Risk factors for early arteriovenous fistula failure in patients undergoing hemodialysis and comparison of diagnostic prediction models.

OBJECTIVE

The aim of this study was to investigate the incidence of early arteriovenous fistula (AVF) failure and identify associated factors using multiple predictive models.

METHODS

A total of 1010 consecutive patients who underwent AVF creation at our institute between 2008 and 2020 were included. Clinical data were retrospectively reviewed, and patients were randomly divided into two training and validation cohorts. The study focused on two main investigations: the cumulative incidence of early AVF failure and the comparison of predictive models, including logistic regression and machine learning techniques such as random forest (RF), support vector machine (SVM), and eXtreme gradient boosting (XGBoost).

RESULTS

Early failure was observed in 46 patients (6.5%), whereas 92 patients (12.8%) experienced AVF failure within 1 year, with early failure accounting for 50% of cases (46/92). The area under the curve (AUC) for logistic regression and RF was 0.802 and 0.828, respectively, whereas SVM and XGBoost both had an AUC of 0.781. Although all models demonstrated high diagnostic performance, RF achieved the highest accuracy. Surgical time was the most important factor for early AVF failure in RF, with a surgical duration of ≥149 minutes significantly increasing the risk.

CONCLUSIONS

Preventing early AVF failure may enhance the 1-year patency rate following AVF creation. Machine learning, especially RF, has proved to be a reliable and accurate method for predicting risk factors, performing comparably to logistic regression. Surgeons should recognize these associated factors and optimize surgical planning to improve AVF outcomes.