Predicting Recurrent Interventions After Radiocephalic Arteriovenous Fistula Creation with Machine Learning and the PREDICT-AVF Web App

Abstract

Radiocephalic arteriovenous fistulas (AVFs) are often criticized for requiring numerous recurrent interventions to achieve and maintain hemodialysis access function. Recent Kidney Disease Outcomes Quality Initiative guidelines discourage ongoing access salvage attempts after two interventions prior to successful use or more than three interventions per year overall. This study's goal was to develop a useful prediction tool for preoperative and postoperative assessment of intervention requirements to help guide decision-making about access appropriateness and futility. The data source was patient-level data from the international multicenter PATENCY-1 and PATENCY-2 randomized controlled trials, which enrolled patients undergoing new radiocephalic AVF creation. Routine ultrasounds were performed at 4 to 6 weeks and 12 weeks from AVF creation and interpreted by a core laboratory. The outcome for the prediction models was the number of recurrent surgical or endovascular interventions required at 1 year. Cox, random survival forest, pooled logistic, and penalized pooled logistic (elastic net) recurrent event survival prediction models were built using a combination of baseline characteristics and postoperative ultrasound measurements. Models were trained with 10-fold cross-validation hyperparameter selection within a randomly split training dataset. Discrimination and calibration performance was assessed on a holdout testing dataset. An interactive web application was built, which generates patient-specific predictions at 1 year, draws cumulative hazard curves, and contextualizes the predictions with the Kidney Disease Outcomes Quality Initiative guidelines for recurrent interventions after access creation. The cohort included 914 patients; mean age was 57 ± 13 years and 22% were female. Radiocephalic AVFs were created at the snuffbox (2%), wrist (74%), or proximal forearm (24%). Patients underwent a median of 1.04 (95% confidence interval [CI], 0.95-1.13) interventions in the first year. Using only baseline characteristics, the random survival forest model performed best, with a C-statistic of 0.72 (95% CI, 0.65-0.78) at 1 year (Table). The addition of ultrasound information to baseline characteristics did not substantially improve performance; however, Cox models using either 4- to 6-week or 12-week postoperative ultrasound information alone had the best discrimination performance, with C-statistics of 0.74 (95% CI, 0.67-0.81) and 0.78 (95% CI, 0.72-0.84) at 1 year (Table). The interactive web application is deployed at PREDICT-AVF.com (Figure). When assessing candidates for radiocephalic AVF in the preoperative setting, baseline patient characteristics can predict future interventions with moderate performance. In the postoperative setting, ultrasound findings alone are sufficient to predict future interventions. A free point-of-care web application can guide patient counseling and guideline-concordant surgeon decision-making as part of a patient-centered end-stage kidney disease life plan.TableDiscrimination performance of each model and feature set, measured by the area under the receiver operating characteristic curve (C-statistic) at 1 year postoperativelyFeaturesCoxRandom survival ForestPooled logistic regressionElastic netBaseline0.704 (0.639-0.769)0.715 (0.647-0.783)0.696 (0.631-0.762)0.703 (0.638-0.769)Baseline + 4- to 6-week US0.704 (0.629-0.779)0.737 (0.666-0.808)0.702 (0.627-0.778)0.703 (0.627-0.778)Baseline + 12 week US0.737 (0.670-0.803)0.766 (0.701-0.832)0.730 (0.663-0.797)0.746 (0.680-0.812)4- to 6-Week US0.743 (0.674-0.813)0.696 (0.620-0.773)0.740 (0.670-0.810)0.742 (0.672-0.812)12-Week US0.782 (0.722-0.842)0.721 (0.652-0.791)0.774 (0.712-0.835)0.775 (0.714-0.837)US, Ultrasound.Perfect discrimination corresponds to an area under the receiver operating characteristic curve of 1. Feature sets included baseline (age, sex, race, smoking status, diabetes, heart failure, renal transplant history, anticoagulation use, antiplatelet use, statin use, hemodialysis status at the time of arteriovenous fistula creation, central venous catheter history, number of prior failed accesses, vein and artery diameter, anastomotic suture technique, access location, and access side), baseline with the addition of postoperative ultrasound measurements (flow volume, access diameter, and luminal stenosis), and ultrasound measurements alone. Open table in a new tab