• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Monday, November 16, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Health

Machine Learning Shows Modest Accuracy in Predicting Early Diabetic Kidney Disease Progression

by
November 14, 2026
in Health
Reading Time: 7 mins read
0
Machine Learning Shows Modest Accuracy in Predicting Early Diabetic Kidney Disease Progression

Machine Learning Shows Modest Accuracy in Predicting Early Diabetic Kidney Disease Progression

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Routine clinical and laboratory data evaluated by machine learning models yielded modest predictive accuracy for short-term kidney disease progression in patients with type 2 diabetes, according to a retrospective single-centre study conducted at Van Yüzüncü Yıl University in Türkiye. In an assessment of 382 hospitalized patients monitored over a fixed three-month horizon, radial-basis support vector machines and regularized logistic regression attained receiver operating characteristic area under the curve values of 0.682 and 0.671 respectively. However, when feature selection procedures were strictly contained within training folds during sensitivity testing to prevent information leakage, cross-validated discrimination dropped to 0.603 for the support vector machine and 0.640 for penalized logistic regression. The investigators underscored that these modest results illustrate substantial methodological optimism in unnested selection workflows, indicating that routine panel data without external prospective validation cannot currently guide clinical intervention or risk stratification.

Diabetic kidney disease represents a major microvascular complication of type 2 diabetes mellitus and remains a primary contributor to end-stage renal disease worldwide. Pathophysiologically, persistent kidney injury involves escalating albuminuria and a progressive loss of glomerular filtration driven by metabolic disruption, altered intrarenal haemodynamics, chronic systemic inflammation, and tubulointerstitial fibrosis. While established risk equations such as the Kidney Failure Risk Equation forecast advanced end-stage outcomes over two- to five-year periods in individuals with stage three to five chronic kidney disease, standard prognostic tools offer limited visibility for individuals presenting with preserved or mildly impaired filtration rates. Routine hospital inpatient evaluations routinely capture broad panels spanning complete blood counts, hepatic panels, inflammatory indicators, lipid profiles, and mineral-bone markers. The research group set out to determine whether combining these routinely collected baseline measurements through supervised machine learning algorithms could reliably forecast early three-month disease worsening across diabetic cohorts, including individuals lacking baseline nephropathy.

The study cohort comprised 382 adults with established type 2 diabetes admitted to the internal medicine inpatient service between January 2024 and December 2025. Across the entire cohort, median diabetes duration was 9 years among non-progressors and 12 years among progressors, reflecting an established rather than newly diagnosed diabetic population. Baseline diabetic kidney disease was identified in 183 participants (47.9%), whereas 199 patients had no documented nephropathy at the index assessment. Researchers excluded individuals presenting with acute kidney injury, end-stage renal disease requiring dialysis or transplantation, non-diabetic primary renal conditions, active systemic inflammatory illnesses, acute infections, malignancies, pregnancy, or chronic liver conditions. Progression was formally defined using KDIGO criteria ascertained at a scheduled three-month routine follow-up: either a decline in estimated glomerular filtration rate category accompanied by at least a 25 percent decrease from baseline, or a categorical increase in urinary albumin-to-creatinine ratio. A fixed three-month observation window was selected to align with KDIGO definitions requiring persistence of renal abnormalities, although total electronic medical record observation extended to a median of 18 months.

At the close of the three-month evaluation, 162 patients (42.4%) experienced kidney disease progression, with 100 meeting the glomerular filtration decline threshold and 62 exhibiting worsening albuminuria categories, with no individual meeting both criteria simultaneously. In descriptive baseline comparisons, patients whose condition progressed were notably older, exhibiting a median age of 60 years compared to 56 years among non-progressors, and had longer documented durations of both diabetes and hypertension. Progressors also demonstrated higher baseline median Charlson comorbidity indices, higher serum creatinine (0.90 mg/dL versus 0.80 mg/dL), blood urea nitrogen (18.00 mg/dL versus 16.00 mg/dL), urinary albumin-to-creatinine ratio (0.08 g/g versus 0.01 g/g), serum uric acid (5.70 mg/dL versus 5.20 mg/dL), haematocrit (44 percent versus 41 percent), C-reactive protein (5.00 mg/L versus 4.00 mg/L), ferritin (73.50 ng/mL versus 57.00 ng/mL), and neutrophil-to-lymphocyte ratios (2.12 versus 1.83). Baseline estimated filtration rates were significantly lower among progressors, recording a median of 80.00 mL/min/1.73m² compared to 89.50 mL/min/1.73m² in non-progressors. In contrast, baseline glycated haemoglobin, total cholesterol, erythrocyte sedimentation rate, procalcitonin, and body mass index showed no statistically significant group differences.

