Abstract: PO1029
Machine Learning for Prediction of Arteriovenous Fistula Failure
Session Information
- Vascular Access Arena: Challenges, Progress, and Prospects
November 04, 2021 | Location: On-Demand, Virtual Only
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 703 Dialysis: Vascular Access
Authors
- Lama, Suman Kumar, Fresenius Medical Care, Waltham, Massachusetts, United States
- Razdan, Rishi, Azura Vascular Care, Malvern, Pennsylvania, United States
- Sor, Murat, Azura Vascular Care, Malvern, Pennsylvania, United States
- Barzel, Eyal, Azura Vascular Care, Malvern, Pennsylvania, United States
- Monaghan, Caitlin, Fresenius Medical Care, Waltham, Massachusetts, United States
- Willetts, Joanna, Fresenius Medical Care, Waltham, Massachusetts, United States
- Chaudhuri, Sheetal, Fresenius Medical Care, Waltham, Massachusetts, United States
- Mclaughlin, Nancy, Azura Vascular Care, Malvern, Pennsylvania, United States
- Zhang, Hanjie, Renal Research Institute, New York, New York, United States
- Kotanko, Peter, Renal Research Institute, New York, New York, United States
- Hymes, Jeffrey L., Fresenius Medical Care, Waltham, Massachusetts, United States
- Larkin, John W., Fresenius Medical Care, Waltham, Massachusetts, United States
- Usvyat, Len A., Fresenius Medical Care, Waltham, Massachusetts, United States
- Maddux, Franklin W., Fresenius Medical Care AG und Co KGaA, Bad Homburg, Hessen, Germany
Background
Over 23% of primary arteriovenous fistula (AVF) placements fail for patients on chronic hemodialysis (HD); interventional procedures can be performed to prevent fistulas from failing completely or to correct the malfunction of fistulas (Al-Jaishi AJKD 2014). We developed a Machine Learning (ML) model to predict the likelihood of an AVF failure within 30 days.
Methods
We used data from a cohort of HD patients with a functioning AVF treated at an integrated kidney disease company from Jan-Dec 2018 to develop an ML model (XGBoost) that predicts AVF failure within 30 days from the last use. AVF failure was defined as a status change (active to permanently/temporarily unusable), or if an interventional procedure (IP) was performed for the first time. Model used approximately 2400 variables, which included baseline and derived lab values, treatment data, clinical notes entered by physicians and dialysis staff, and care providers. The cohort was randomly split into 60% training, 20% validation, and 20% test datasets.
Results
We identified a cohort of 15,449 HD patients actively using an AVF to develop the ML model. We achieved area under the receiver operating characteristic curve as 0.76 in the test dataset. When using a 0.5 probability threshold for classifying predictions as positive or negative for AVF failure in the next 30 days, the model showed suitable performance with precision as 0.57, and recall as 0.29. Variables such as prediction score from clinical notes, days since last treatment, and days since fistula was placed had a positive relationship with the fistula failure prediction whereas comorbidity counts had a negative relationship with the failure of fistula (Figure 1).
Conclusion
Our AVF failure prediction model appears to have the potential to provide an early identification of an access that is likely to malfunction. Further evaluation and clinical testing is warranted to validate the ML model.
Funding
- Commercial Support –