Predicting radiocephalic arteriovenous fistula success with machine learning

Hundreds of thousands of patients in the United States with end-stage kidney disease (ESKD) require functional vascular access for chronic intermittent hemodialysis. Efforts to promote autogenous hemodialysis access, like the National Kidney Foundation's Kidney Disease Outcomes Quality Initiative (KDOQI), have shifted towards arteriovenous fistulae (AVFs). However, a significant proportion of AVFs are never used successfully, often leading to reliance on central venous catheters (CVCs). Determining access readiness for use is challenging, even with the augmentation of physical exams with ultrasound. Existing criteria, like the KDOQI and University of Alabama at Birmingham (UAB) criteria, utilize static thresholds, potentially losing valuable information on access maturation. This study aims to develop a more nuanced prediction tool using machine learning to improve the prediction of successful radiocephalic AVF use, thereby guiding clinical decisions related to CVC use, interventions, and hemodialysis initiation.

Previous studies, such as the Hemodialysis Fistula Maturation (HFM) study, have explored predicting AVF use. However, these studies often included a minority of forearm accesses and utilized smaller, heterogeneous cohorts. The existing KDOQI and UAB ultrasound threshold criteria, while valuable, are based on smaller datasets and might not be optimal for radiocephalic AVFs. This study addresses these limitations by focusing exclusively on radiocephalic AVFs and utilizing a larger, more homogenous cohort from the PATENCY-1 and PATENCY-2 randomized controlled trials.

This study performed a post-hoc analysis of pooled patient-level data from the PATENCY-1 and PATENCY-2 randomized controlled trials (2014-2019), which included 704 patients undergoing new radiocephalic AVF creation. The primary outcome was successful unassisted AVF use within one year, defined as ≥90 days of two-needle cannulation without intervention. Several machine learning models were built and tested using a training, tuning, and testing paradigm. These included logistic regression, penalized logistic regression (lasso and elastic net), decision tree, random forest, and boosted tree models. The models used a combination of baseline clinical characteristics (age, sex, race, comorbidities, etc.) and 4-6 week ultrasound parameters (outflow vein diameter, flow volume, stenosis). Model performance was assessed using receiver operating characteristic (ROC) curves, precision-recall curves, calibration plots, and decision curves. Missing data was handled using K-nearest neighbors imputation. The Lasso model was chosen as the optimal prediction model due to its combination of performance and parsimony.

The Lasso model demonstrated excellent discrimination performance (AUROC 0.794, AUPRC 0.719, accuracy 72.5%), outperforming both the UAB and KDOQI criteria. The model retained only three predictors: larger outflow vein diameter (per mm, odds ratio [OR] 1.95), higher flow volume (per 100 mL/min, OR 1.08), and the absence of >50% luminal stenosis (OR 2.74). No preoperative or intraoperative characteristics were retained. All modeling approaches except the decision tree showed similar discrimination performance and comparable net benefit (AUROC 0.78–0.81, accuracy 69.1–73.6%). The model's superior performance compared to existing criteria highlights the potential for improved clinical decision-making using this machine learning approach. Analysis of ultrasound parameters showed that flow volume and cephalic vein diameter were significantly different between the three AVF use categories at both 4-6 and 12 weeks. Patients with successful unassisted AVF use were more likely to meet UAB and KDOQI criteria.

This study successfully developed and validated a machine learning model to predict successful unassisted radiocephalic AVF use. The model's superior performance over existing criteria, coupled with its simplicity (only three ultrasound parameters), makes it a practical tool for point-of-care applications. The availability of an online calculator further enhances its accessibility and usability. The findings underscore the potential of machine learning to improve the accuracy of AVF readiness assessments, potentially leading to better patient outcomes and reduced reliance on CVCs. The model's focus on readily available ultrasound parameters strengthens its clinical applicability and avoids the need for additional testing.

This study demonstrates the effectiveness of a machine learning-based prediction model for successful unassisted radiocephalic AVF use. The Lasso model, incorporating three easily obtainable ultrasound parameters, surpasses existing criteria in predictive accuracy. An accessible online calculator makes this model readily applicable in clinical practice, potentially improving patient outcomes and reducing complications associated with AVF maturation. Future research should investigate the model’s performance in diverse populations and explore strategies to integrate predicted probabilities into clinical decision-making.

The study's analysis was purely predictive, not causal. Competing events (death, transplant, loss to follow-up) were treated as non-events, meaning the model predicts the probability of *observed* successful AVF use, not AVF use itself. The lack of AVF depth information in the ultrasound data limited the complete approximation of the KDOQI criteria. The study population was limited to North American patients, and caution should be exercised when generalizing the findings to other settings. Finally, the clinical implications of the predicted probabilities require further investigation.