Explainable machine learning characterization of routine clinical variables associated with arteriovenous fistula use in hemodialysis patients

Background For suitable patients, an arteriovenous fistula (AVF) is the favored hemodialysis vascular access for its long-term patency and lower complication rate. However, the clinical characteristics associated with prevalent AVF use, including whether their relationships are nonlinear, involve interactions, or vary across patient groups, remain incompletely understood. We aimed to use explainable machine learning to identify routine clinical variables associated with AVF use and explore their contributions to model output. Methods In this retrospective cross-sectional study, 2,106 maintenance hemodialysis patients were included. Five machine learning algorithms: Logistic Regression, Random Forest (RF), Extreme Gradient Boosting (XGBoost), Adaptive Boosting, and Support Vector Machines, were developed to separate AVF from non-AVF use. Each model used model-specific feature selection from 25 candidate variables and training-fitted preprocessing. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and three repeats of five-fold stratified cross-validation. Shapley Additive Explanations analysis was used to identify important features and quantify their contributions. Separate logistic analyses examined nonlinear terms, selected interactions, and adjusted disease-by-creatinine associations. Results Among the tested models, XGBoost and RF achieved internal-holdout AUCs of 0.832 (95%CI, 0.793–0.867) and 0.820 (95%CI, 0.780–0.859), respectively. Their mean repeated-cross-validation AUCs were 0.796 and 0.790. Serum creatinine, age, primary disease, dialysis vintage, and gender were the most important SHAP features. The holdout top-five features remained among the top five in all assessed validation folds in both models. Selected interactions improved cross-validation discrimination: the AUC difference beyond nonlinear additive terms was 0.0088 (0.0051–0.0128), while 0.0006 (-0.0069–0.0080) on the internal holdout. Calibration results varied across models. Conclusion Routine clinical variables helped distinguish prevalent AVF from non-AVF use. SHAP characterized feature contributions across models and validation folds. These findings may inform clinical assessment of AVF use and support the development of individualized vascular-access guidance.

Authors

Institutions

Publication Details

Journal
Frontiers in Medicine
Published
2026-11-18
DOI
https://doi.org/10.3389/fmed.2026.1961768
Primary Topic
Central Venous Catheters and Hemodialysis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Explainable machine learning characterization of routine clinical variables associated with arteriovenous fistula use in hemodialysis patients

Shizhi Fei, Li Cheng, Alex Zhang, Ying Huang
5 links
Frontiers in Medicine
Central Venous Catheters and Hemodialysis
article

Explainable machine learning characterization of routine clinical variables associated with arteriovenous fistula use in hemodialysis patients

Shizhi Fei, Li Cheng, Alex Zhang, Ying Huang
article en

Abstract

Background For suitable patients, an arteriovenous fistula (AVF) is the favored hemodialysis vascular access for its long-term patency and lower complication rate. However, the clinical characteristics associated with prevalent AVF use, including whether their relationships are nonlinear, involve interactions, or vary across patient groups, remain incompletely understood. We aimed to use explainable machine learning to identify routine clinical variables associated with AVF use and explore their contributions to model output. Methods In this retrospective cross-sectional study, 2,106 maintenance hemodialysis patients were included. Five machine learning algorithms: Logistic Regression, Random Forest (RF), Extreme Gradient Boosting (XGBoost), Adaptive Boosting, and Support Vector Machines, were developed to separate AVF from non-AVF use. Each model used model-specific feature selection from 25 candidate variables and training-fitted preprocessing. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and three repeats of five-fold stratified cross-validation. Shapley Additive Explanations analysis was used to identify important features and quantify their contributions. Separate logistic analyses examined nonlinear terms, selected interactions, and adjusted disease-by-creatinine associations. Results Among the tested models, XGBoost and RF achieved internal-holdout AUCs of 0.832 (95%CI, 0.793–0.867) and 0.820 (95%CI, 0.780–0.859), respectively. Their mean repeated-cross-validation AUCs were 0.796 and 0.790. Serum creatinine, age, primary disease, dialysis vintage, and gender were the most important SHAP features. The holdout top-five features remained among the top five in all assessed validation folds in both models. Selected interactions improved cross-validation discrimination: the AUC difference beyond nonlinear additive terms was 0.0088 (0.0051–0.0128), while 0.0006 (-0.0069–0.0080) on the internal holdout. Calibration results varied across models. Conclusion Routine clinical variables helped distinguish prevalent AVF from non-AVF use. SHAP characterized feature contributions across models and validation folds. These findings may inform clinical assessment of AVF use and support the development of individualized vascular-access guidance.

Frontiers in MedicineVol. 13
The Ohio State University Wexner Medical Center (US), Wuhan City Chinese Medicine Hospital (CN), Hubei 672 Orthopedics Hospital of Traditional Chinese Medicine and Western Medicine (CN), Wuhan No.1 Hospital (CN), The Ohio State University (US)
Good health and well-being
Central Venous Catheters and Hemodialysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon