Title : Accuracy of artificial intelligence-assisted diagnosis for early esophageal cancer in white light endoscopy: A meta-analysis
Abstract:
Objective The artificial intelligence(AI)-assisted detection reduces underdiagnosed or overlooked early esophageal cancers. We performed a meta-analysis to determine the diagnostic accuracy of AI used in white light endoscopy of patients with early esophageal cancer.
Methods A systematic search of PubMed, Embase, and Scopus until December 2025 was performed. We included studies that use AI in the white light endoscopy for diagnosis in patients with early esophageal carcinoma, using pathologic examination as the reference standard. The primary outcome was the sensitivity and specificity of 17 studies included. Secondary outcomes were pooled as positive likelihood ratio and negative likelihood ratio
Results A total of 17 studies using AI to detect early esophageal neoplasia under white light endoscopy were included, involving 2808 patients and 7070 images. In the 17 studies with the least biased estimates, sensitivity ranged from71.8% to99.4% and specificity ranged from76.5% to99.3%. the meta-analytical pooled sensitivity and specificity (95% CI) of early esophageal cancer detection by AI were 90.8% (83.6%-95.1%) and 88.6% (76.1%-95.0%), the positive likelihood ratio and negative likelihood ratio (95% CI) were 7.985(3.589-17.766) and 0.103(0.055-0.192).
Conclusion As a white light endoscopic aid, AI technology can help accurately detect esophageal malignancies. It is expected to significantly reduce the missed diagnosis rate of early cancer in clinical practice. The value of accuracy estimates is considerably undermined by the small number of included studies, and concerns about risk of bias relating to the index test and the reference standard.
