Predicting radiocephalic arteriovenous fistula success with machine learning

Functional vascular access is essential for patients with end stage kidney disease (ESKD) on hemodialysis. Despite initiatives such as KDOQI that promote autogenous AVFs over prosthetic access or catheters, nearly half of AVFs are never successfully used and many patients start dialysis with a central venous catheter. Determining AVF readiness for use after creation is challenging; maturation ideally occurs over about 6 weeks but may be delayed or require intervention. Skilled physical examination is variable in availability, so ultrasound-based rules (e.g., KDOQI Rule of 6s: flow ≥600 mL/min, diameter ≥6 mm, depth ≤6 mm; UAB: flow ≥500 mL/min, diameter ≥4 mm) are used to dichotomize readiness. However, these static thresholds may discard information and were developed in heterogeneous cohorts with few forearm accesses, limiting applicability to radiocephalic AVFs. Prior prediction models (e.g., HFM study) included a minority of forearm accesses. Research question: Can machine learning using standardized 4–6 week duplex ultrasound and baseline clinical data improve prediction of successful unassisted use of radiocephalic AVFs at 1 year compared with existing ultrasound threshold criteria, and provide individualized probabilities to aid point-of-care decision-making?

Study design and data source: Post hoc analysis of pooled patient-level data from the multicenter phase III randomized controlled trials PATENCY-1 (NCT02110901) and PATENCY-2 (NCT02414841), conducted in the US and Canada (2014–2019), with prospective follow-up up to 3 years and a registry. The study drug vonapanitase had limited clinical impact at 1 year; data were used for prediction modeling. Eligibility and cohort construction: All advanced CKD patients undergoing new radiocephalic AVF creation were trial-eligible except those with life expectancy <6 months, active malignancy, or prior vonapanitase exposure. For prediction modeling, inclusion required being at risk for AVF use during follow-up (on hemodialysis) and having complete 4–6 week ultrasound data. Patients who did not progress to hemodialysis during follow-up were excluded. From 914 registry participants, 704 were eligible for hemodialysis; 591 had complete 4–6 week ultrasound data for model building (training/testing). A 70/30 random split created training (n=413) and testing (n=178) sets. Ultrasound protocol: Standardized duplex ultrasounds at 4–6 weeks and 12 weeks post-AVF creation. Outflow cephalic vein diameter: two measurements at three locations (3 cm proximal to anastomosis, mid-forearm, below antecubital fossa) averaged. Flow volume: three measurements in cephalic vein 5 cm proximal to anastomosis averaged. Stenosis: dichotomized as presence/absence of ≥50% luminal narrowing anywhere along the access. Depth not assessed. All studies interpreted by a blinded core lab (VasCore, Boston, MA). Outcome definition: Successful unassisted AVF use within 1 year, defined as 2-needle cannulation for hemodialysis for ≥90 consecutive days without a preceding intervention. Patients who did not achieve successful use by 1 year or before a terminal event (death, transplant, access abandonment, or loss to follow-up) were categorized as not having successful use. For prevalent HD patients, the 1-year window started on surgery date. For pre-dialysis patients who initiated HD later, successful use required 2-needle cannulation for all prescribed HD for 90 consecutive days starting within 6 weeks of HD initiation. Covariates: Baseline variables included age, sex, race, ethnicity, BMI, smoking status, comorbidities (e.g., diabetes, hypertension, heart failure, CAD, PAD, cerebrovascular disease), hemodialysis status at AVF creation, history of central venous catheter, CKD etiology, intraoperative baseline vein and artery diameters (after anesthesia induction), AVF location (wrist, forearm, snuffbox), anesthesia modality, anastomotic suture technique, statin use, antithrombotic use, and enrolling site volume. Ultrasound covariates from 4–6 weeks included cephalic vein diameter (mm), AVF flow volume (mL/min), and presence/absence of ≥50% stenosis. Analysis restricted to complete 4–6 week US data. Data preprocessing and missing data: Continuous variables centered and scaled. K-nearest neighbors imputation for 5 missing values (BMI n=1; intraoperative vein diameter n=2; intraoperative artery diameter n=2). Modeling approaches: Compared multiple classifiers—multivariable logistic regression (all covariates), penalized logistic regression with Lasso (variable selection) and Elastic Net, classification and regression tree (CART) with pruning, Random Forest (1,000 trees), and XGBoost (boosted trees) with logistic loss. Hyperparameters tuned using grid or maximum-entropy searches with nested 10-fold cross-validation within the training set. Variable importance assessed per model (scaled coefficients for penalized models; Gini impurity reduction for trees/random forest; information gain for XGBoost). Evaluation: After tuning, final models refit on the entire training set and evaluated on the hold-out testing set. Classification threshold set at 0.5 for reporting sensitivity, specificity, PPV, NPV, and accuracy. Discrimination assessed via AUROC and AUPRC; calibration assessed via plots and logistic calibration slope/intercept. Decision curve analysis computed net benefit across threshold probabilities. Models compared to static ultrasound thresholds approximating UAB (flow >500 mL/min and diameter >4 mm) and KDOQI (flow >600 mL/min and diameter >6 mm). Statistical analyses: Descriptive statistics with means (SD), medians [IQR], counts (%). ANOVA with Tukey post hoc for group differences in ultrasound variables; paired t-tests for repeated US; Pearson’s Chi-squared for categorical comparisons; two-tailed alpha=0.05. Implemented in R 4.0.5 with tidyverse, tidymodels, glmnet, rpart, ranger, and xgboost packages.

