Artificial intelligence-assisted endoscopic diagnosis of upper gastrointestinal neoplasia: A systematic review and meta-analysis of randomized controlled trials and diagnostic test accuracy studies.

338 Background: Upper gastrointestinal (GI) cancers represent a major global health burden, often diagnosed at advanced stages due to limitations in early detection. Endoscopy remains the gold standard for diagnosis. Artificial intelligence (AI)-assisted endoscopy (AIE) has emerged as a potential tool to enhance diagnostic performance. This study systematically evaluated the effectiveness and diagnostic accuracy of AIE in detecting gastric and esophageal cancers. Methods: A comprehensive search of PubMed, Scopus, Cochrane Library, and Web of Science was conducted up to May 2025. Eligible studies included randomized controlled trials (RCTs) and diagnostic test accuracy (DTA) studies assessing AIE. Data were extracted on study design, patient population, lesion type, machine learning (ML) models, and diagnostic outcomes. Primary outcomes included blind spot rate, miss rate, neoplasm detection rate, inspection time, biopsy rate, and diagnostic accuracy measures (sensitivity, specificity, diagnostic odds ratio, and area under the curve [AUC]). Risk of bias was assessed with RoB 2 for RCTs and QUADAS-2 for DTA studies. Meta-analysis was performed using random-effects models in RevMan 5.4 and Meta-DiSc. Results: From 3240 records, 18 studies met inclusion criteria (7 RCTs, 9 DTA studies, and 2 hybrid RCTs). The RCTs enrolled 17,125 patients. Five DTA studies focused on early gastric cancer or metaplasia, three on Barrett’s esophagus-related neoplasia, and three on SESCC. Convolutional neural networks (CNNs) were the most widely used ML architecture. Compared with conventional endoscopy, AIE significantly reduced blind spots (mean difference [MD] = -3.44; 95% CI: -4.56 to -2.32; p < 0.00001) and miss rates (risk ratio [RR] = 0.29; 95% CI: 0.11–0.76; p = 0.01). No significant differences were observed for neoplasm detection rate (RR = 1.09; p = 0.28), inspection time (MD = 0.07; p = 0.50), or biopsy rate (RR = 1.01; p = 0.79). Pooled diagnostic performance showed excellent accuracy, with sensitivity and specificity of 99%, a diagnostic odds ratio of 1038.29, and an AUC of 0.99. Quality assessment indicated that most RCTs were at low risk of bias, although two raised concerns regarding randomization and outcome assessment. Conclusions: AI-assisted endoscopy enhances upper GI lesion detection by significantly reducing blind spots and missed diagnoses while achieving near-perfect diagnostic accuracy. Although current evidence supports its clinical utility, heterogeneity in ML models and study methodologies highlights the need for large, multicenter RCTs with standardized protocols to validate its integration into routine clinical practice.