An AI photo classifier flagged 25 of 26 advanced fistula and graft aneurysms in a first test against bedside exams

The problem

Aneurysms of arteriovenous fistulas and grafts are common, and a ruptured one can cause rapid, life-threatening bleeding. The authors cite estimates that fatal vascular access hemorrhage accounts for about 0.4% to 0.6% of hemodialysis deaths, and in one US series, 44% of these deaths occurred at home and 37% on the way to the hospital. There is no standard classification of aneurysms, and access expertise varies between dialysis units.

What the study did

A team led by Zijun Dong, with Dr. Peter Kotanko as senior author, at the Renal Research Institute in New York tested a cloud-based image classifier that returns, within about a second, the probability that an aneurysm is "advanced" (needing prompt attention) or "not advanced". In a masked study at two New York vascular access centers in 2023, trained research staff took six tablet photos of each patient's access, from three angles, against a plain white background. Interventional nephrologists who did not see the app's result examined each patient in person, with history and, at their discretion, ultrasound, and classified the aneurysm. Their classification served as the reference.

What they found

• 120 patients were analyzed (81% fistulas, 19% grafts); the physicians classified 26 (22%) as advanced.
• The area under the ROC curve was 0.908 (95% CI 0.843 to 0.963), using the average of the six images.
• At the cutoff chosen on this data set, the app identified 25 of 26 advanced aneurysms (sensitivity 96%) and correctly classified 77 of 94 others (specificity 82%): one missed case and 17 false alarms.
• Performance was similar in Black and non-Black patients, in fistulas and grafts, and across camera angles.

Why it matters

The authors suggest that a quick photo at the chair, or at home, could make aneurysm checks more frequent and consistent and prompt earlier referral. The authors describe the app as a screening aid, not a stand-alone diagnosis, and suggest uses including home hemodialysis.

Caveats

The authors call this a derivation-stage, proof-of-concept study. It involved 120 patients from one dialysis provider's clinics, seen at two access centers, with photos taken by trained staff under standardized conditions. Our caution: the cutoff was chosen on the same data, so performance may be lower in new settings. Two of the seven examining physicians also helped label the images used to train the model, which the authors acknowledge, arguing that the in-person and photo-only assessments differ enough to limit circularity. There are no outcome data yet showing that flags prevent ruptures.

Ties: the study was funded by the Renal Research Institute, which authors' disclosure forms describe as "a wholly owned subsidiary of Fresenius Medical Care," and all patients were treated in Fresenius Kidney Care clinics. Several authors are employees of the institute or of Fresenius, and Dr. Kotanko reports holding Fresenius Medical Care stock.

Disclosure: Dr. Wael Hussein serves on the Fresenius Medical Care US Home Medical Advisory Board.

The source: Dong Z, Zhang H, Wang L-C, Nandorine Ban A, Ren S, Tisdale L, Han M, Douglas D, Starakiewicz P, Shtaynberg N, Fuca N, Kozyra A, Prakash-Polet S, Bittencourt VG, Rosales M L, Thijssen S, Preddie DC, Kotanko P. Cloud-Based AI Application for Classifying Arteriovenous Hemodialysis Vascular Access Aneurysms: Findings from a Masked Multicenter Derivation Cohort Study. Clin J Am Soc Nephrol. Published online ahead of print 2026 (open access). doi:10.2215/CJN.0000001231 Read the original

🎧 Listen: Weekly · Oct 5, from 19:21.

Physician-led, AI-assisted. Dr. Wael Hussein chooses every item NephSpace covers and reviews and edits every story before it is published. The first draft was written in our own words with AI (Claude, by Anthropic) from the full text of the paper and checked against our reading notes; Dr. Hussein then reviewed and edited it. For education and information only, not medical advice. Please talk to your own physician or care team about your health. Clinicians: read the original before changing practice. See our disclaimer and disclosures. Spotted an error? Tell us.