A Machine Learning Model for Predicting Mortality within 90 Days of Dialysis Initiation

Area of Science:

• Nephrology
• Data Science
• Biostatistics

Background:

• The initial 90 days post-dialysis initiation present significant morbidity and mortality risks for end-stage kidney disease (ESKD) patients.
• Accurate mortality prediction is crucial for informed patient-clinician shared decision-making regarding dialysis initiation or alternative medical management.

Purpose of the Study:

• To develop and validate a machine learning model for predicting 90-day mortality risk in patients initiating dialysis.
• To assess the model's performance across diverse patient subgroups.

Main Methods:

• Utilized the eXtreme Gradient Boosting (XGBoost) algorithm on a large, nationally representative US Renal Data System cohort (2008-2017).
• Included 188 predictors of early mortality known before dialysis initiation.
• Evaluated model performance using c-statistics, stratified by age, sex, race, and dialysis modality, on both complete-case and imputed datasets.

Main Results:

• Analyzed 1,150,195 ESKD patients; 8% (86,083) died within 90 days of dialysis initiation.
• XGBoost models demonstrated strong discrimination (c-statistics of 0.826-0.827) and performed well across all evaluated subgroups (c > 0.75).
• Model calibration was excellent, accurately estimating mortality probability across predicted risk ranges.

Conclusions:

• The developed XGBoost model effectively predicts early mortality risk post-dialysis initiation.
• The model exhibits excellent calibration and robust performance across key demographic and clinical subgroups.
• This tool can support clinical decision-making for patients with ESKD facing dialysis initiation.