Chapter 9: Artificial Intelligence (AI) in Dialysis Access

Dialysis access care in end-stage renal disease (ESRD) represents a classic complex system problem: high comorbidity burden, multiple stakeholders, pronounced practice variation, and cumulative decision-making across patient lifespans¹². Paradoxically, ESRD is also uniquely suited for data-driven approaches because patients undergo frequent observation, typically three hemodialysis sessions weekly, generating dense longitudinal datasets amenable to computational analysis and real-time clinical feedback³⁴⁵⁶.

Artificial intelligence (AI) offers four evidence-supported applications in dialysis access. First, pre-procedure planning: machine learning models trained on registry data achieve area under the receiver operating characteristic curve (AUROC) of 0.90 for predicting 1-year arteriovenous (AV) access success, substantially outperforming traditional logistic regression (AUROC 0.70) and enabling patient-centered access type and site selection within a comprehensive KDOQI ESRD Life Plan framework²⁵²⁶. Second, early post-creation assessment: point-of-care machine learning tools integrating ultrasound parameters with clinical variables (AUROC 0.78–0.81) can complement physical examination in predicting cannulation readiness and identifying patients requiring targeted salvage interventions²⁶. Third, surveillance and triage: deep learning analysis of blood flow sounds using vision transformers achieves stenosis screening performance comparable to nephrologist physical examination while enabling scalable, automated monitoring²⁷²⁹. Fourth, intradialytic safety: recurrent neural network models analyzing hemodialysis session data achieve AUROC 0.94 for real-time prediction of intradialytic hypotension, enabling preemptive prescription adjustments²⁹³⁰.

Despite momentum, systematic review confirms that most vascular access AI evidence remains retrospective, single-center, and methodologically vulnerable to bias, label noise, and limited external validation—mandating prospective evaluation, rigorous calibration, and human-in-the-loop deployment before clinical adoption³¹. Consistent with the KidneyAcademy mission, this chapter delineates where AI demonstrates implementation readiness versus where it remains investigational, and how to deploy these tools in ways that strengthen—rather than supplant—structured clinical reasoning¹⁷. This chapter functions as a living resource with linked multimedia content and pre/post self-assessment questions to reinforce applied learning.

Chapter content coming soon.