What question did this study set out to answer?

Evaluate the predictive power of AI-derived endoscopic scores on histological remission states in ulcerative colitis.

January 23, 2026

P0554AI assessment of Ulcerative Colitis from endoscopy videos using MES and UCEIS scoring systems

Key Points

Evaluate the predictive power of AI-derived endoscopic scores on histological remission states in ulcerative colitis.
Developed a convolutional-recurrent neural network for video analysis.
Trained on 219 endoscopy videos from 64 ulcerative colitis patients.
Eliminated low-quality frames through automated quality control.
Tested model on 100 videos, comparing AI results with expert assessments using AUC and QWK statistics.
Achieved QWK agreement of 0.78 for MES and 0.84 for UCEIS assessments.
Demonstrated 0.89 accuracy for MES and 0.92 accuracy for UCEIS.
Showed AUC of 0.88 for MES and 0.92 for UCEIS, indicating strong diagnostic performance.

Abstract

Abstract Background Ulcerative colitis trials commonly pair endoscopic and histological assessments to define disease activity. Because endoscopic remission often predicts histological remission, endoscopybased AI that scores disease severity can also infer histological state. Demonstrating this capability would validate model robustness and enable biopsysparing endpoints for clinical trials. Our aim is to evaluate how AI-derived endoscopic Mayo Endoscopic Subscore (MES) and Ulcerative Colitis Endoscopic Index of Severity (UCEIS) metrics predict histological Nancy Index (NI), Robarts Histopathology Index (RHI) and Simplified Geboes Score (SGS) pathologist scores. Methods A convolutional–recurrent (CNN + RNN) architecture was trained on 219 videos with 1.1M frames from 64 UC patients from the WHITS study (NCT05000242). Automated quality control eliminated approximately 0.5M frames from the dataset that contained high artefact load or were otherwise not scorable. Ground truth data was provided by a consensus of 2 expert endoscopists with a mean of 15 years’ experience. Testing was carried out on a held-out dataset of 100 videos, and model performance was evaluated against expert-graded MES/UCEIS for diagnostic accuracy using area under the receiver operating characteristic curve (AUC), and agreement across categories was assessed using quadratic weighted kappa (QWK) statistics. Patient demographics: mean age 43 years, 42% female, with varied treatments. Results The model achieved substantial agreement with endoscopist MES and UCEIS assessment of active/remission states. For MES, agreement was QWK 0.78, with an accuracy of 0.89, AUC 0.88 (95% CI 0.82–0.94). For UCEIS, agreement was QWK 0.84, with an accuracy of 0.92, AUC 0.92 (95% CI 0.84–0.98). Conclusion The AI model provides an objective and accurate assessment of active and remission states in UC. This demonstrates the model’s potential for providing scoring assistance, improving data quality and generating new insight for UC clinical trials. Conflict of interest: Mr. Bogush, Alexander: I disclose the following financial relationship(s) with a commercial interest: Perspectum Ltd Ralli, George: No conflict of interest Windell, Dylan: I disclose the following financial relationship(s) with a commercial interest: Perspectum Ltd Wakefield, Phil: No conflict of interest Suzuki, Noriko: No conflict of interest Toskas, Alexandros: No conflict of interest Landy, Jonathan: No conflict of interest

Bookmark

Cite This Study

Bogush et al. (Thu,) studied this question.

synapsesocial.com/papers/69730f34c8125b09b0d1f03f https://doi.org/https://doi.org/10.1093/ecco-jcc/jjaf231.735

Bookmark