S3034 Transforming IBD Diagnostics: A Multicenter Validation Study on Artificial Intelligence in Capsule Endoscopy

Introduction: Capsule endoscopy (CE) is a key diagnostic tool for evaluating small bowel conditions, particularly Crohn’s disease. Detecting ulcers and erosions is essential for assessing disease activity, guiding treatment, and monitoring response. However, CE interpretation is time-consuming and prone to inter- reader variability. Artificial intelligence (AI), especially convolutional neural networks (CNNs), offers a promising solution by improving lesion detection, reducing variability, and shortening reading times. This study aimed to develop and validate an AI model for detecting and differentiating small bowel ulcers and erosions across multiple CE devices and clinical centers. Methods: A multicenter, prospective study was conducted between January 2021 and April 2025 across 9 centers in Europe (Portugal and Spain), South America (Brazil), Australia, and the USA. Three CE systems (PillCam SB2, SB3 and Olympus EC-10) were used to perform 398 anonymized exams. AI-assisted readings generated by a deep learning model were compared with standard-of- care (SoC) readings, using expert board consensus as the gold standard. Diagnostic performance was assessed using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUROC). Results: Ulcers and erosions were identified in 133 patients (33.4%) during expert review. SoC readings showed 68.9% sensitivity, 92.8% specificity, 82.7% PPV, 85.7% NPV, and 84.8% accuracy. AI-assisted readings achieved 87.9% sensitivity, 81.4% specificity, 70.3% PPV, 93.0% NPV, and 83.5% accuracy. The AI model detected 165 lesions versus 110 with SoC, performing consistently across CE platforms and centers. Mean AI-assisted reading time was 239 seconds (SD 138 seconds) per exam. Conclusion: The AI model demonstrated higher sensitivity and reduced reading time compared to SoC. This is the first multicenter, interoperable validation of an AI tool for detecting and differentiating ulcers and erosions using data from multiple CE systems. These findings support AI integration as a reliable and efficient approach to enhance CE-based assessment in inflammatory bowel disease.