Plasma proteomics based on machine learning predict the early risk of kidney outcomes in patients with DKD: a prospective cohort study from UK Biobank

Background

Diabetic kidney disease (DKD) is the predominant cause of end-stage kidney disease worldwide. Early identification of patients at heightened risk for adverse kidney outcomes remains a critical unmet clinical need.

Methods

We conducted a prospective cohort study of 918 DKD participants from the UK Biobank. Baseline plasma proteomic profiling quantified 1,463 proteins using Olink proximity extension assays. Three nested Cox proportional hazards models with incremental covariate adjustment identified proteins associated with composite kidney outcomes. To prevent information leakage, all feature selection, hyperparameter optimization, and model development were performed exclusively within the training cohort (70% random split) via a multi-stage pipeline integrating LASSO-Cox regression, Random Survival Forest, Boruta algorithm, and Sequential Forward Selection. XGBoost Cox modeling with SHAP interpretability analysis quantified variable contributions. Predictive performance was validated through Kaplan–Meier survival analysis, 15-year longitudinal trajectory modeling, ROC benchmarking, 10-fold nested cross-validation, and sensitivity analyses restricted to KDIGO-defined DKD. An interactive web application was developed for clinical translation.

Results

Of 1,463 proteins examined, 633 demonstrated significant associations with kidney outcomes across all three Cox models. Functional enrichment highlighted immune-inflammatory pathways, PI3K-Akt signaling, and chemokine cascades as central to DKD progression. The machine learning framework identified an 11-protein signature, which was refined to 10 core biomarkers (HLA-E, EFNA1, GPR158, FSTL3, ART3, GM2A, CLEC1A, CKAP4, IFNGR1, EPHA2) following survival and longitudinal trajectory validation. The protein-only model achieved robust discrimination (AUC = 0.808 [95% CI 0.767–0.848] for composite outcomes; AUC = 0.807 [95% CI 0.765–0.848] for renal death), comparable to fully integrated models incorporating demographics and metabolic variables. Systematic benchmarking across 101 algorithm combinations identified LASSO-RSF as optimal (C-index = 0.94 in training; 0.73 in independent testing), with reliable calibration through mid-term follow-up (Integrated Calibration Index = 0.019–0.022). The web-based tool enables real-time, personalized risk stratification.

Conclusions

This study establishes a validated 10-protein signature for early prediction of kidney outcomes in DKD. The systematic machine learning framework and deployable web application provide accessible, interpretable risk assessment to support precision nephrology and preemptive clinical intervention.