Abstract
INTRODUCTION: The global burden of chronic kidney disease (CKD) in elderly patients is rising, presenting significant management challenges. These patients often experience rapid cognitive and physical decline following dialysis initiation, complicating treatment decisions. Given that age alone is an insufficient criterion for identifying candidates likely to benefit from dialysis, an optimal shared decision-making model should integrate clinical parameters related to comorbid conditions, as well as patient-centered factors such as individual values and treatment preferences. Traditional models, such as the REIN and Wick scores, use predefined linear relationships and are often based on Western populations, limiting their generalizability. Machine learning (ML) offers an innovative approach by identifying complex, nonlinear relationships in diverse datasets, potentially improving predictive accuracy. METHODS: This study aimed to develop ML models to predict six-month mortality in elderly patients initiating haemodialysis. We used data from a single-center cohort of 1,606 patients with advanced CKD, aged ≥65 years, who initiated haemodialysis at Changi General Hospital, Singapore, between January 2015 and October 2023. Specifically, random forest and balanced random forest models were developed and their predictive performance was compared with existing prognostic tools. Both models incorporated feature importance rankings determined by SHapley Additive exPlanations (SHAP). RESULTS: The normal random forest model marginally outperformed the balanced random forest model, achieving an area under the receiver operating characteristic curve (ROC-AUC) of 0.83 (95% CI 0.775 - 0.874) versus 0.82 (95% CI: 0.772 - 0.873). Both models demonstrated superior performance and calibration compared to conventional tools. CONCLUSION: This study demonstrates that ML-based models may offer modest improvements in predicting six-month mortality in elderly patients with end-stage renal disease (ESRD). These tools could support, but not replace, shared decision-making by providing additional prognostic insights alongside clinical judgment. CLINICAL TRIAL NUMBER: Not applicable.