Enhanced Dialysis Access Patency Using Machine Learning

Abstract

Arteriovenous fistulas (AVFs) and grafts (AVGs) are associated with an overall patency of 64% to 78% at 1 year and typically require multiple primary and secondary interventions to achieve those results. Improved surveillance, routine ultrasound, medical management, and better care coordination can improve patency but require greater resources and complex process improvements. The purpose of this paper was to describe an innovative data-driven approach to significantly improve access patency through machine learning. Patients who had creation of an AVF or AVG for end-stage renal disease between May 2016 and March 2019 were monitored as part of an enhanced surveillance program intended to improve access-related patency. Patient demographics, hospital covariates, and procedural details were collected. Flow velocities were measured on a weekly basis, and ultrasound velocities were compiled on a 3-month basis during clinic visits. All data were collected within the purpose-built QuartzClinical (Chicago, Ill) data registry prospectively. A machine learning model was used to assist with clinical decision-making on access-related interventions from May 2016 through March 2018. A total of 68 AVFs and 44 AVGs were monitored by the machine learning algorithm. The algorithm recommended interventions on 29 AVFs (42.6%) and 10 AVGs (22.7%) during a 1-year period. The primary assisted patency was 77.9% for AVFs and 81.8% for AVGs. There were four secondary interventions for AVFs and three secondary interventions for AVGs. The overall patency was 83.8% for AVFs (P < .001 vs baseline) and 88.6% for AVGs (P < .01 vs baseline). The rate of interventions was significantly less after using the machine learning algorithm to guide treatment (observed-expected ratio = 0.76 [P < .00]1; observed-expected ratio = 0.72 [P < .01]). Through the decrease in number of procedures, avoidance of central venous catheters for access failure, and decreased rate of complications from secondary interventions, the total cost of care was decreased by $242,859. Machine learning offers the potential to more precisely guide clinical care in complex cases. A data-driven approach combined with sophisticated predictive analytics can reduce complications and failure rates in dialysis access while significantly decreasing the cost of care.