BACKGROUND: Introducing Arges, an AI tool that estimates a novel continuous disease severity score in ulcerative colitis (UC) clinical trials, distinct from the traditional categorical Mayo Endoscopic subscore (MES). The novel score ranges continuously from 0 to 3 and can be computed over the entire colon (Arges-CMES_ALL) or as an average across multiple colon segments (Arges-CMES_SEG), providing enhanced sensitivity to subtle changes in disease state. Using a trained (locked) model, we evaluated treatment effect sizes and the sensitivity of Arges-CMES changes to treatment response in a prospective UC clinical trial.
METHODS: Arges-CMES was trained on UC clinical trial (UNIFI: NCT02407236, Phase 3, 965 subjects, 3128 videos) to estimate the mean MES between local and central readers to model the inter-rater variability. Training involved pre-training a transformer network with self-supervised learning, then training a smaller transformer network. Arges can output a continuous MES value per colon (Arges-CMES_ALL) or a segment-average score (Arges-CMES_SEG) computed by averaging scores from colon segments (descending colon, sigmoid colon, and rectum) labeled by endoscopists during the endoscopy. To evaluate utility, we used baseline and week 12 visit data from a prospective clinical trial (QUASAR: NCT04033445, Phase 2b induction study, 313 subjects, 615 videos), which evaluated guselkumab IV (GUS) vs placebo (PBO) for UC treatment. To determine whether a treatment effect was detected with the novel scores, change in Arges-CMES_ALL or Arges-CMES_SEG was compared by treatment groups using ANOVA and a pair-wise comparison between GUS and PBO. To evaluate potential impact on trial sample size, standardized effect size and power for a given sample size was computed for novel scores and for MES.
RESULTS: A significant treatment effect was observed using the novel scores (Arges-CMES_ALL and Arges-CMES_SEG) with mean change from baseline to week 12 significantly higher in GUS compared to PBO (Table 1(a), p<0.001 for both). Also, the effect sizes, measured by Hedge's g, were larger for the change in AI novel scores than the change in MES (Table 1(b)). This implies that if a trial were designed to detect a difference in AI novel score instead of MES, a much smaller sample size would be required to achieve the same power (Figure 1). For example, for Arges-CMES_ALL, the potential reduction in required sample size is 24% (from 100 to 76 patients), and for Arges-CMES_SEG, it is 60% (from 100 to 40 patients), to achieve 80% power in detecting a treatment difference at a significance level of 0.05 (Figure 1).
CONCLUSION: The Arges-CMES novel scoring system effectively detects significant treatment differences and demonstrates a larger effect size than MES. As a result, Arges-CMES has the potential to reduce the required sample size for a given power threshold compared to MES.

John Kwon

Appearances