Remnant Cholesterol for Diabetic Kidney Disease Risk Stratification in Type 2 Diabetes: A Machine Learning-Based Prevention Tool

Yuehong Dai¹Qi Pan²Yujie Yu³Yongjun Ma¹Guangming Chen¹Huabin Wang^1*

• ¹Affiliated Jinhua Hospital, Zhejiang University School of Medicine, Jinhua, China
• ²Nottingham Ningbo China Beacons of Excellence Research and Innovation Institute, Ningbo, China
• ³Wenzhou Medical University, Wenzhou, China

Background: Disturbances in lipid metabolism play a critical role in the onset and progression of diabetic kidney disease (DKD). Remnant cholesterol (RC), a marker of remnant lipoprotein metabolism, is an established cardiovascular residual risk factor. However, evidence linking RC to the risk of incident DKD is limited. This study aimed to investigate the association between RC and incident DKD and to develop a risk prediction model incorporating RC and other clinical variables in patients with type 2 diabetes (T2D). Methods: A retrospective cohort study of 2,122 patients with T2D and without baseline DKD was conducted. The association between RC and DKD risk was examined using multivariable Cox regression and restricted cubic spline (RCS) analysis. A random survival forest (RSF) algorithm was applied to identify potential predictors, followed by multicollinearity assessment. A RSF-based prediction model was developed and evaluated for discrimination, calibration, and clinical utility. Results: During a median follow-up of 4.22 years, 435 participants (20.5%) developed DKD. Higher RC quartiles were associated with an increased risk of DKD across all models; however, the hazard ratios for Q2 to Q4 were numerically similar, indicating the absence of a clear linear dose–response pattern. RCS analysis revealed a nonlinear association between RC and DKD risk (P for nonlinearity = 0.031), characterized by a steep initial increase followed by a plateau at higher RC levels. RSF identified 14 predictors (including ACR, RC) with no significant multicollinearity ( all the variance inflation factors < 3). The model exhibited strong discrimination (3-year AUC=0.86, 5-year AUC=0.91) and calibration (3-year mean absolute error=0.011, 5-year mean absolute error=0.026), and outperformed “treat-all”/“treat-none” strategies in decision curve analysis. Conclusion: RC was independently and nonlinearly associated with DKD risk in T2D. The RSF model demonstrated good predictive performance and may assist individualized risk assessment and management.