XGBOOST MACHINE LEARNING MODEL FOR PLANNING FOLLOW-UP ULTRASOUND TO REDUCE VASCULAR ACCESS THROMBOSIS RISK

Percutaneous transluminal angioplasty (PTA) extends lifespan of vascular accesses; however, optimal planning of next ultrasound after treatment currently relies only on empirical judgment and occasionally results in unexpected thrombosis. If risk of failure can be estimated after PTA, incidence of thrombosis is potentially reduced by proactive surveillance of vascular access and timely intervention.

XGBoost based machine learning model was developed in the R environment. A total of 3,469 PTA cases performed between June 1, 2019 and May 31, 2025 were reviewed. Patients were divided into those who experienced thrombotic occlusion during the study period (n=168) and those who did not (n=490). For thrombosis cases, the earliest thrombosis event in the period was extracted and its most recent PTA prior to event was used if it was done during the period; for non-thrombosis patients, the first PTA during the period was selected. In that way, only one record per patient was used to minimize overfitting and selection bias. The study endpoint was thrombotic occlusion occurring within 90 days after PTA. Cases not observable for 90 days after PTA due to planned re-PTA, death, transfer, or loss to follow-up were excluded, resulting in 459 eligible records (131 and 328 records from thrombosis and non-thrombosis cases, respectively).

Twenty-one clinical, ultrasonographic, and procedural variables, including balloon diameter, flow volume (FV), laboratory data, and demographics were analyzed. The XGBoost model was trained using PTAs performed on odd-numbered dates (n = 222) via five-fold cross-validation, optimized for the area under the precision-recall curve (PR-AUC). Patients were classified into high- and low-risk groups for 90-day thrombosis using a threshold that provided the highest sensitivity while maintaining a positive predictive value (PPV) above 40%. The model was externally tested using PTAs performed on even-numbered dates (n = 237), and performance metrics were recorded.

Using eight variables－implementation of thrombolysis, type of access (AVF/AVG), age, sex, presence of intradialytic hypotension, FV, balloon diameter applied to the narrowest stenosis, and diameter ratio (>1.1) compared with the adjacent normal vessel－ the model achieved a test PR-AUC of 0.459. With a probability cutoff of 0.42, accuracy, PPV, sensitivity, specificity, and AUROC were 0.882, 0.464, 0.500, 0.929, and 0.747, respectively. Kaplan–Meier survival analysis (figure below) demonstrated significantly longer thrombosis-free intervals in the low-risk group (log-rank p < 0.0001).

XGBoost machine learning model can be recognized as a potential risk-stratification tool for planning early ultrasound follow-up for those who are uncertain for thrombosis risk including recurrent attack. Despite its limited sensitivity, the model may still help guide proactive ultrasound follow-up for high-risk patients. Assuming that earlier ultrasound follow-up may help identify pre-occlusive stenosis and enable timely PTA, this tool could support directing further effort to reduce real-world thrombosis. Further validation using prospective cohorts is warranted.