Abstract Background The clinical heterogeneity of acute Ulcerative Colitis (UC) poses a notable challenge to effective risk stratification and timely intervention. Existing severity indices derived in the 20th century may be obfuscated by the effects of novel therapeutics and fail to capture disease complexity leading to suboptimal treatment strategies. This study aimed to leverage machine learning to identify and characterise distinct UC clinical sub-phenotypes, to facilitate early and accurate acute UC risk stratification and personalisation of care. Methods A retrospective cohort analysis was conducted on 753 acute UC admissions to a single tertiary referral centre between January 1998 and February 2025. Using 13 routinely collected baseline variables, encompassing demographics, inflammatory markers, disease extent and prior therapeutic exposure, an unsupervised hierarchical clustering analysis was performed to identify novel UC clinical sub-phenotypes without preconceived classifications. Results Unsupervised clustering successfully identified six discrete acute UC phenotypes with distinct underlying disease severity, systemic inflammatory response, therapeutic efficacy, and prognosis, including risk of colectomy. One-year colectomy rate varied more than threefold across these sub-phenotypes, ranging from 19% in “Favourable Left-Sided Disease” (P1) to 58% in the “Multi-Drug Refractory” (P6) group. Notably, the “Hyper-Inflammatory” (P3) phenotype with the highest median admission C-Reactive Protein (CRP, 92 mg/L) had an intermediate one-year colectomy risk (42%), while inflammation was deceptively low (CRP 29 mg/L) in the high-risk “Multi-Drug Refractory” (P6) group. This led to identification of a “CRP paradox”, whereby a patient’s treatment history may be a more robust predictor of colectomy risk than their acute phase reactants. Phenotypic analysis demonstrated two distinct pathways to treatment failure: pharmacokinetic failure in highly inflamed, treatment-naïve patients and pharmacodynamic resistance in extensively pre-treated individuals. Conclusion This data-driven approach deconstructs the heterogeneous nature of acute UC into six prognostically significant clinical sub-phenotypes. These findings provide a rational, mechanism-based roadmap for personalising acute UC management, from identifying low-risk patients suitable for standard care, to high-risk individuals that require prompt therapeutic escalation, with upfront dose-intensified therapies, novel agents, or early surgical intervention. Future studies that integrate molecular data such as mucosal transcriptomics to define endotype-phenotype relationships may confirm the translational utility of this approach. Conflict of interest: Dr. Etchegaray, Amirah: No conflict of interest Goetz, Naeman: No conflict of interest Okano, Satomi: No conflict of interest Hartel, Gunter: No conflict of interest Hanigan, Katherine: No conflict of interest Phillips, Jennifer: No conflict of interest Kumar, Rina: No conflict of interest Tambakis, George: No conflict of interest Brown, Allison: No conflict of interest Radford-Smith, Graham: No conflict of interest Walker, Gareth: In the last 24 months, Dr Walker has received investigator grants or served as a speaker, a consultant or an advisory board member for: Janssen AbbVie Takeda Ferring Dr Falk Pharma Georgiamune Croft, Anthony: No conflict of interest
Etchegaray et al. (Thu,) studied this question.