Learning Curve of Fenestrated-Branched Endovascular Aortic Repair (F-BEVAR) Using Machine Learning: A Prospective National Multicenter Registry Study

Abstract

Objective: Fenestrated-branched endovascular aortic repair (F-BEVAR) is a complex procedure that requires significant experience and advanced technical proficiency. This study leverages machine learning methods to analyze the learning curve of F-BEVAR in the treatment of complex abdominal and thoracoabdominal aortic aneurysms (TAAA).

Methods: Patients undergoing 3- and 4-vessel F-BEVAR for intact complex abdominal aortic aneurysms (cAAA) and TAAA between January 2014 and September 2024 were identified in the Society for Vascular Surgery Vascular Quality Initiative; a prospective, nationwide, multicenter registry. cAAA were defined as juxta-, para-, and supra-renal abdominal aortic aneurysms. F-BEVAR procedures performed by each individual surgeon were chronologically ordered and numbered to represent each physician's experience with F-BEVAR at the time of each operation. A deep learning neural network model was developed to quantify the learning curve by predicting outcome rates based on physician experience. Primary outcomes were perioperative mortality, procedural technical success, and major adverse events (MAE) defined as the composite outcome of mortality, conversion to open surgery, spinal cord ischemia, visceral ischemia, renal ischemia, myocardial infarction, stroke, dialysis, pneumonia, and blood loss >1,000 mL. Secondary outcomes included perioperative aortic reintervention, procedural metrics (operative time, fluoroscopy time, blood loss, and contrast volume), and the individual components of the MAE composite measure.

Results: A total of 5,540 patients underwent F-BEVAR by 539 unique physicians, of which 2,956 patients underwent 3- and 4-vessel F-BEVAR. Of those, 64.4% (n=1,901) were treated for cAAA and 35.6% (n=1,055) for TAAA. The incidence of MAE gradually decreased from 31.0% (95% CI, 30.7%-31.1%) in initial procedures to a low of 18% (95% CI, 17.7%-18.1%) with increased physician experience, with a learning plateau in MAE rates (23% [95% CI, 22.5%-23.2%]) occurring between 40-70 procedures. Procedural technical success increased substantially from 91% (95% CI, 90.8%-91.5%) to 97% (95% CI, 96.5%-97.0%) with increasing physician experience, with a learning plateau observed once again between 40 and 70 procedures when a success rate of 96.0% (95% CI, 95.7%, 96.2%) was achieved. Increasing physician experience was associated with a decline in mortality from 4.4% (95% CI, 4.3%-4.6%) to 2.0% (95% CI, 1.9%-2.1%) and aortic reintervention rates from 8% (95% CI, 6.9%-8.1%) to 5% (95% CI, 4.9%-5.3%). Marked improvements were observed in procedural metrics with increasing physician experience, including blood loss declining from 551 mL (95% CI, 508-594 mL) to 179 mL (95% CI, 136-222 mL), fluoroscopy time decreasing from 91 mins (95% CI, 87-95 mins) to 55 mins (95% CI, 51-59 mins), and operative time decreasing from 320 mins (95% CI, 306-331 mins) to 242 mins (95% CI, 230-255 mins).

Conclusion: The learning curve for F-BEVAR demonstrates a proficiency plateau achieved after 40-70 cases, providing a valuable benchmark for surgeons and institutions adopting this complex procedure. As the use of F-BEVAR continues to grow in the treatment of cAAA and TAAA, these findings can help inform practice guidelines and support data-driven decision-making to ensure a safe and effective nationwide expansion of this procedure.

Keywords: Deep Learning; Fenestrated Branched Endovascular Aortic Repair; Learning Curve; Machine Learning; Physician Modified Endograft.