Method for Predict Stenosis of Arteriovenous Fistula Patients Based on Machine Learning

Yunghsin Chen; Wei-Tse Hsu; Christopher Chen; Wei-Ta Chen

doi:10.1111/sdi.70001

Method for Predict Stenosis of Arteriovenous Fistula Patients Based on Machine Learning

Semin Dial. 2025 Jul 1. doi: 10.1111/sdi.70001. Online ahead of print.

Authors

Yunghsin Chen¹, Wei-Tse Hsu², Christopher Chen¹, Wei-Ta Chen^{3

4

5}

Affiliations

¹ Above Care Inc., San Jose, California, USA.
² Sydney School of Veterinary Science, Faculty of Science, University of Sydney, Sydney, New South Wales, Australia.
³ Division of Cardiovascular Medicine, Department of Internal Medicine, Taipei Medical University Hospital, Taipei, Taiwan (ROC).
⁴ Division of Cardiovascular Medicine, Department of Internal Medicine, School of Medicine, College of Medicine, Taipei Medical University, Taipei, Taiwan (ROC).
⁵ Taipei Heart Institute, Taipei Medical University, Taipei, Taiwan (ROC).

PMID: 40598855
DOI: 10.1111/sdi.70001

Abstract

Objectives: Arteriovenous fistula (AVF) is the most ideal vascular access for hemodialysis. People with AVF have a longer vascular access survival rate and a lower complication rate. Thrombosis and stenosis are the most common complications of AVF. The annual thrombosis event rate is 10%-50%. Appropriate identification of AVF stenosis and management could reduce the risk of thrombosis and access loss. Guidelines recommended physical examinations as the first line of AVF stenosis monitoring. However, even for health professionals, the diagnosis rate by hearing the bruit varied. The sound waves of AVF can be recorded by electronic stethoscopes and the analysis of the digitalized signal may help predict stenosis of AVF and trigger the next step of management.

Methods: From January 1, 2019, to December 31, 2019, all dialysis patients with AVF referred to our angiography laboratory for AVF angiography were enrolled. Significant stenosis was defined as stenosis severity > 70% on angiography. The stenosis severities were measured before and after the angioplasty. Before and after the angioplasty/angiography, the sounds of AVF were digitally recorded by an electrical stethoscope. Two sections longer than 10 s were obtained at different sites for each recording. Seventy percent of all the data was used to train the machine learning algorithm. The other 30% was used for testing. For the output of the algorithm, the AVF stenosis severity was classified into significant stenosis or non-significant stenosis.

Results: One hundred ninety-nine patients were enrolled. Ninety-six patients were with significant stenotic AVF and the other 103 patients were with insignificant stenosis. One hundred eighty-nine recording sections for significant stenosis and 511 recording sections for insignificant stenosis were obtained. The machine learning artificial intelligence can classify the input sound waves as significant or insignificant stenosis with a 94.1% sensitivity rate and an 81.7% specificity rate.

Conclusions: Artificial intelligence can help predict AVF stenosis by analyzing the digitalized sound waves of AVF. This analysis is convenient and non-invasive. Moreover, this technique can help the development of a remote monitor of AVF stenosis, which is especially important in the era of the COVID-19 pandemic.