#5088 Artificial Intelligence Applied to Vascular Doppler Ultrasound Models for Predicting Underdevelopment of Vascular Access in Nephrology

Abstract

Background and Aims The vascular access of choice for hemodialysis patients is the arteriovenous fistula (AVF). There is a high rate of early primary failure and loss of primary AVF patency. Monitoring of vascular access is essential for early diagnosis of complications and prolonging survival. Models based on Artificial Intelligence (AI) and Machine Learning (ML) can be used for this. Method Retrospective descriptive study of the Vascular Doppler Ultrasound (VDU) in adults carried out since January 2019 to January 2022 in our AVF follow-up nephrology clinic. We analyze the results and create AI-based AVF underdevelopment prediction models. We included clinical, demographic and ultrasound variables. Patients were undergoing AVF post-surgery follow-up (VDU by protocol at 3-4 weeks after AVF surgery) or were referred to the clinic with signs of AVF dysfunction. The insufficient development of the vascular access is established as an objective variable. SPSS 20 Statistical Package. Automated Learning Analysis (ML) with Orange ML and BigML. Results 243 VDU were performed. Of the total, 139 (57%) were follow-up post-surgical VDU per protocol and 104 (43%) were AVF dysfunction VDU. Using supervised ML Analysis techniques with random sampling of 80% of the instances for Training and 20% for Test, we obtain prediction models for the underdevelopment (UD) attribute of FAV: Decision tree algorithm, Area under the curve (AUC) 89%, Classification accuracy (CA) 90%, Precision 90%. Random Forest Algorithm (RF) (AUC) 95%, (CA) 86%, Accuracy 81%. Near Neighbor Algorithm (K-NN) (AUC) 88%, CA 82%, Accuracy 78%. Convolutional Neural Networks (NNC) (AUC) 82%, CA 74%, Accuracy 60%. Algorithm with unsupervised technique of clustering in k-Means 3 clusters are obtained. The variables that best correlate with the objective variable are access flow, vein diameter, resistance index (RI) proximal, (RI) distal, and diameter of the anastomosis. Conclusion The vascular ultrasound systematized by the nephrologist facilitates the early diagnosis of complications that lead to early intervention. Analysis of the data with techniques (ML) can facilitate early diagnosis AVF poor development requiring close monitoring or intervention. The development of a nephrology clinic for monitoring vascular access could avoid invasive and unnecessary procedures for the patient.