBoost). Methods: A retrospective cross-sectional study was conducted at a single endocrinology centre (La Rabta University Hospital, Tunis, Tunisia). Patients with T2D treated with metformin for at least three years were included (n = 257); those with conditions independently affecting vitamin B12 metabolism were excluded. Vitamin B12 deficiency was defined as a serum B12 level below 150 pmol/L or a borderline level (150-221 pmol/L) with concurrent hyperhomocysteinemia (>15 μmol/L). XGBoost was selected after comparison with Logistic Regression (L2), Random Forest, and Support Vector Machine on the same 5-fold stratified cross-validated pipeline. Hyperparameters were optimized via Bayesian search (100 iterations × 5-fold stratified cross-validation), with the Matthews correlation coefficient (MCC) as the primary optimization metric to account for class imbalance. Model interpretability was achieved using SHapley Additive exPlanations (SHAP). Discrimination and calibration were assessed on an independent test set using bootstrap 95% confidence intervals (2000 resamples). Results: Of 257 patients, 95 (37.0%) presented with vitamin B12 deficiency. On the independent test set (n = 52), the optimized XGBoost model achieved an ROC-AUC of 0.671 [95% CI: 0.514-0.818], sensitivity of 0.737 [95% CI: 0.533-0.938], specificity of 0.545 [95% CI: 0.375-0.710], MCC of 0.273 [95% CI: 0.018-0.517], and a Brier Score of 0.259. SHAP analysis identified HbA1c, microalbuminuria, autonomic neuropathy, BMI, DN4 score, and fasting glucose as the most influential predictors. Nonlinear SHAP interaction plots revealed an increased predicted risk in patients with low HbA1c combined with a high cumulative metformin dose. Conclusions: The XGBoost-SHAP framework provided interpretable predictions of vitamin B12 deficiency in patients with T2D on metformin, identifying key clinical profiles for targeted screening. External multi-centre validation is required before clinical deployment.

Keywords:  SHAP; XGBoost; artificial intelligence; metformin; type 2 diabetes; vitamin B12 deficiency

DOI:  https://doi.org/10.3390/metabo16040227
Comput Med Imaging Graph. 2026 Apr 22. pii: S0895-6111(26)00074-1. [Epub ahead of print]131 102771

AI-driven retinal image quality monitoring in diabetic retinopathy screening: A retrospective study identifying actionable insights to improve imaging protocols.

Imanol Pinto, Álvaro Olazarán, David Jurio, Borja de la Osa, Miguel Sainz, Aritz Oscoz, Jerónimo Ballaz, Javier Gorricho, Mikel Galar, José Andonegui.

   OBJECTIVE: To verify the applicability of retinal image quality (RIQ) models in real-world diabetic retinopathy screening for uncovering actionable insights to improve imaging protocols.
MATERIALS AND METHODS: NaIA-RD, a custom AI system developed by the University Hospital of Navarre (Spain) for diabetic retinopathy screening, was employed to monitor retinal image quality (RIQ) across multiple imaging sites within the hospital. A large retrospective dataset consisting of 55,801 routine retinal images collected over a period of 3.6 years was compiled for this purpose. Additionally, two convolutional neural networks, trained on external public datasets (EyeQ and DeepDRiD), were used as independent comparators. The longitudinal RIQ outputs from NaIA-RD, EyeQ, and DeepDRiD models were then analyzed to assess their alignment with clinical decisions.
RESULTS: All three models identified similar differences in RIQ across imaging sites, camera models, and imaging technicians. Ungradable rates varied widely among sites, ranging from 2.23% to 28.23%. These differences evolved over time due to changes in data distribution, or data drifts. Among the models, the one trained with DeepDRiD demonstrated the highest agreement with clinicians, achieving an Average Precision of 0.431, compared to 0.389 for NaIA-RD and 0.392 for EyeQ.
DISCUSSION: Monitoring RIQ revealed actionable insights, such as identifying differences related to camera models and technician experience, suggesting potential benefits from targeted training and imaging protocol standardization. Comparing outputs from multiple models strengthened the reliability of observed trends.
CONCLUSION: AI tools with modular design and detailed RIQ scoring can effectively monitor clinical imaging workflows, enabling data-driven healthcare quality improvement initiatives.

