Machine Learning to Identify Dialysis Patients at High Death Risk

Area of Science:

• Nephrology
• Data Science
• Biomedical Informatics

Background:

• High mortality rates post-dialysis initiation necessitate improved risk prediction.
• Electronic health records (EHRs) offer complex data for predictive modeling.
• Accurate short-term mortality estimation supports informed clinical and patient decision-making.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting short-term mortality in patients initiating dialysis.
• To leverage comprehensive EHR data for enhanced predictive accuracy.
• To identify patients at high risk for post-dialysis mortality.

Main Methods:

• A random forest ML algorithm was applied to a cohort of 27,615 US veterans with incident end-stage renal disease (ESRD).
• Forty-nine pre-dialysis variables from EHRs were used to predict 30-, 90-, 180-, and 365-day all-cause mortality.
• Model performance was evaluated using C-statistics and validated across diverse patient subgroups.

Main Results:

• The random forest model achieved C-statistics ranging from 0.7185 to 0.7504 across the four mortality prediction windows.
• The model demonstrated good internal validity and consistent performance in various demographic and clinical subgroups.
• Performance was comparable or superior to other ML algorithms, though generalizability to non-veterans is limited.

Conclusions:

• An ML-based approach effectively predicts short-term post-dialysis mortality in incident ESRD patients.
• These predictive models can significantly assist clinicians and patients in making informed decisions regarding dialysis initiation and management.
• The study highlights the utility of EHR data and ML in improving outcomes for ESRD patients.