S1353 Multicenter Validation of an AI Tool for Detecting Small Bowel Vascular Lesions in Capsule Endoscopy

Introduction: Capsule endoscopy (CE) plays a central role in evaluating small bowel pathology, particularly in patients with iron-deficiency anemia or obscure gastrointestinal (GI) bleeding. Vascular lesions—such as angiectasias and red spots—are among the most frequent findings in CE and a leading cause of bleeding in these settings. Accurate identification of these lesions is crucial for diagnosis and therapeutic planning. However, CE interpretation remains time-consuming and prone to reader variability. Artificial intelligence (AI), especially convolutional neural networks (CNNs), offers a promising solution to enhance lesion detection, improve consistency, and reduce interpretation time. This study aimed to develop and validate an AI model for detecting and classifying vascular lesions across multiple CE platforms and clinical environments. Methods: A prospective, multicenter study was conducted from January 2021 to April 2025, involving 9 centers across Portugal, Spain, Brazil, Australia, and the USA. A total of 398 anonymized CE exams were analyzed, acquired using 3 different systems (PillCam SB2, SB3, and Olympus EC-10). Each video underwent standard-of-care (SoC) reading, followed by an AI-assisted interpretation. Discrepancies were adjudicated by an independent expert panel to establish the gold standard. Diagnostic performance was evaluated using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy. Results: Vascular lesions were confirmed in 234 patients (58.9%) by expert consensus. AI-assisted reading achieved higher sensitivity (94.4% vs 64.5%) and overall accuracy (83.8% vs 71.6%) than SoC. Specificity was slightly lower with AI (68.3% vs 82.0%). The AI model’s superiority over SoC was statistically significant (P < 0.001). Average reading time was reduced to 316 seconds per video using AI assistance. Conclusion: This large multicenter study shows that AI-assisted CE significantly improves detection of vascular lesions with substantial time savings. The use of multiple CE systems and international clinical sites supports the model’s robustness, interoperability, and real-world applicability. These results reinforce AI’s role in enhancing CE-based diagnosis of GI bleeding and anemia-related conditions.