Keywords:  Artificial intelligence; Computer-assisted image processing; Diabetic retinopathy; Health care quality assurance

DOI:  https://doi.org/10.1016/j.compmedimag.2026.102771
Front Public Health. 2026 ;14 1778156

Central adiposity outperforms disordered eating behavior in predicting type 2 diabetes risk in young adults: an explainable machine learning approach.

Hacer Alatas, Nurgul Arslan, Harun Kurt, Serkan Bentli.

   Background: Disordered eating behaviors are often associated with adverse metabolic outcomes, yet their relationship with type 2 diabetes mellitus(T2DM) risk in young adults is less clear. This study aimed to evaluate the impact of disordered eating behaviors on diabetes risk among university students, using both traditional statistical methods and machine learning approaches.
Methods: A total of 1,302 university students participated in this cross-sectional study. Disordered eating behavior was assessed using the Eating Attitudes Test-40 (EAT-40), a validated screening tool for abnormal eating attitudes, while type 2 diabetes risk was estimated using the Finnish Diabetes Risk Score (FINDRISK), a widely used non-invasive instrument designed to estimate the 10-year risk of developing T2DM. Anthropometric measures were recorded according to standardized protocols. Bivariate associations were examined using correlation analysis, while multivariable regression and machine learning models (XGBoost) were applied to determine predictors of diabetes risk.
Results: Of the 1,302 participants, 90.8% were classified as low/mild risk, 6.2% as moderate risk, and 3.0% as high/very high risk according to FINDRISK. No significant correlation was found between EAT-40 scores and FINDRISK (r = 0.01, p = 0.755). In multivariable regression, waist-to-height ratio (β = 1.42 per 0.05 increase, p < 0.001) and body mass index (β = 0.31, p < 0.001) were the strongest predictors of diabetes risk. Machine learning models, particularly XGBoost (AUROC = 0.87), highlighted waist-to-height ratio as the most influential predictor.
Conclusion: In young adults, central adiposity specifically waist-to-height ratio was the most significant predictor of T2DM risk, while disordered eating behavior had minimal independent impact. These findings suggest that simple anthropometric measures could be prioritized for early diabetes risk assessment over eating attitude screening.

Keywords:  FINDRISK; disordered eating; machine learning; type 2 diabetes; university students

DOI:  https://doi.org/10.3389/fpubh.2026.1778156
Front Digit Health. 2026 ;8 1768843

ChatGPT for diabetes education: potential, accuracy, and accessibility in patient support.

Mohammad Abuzar, Sandeep Rai, Tabreskhan Pathan, Foorkan Fakki, Shariq Syed.

