Abstract: SA-PO0446
Single-Session Prediction of Arteriovenous Fistula Failure
Session Information
- Dialysis: Vascular Access
November 08, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 803 Dialysis: Vascular Access
Authors
- Bregoli, Alessandro, Alessandro Bregoli, Brescia, Italy
- Bellocchio, Francesco, Renal Research Institute, New York, New York, United States
- Maierhofer, Andreas, Fresenius Medical Care AG, Bad Homburg, HE, Germany
- Kopperschmidt, Pascal, Fresenius Medical Care AG, Bad Homburg, HE, Germany
- Usvyat, Len A., Renal Research Institute, New York, New York, United States
- Neri, Luca, Renal Research Institute, New York, New York, United States
Background
Arteriovenous fistulae (AVFs) are essential for hemodialysis in patients with chronic kidney disease, and ensuring their patency is critical. While traditional monitoring methods—such as physical examinations and technical surveillance—are valuable by highly operator dependent and time consuming, integrating AI-based models may enhance predictive capabilities. In this study, we present a novel approach to assess the risk of AVF failure using only data collected from a single dialysis session via the dialysis machine.
Methods
We extracted data from the Nephrocare European Clinical Database (EuCliD), selecting adult Portuguese dialysis patients treated between January 1, 2023, and January 31, 2024. The dataset included 523,061 dialysis sessions from 4,257 unique individuals, with 1,581 documented AVF failures in 945 patients. The model input consisted of the following four metrics derived from machine data: sum of arterial and venous pressure, clearance, clearance-reference clearance ratio, and clearance-blood flow ratio The prediction target was defined as AVF failure occurring within one month after a given dialysis session.
To predict fistula failure, we used LightGBM, a gradient boosting machine learning algorithm.
Results
We assessed model performance using the Area Under the Receiver Operating Characteristic curve (AUROC) with 3-fold cross-validation and evaluated its stratification capability across four risk categories (Table 1). The model achieved a mean AUROC of 0.74.
Conclusion
This study demonstrates the feasibility of using a machine learning model to predict AVF failure risk from a single dialysis session. Despite the limited number of input features, the model achieved promising predictive performance. These findings suggest the potential for direct integration of such models into dialysis machines, enabling more timely and effective AVF monitoring.
| Risk Class | Prevalence od Dialysis Treatments | Incidence of Failures within 1 month |
| Low (<52%) | 90.0% | 2% |
| Medium (52-69%) | 9.0% | 9% |
| High (69-74%) | 0.9% | 26% |
| Very High (>74%) | 0.1% | 60% |
Risk stratification based on predicted AVF failure probability