Cohort: Of 591 patients with complete 4–6 week ultrasound data, mean age 57 (SD 13), 22% female, 65% white; 55% on hemodialysis at AVF creation. AVF locations: wrist 75.3%, proximal forearm 22.3%, snuffbox 2.4%. Intraoperative mean vein diameter 3.37 mm (SD 0.82), artery 2.75 mm (SD 0.67). Median follow-up 719 days (IQR 458–1068). Unassisted AVF use within 1 year achieved by 277/591 (46.8%). Ultrasound associations: At 4–6 weeks, unassisted-use group had higher flow and larger diameter vs no-use group—flow mean difference 250 mL/min (95% CI 175–326); diameter mean difference 0.69 mm (95% CI 0.48–0.90). At 12 weeks, flow mean difference 235 mL/min (95% CI 136–334); diameter mean difference 0.95 mm (95% CI 0.70–1.2). From 4–6 to 12 weeks, flow increased by 67.8 mL/min (95% CI 39.5–96.1) and diameter by 0.43 mm (95% CI 0.37–0.49). Meeting UAB and KDOQI criteria at 4–6 weeks was more common among those with unassisted use (UAB χ²[2]=53.0, p<0.001; KDOQI χ²[2]=31.6, p<0.001). Notably, among those with successful unassisted use, 73% (4–6 weeks) and 52% (12 weeks) did not meet KDOQI criteria. Model performance (testing set): Penalized/logistic and ensemble models had similar discrimination (AUROC ≈0.78–0.81) and accuracies ≈69–74%, outperforming UAB and KDOQI thresholds. Table 3 highlights: Lasso AUROC 0.794, AUPRC 0.719, accuracy 72.5%, sensitivity 66.7%, specificity 77.0%, PPV 69.3%, NPV 74.8%; Elastic Net AUROC 0.807, AUPRC 0.737, accuracy 71.3%; Logistic Regression AUROC 0.786, accuracy 73.6%; Random Forest AUROC 0.791, accuracy 69.1%; Boosted Trees AUROC 0.779, accuracy 70.2%; Pruned Tree AUROC 0.730, accuracy 66.9%. Threshold strategies: UAB accuracy 65.2% (sensitivity 78.2%, specificity 55.0%); KDOQI accuracy 61.8% (sensitivity 23.1%, specificity 92.0%). Final model and predictors: Lasso selected 3 ultrasound predictors measured at 4–6 weeks: larger outflow vein diameter (per mm, OR 1.95; 95% CI 1.48–2.60), higher flow volume (per 100 mL/min, OR 1.08; 95% CI 1.00–1.17), and absence of ≥50% stenosis (OR 2.74; 95% CI 1.65–4.60). No pre- or intraoperative clinical variables were retained. Calibration acceptable for all but pruned tree; model-based strategies showed higher net benefit than UAB/KDOQI across plausible thresholds. A point-of-care calculator was deployed online (https://patrickheindel.shinyapps.io/predict-avf/).

The study demonstrates that machine learning models using standardized 4–6 week duplex ultrasound measurements can accurately and more usefully predict 1-year successful unassisted use of radiocephalic AVFs than static ultrasound threshold criteria. By providing calibrated individual probabilities rather than binary classifications, the tool supports nuanced clinical decision-making regarding catheter dependence, timing of cannulation, and need for interventions. Among multiple modeling strategies, the Lasso model balanced interpretability and performance, achieving discrimination comparable to more complex ensemble methods while retaining only three ultrasound-based predictors (flow, diameter, and ≥50% stenosis). These findings align with and extend prior work from the HFM study, confirming the primacy of flow and diameter as predictors of AVF use, while showing that adding numerous baseline clinical variables did not improve prediction. Decision curve analysis indicates higher net benefit of the model-based approach over UAB and KDOQI across a range of thresholds, suggesting clinical advantage without additional testing burden. The outputs can be tailored to clinician or patient preferences by selecting thresholds that emphasize sensitivity or specificity according to the clinical context, within a shared decision-making framework.

Using prospectively collected multicenter RCT data, the authors developed and validated a practical, point-of-care prediction tool (PREDICT-AVF) for estimating individual probability of successful unassisted radiocephalic AVF use at 1 year. A parsimonious Lasso model using three 4–6 week ultrasound variables (vein diameter, flow volume, and absence of ≥50% stenosis) outperformed existing UAB and KDOQI threshold criteria in discrimination and decision net benefit, with acceptable calibration. The tool is accessible via an online calculator, cross-table, and nomogram to facilitate clinical use. Future work should evaluate clinical integration strategies, assess impact on decision-making and patient outcomes, perform external validation in diverse geographic and practice settings, incorporate additional ultrasound parameters (e.g., depth), and extend modeling to other AV access configurations.

• Prediction-focused analysis without a causal framework; individual predictor effects should not be interpreted causally.
• Competing events (loss to follow-up, death, transplant) treated as non-events; the model predicts being observed with successful use rather than use itself.
• Ultrasound protocol did not capture access depth, necessitating approximations of KDOQI criteria using only flow and diameter.
• Conducted in North American centers; cannulation practices vary internationally, limiting generalizability.
• Unclear how specific predicted probabilities should guide interventions; clinical decision pathways require further study.