Implementation of a prospective video databank for artificial intelligence model development, validation and clinical outcomes research in gastrointestinal endoscopy.

内容摘要

High-quality, annotated datasets are fundamental to clinical research and artificial intelligence (AI) model development. Existing endoscopic databanks are predominantly retrospective, image-based, or lacking patient consent and structured prospective clinical annotations, limiting scientific reproducibility and regulatory-compliant AI validation. This study establishes a prospective endoscopy video databank with standardized annotations enabling AI development and clinical outcomes research. Since 2022, we have prospectively enrolled consecutive patients undergoing esophagogastroduodenoscopy (EGD) or colonoscopy at a tertiary center (NCT06822616). Endoscopies were recorded in full (1920×1080 resolution, 60 frames per second, 10-bit color depth), and de-identified at acquisition. Trained staff documented timestamped anatomical landmarks, lesion characteristics, optical diagnoses, interventions, and disease scores in structured case report forms during procedures. Pathology results and metadata were linked to all events observed during the endoscopies and corresponding timestamped video frames. Up to March 2026 8658 patients (mean age 59.3 years; 52.0% female) were enrolled contributing 10831 procedures (7909 colonoscopies, 2922 EGDs) performed by 55 endoscopists. Each colonoscopy and EGD captured up to 747 and 455 structured variables, respectively. Individual polyps (n=10057) and biopsies (n=11346) were annotated with 49 and 45 variables, respectively. Inflammatory bowel disease-specific documentation comprised up to 188 variables per procedure. Pathological findings were present in 83.7% of records. This prospective databank provides the infrastructure for endoscopy research requiring traceable data sourcing, whether for clinical outcomes research or AI development and validation. The integration of full-length videos, real-time procedural annotation, and histopathological correlation addresses critical gaps in optimizing clinical outcomes and AI research.