Abstract: TH-PO299
Predicting Hemodialysis Arteriovenous (AV) Access Stenosis Using Artificial Intelligence
Session Information
- Vascular Access: From Biology to Managing Complications
November 03, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 703 Dialysis: Vascular Access
Authors
- Elbita, Omar Souleman Salem, University of Louisville, Louisville, Kentucky, United States
- Brier, Michael E., University of Louisville, Louisville, Kentucky, United States
- Gaweda, Adam E., University of Louisville, Louisville, Kentucky, United States
- Alsauskas, Zygimantas C., University of Louisville, Louisville, Kentucky, United States
Background
AV access flow dysfunction due to underlying stenosis or thrombosis is associated with significant morbidity and cost to the healthcare system. In this proof of concept study we have used artificial intelligence (AI) to reliably predict the occurrence of AV access stenosis using intradialytic data routinely collected and stored in the electronic medical record (EMR).
Methods
Data were obtained from the dialysis and interventional nephrology electronic medical record for the years 2018 to 2020. Routinely collected information were obtained and segmented by patient and treatment day into one-month sequences. A Recurrent Neural Network was trained to predict one week ahead of time for the following stenosis events: central venous, arterial anastomosis, venous anastomosis, or thrombosis, as obtained from interventional procedure records. The model was trained in Matlab/Simulink (MathWorks, Natick, MA) using 80% of the data for training and 20% for validation.
Results
Twenty-seven patients were included in the analysis with a total of 177 events. Two models were developed, a reduced model (Model 1) using mean systolic and diastolic blood pressure and mean venous and arterial pressure recorded during dialysis as predictors. Model 2 contained the variables in Model 1 plus the gradients of those parameters during dialysis. The ROC for the predictions are shown in Figure 1 (ROC=0.859). 75% of the events were identified one week in advance with a false positive rate of about 5%.
Conclusion
An AI-based approach using data from the EMR successfully predicted AV access stenosis. The strength of the approach is that predictions were trained using data from patients that had an event leading to discontinuation of dialysis and referral to interventional nephrology unlike procedures developed for access surveillance. This could be used as a supplementary tool to routine AV access monitoring.