Abstract
Background:
In hemodialysis patients, arteriovenous fistulas (AVF) are the preferred vascular access for most hemodialysis patients. However, aneurysms of the AVF can lead to severe complications, including rupture and life-threatening exsanguination. To address this risk, we recently developed, trained, and validated a convolutional neural network capable of classifying digital images of arteriovenous aneurysms captured via mobile devices as showing either “Advanced” or “Not Advanced” aneurysms. This cloud-based aneurysm classification application (ACA) aims to facilitate early detection of advanced aneurysms, potentially reducing morbidity and improving outcomes. Our study aimed to compare the diagnostic performance of ACA with in-person aneurysm assessments by physicians specialized in vascular access care.
Methods:
We conducted a single-masked, multicenter, observational, cross-sectional study at two vascular care centers with a catchment area of 21 hemodialysis clinics. Using mobile devices, research staff captured images of arteriovenous accesses and uploaded them to the cloud. The ACA then computed the probability of “Advanced” and “Not Advanced” aneurysms. Separately, physicians masked to the ACA results performed thorough physical and ultrasound examinations of the AVF and classified aneurysms as “Advanced” or “Not Advanced”. Physician classifications served as the ground truth for ACA performance analysis.
Results:
We analyzed 720 images from 120 incenter hemodialysis patients (six images per patient). Mean patient age was 62 years, 66 were male. Physicians classified 26 patients (22%) having “Advanced” aneurysms and 94 (78%) having “Not Advanced” aneurysms. The ACA achieved an AUROC of 0.908 (95% confidence interval: 0.843–0.963). At the maximum Youden index, ACA classified 42 aneurysms (35%) as “Advanced” and 78 (65%) as “Not Advanced,” yielding a sensitivity of 96% and a specificity of 82%.
Conclusion:
The ACA demonstrated promising accuracy in this initial masked evaluation. Further validation in larger, independent cohorts and under real-world conditions is warranted before broader clinical deployment.