Time-updated explainable machine learning predicts short-term mortality in peritoneal dialysis patients

Area of Science:

• Nephrology
• Artificial Intelligence
• Data Science

Background:

• Continuous ambulatory peritoneal dialysis (CAPD) patients face significant short-term mortality risks.
• Accurate and timely risk stratification is crucial for effective patient management.

Purpose of the Study:

• To develop and validate a time-updated, explainable machine learning (ML) early-warning system for predicting 6-month mortality in CAPD patients.
• To identify key predictive factors for mortality risk.

Main Methods:

• Retrospective analysis of 1,484 CAPD patients.
• Validation of multiple supervised ML techniques, including light gradient boosting machine (lightGBM).
• Utilized time-updated clinical and laboratory data; performance assessed by AUC and accuracy; SHapley Additive exPlanation (SHAP) for interpretability.

Main Results:

• The lightGBM model achieved high performance in both internal (AUC 0.888, accuracy 0.879) and temporal validation (AUC 0.850, accuracy 0.874) cohorts.
• SHAP analysis identified key features contributing to accurate mortality risk prediction.
• The developed system is explainable and uses easily accessible, time-updated data.

Conclusions:

• A time-updated lightGBM early-warning system effectively identifies CAPD patients at high risk of 6-month mortality.
• The system provides clinicians with actionable insights for personalized treatment strategies.
• This ML approach enhances early risk detection in CAPD patients.