Predicting mortality in elderly chronic kidney disease patients using algorithmic risk assessment: a Southeast Asian...

Area of Science:

• Nephrology
• Geriatrics
• Data Science
• Artificial Intelligence

Background:

• The global burden of chronic kidney disease (CKD) in elderly patients is increasing, posing significant challenges for treatment.
• Elderly patients often experience decline after dialysis initiation, complicating treatment decisions.
• Current prognostic models have limitations in generalizability and capturing complex patient factors.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting six-month mortality in elderly patients initiating hemodialysis.
• To compare the performance of ML models against traditional prognostic tools.
• To identify key predictors of mortality using SHapley Additive exPlanations (SHAP).

Main Methods:

• Development of random forest and balanced random forest ML models.
• Utilized a single-center cohort of 1,606 elderly patients (≥65 years) with advanced CKD initiating hemodialysis.
• Compared ML model performance (ROC-AUC) with existing prognostic tools, incorporating SHAP for feature importance.

Main Results:

• The normal random forest model achieved an ROC-AUC of 0.83, marginally outperforming the balanced random forest model (0.82).
• Both ML models demonstrated superior performance and calibration compared to conventional prognostic tools.
• SHAP analysis provided feature importance rankings for model interpretability.

Conclusions:

• ML-based models offer modest improvements in predicting six-month mortality for elderly end-stage renal disease (ESRD) patients.
• These predictive tools can supplement clinical judgment in shared decision-making.
• ML models provide valuable prognostic insights to support, not replace, clinical expertise.