S870 A Ubiquitous and Interoperable Deep Learning Model for Automatic Detection of Pleomorphic Gastroesophageal Lesions

Introduction: Artificial intelligence (AI) has been widely explored recently to enhance capsule endoscopy (CE), with the goal of improving the reading process efficiency. The majority of AI models focus on analyzing the small intestine and colon, while the development of esophagogastric (EG) models has been limited by the lack of frames, posing a challenge to building models with robust performance. This study aims to develop an interoperable ubiquitous model capable detecting pleomorphic lesions in the EG tract. Methods: A total of 59,482 endoscopic frames were included from 774 CE procedures conducted across 5 medical centers to train a convolutional neural network (CNN). The data were separated following an exam-based split, with 90% assigned to training set—including a 5-fold cross-validation—and the remaining 10% reserved for testing. The main performance metrics evaluated included sensitivity, specificity, accuracy, and the area under the receiver operating characteristic curve (AUC-ROC). Results: In training set, the average sensitivity was 85.0% (95% confidence interval [CI] 76.5-93.5), specificity was 96.3% (95% CI 93.9-98.7) and accuracy was 93.6% (95% CI 91.7-95.4), with an AUC-ROC of 0.98 (95% CI 0.97-0.98). In the testing phase, the sensitivity was 92.2%, specificity was 95.1%, and accuracy was 94.6%. Conclusion: AI models for CE that can evaluate the esophagus and stomach represent an important advancement toward creating a comprehensive panendoscopic system. This ubiquitous and interoperable model has shown robust accuracy in identifying lesions in both regions. Nonetheless, further prospective studies in real-world clinical settings are essential to confirm its effectiveness compared to conventional upper endoscopy.