Predicting and explaining poor prognosis in diabetic kidney disease using SHAP-based interpretable machine learning

Prognostic assessment of diabetic kidney disease (DKD) is essential for personalized management. This study developed eight machine learning models using data from 180 biopsy-proven patients with DKD to predict a composite endpoint of all-cause mortality, dialysis initiation, or renal transplantation. Internally, the Naive Bayes (NB) model achieved the highest accuracy of 82.3%, while the logistic regression (LR), support vector machine (SVM), and NB models shared the highest AUC of 0.788. An independent external validation confirmed robust generalizability, yielding an AUC of 0.834. SHAP analysis identified eGFR, serum albumin, C3, serum creatinine, and urinary red blood cell count (URBC) as the most impactful features. Feature stability was confirmed via a "leave-top1-out" sensitivity analysis. The models highlighted the predictive value of C3 and URBC by capturing non-linear patterns often missed by traditional linear methods, providing granular insights for personalized prognosis evaluation.