To model progression outcomes, investigators gathered 52 baseline demographic, anthropometric, and biochemical candidate predictors. Variables were specified a priori according to biological plausibility and clinical availability, without exploratory pre-screening. Post-baseline follow-up values, invariant assay constants, free-text annotations, and medication classes were removed to eliminate target leakage and confounding by indication. To manage the constrained ratio of 3.1 events per candidate predictor, least absolute shrinkage and selection operator (LASSO) logistic regression with internal ten-fold cross-validation was deployed to penalize uninformative weights. This regularized feature selection pruned the feature space from 52 down to 22 non-zero variables, which retained duration of diabetes, baseline urinary albumin-to-creatinine ratio, blood urea nitrogen, uric acid, haematocrit, erythrocyte sedimentation rate, C-reactive protein, procalcitonin, and ferritin. Spearman correlation analysis across selected parameters showed strong collinearity between haematocrit and haemoglobin (rho = 0.93) and between creatinine and blood urea nitrogen (rho = 0.72), alongside moderate correlations between erythrocyte sedimentation rate and haematocrit.

The study compared seven supervised machine learning architectures trained across three repeats of stratified five-fold cross-validation: LASSO logistic regression, radial-basis support vector machines, random forests, k-nearest neighbours, extreme gradient boosting (XGBoost), artificial neural networks (multilayer perceptrons), and single decision trees. Missing laboratory entries, which affected only seven values across nearly twenty thousand fields, were handled by median imputation fitted strictly within training partitions. The radial-basis support vector machine achieved the highest mean cross-validated area under the receiver operating characteristic curve at 0.682 (95 percent fold interval 0.612 to 0.775), yielding sensitivity of 0.765, specificity of 0.541, positive predictive value of 0.551, and negative predictive value of 0.758 at its descriptive Youden threshold. LASSO logistic regression yielded an area under the curve of 0.671 (95 percent fold interval 0.588 to 0.754), an area under the precision-recall curve of 0.593, sensitivity of 0.648, specificity of 0.650, positive predictive value of 0.577, and negative predictive value of 0.715. Paired bootstrap resampling between the support vector machine and LASSO models revealed an AUC difference of 0.010 with a p-value of 0.49, demonstrating no statistically significant performance separation.

Because transparent linear coefficients permit direct clinical audit, researchers prioritized LASSO logistic regression for in-depth model interpretation and benchmarking. When evaluated against a conventional seven-variable clinical comparator comprising age, sex, diabetes duration, hypertension duration, glycated haemoglobin, filtration rate, and urinary albumin-to-creatinine ratio, the full 22-variable LASSO model did not deliver a statistically significant improvement. In a matched evaluation re-estimated across identical cross-validation folds, the 22-variable LASSO achieved an AUC of 0.671 versus 0.632 for the seven-variable model, reflecting a modest paired bootstrap difference of 0.038 (95 percent confidence interval -0.005 to 0.082; p = 0.085). Furthermore, Platt probability scaling shifted the LASSO model calibration intercept from -0.28 to 0.00 and adjusted the Brier score from 0.227 to 0.225. However, the post-scaling calibration slope reached 1.39, remaining substantially above 1.0 and indicating that model probability estimates were insufficiently dispersed. Although decision-curve analysis demonstrated positive net clinical benefit over treat-all or treat-none strategies between threshold probabilities of 0.25 and 0.60, the investigators cautioned that internal decision curves cannot verify patient benefit in practice.

To interrogate model behavior, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) were deployed to track predictor contributions. Global SHAP analyses identified haematocrit, diabetes duration, total cholesterol, erythrocyte sedimentation rate, ferritin, and C-reactive protein as the dominant contributors to model outputs. Elevated haematocrit, longer diabetes duration, and elevated concentrations of cholesterol, ferritin, and C-reactive protein shifted predictions toward higher risk, whereas higher serum total protein pushed estimates downward. Counter-intuitively, elevated erythrocyte sedimentation rate contributed toward lower risk in the model, an artifact likely induced by collinearity with haematocrit and haemoglobin. In an exploratory multivariable logistic regression incorporating all 22 LASSO variables, adjusted odds ratios per one standard deviation increase were statistically significant for diabetes duration (OR 1.37, 95% CI 1.06 to 1.79; p = 0.018), total cholesterol (OR 1.35, 95% CI 1.05 to 1.72; p = 0.017), and ferritin (OR 1.34, 95% CI 1.01 to 1.76; p = 0.039). Haematocrit (OR 2.19, 95% CI 0.99 to 4.83; p = 0.053) and C-reactive protein (OR 1.26, 95% CI 0.92 to 1.72; p = 0.145) did not achieve statistical significance upon mutual adjustment.

Local explanation analyses illustrated how individual predictions operated under extreme conditions. For a correctly classified high-risk progressor whose uncalibrated apparent probability reached 0.99, elevated glycated haemoglobin provided the dominant upward contribution of approximately 6.1 log-odds, while LIME and SHAP shared seven of eight leading local features. For a correctly classified non-progressor with an apparent probability of 0.001, elevated total protein supplied a downward contribution of negative 9.2 log-odds, counteracting upward contributions from blood urea nitrogen and creatinine. However, Platt-scaled cross-validated probabilities for these identical patients were 0.48 and 0.06 respectively, illustrating how raw linear explanations can produce overly extreme apparent outputs. The authors stressed that feature attribution tools quantify internal algorithmic weighting rather than underlying biological mechanisms, proving incapable of substituting for empirical causal validation or establishing independent prognostic biomarkers.

