Optimization and Validation of an AI Model that Screens for Arteriovenous Fistula Stenosis in Dialysis Patients using Sound Files from a Digital Stethoscope

Weill Medical Coll Of Cornell Univ, New York NY

Investigators

Abstract

PROJECT SUMMARY The overarching goal of this proposal is to determine if screening for stenosis in dialysis patients with an arteriovenous fistula (AVF) can be performed using an AI model that processes sound files from a digital stethoscope. This proposal will optimize an AI model by training it on an expanded dataset of prospectively acquired measurements and validating its performance against angiography, which is the diagnostic gold standard. Lastly, the clinical translatability of this novel screening approach will be characterized by evaluating its accuracy when performed by a trained technician versus the patients themselves. The AVF is often touted as the “lifeline” for dialysis patients. According to the National Kidney Foundation (NKF), vascular access is globally ranked as a top priority for dialysis patients, health care providers, and clinical research. Preserving dialysis access is a high priority for providers and patients because the consequences of AVF dysfunction and subsequent access failure significantly contributes to patient morbidity and healthcare costs. Unfortunately, AVF dysfunction is not uncommon and requires treatment via surgery or radiologic intervention. However, patient referral for treatment typically does no come from the surgeon or interventionalist that are trained to perform diagnostic imaging, but instead relies on technicians or physicians at the dialysis centers that are burdened by large patient volumes and often resulting in referrals not based on the patient’s actual disease state. This leads to many patients being sent to treatment pre-maturely and other patients not being sent at all. Therefore, low-cost and easy-to-use screening tools would mitigate these referral issues by providing patients an objective assessment of their AVF state. Auscultation (i.e., listening for internal body sounds) is a noninvasive method, compared to digital subtraction angiography or venous cannulation, and more convenient compared to ultrasound for detecting abnormal blood flow. Additionally, a change in access bruit or thrill may be one of the earliest clinical indicators that a stenosis is developing and can be measured using a low-cost and widely available digital stethoscope. However, the reality is that auscultation is a highly subjective physical exam technique and largely depends on the skill of the listener. Since the timely diagnosis of stenosis is crucial for maintaining dialysis access, applying deep learning to AVF blood flow sounds can enhance the ability of healthcare providers to screen for AVF stenosis both reliably and efficiently. The short-term impact of this proposal will be to optimize the AI model and validate the feasibility of this novel screening tool for AVF stenosis in patients at dialysis centers. The long-term impact of this proposal will be realized in a subsequent randomized clinical trial that will validate this AI model can effectively screen for a stenotic AVF in a dialysis clinic, in order to improve the efficiency of AVF maintenance for both the health of the patient and the overall burden to the healthcare system.