   Background: Diabetes mellitus is a chronic metabolic disease with rising global prevalence. Adequate patient education is essential to encourage self-management and reduce complications. Artificial intelligence applications such as ChatGPT have emerged as potential supplementary resources for patient education alongside the broader integration of technology in healthcare.
Methods: A cross-sectional evaluation was conducted using ten frequently asked questions (FAQs) on diabetes, selected from the Diabetic Association of India and the International Diabetes Federation. ChatGPT-4o (accessed via the web interface in March 2025) generated responses to each question in separate, stand-alone chat sessions to simulate typical patient interactions. Five board-certified endocrinologists (diabetologists) with a mean clinical experience of ≥10 years independently evaluated the responses using a 4-point Likert scale across five domains: overall quality, content accuracy, clarity, relevance, and trustworthiness. Final domain scores were computed as the mean of all five raters' scores. Readability was assessed using the Flesch Reading Ease Score (FRES) and Flesch-Kincaid Grade Level (FKGL). All readability analyses apply exclusively to the English-language outputs generated in this study.
Results: The mean FRES was 38.19 and the mean FKGL was 16.87, indicating a reading level appropriate for college-educated individuals and substantially above the recommended sixth-grade benchmark for patient health materials. Mean response length was 300 ± 100 words across the ten prompts. Expert ratings were generally high: aggregated mean scores (±SD) were 4.0 (±0.0) for content accuracy and overall quality, 3.98 (±0.10) for relevance, and 3.9 (±0.20) for clarity and trustworthiness. No clinically inaccurate statements were identified by the raters; however, the high scores and narrow score range indicate a potential ceiling effect that limits discrimination between responses. Raters expressed concern about linguistic complexity, which may impede comprehension among patients with limited health literacy.
Conclusions: ChatGPT-4o generated generally accurate and relevant diabetes education content, suggesting potential as a supplementary tool in diabetes care. However, the high reading-level complexity, small evaluation scope (ten prompts, one model, one session), and English-only assessment limit the generalisability of these findings. AI-generated content should supplement, not replace, clinician-led education. Future work should address language simplification, multilingual evaluation, and longitudinal assessment of patient outcomes.

Keywords:  AI in healthcare; ChatGPT; artificial intelligence; blood glucose control; diabetes mellitus; patient education

DOI:  https://doi.org/10.3389/fdgth.2026.1768843
JMIR Med Inform. 2026 04 28. 14 e98668

Retraction: Artificial Intelligence-Based Neural Network for the Diagnosis of Diabetes: Model Development.

JMIR Publications Editorial Office

DOI: https://doi.org/10.2196/98668
J Diabetes Sci Technol. 2026 Apr 26. 19322968261441637

MetaboNet: The Largest Publicly Available Consolidated Data Set for Type 1 Diabetes Management.

Miriam Kopperstad Wolff, Peter Calhoun, Eleonora Maria Aiello, Yao Qin, Sam Fox Royston.

   BACKGROUND: Progress in type 1 diabetes (T1D) algorithm development is limited by the fragmentation and lack of standardization across existing T1D management data sets. Current data sets differ substantially in structure and are time-consuming to access and process, which impedes data integration and reduces the comparability and generalizability of algorithmic developments. This work aims to establish a unified and accessible data resource for T1D algorithm development.
METHOD: Multiple publicly available T1D data sets were harmonized into a unified resource, termed the MetaboNet data set. Inclusion required the availability of both continuous glucose monitoring (CGM) data and corresponding insulin pump dosing records. In addition, auxiliary information such as reported carbohydrate intake and physical activity was retained when present.
RESULTS: The MetaboNet data set comprises 3135 subjects and 1228 patient-years of overlapping CGM and insulin data, making it substantially larger than existing standalone benchmark data sets. The resource is distributed as a fully public subset available for immediate download at https://metabo-net.org/, and with a data use agreement (DUA)-restricted subset accessible through their respective application processes. For the data sets in the latter subset, processing pipelines are provided to automatically convert the data into the standardized MetaboNet format.
CONCLUSIONS: A harmonized public data set for T1D research is presented, and the access pathways for both its unrestricted and DUA-governed components are described. The resulting data set covers a broad range of glycemic profiles and demographics and thus can yield more generalizable algorithmic performance than individual data sets.

Keywords:  data set; glucose prediction; machine learning; type 1 diabetes

DOI:  https://doi.org/10.1177/19322968261441637
Food Sci Nutr. 2026 May;14(5): e71828

Machine Learning Unveils Dietary Antioxidants as Influential Factors for Diabetes-Cancer Comorbidity: Insights From National Health and Nutrition Examination Survey.

Ming Xu, Ru Li, Shuqing Jin, Qinhao Liu, Hongxia Wei, Zhe Li, Yuchen Sun, Yi Zhang, Yunfeng Liu.

