A framework for evaluating data quality and generalizability in artificial intelligence for gastrointestinal endoscopy

Background Artificial intelligence (AI) systems for gastrointestinal (GI) endoscopy have advanced rapidly in recent years, driven by deep learning architectures and availability of large image and video datasets. These developments have enabled automated polyp detection, lesion characterization, and disease activity scoring. Contemporary studies increasingly report models trained on hundreds of thousands or even millions of endoscopic images and video frames. However, dataset scale alone does not guarantee robustness or clinical generalizability. Methods In this narrative review, we examine the characteristics of datasets used to develop AI systems in GI endoscopy and discuss how dataset structure influences model performance and generalizability. Results We highlight key challenges including acquisition heterogeneity, label noise, and inappropriate data partitioning. We also examine the implications of emerging foundation models and large-scale pretraining strategies for endoscopic imaging. Finally, we outline practical considerations for interpreting dataset claims and propose the data quality and generalizability in endoscopy AI (DQG-EAI) framework for evaluating whether reported datasets are suitable for building clinically robust AI systems. Conclusion Understanding the relationship between dataset design and model performance is essential as AI moves toward routine clinical deployment. Emphasizing data diversity, transparent reporting, and rigorous validation will be critical for developing endoscopy AI systems that generalize reliably across clinical settings.