Artificial intelligence (AI) methods have been developed for automated assessment of ulcerative colitis (UC) endoscopic disease activity from still images and, increasingly, full-length videos. Multiple systems have achieved substantial agreement with expert readers and central reading paradigms for Mayo Endo-scopic Subscore (MES/MCES), Ulcerative Colitis Endoscopic Index of Severity (UCEIS), and related remission definitions, and several approaches have linked endoscopic AI outputs with histology and clinical outcomes such as relapse risk. Video-based pipelines have also been positioned for clinical trial central reading and have included quality gating, temporal aggregation, and, in some cases, real-time use. Continuous and extent-sensitive metrics have been proposed to overcome limitations of “worst-segment” ordinal scoring and have been evaluated in trial contexts. Despite these advances, the literature has left a translation gap between high-performing scoring models and health-informatics requirements for routine clinical deployment: interoperable representation of AI outputs in electronic health records (EHRs), uncertainty-aware escalation and adjudication, auditability, workflow-centered interfaces, and post-deployment governance. This narrative review has synthesized the technical and clinical trajectory of UC endoscopic AI and has proposed an informatics blueprint that has operationalized model outputs as standardized, provenance-rich “evidence bundles” suitable for integration into clinical decision support systems (CDSS) and trial workflows. The blueprint has included (i) a reference architecture for video in-gestion through scoring and longitudinal monitoring; (ii) a minimal data object schema for endoscopic AI readouts (scores, continuous indices, extent maps, quality, uncertainty, explanations, and provenance); (iii) a human–AI interaction pattern for “second-opinion triggering” and adjudication; and (iv) an evalua-tion and governance framework spanning performance, calibration, usability, equity and domain shift monitoring. The proposed blueprint has aimed to support evidence-based adoption while preserving clinical accountability and regulatory feasibility.
Taabzadeh et al. (Mon,) studied this question.