From algorithms to action: Risk stratification for arteriovenous fistula failure in hemodialysis

Area of Science:

• Nephrology
• Vascular Surgery
• Data Science

Background:

• Arteriovenous fistula (AVF) is the preferred hemodialysis access, but primary failure and dysfunction are common.
• Existing prediction models for AVF failure have limited accuracy.
• Machine learning (ML) offers potential for improved prediction of AVF outcomes.

Purpose of the Study:

• To identify predictors of AVF failure in a Turkish hemodialysis cohort.
• To compare the performance of ML models against logistic regression for AVF failure prediction.
• To develop a clinically applicable risk score for AVF failure.

Main Methods:

• Retrospective analysis of 385 adult patients with native AVFs.
• Extracted demographic, clinical, and laboratory variables from electronic health records.
• Fitted and evaluated logistic regression (GLM, LASSO), Random Forest, and XGBoost models using cross-validation and an independent test set.

Main Results:

• Older age, lower BMI, albumin, and creatinine were associated with AVF failure.
• ML models, especially Random Forest and XGBoost, demonstrated superior predictive performance (AUCs up to 0.950) compared to logistic regression (AUC 0.906).
• A derived clinical risk score effectively stratified patients into low, intermediate, and high-risk categories for AVF failure.

Conclusions:

• Machine learning models, particularly ensemble methods, significantly outperform logistic regression in predicting AVF failure.
• The developed clinical risk score is a simple and usable tool for stratified surveillance of hemodialysis patients.
• These findings support the integration of ML into clinical practice for optimizing vascular access management.