Artificial intelligence in endoscopic prediction of submucosal invasion in gastric neoplasms: systematic review and...

Accurate assessment of Submucosal (SM) invasion in gastric neoplasms is vital for curative endoscopic management, yet conventional endoscopy remains moderately accurate and operator dependent. Whether AI can deliver reliable, generalizable depth prediction is unknown.

This investigation establishes the diagnostic capability of AI-driven endoscopic systems for identifying SM invasion.

A PRISMA-DTA systematic review and meta-analysis identified studies that used machine-learning or deep-learning models applied to endoscopic imaging with histopathology as the reference standard. Validation-set 2 × 2 data were pooled using hierarchical models to estimate sensitivity, specificity, and AUC, and heterogeneity was assessed through subgroup analyses.

Eight studies were included in the systematic review; seven studies provided sufficient data (including ten cohorts for the meta-analysis). AI systems demonstrated a pooled sensitivity of 0.75 (95% CI 0.69-0.81) and a specificity of 0.84 (95% CI 0.79-0.87). Discriminatory ability was high (AUC = 0.87, 95% CI 0.84-0.90), providing a conservative, transportable estimate of performance based on validation cohorts. Heterogeneity was substantial. Externally validated models showed lower sensitivity than internally validated models (0.72 vs. 0.79; p < 0.001), revealing overfitting. Korean cohorts demonstrated higher specificity than studies from other regions (0.87 vs. 0.79; p < 0.001). Fagan's analysis showed that a positive AI result increased post-test probability from 25% to 61%, while a negative result reduced it to 9%.

AI-assisted endoscopic assessment establishes a high-accuracy framework for identifying SM invasion and opens the door to more consistent treatment selection in early gastric cancer. International external validation and unified invasion definitions remain essential for dependable clinical use.