Brigham Research Institute Poster Session Site logo-1
Search
Close this search box.

Christopher Holden-Wingate

Pronouns

He/Him/His

Job Title

Clinical Research Assistant

Academic Rank

Department

Surgery

Authors

C. Holden-Wingate, P. Heindel, A. Hart, J. Feliz, E. Rouanet, N. Rydberg, R. Chamberlain, M. Hussain, D. Hentschel, C.K. Ozaki

Principal Investigator

C.K. Ozaki

Research Category: Cardiovascular, Diabetes, and Metabolic Disorders

Tags

Auscultatory Machine Learning for Early Detection of Hemodialysis Access Dysfunction

Scientific Abstract

Background: Most end stage kidney disease (ESKD) patients require hemodialysis through a surgically created arteriovenous fistula (AVF) or arteriovenous prosthetic graft (AVG). Early identification of access dysfunction can prevent costly and dangerous access failures; however, effective access monitoring requires substantial skill and time. We hypothesized that an auscultatory system based on machine learning could assist non-experts in access monitoring by accurately predicting access dysfunction.

Methods: In this pilot project (2021-2022), we used a digital stethoscope to auscultate access bruits at pre-defined clinically relevant locations as part of physical examination during hemodialysis access clinic visits. We simultaneously recorded expert clinician assessments of access dysfunction and the need for procedural intervention. Audio recordings were translated into Mel spectrograms. Using a training and testing paradigm, we built a neural network prediction model and assessed its performance.

Results: The dataset consisted of recordings from 54 AVFs and AVGs. The prediction model demonstrated a 90% accuracy in detecting the need for intervention and a 75% accuracy in detecting the presence of a clinically significant stenosis.

Conclusions: We anticipate deployment of this auscultatory prediction technology in an easy-to-use platform that will standardize access monitoring and increase patient autonomy in management of their hemodialysis access.

Lay Abstract

Background: Most patients with kidney failure require replacement of their kidney’s function through hemodialysis (cleaning of the blood). Hemodialysis usually occurs through a surgically created hemodialysis access, which strengthens and increases blood flow to a vein. Early identification of access dysfunction can prevent costly and dangerous access failures; however, effective access monitoring requires substantial skill and time. We believe a sound recording-based system that uses machine learning could assist non-experts in monitoring accesses by predicting their dysfunction.

Methods: We used a digital stethoscope to listen to access flow murmurs as part of physical examination during hemodialysis access clinic visits. We also recorded expert clinician assessments of access dysfunction and the need for a corrective procedure. Audio recordings were translated and used by machine learning to predict access dysfunction.

Results: The dataset consisted of recordings from 54 dialysis accesses. The prediction model showed a 90% accuracy in detecting the need for a corrective procedure and a 75% accuracy in detecting the presence of a significant access narrowing.

Conclusions: We envision this prediction technology to be expanded as an easy-to-use platform that standardizes monitoring and increases patient involvement in management of their hemodialysis access.

Clinical Implications

After point-of-care deployment, our auscultatory machine learning system will standardize early detection of hemodialysis access dysfunction, increase patient engagement with their access, and facilitate timely referral to hemodialysis access specialists.