  This study aims to assess the predictive value of dietary antioxidants in diabetes-cancer comorbidity using interpretable machine learning (ML) models and to identify key clinical factors. Data were sourced from the National Health and Nutrition Examination Survey (NHANES) 2007-2010 and 2017-2018 cycles, including 44 dietary antioxidants, as well as demographic, lifestyle, and health-related features. 8 ML models (Random Forest, light Gradient Boosting Machines [LightGBM], Logistic Regression, Decision Tree, Multilayer Perceptron, Naïve Bayes, Kernel k-Nearest Neighbors, and Support Vector Machine with Radial Basis Function) were trained, with preprocessing steps for multicollinearity, class imbalance (SMOTE), and data normalization. Model performance was evaluated using AUC, accuracy, Brier scores, and calibration plots. SHapley Additive exPlanations (SHAP) values were applied to interpret feature importance. Data from 8644 participants were analyzed, including 272 individuals with confirmed diabetes-cancer comorbidity. After removing collinear features, the ML model included 30 dietary antioxidant features and 10 baseline features. The Random Forest model achieved optimal performance (AUC = 0.996, accuracy = 0.978, brier score = 0.0241), followed by LightGBM (AUC = 0.993). SHAP analysis revealed that while advanced age, cardiovascular disease, and hypertension were the primary drivers of comorbidity probability, dietary antioxidants are also influential factors. Specifically, polyphenols (daidzein, malvidin, pelargonidin, cyanidin) and essential minerals (magnesium) emerged as the most influential nutritional features. The high accuracy of the Random Forest and LightGBM models underscores their clinical utility in risk stratification for diabetes-cancer comorbidity. While advancing age and cardiometabolic dysfunction primarily drives the probability of diabetes-cancer comorbidity. This study establishes dietary antioxidants, particularly polyphenols such as daidzein and malvidin, as predictive factors for diabetes-cancer comorbidity.

Keywords:  National Health and Nutrition Examination Survey; cancer, machine learning; diabetes mellitus type 2; dietary antioxidants

DOI:  https://doi.org/10.1002/fsn3.71828
Medicine (Baltimore). 2026 May 01. 105(18): e48610

Identification of COL5A1 and TGFBI as key basement membrane genes and their molecular mechanisms in gestational diabetes mellitus.

Xiaofang Weng, Xingdi Guo, Chunmei Pan.

  Gestational diabetes mellitus (GDM) is a common complication during pregnancy, but the role of the basement membrane (BM) in GDM is not well understood. This study aims to investigate BM-related genes in GDM to provide new insights for diagnosis and treatment. Differentially expressed genes were identified from the GSE203346 dataset in the Gene Expression Omnibus and intersected with BM-related genes to identify BM-related differential genes. Machine learning and gene expression validation were used to identify key genes, which were further validated using an artificial neural network. Additional analyses included gene set enrichment analysis, immunoprecipitation, drug prediction, gene localization, and the construction of lncRNA-miRNA-mRNA and transcription factor-mRNA regulatory networks to explore underlying mechanisms. Among 801 differentially expressed genes, 24 BM-related differential genes were identified. COL5A1, TGFBI, AGRN, TNC, and ITGB6 were identified as candidate genes, with COL5A1 and TGFBI showing consistent low expression across datasets and being designated as key genes. The artificial neural network demonstrated that these key genes effectively distinguished GDM from control samples. Gene set enrichment analysis revealed the involvement of these genes in pathways such as systemic lupus erythematosus and cytokine-cytokine receptor interaction. TGFBI showed a significant positive correlation with CD4+ memory T cells, common lymphoid progenitors, hematopoietic stem cells, and smooth muscle, while COL5A1 was positively correlated with common lymphoid progenitors and smooth muscle. Six drugs were identified as interacting with both key genes. Our study suggests that COL5A1 and TGFBI offer the possibility of personalized treatment strategies for GDM in the future.

