Development and Validation of Machine Learning Models for Predicting Initiation of Emergency Dialysis in Advanced Chronic Kidney Disease

Background: Initiation of emergency dialysis, often requiring temporary catheter owing to unprepared definitive vascular access, is associated with infectious and vascular complications and suggests advanced chronic kidney disease (CKD) care gaps. Previous studies focused on kidney failure or dialysis timing. This study aimed to predict initiation of emergency dialysis using machine learning and baseline data. Methods: This retrospective cohort study used the Japan Medical Data Center claims data (2014-2022). Adults with an estimated glomerular filtration rate (eGFR) <15 mL/min/1.73 m2 were included. The primary outcome was initiation of emergency dialysis (temporary catheter code without evidence of previous access preparation). Participants were randomly divided into derivation (80%) and validation (20%) cohorts. Logistic regression, support vector machine, XGBoost, LightGBM, and random forest models were evaluated using internal cross-validation, post-hoc calibration of the selected model, and bootstrap confidence intervals. Results: The cohort included 3,062 individuals (derivation; n=2,449, validation; n=613). Emergency dialysis was initiated in 237 participants (7.7%); 185 (7.6%) and 52 (8.5%) in the derivation and validation cohorts, respectively. Validation area under the receiver operating characteristic curve ranged from 0.781-0.799, with the highest value observed for random forest (0.799, 95% confidence interval; 0.740-0.850). Risk stratification showed clear event enrichment in higher predicted risk categories. SHAP analyses identified hemoglobin, proteinuria, baseline eGFR, diabetes history, and diuretic use as key predictors. Decision curve analysis showed greater net benefit than eGFR alone at lower threshold probabilities. Conclusions: Baseline machine learning models showed moderate discrimination for initiation of emergency dialysis and identified clinically plausible predictors. These findings support potential use for risk stratification, although external validation and evaluation within pre-specified care pathways are needed before implementation.