Kidney Week

Abstract: TH-PO023

Validation of a Cloud-Based Convolutional Neural Network Classifying Arteriovenous Access Aneurysms in a Multicenter Study

Session Information

• Augmented Intelligence for Prediction and Image Analysis
  October 24, 2024 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Augmented Intelligence, Digital Health, and Data Science

• 300 Augmented Intelligence, Digital Health, and Data Science

Authors

• Dong, Zijun, Renal Research Institute, New York, New York, United States

Zijun Dong

• Zhang, Hanjie, Renal Research Institute, New York, New York, United States

Hanjie Zhang, PhD

• Wang, Lin-Chun, Renal Research Institute, New York, New York, United States

Lin-Chun Wang, MS

• Ren, Sarah, Renal Research Institute, New York, New York, United States

Sarah Ren

• Han, Maggie, Renal Research Institute, New York, New York, United States

Maggie Han, MS, MSc

• Tisdale, Lela, Renal Research Institute, New York, New York, United States

Lela Tisdale

• Gusmao Bittencourt, Valeria, Renal Research Institute, New York, New York, United States

Valeria Gusmao Bittencourt

• Rosales M., Laura, Renal Research Institute, New York, New York, United States

Laura Rosales M., MD, MPH

• Starakiewicz, Piotr, Azura Vascular Care, New York, New York, United States

Piotr Starakiewicz, MD

• Douglas, Denzil, Azura Vascular Care, New York, New York, United States

Denzil Douglas, MD

• Shtaynberg, Norbert, Azura Vascular Care, New York, New York, United States

Norbert Shtaynberg, DO, FASN

• Kozyra, Andrzej, Azura Vascular Care, New York, New York, United States

Andrzej Kozyra, MD

• Fuca, Nicholas, Azura Vascular Care, New York, New York, United States

Nicholas Fuca, MD

• Prakash-Polet, Sindhuri, Azura Vascular Care, New York, New York, United States

Sindhuri Prakash-Polet, MD, PhD

• Thijssen, Stephan, Renal Research Institute, New York, New York, United States

Stephan Thijssen, DrMed

• Preddie, Dean C., Azura Vascular Care, New York, New York, United States

Dean C. Preddie, MD, MA

• Kotanko, Peter, Renal Research Institute, New York, New York, United States

Peter Kotanko, MD, FASN

Background

Arteriovenous (AV) access aneurysms may become life-threatening, e.g., in the event of ruptures. We developed an artificial intelligence classification application (ACA) that categorizes aneurysms using AV access images. This study prospectively evaluates ACA’s classification of AV aneurysms against physicians specializing in access care.

Methods

The study took place at two vascular access centers in New York, NY, USA. Subjects were assessed by physicians and classified as having “Advanced” or “Not Advanced” AV aneurysms. Six images, two from above the access and two on each side, were taken (Figure. 1c), uploaded to the cloud, and classified by a convolutional neural network within seconds (Figure. 1b). ACA classifications were compared to those made by physicians, and the area under the receiver operating characteristics (AUROC) curves were computed (Figure. 1a). Physicians were blinded to ACA results. Their assessment served as the ground truth.

Results

We included 720 images from 120 subjects (Figure. 1d). Physicians identified a 22% prevalence of advanced aneurysms. ACA classifications resulted in a mean AUROC of 0.907 (95% CI: 0.840-0.952), a median AUROC of 0.893 (95% CI: 0.823-0.942), a minimum AUROC of 0.866 (95% CI: 0.792-0.921), and a maximum AUROC of 0.910 (95% CI: 0.844-0.955) (Figure. 1e).

Conclusion

ACA demonstrated actionable accuracy in classifying AV aneurysms, even with angled images. Our results align with Zhang (CKJ, 2021), where ACA achieved an AUROC of 0.96 (n=402). Importantly, in the 2021 study physicians evaluated only images, whereas in the current study detailed physical exams were performed.
ACA has potential to support aneurysm monitoring, expediting detection and interventions.