Dialysis access monitoring using a digital stethoscope-based deep learning system

This SBIR Phase I project will develop a deep learning-based algorithm to analyze the sound of blood flow in newly created arteriovenous fistulas (AVFs) used for hemodialysis access. This monitoring tool can help to identify fistulas that are unlikely to mature in patients who need surgical intervention to achieve successful maturation. The specific aims of the study are (1) to create the world’s first deep learning-scale database of newly created AVF sounds from hemodialysis patients, and (2) develop and evaluate the performance of a deep learning classification model trained via semi-supervised learning to discriminate between patients with AVFs likely to mature and patients with AVFs unlikely to mature. By integrating this deep learning algorithm into Eko’s mobile and cloud software platform, we anticipate this algorithm will enable better monitoring of the maturation process for newly created fistulas. During Phase I of the project, we will recruit study subjects in access centers at the University of North Carolina (UNC).SBIR Project Narrative In creating the world’s largest annotated dataset of mature AV fistula sounds and corresponding ultrasounds, Eko will gain unprecedented insights into fistula blood flow. We expect this to advance our scientific understanding of the fluid dynamics of fistulas, and help explain how and why hemodialysis access fails at such a high rate over time. A clinical decision support algorithm trained from this dataset can help healthcare systems detect fistula failures earlier, leading to more timely interventions that can avert costly complications, prolong life expectancy, and improve quality of life for hemodialysis patients as their care transitions from clinic to home care.