Zehua Dong, Lianlian Wu, Hongliu Du, Xiao Tao, Junxiao Wang, Xiaoquan Zeng, Yijie Zhu, Si Tan, Jiamin Wang, Mei Deng, Tian Yang, Chaoqi Chen, Liming Wen, Ying Li, Congyao Jiang, Chuang Tang, Yulin Wang, Zhifeng Zhao, Qinghua Wang, Haodong Yu, Haodong Yu, Chaoyuan Chen, Yaowei Ai, Jie Pan, Fengsong Chen, Qiang Guo, Xinli Mao, Junhong Li, Huifang Pang, Jianyu Hao, Yan Jiao, Qizhi Wang, Heng Zhang, Shicai Ye, Tong Dang, Shen Hu, Liangliang Yu, Shan Hu, Wei Zhou, Yunchao Deng, Yanxia Li, Boru Chen, Bing Xiao, Jun Liu, Jiamin Qin, Ting Zhang, Tong Bu, Jiayi Shi, Yumei Wang, Hong Zhang, Jiejun Lin, Lei Jin, Jianfeng Yu, Bofu Tang, Xia Tan, Lilei Yu, Lilei Yu, Honggang Yu, Honggang Yu
BACKGROUND & AIMS: Evidence about the effect of artificial intelligence (AI) on upper endoscopy in multicenter, randomized controlled trials is lacking. We aimed to explore whether AI can enhance gastric neoplasm detection. METHODS: Participants from 24 hospitals in China from December 21, 2021, to November 11, 2023, were randomized to AI-assisted or nonassisted esophagogastroduodenoscopy. Primary outcome was detection rate of gastric neoplasms after pathologic review. Secondary outcomes included detection rate of gastric neoplasms before review, relative early gastric cancer detection ratio, detection rate of intestinal metaplasia and/or gastric atrophy before or after review, biopsy rate, number of blind spots, and procedure/inspection time. We did intention-to-treat (ITT), per-protocol, and exploratory subgroup analysis. RESULTS: In the ITT cohort, 29,514 patients were enrolled. AI did not improve detection rate of gastric neoplasm after pathological review (RR, 1.13; 0.92-1.38; 1.42 vs 1.25%; P = .25). However, based on original pathology, an improvement with AI was observed (RR, 1.14; 1.0-1.28; 4.06 vs 3.57%; P = .03). AI reduced blind spots number from 2.52 to 1.07 (P < .001) and prolonged procedure and inspection time. No significant differences were observed for relative early gastric cancer detection ratio or detection rate of intestinal metaplasia and/or gastric atrophy before/after pathologic review in ITT cohort. Subgroup analysis suggested potential benefit among less experienced endoscopists and during fatigue periods. In the experimental group, AI diagnosed 100%, 91.9%, and 57.1% of pathologically confirmed gastric adenocarcinoma, high-grade, and low-grade intraepithelial neoplasia, respectively. CONCLUSIONS: AI did not improve the detection rate of gastric neoplasms. Further real-world studies are needed to fully address the adaptability of AI. (Chinese Clinical Trial Registry, ChiCTR2100054449.).