Machine Learning Models for the Assessment of the Mayo Endoscopic Score in Ulcerative Colitis Trial Endpoints: A Systematic Review

BACKGROUND: The Mayo endoscopic score (MES) provides a criterion-based, but still subjective, human assessment of endoscopy and related endpoints in therapeutic clinical trials in ulcerative colitis (UC). A novel solution to address issues of reproducibility is the use of machine learning (ML) models to standardize MES evaluations. Broader applicability of this solution requires an understanding of the models and related performance characteristics. The objective of this study is to provide a systematic review on training and testing of ML MES prediction models on full-length endoscopic videos from patients with UC. METHODS: PubMed/MEDLINE, EMBASE, and Web of Science were systematically searched on December 31, 2024, and supplemented by reference checks and Google search to identify studies on training or testing of ML models to produce an automated MES grade on endoscopic procedure videos in UC. RESULTS: A total of 7 studies met the inclusion criteria, and of those, 5 were eligible for reporting on model performance. Accuracy in predicting ordinal MES grades (0, 1, 2, 3) ranged from 56.8% to 83.3%. Accuracy in predicting MES 0, 1 vs 2, 3 and MES 0 vs 1, 2, 3 (each aligned with a definition of endoscopic improvement and remission in trials) ranged from 84% to 90.2% and from 90% to 95.5%, respectively. CONCLUSIONS: Our review demonstrates strong performance characteristics of ML models to assess the MES on endoscopic videos in UC, potentially offering a standardized and reproducible solution to measure endoscopic severity. Further research will investigate the impact of this technology on clinical trial outcomes.