TyHGB combined with routine clinical variables for identifying early diabetic kidney disease: comparison and temporal validation of eight machine-learning models
Background Early diabetic kidney disease (DKD) can be difficult to recognize while glomerular filtration remains preserved. We examined whether the triglyceride-high-density lipoprotein cholesterol-glucose-body mass index (TyHGB), combined with routine clinical information, could help identify early DKD in adults with type 2 diabetes mellitus (T2DM) and assessed the relative contribution of TyHGB within a multivariable model. Methods We retrospectively analyzed 542 adults with T2DM from a single center. Patients seen from January 2023 through December 2024 formed the development cohort ( n = 411), whereas those seen from January 2025 through June 2026 formed a subsequent temporal validation cohort ( n = 131). Eight supervised-learning algorithms were compared using nested cross-validation, with least absolute shrinkage and selection operator (LASSO) predictor selection within each outer training fold. Development performance was estimated from out-of-fold (OOF) predictions; final models were fitted using the full development cohort and evaluated without retuning in the temporal cohort. Discrimination, calibration, prediction error, and decision-curve net benefit were assessed. SHapley Additive exPlanations (SHAP) quantified predictor contributions, and restricted cubic splines (RCS) examined the TyHGB-DKD association. Results SVM achieved the highest development OOF AUROC (0.733), followed by logistic regression (LR; 0.726). In temporal validation, RF achieved the highest AUROC (0.759), vs. 0.709 for LR. The calibration intercept and slope for LR in temporal validation were −0.139 and 0.717, respectively. After Holm adjustment, none of the alternative algorithms showed a significant AUROC advantage over LR. TyHGB ranked first by mean absolute SHAP value (0.681), although the improvement in discrimination associated with adding TyHGB in the development cohort was not reproduced in temporal validation. The adjusted RCS analysis showed an overall positive association with early DKD ( P -overall <0.001), with no significant evidence of nonlinearity ( P -nonlinea r = 0.423). Across clinically relevant threshold probabilities, LR also yielded positive net benefit on decision-curve analysis. Conclusion A parsimonious LR model combining TyHGB with routine clinical variables showed moderate discrimination and maintained similar performance in temporal validation. TyHGB contributed strongly to model output, but its SHAP-based importance did not establish incremental predictive benefit, which was not confirmed in temporal validation. The model's calibration and clinical utility should be confirmed in independent multicenter populations.
Authors
- Chengyun Xia (ORCID: https://orcid.org/0009-0008-8203-6248)
- Xinyu Chen (ORCID: https://orcid.org/0009-0008-8584-3352)
- Sheng Xie (ORCID: https://orcid.org/0009-0008-2072-4889)
- Mei Wu
- Yue Feng
- Jiawei Qi
- Chuanyang Tang
- Ling Liu
Institutions
- North Sichuan Medical University (CN)
- Affiliated Hospital of North Sichuan Medical College (CN)
- Chengdu University (CN)
Publication Details
- Journal
- Frontiers in Medicine
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3389/fmed.2026.1980627
- Primary Topic
- Chronic Kidney Disease and Diabetes
- Type
- article
- Field-Weighted Citation Impact
- 0.00