Development and internal validation of an interpretable machine learning model for predicting dialysis risk in patients with stage 3-4 chronic kidney disease.

Background: Clinicians need practical tools to identify chronic kidney disease (CKD) patients at highest short-term risk of dialysis using only routine clinical data. Methods: We retrospectively analyzed 400 adults with CKD stages 3-4 treated at The Central Hospital of Wuhan (2022-2024). Incident hemodialysis during follow-up was the outcome. From 64 candidate variables, LASSO logistic regression embedded within 10-fold cross-validation selected predictors spanning renal, hematologic, and metabolic domains. Ten machine learning models were trained and evaluated using nested cross-validation; temporal validation was performed on a 2024 hold-out set. Performance was summarized as mean ± SD with 95% confidence intervals. Results: After correcting for data leakage, the Random Forest model demonstrated excellent discrimination with an AUC of 0.988 (95% CI: 0.974-1.003), accuracy of 0.965 (95% CI: 0.941-0.989), and recall of 0.970 (95% CI: 0.926-1.015). XGBoost and ANN showed comparable AUCs (0.987 and 0.985, respectively). Temporal validation yielded perfect discrimination (AUC = 1.000, recall = 1.000). Subgroup analysis showed consistent performance across sex, age, and diabetes strata. SHAP analysis identified creatinine, urine microalbumin, and eGFR as key predictors, with evidence of interaction between proteinuria and erythropoietic dysfunction. Conclusion: A model based on widely available clinical tests accurately predicts 12-month dialysis risk in stage 3-4 CKD patients. Its high performance and interpretability support potential use for early risk stratification in real-world nephrology practice, without requiring novel biomarkers or longitudinal monitoring.