Keywords:  basement membrane; bioinformatics; gene regulatory network; gestational diabetes mellitus; machine learning

DOI:  https://doi.org/10.1097/MD.0000000000048610
Placenta. 2026 Apr 21. pii: S0143-4004(26)00139-6. [Epub ahead of print]181 1-13

Deep learning-based early prediction of gestational diabetes mellitus through first-trimester placental texture analysis.

Yao Peng, Xiaofeng Zhang, Ming Fang, Yang He, Huating Yuan, Wenyang Du, Wei Feng, Ling Gan, Jiaqi Zhang.

   INTRODUCTION: This study aimed to develop a multi-parameter fusion model for early GDM risk prediction and validate its performance through external multicenter testing.
METHODS: A total of 628 pregnant women at 11+0-13+6 weeks were enrolled from two medical centers. The Center I cohort was divided into training (n = 356) and testing sets (n = 153). Radiomic features (1,289) and deep learning features (2,048) were extracted from placental ultrasound images. Feature-level fusion resulted in 3337 features, which were selected using Spearman correlation, mRMR, and LASSO. Five models were built: Rad Model, DTL Model, DLR Model, Clinic Model, and Combined Model. Performance was assessed using ROC analysis, DCA, and calibration curves.
RESULTS: The Combined Model achieved the best overall performance, with an area under the ROC curve (AUC) of 0.879 in the internal validation, significantly outperforming any single-modality model (P < 0.05). DCA demonstrated that the fusion-based model provided higher net clinical benefit across a wide range of threshold probabilities compared with both "treat-all" and "treat-none" strategies. The calibration curve showed excellent agreement between predicted and observed probabilities (Hosmer-Lemeshow test, P > 0.05).
DISCUSSION: The multimodal fusion model enhanced early GDM prediction by detecting subtle placental changes in first-trimester, enabling timely intervention and personalized decision-making.

Keywords:  Deep learning features; Early prediction; Gestational diabetes mellitus; Multimodal fusion model; Placental ultrasound radiomics

DOI:  https://doi.org/10.1016/j.placenta.2026.04.016
Diabetes Res Clin Pract. 2026 Apr 25. pii: S0168-8227(26)00199-3. [Epub ahead of print] 113279

Evaluation of large language models for carbohydrate counting in type 1 diabetes mellitus: a comparative study with patient estimations.

Ahmet Kürşat Soyer, Gamze Çelen Akarsu, Belgin Demir Yelli, Belma Özlem Tural Balsak, Didem Özdemir, Oya Topaloğlu, Reyhan Ersoy, Bekir Çakır.

   AIMS: To evaluate large language models (LLMs) accuracy in carbohydrate (CHO) counting (CC) and compare their performance with estimations by patients with type-1 diabetes mellitus (T1DM).
METHODS: This cross-sectional study included 14 adults with T1DM with advanced CC training and > 70% accuracy on the AdultCarbQuiz. Eighteen main meals and four snacks were prepared by dietitians using standardized descriptions and two-angle photographs. Dietitian-determined values served as the reference. CHO estimations by patients and LLMs (ChatGPT-5.2, Gemini-3 Pro, and Claude Sonnet-4.5) were compared. LLM performance was evaluated in two-stages: image-only (stage-1) and image-plus-detailed-description (stage-2) inputs. Performance was assessed using mean absolute error (MAE), mean absolute percentage error (MAPE), counting accuracy (CA; ≤10 g), Wilcoxon signed-rank test, and Bland-Altman analysis.
RESULTS: Patients showed MAE 8.31 g, MAPE 13.82%, and CA 73.74%. In Stage-1, Gemini had the lowest MAE (6.55 g), and MAPE (12.54%), while Claude achieved the highest CA (81.8%). In Stage-2, Gemini outperformed patients (p < 0.001) with MAE 2.14 g, MAPE 3.21%, CA 100%, and mean bias - 1.05 g. Other model-stage combinations showed performance comparable to patients.
CONCLUSION: LLMs, particularly Gemini-3 Pro, may support CC when detailed inputs are provided; however, findings should be interpreted cautiously given the small, highly selected sample.