The authors contextualized their modest predictive findings against published machine learning literature in nephrology, where several cross-sectional diagnostic studies have claimed classification accuracies exceeding 90 percent and receiver operating characteristic areas above 0.97. The authors explained that such reports rely on concurrent urinalysis markers, such as pus cells, bacteria, and heavy proteinuria, to differentiate established chronic kidney disease from healthy baselines. Detecting prevalent kidney damage using near-diagnostic contemporaneous criteria represents a fundamentally simpler task than forecasting future functional deterioration among individuals with established diabetes using non-specific blood-based routine markers. In secondary analyses within the current cohort, a random forest model separating prevalent diabetic nephropathy from diabetes without kidney disease attained an AUC of 0.770, while an ancillary separation of healthy controls from diabetic patients reached an AUC above 0.99. These auxiliary analyses confirm that cross-sectional diagnostic discrimination cannot serve as a meaningful benchmark for prospective longitudinal progression.

Multiple methodological constraints temper the clinical interpretation of the findings. Because the study was conducted retrospectively within a single tertiary medical center in eastern Türkiye, inpatient characteristics, referral patterns, and severe co-pathologies may limit direct generalization to outpatient primary-care environments. Furthermore, kidney progression was ascertained from a single follow-up measurement at three months, precluding verification of whether renal declines were sustained over years or represented transient physiological fluctuations. The investigation utilized the Modification of Diet in Renal Disease formula rather than newer CKD-EPI equations, lacked tracking of renoprotective drug adherence, and did not examine longitudinal biomarker trajectories. Above all, the significant drop in model discrimination when feature selection was nested inside cross-validation partitions highlights the risk of overfitting in modest datasets. The authors concluded that current machine learning models trained on routine laboratory panels do not warrant clinical adoption for guiding treatment intensification or specialist referral without rigorous prospective external validation.

Subject of Research: Machine learning prediction of short-term kidney disease progression in type 2 diabetes

Article Title: Explainable machine learning for predicting kidney disease progression defined by changes in estimated glomerular filtration rate in type 2 diabetes: a single-centre retrospective study

Article References: Erdem, M., & Aydemir, O. (2026). Explainable machine learning for predicting kidney disease progression defined by changes in estimated glomerular filtration rate in type 2 diabetes: a single-centre retrospective study. Frontiers in Endocrinology, 17, Article 1983789. https://doi.org/10.3389/fendo.2026.1983789

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1983789

Keywords: machine learning, diabetic kidney disease, type 2 diabetes, disease progression, eGFR, albuminuria, LASSO regression, support vector machine, SHAP, LIME, biomarkers, ferritin, haematocrit, cross-validation, predictive modeling, clinical decision support, risk stratification

News Source: Jerry Hayes. (November 14, 2026). Machine Learning Shows Modest Accuracy in Predicting Early Diabetic Kidney Disease Progression. Scienmag.

Tags: albuminuriabiomarkersClinical decision supportcross-validationDiabetic kidney diseasedisease progressionEGFRferritinhaematocritLASSO regressionLIMEMachine Learningpredictive modelingrisk stratificationSHAPsupport vector machineType 2 diabetes
Share12Tweet7Share2ShareShareShare1

Related Posts

Information-Based Prescription Review Systems Evolve from Manual Screening to Intelligent Interception

Information-Based Prescription Review Systems Evolve from Manual Screening to Intelligent Interception

November 16, 2026
Study Reveals Incidence of Rare Pediatric Hip Disorders in Japanese Children

Study Reveals Incidence of Rare Pediatric Hip Disorders in Japanese Children

November 16, 2026

Children With Nephrotic Syndrome Exhibit Severe Disruption Across Gut Microbial Kingdoms

November 16, 2026

Low-Molecular-Weight Heparin Linked to Lower Clot Odds in Blunt Trauma

November 16, 2026

POPULAR NEWS

  • Information-Based Prescription Review Systems Evolve from Manual Screening to Intelligent Interception

    Information-Based Prescription Review Systems Evolve from Manual Screening to Intelligent Interception

    29 shares
    Share 12 Tweet 7
  • Study Reveals Incidence of Rare Pediatric Hip Disorders in Japanese Children

    29 shares
    Share 12 Tweet 7
  • Children With Nephrotic Syndrome Exhibit Severe Disruption Across Gut Microbial Kingdoms

    29 shares
    Share 12 Tweet 7
  • Automated Platform Maps Genomic Hotspots in CHO Cells for Antibody Production

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Cognitive Grounding Improves Large Language Models in Simulating Human Collective Behavior

Bowl-Shaped Nanocarbon Successfully Integrates Aromatic Cyclopentadienide Ring into Geodesic Framework

Non-coding RNA networks in tumor fibroblasts drive cancer progression, review finds

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.