An interpretable XGBoost model for predicting arteriovenous fistula dysfunction in end stage renal disease.

Background: Arteriovenous fistula (AVF) dysfunction remains a major challenge in patients with end-stage renal disease (ESRD) undergoing hemodialysis. This study aimed to develop and evaluate an interpretable machine learning model based on clinical variables and preoperative laboratory parameters to predict AVF dysfunction. Methods: This retrospective study included patients with ESRD who underwent creation of a new autogenous AVF between January 2021 and December 2023. AVF dysfunction was defined as clinically relevant inadequate access function caused by AVF stenosis, occlusion, or thrombosis within 1 year after AVF creation. Candidate predictors included preoperative clinical variables, routine laboratory parameters, and derived composite indices. The overall cohort was randomly divided into train and test cohorts. Five machine learning models, including XGBoost, random forest, Naive Bayes, support vector machine, and logistic regression, were developed and compared. Results: Among the 696 patients, 130 (18.7%) developed AVF dysfunction within 1 year. Through recursive feature elimination-based feature selection, six predictors were selected for the final model, including hemoglobin (HB), aggregate index of systemic inflammation (AISI), dialysis vintage, calcium-phosphorus product, triglycerides, and C-reactive protein-to-albumin ratio. The XGBoost model showed favorable discrimination, with AUCs of 0.915 in the train cohort and 0.912 in the test cohort, as well as favorable overall performance in terms of area under the precision-recall curve, sensitivity, specificity, positive and negative predictive values, accuracy, and Conclusion: An interpretable XGBoost model based on six preoperative variables was developed to predict AVF dysfunction within 1 year after AVF creation in patients with ESRD.