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
- 1Department of Internal Medicine, Faculty of Medicine, Van Yüzüncü Yıl University
Abstract
Background and objectives: Routine clinical data may contain information on the future course of kidney disease in type 2 diabetes mellitus (T2DM). This study assessed the predictive performance and interpretability of seven machine learning (ML) algorithms applied to baseline measurements.
Materials and methods: This single-centre retrospective study included 382 patients with T2DM evaluated between January 2024 and December 2025. Progression was ascertained at a fixed three-month reassessment and defined according to KDIGO criteria as a decline in estimated glomerular filtration rate (eGFR) category with a fall of at least 25% from baseline or an increase in albuminuria category. Least absolute shrinkage and selection operator (LASSO) regression selected 22 of 52 candidate predictors before repeated stratified five-fold cross-validation, and a sensitivity analysis repeated the selection within training folds. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) were used to interpret the models.
Results: Progression occurred in 162 patients (42.4%). The area under the receiver operating characteristic curve (AUC) was 0.682 for the radial-basis support vector machine (SVM) and 0.671 for LASSO logistic regression; when feature selection was repeated within training folds, the AUCs were 0.603 and 0.640, respectively. LASSO logistic regression was retained for interpretation because its coefficients are transparent and its AUC did not differ significantly from that of SVM. At the Youden threshold, its sensitivity, specificity, positive predictive value and negative predictive value were 0.648, 0.650, 0.577 and 0.715. The best full model did not differ significantly from the best seven-variable model (ΔAUC 0.046; p = 0.052). After Platt scaling, the Brier score of the LASSO model was 0.225 and the calibration slope 1.39. Diabetes duration, total cholesterol and ferritin were associated with progression in an exploratory multivariable analysis.
Conclusions: Routine baseline data provided modest discrimination, part of which reflected optimism from feature selection performed before cross-validation. The incremental clinical value of the model has not been established, and external, preferably prospective, validation is required before clinical use.
Related
- Association of Finerenone Dose and Estimated Glomerular Filtration Rate (eGFR) Slope Changes in Patients With Diabetic Kidney Disease: A Real-World Retrospective Study.
- Non-invasive assessment of estimated glomerular filtration rate using lens autofluorescence in a population without chronic kidney disease
- A randomized trial of high versus low serum bicarbonate targets and glomerular filtration rate decline in stage 3–4 chronic kidney disease
- Molecular anatomy of the glomerular filtration barrier reveals topology-encoded albumin sieving
- Anemia and Red-Cell Size Across Estimated Glomerular Filtration Rate in a Hospital Laboratory Cohort
BibTeX
@article{erdem2026explainable,
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}},
author = {Mehmet Erdem and Ozan Aydemir},
journal = {Frontiers in Endocrinology},
year = {2026},
month = nov,
doi = {10.3389/fendo.2026.1983789},
url = {https://www.frontiersin.org/articles/10.3389/fendo.2026.1983789/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - 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
AU - Mehmet Erdem
AU - Ozan Aydemir
PY - 2026
DA - 2026/11/09
JO - Frontiers in Endocrinology
DO - 10.3389/fendo.2026.1983789
UR - https://www.frontiersin.org/articles/10.3389/fendo.2026.1983789/full
AB - Background and objectives: Routine clinical data may contain information on the future course of kidney disease in type 2 diabetes mellitus (T2DM). This study assessed the predictive performance and interpretability of seven machine learning (ML) algorithms applied to baseline measurements. Materials and methods: This single-centre retrospective study included 382 patients with T2DM evaluated between January 2024 and December 2025. Progression was ascertained at a fixed three-month reassessment and defined according to KDIGO criteria as a decline in estimated glomerular filtration rate (eGFR) category with a fall of at least 25% from baseline or an increase in albuminuria category. Least absolute shrinkage and selection operator (LASSO) regression selected 22 of 52 candidate predictors before repeated stratified five-fold cross-validation, and a sensitivity analysis repeated the selection within training folds. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) were used to interpret the models. Results: Progression occurred in 162 patients (42.4%). The area under the receiver operating characteristic curve (AUC) was 0.682 for the radial-basis support vector machine (SVM) and 0.671 for LASSO logistic regression; when feature selection was repeated within training folds, the AUCs were 0.603 and 0.640, respectively. LASSO logistic regression was retained for interpretation because its coefficients are transparent and its AUC did not differ significantly from that of SVM. At the Youden threshold, its sensitivity, specificity, positive predictive value and negative predictive value were 0.648, 0.650, 0.577 and 0.715. The best full model did not differ significantly from the best seven-variable model (ΔAUC 0.046; p = 0.052). After Platt scaling, the Brier score of the LASSO model was 0.225 and the calibration slope 1.39. Diabetes duration, total cholesterol and ferritin were associated with progression in an exploratory multivariable analysis. Conclusions: Routine baseline data provided modest discrimination, part of which reflected optimism from feature selection performed before cross-validation. The incremental clinical value of the model has not been established, and external, preferably prospective, validation is required before clinical use.
SN - 1664-2392
ER -