Abstract: FR-PO523
Arteriovenous Fistula Failure Prediction Using Single Treatment Information
Session Information
- Dialysis Vascular Access
October 25, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 803 Dialysis: Vascular Access
Authors
- Bregoli, Alessandro, Universita degli Studi di Milano-Bicocca, Milano, Italy
- Bellocchio, Francesco, Fresenius Medical Care Italia SpA, Palazzo Pignano, Lombardia, Italy
- Maierhofer, Andreas, Fresenius Medical Care AG, Bad Homburg, Hessen, Germany
- Hymes, Jeffrey L., Fresenius Medical Care Holdings Inc, Waltham, Massachusetts, United States
- Usvyat, Len A., Fresenius Medical Care Holdings Inc, Waltham, Massachusetts, United States
- Neri, Luca, Fresenius Medical Care Italia SpA, Palazzo Pignano, Lombardia, Italy
- Stella, Fabio, Universita degli Studi di Milano-Bicocca, Milano, Italy
Background
Predicting the natural course of AVF complications with technical surveillance has been an elusive task. We developed a machine learning algorithm assessing the risk of failure of AVF using information collected by the dialysis machine in a single treatment.
Methods
We included all patients undergoing hemodialysis with AVF from the clinics belonging to Nephrocare France between 2021 and 2022. We extracted data from the European Clinical Database (EuCliD), and we combine it with the data sampled every minute from the sensors of the dialysis machine. We developed a model combining a continuous time naive Bayes and a logistic regression to access the risk of failure of a fistula within a month using as input variables the number of previous failures and the machine data of a single dialysis session such as: venous pressure, arterial pressure, blood flow and clearance.
Results
We included 171,969 hemodialysis sessions performed with an AVF over 1,845 French patients in the period from June 1, 2021, to May 31, 2022. During this period, 164 AVF failures were observed among 90 distinct patients.
We evaluated the performances of the classifiers calculating the Area Under the Curve (AUC) over a 5-fold cross validation. Figure 1 shows that the classifier combining machine data and number of previous failures achieves the best performance. This result's statistical significance was accessed with a Wilcoxon signed-rank test (p<0.1).
Conclusion
Considering the performances obtained, the simplicity of the model and the prevalent use of machine data, this might suggest that it would be possible to install such a model directly on dialysis machines to have a timely assessment of the risk of failure at each treatment without performing any specific time-consuming test.
Performance obtained in cross validation by the 3 classifiers