Machine Learning and Artificial Intelligence Approaches for Predicting Arteriovenous Fistula Dysfunction in Hemodialysis: A Systematic Review.

Arteriovenous fistula (AVF) is the preferred vascular access for hemodialysis patients; however, AVF dysfunction remains a common complication that compromises dialysis adequacy and patient outcomes. Traditional risk prediction methods have limited ability to capture complex, multifactorial interactions. Machine learning (ML) and artificial intelligence (AI) offer promising approaches for enhancing predictive accuracy. This systematic review aims to critically synthesize current evidence on AI and ML approaches for predicting AVF dysfunction in hemodialysis patients. A systematic literature search was conducted in PubMed, Embase, Scopus, and Web of Science for original peer-reviewed studies published in English between January 2021 and December 2025. Studies were included if they applied AI or ML techniques to predict AVF dysfunction (including stenosis, thrombosis, occlusion, patency failure, or access dysfunction) in adult hemodialysis patients. Data extraction covered study characteristics, AI/ML models, input variables, validation methods, and performance metrics. Risk of bias was assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST). A narrative synthesis was performed due to substantial methodological heterogeneity. Ten studies met the inclusion criteria. Most studies were retrospective cohort designs, originating predominantly from China and the USA, with sample sizes ranging from 150 to nearly 60,000 patients. Predicted outcomes included thrombosis, stenosis, occlusion, and patency failure. Ensemble tree-based models (random forest, XGBoost, and LightGBM) consistently outperformed conventional statistical and regression-based approaches, including logistic regression and Cox proportional hazards models, achieving AUC-ROC values between 0.80 and 0.98. Key predictors included prior surgeries, inflammatory markers, imaging parameters, and, in one study, acoustic features from AVF sounds. PROBAST assessment indicated low risk of bias for eight studies and some concerns for two studies, primarily related to incomplete reporting of calibration or sample size. AI and ML models, particularly ensemble tree-based methods, demonstrate good to excellent discrimination for predicting AVF dysfunction in hemodialysis patients. However, external and prospective validation remain lacking, and heterogeneity in outcome definitions limits direct comparisons. Future research should focus on externally validated, clinically implementable models with standardized reporting of calibration and decision curve analysis.