Keywords:  Artificial intelligence in diabetes; Carbohydrate counting; Type 1 diabetes mellitus

DOI:  https://doi.org/10.1016/j.diabres.2026.113279
J Med Phys. 2026 Jan-Mar;51(1):51(1): 136-144

Transfer Learning-driven Biofield Image Analysis for Predictive Modeling of Diabetes.

Shivanand S Gornale, Supriya Shankar Patil, Anup Waman Deo, Abdulgaphur Athani.

   Background: The human biofield serves as an indicator of an individual's physical and emotional health status. Biofield-based therapeutic techniques, also known as complementary and alternative medicine (CAM) techniques such as Reiki, Therapeutic Touch, and Pranic Healing, leverage this information in the preliminary assessment phase before treatment initiation. These modalities are increasingly integrated as complementary methods within health diagnostic frameworks. Among the techniques employed for biofield visualization, gas discharge visualization (GDV) and polycontrast interference photography (PIP) are the predominant imaging methodologies. Notably, the majority of scientific investigations and empirical studies have primarily utilized GDV-derived images, with comparatively fewer studies focusing on PIP-based data.
Purpose: The primary objective of this study is to identify energy imbalances within the pancreatic region using biofield imaging and to utilize these patterns for classifying subjects as diabetic or nondiabetic. This work emphasizes the relevance of biofield information in health assessment and evaluates its potential for supporting energy-based diagnostic approaches.
Materials and Methods: Color-based clustering methods were applied for segmentation. A transfer learning-based ensemble framework was developed using pretrained convolutional neural network (CNN) architectures ConvNeXtBase and ResNet50 to classify biofield images into diabetic and nondiabetic categories. Grid search optimization identified the optimal hyperparameters, which were applied during fine-tuning to improve feature learning. Ensemble model was evaluated, with the ConvNeXtBase + ResNet50 combination achieving the highest accuracy of 99.12%. Robust performance validation was ensured using 5-fold cross-validation to minimize sampling bias and enhance generalization.
Results: The ensemble of ResNet50 and ConvNeXtBase achieved the highest accuracy of 99.12%, outperforming individual models (ConvNeXtBase: 97.93% and ResNet50: 96.28%). Receiver operating characteristic analysis confirmed strong reliability with area under the curve values above 0.99 for both classes (diabetic and nondiabetic). The 5-fold cross-validation analysis further demonstrated the robustness of the proposed ensemble model, achieving a mean accuracy of 97.45%, indicating highly consistent performance across different dataset partitions.
Conclusions: The CNN-based models can be trained to classify the biofield images, and this approach can enable automated analysis of biofield images. The approach of using clustering, deep learning, and ensemble modeling as analyzed and described in this study seems to be highly effective. The overall system of biofield imaging and automated clustering can act as a potential noninvasive diagnostic support tool, though further testing with larger datasets and expert validation is necessary for clinical application.

Keywords:  Biofield image; clustering image segmentation; deep learning techniques; gas discharge visualization; hyperparameter tuning; pancreas energy; polycontrast interference photography; transfer learning techniques

DOI:  https://doi.org/10.4103/jmp.jmp_245_25
Gait Posture. 2026 Apr 22. pii: S0966-6362(26)00106-2. [Epub ahead of print]128 110196

Multi-modal sensor fusion in smart footwear to capture altered gait dynamics in patients with diabetic peripheral neuropathy: System validation and clinical applications.

Xiaohui Wang, Anqi Xie, Jiaxing Zhang, Ji Huang, Xiang Geng, Weijing Zhao, Jun Yin, Xin Ma, Wen-Ming Chen.

   BACKGROUND: Individuals with diabetic peripheral neuropathy (DPN) are confronted with significantly elevated risk of falling and diabetic foot ulcers. There is an increasing need for clinically viable wearable technologies for routine biomechanical assessment of DPN patients.
RESEARCH QUESTION: Can a footwear-based wearable system ("Lab-in-Shoe") achieve gold-standard accuracy and effectively distinguish gait characteristics among patients with type 2 diabetes (T2DM) under various neuropathic conditions?
METHODS: The spatiotemporal accuracy of the system was first validated against an 8-camera Vicon motion capture system using a single-marker strategy. For clinical evaluation, thirty T2DM patients were classified into G1 (no neuropathy), G2 (subclinical), and G3 (confirmed DPN). All patients performed a 10 m walking test, utilizing the "Lab-in-Shoe" to simultaneously capture spatiotemporal parameters and plantar pressure to evaluate their effectiveness in distinguishing degrees of neuropathy. Additionally, a Diabetic Gait Index (DGI) was formulated based on these fused features.
RESULTS: The system showed excellent agreement with Vicon across all seven evaluated spatiotemporal parameters, yielding mean absolute errors (MAE) between 2% and 8% and intraclass correlation coefficients (ICC) ranging from 0.812 to 0.993. The severe neuropathy group (G3) exhibited a 14% reduction in step length (p < 0.001) and a 14% reduction in step height (p < 0.01). Concurrently, step frequency was significantly increased by 8% (p < 0.001) and swing phase duration was shorter by 9% (p < 0.01). Regarding plantar loading, G3 showed significantly higher load under the second metatarsal and lateral heel compared to G1 (all p < 0.05). Furthermore, the proposed DGI achieved an 86.7% accuracy in classifying the three neuropathy severity groups.
SIGNIFICANCE: This wearable system effectively detects subtle gait changes associated with the progression of DPN. It is well-suited for routine clinical practice, offering an objective tool for early risk screening for falls and ulceration in DPN patients.

Keywords:  Diabetic Foot; Diabetic Neuropathy; Digital Health; Gait Analysis; Wearable Devices

DOI:  https://doi.org/10.1016/j.gaitpost.2026.110196
Biosensors (Basel). 2026 Apr 10. pii: 214. [Epub ahead of print]16(4):

A Flexible Wearable Glucose Sensor for Noninvasive Diabetes Screening: Functional Equivalence and Model Interpretability.

Wenhan Xie, Jinqi Wang, Hao Liu, Shuo Chen, Peng Wang, Yumei Han, Xianxiang Chen, Zhen Fang, Zhan Zhao, Guohong Zhang, Xiuhua Guo.

  Real-world evidence for wearable noninvasive glucose monitoring (NIGM) remains limited. To evaluate the functional equivalence of a wearable NIGM device and explore its utility for T2DM and prediabetes screening. In this multicenter study, 12-h daytime glucose profiles obtained by a flexible reverse iontophoresis-based electrochemical sensor were compared with capillary glucose using functional equivalence. Subgroup analyses were conducted. Screening models of T2DM and prediabetes were developed using elastic net and Logistic regression. A total of 135 participants (mean age 35.3 years; 60.0% female) were included, and no serious device-related adverse events were reported. Compared to the capillary measurements, functional equivalence was confirmed (T = -6.537 < threshold = -2.081) in the general population but not in older adults or T2DM patients. The T2DM noninvasive screening model demonstrated discrimination and reclassification performance comparable to those of the capillary-based model (AUC: 0.906 vs. 0.850, NRI: 0.044, IDI: -0.078, p > 0.05). Functional principal component scores facilitated the identification of prediabetes (AUC = 0.760). The device demonstrated acceptable accuracy and functional equivalence with reference methods. Its capability to detect T2DM and early glycemic anomalies supports its feasibility as a wearable, interpretative adjunct tool for large-scale screening in free-living populations.

Keywords:  machine learning; noninvasive glucose monitoring; population screening; type 2 diabetes mellitus; wearable device

DOI:  https://doi.org/10.3390/bios16040214