Ó³»­´«Ã½

BACKGROUND: Disease course following hospitalization for severe ulcerative colitis (UC) is poorly understood. Here, we evaluate the outcomes of hospitalized patients with severe UC defined using endoscopic activity across the Mass General Brigham (MGB) health system over a 24-year period.METHODS: We developed and internally validated a natural language processing (NLP) algorithm to derive endoscopic disease severity. We included all hospitalized patients with UC across MGB from 1997 to 2021 and used the NLP algorithm to identify those with severe disease activity, defined using the Mayo Endoscopic Score. We used Cox proportional hazards modeling to estimate the risk of 90-day colectomy and rehospitalization across three calendar periods (1997-2005, 2006-2015, and 2016-2021) for patients hospitalized with severe UC.FINDINGS: The NLP algorithm had excellent performance for identifying endoscopic features, with high Cohen's Kappa scores between the algorithm's prediction and the reviewers' consensus for derived Mayo endoscopic score (k = 0.92). We identified 279 independent hospitalizations for severe UC using our NLP algorithm and manual validation methods from 1997 to 2021. The 90-day cumulative incidence of colectomy were 17.9% in 1997-2005, 18.3% in 2006-2015 and 24.4% in 2016-2021. In contrast, rates of biologic medication use increased from 10.7% in 1997-2005 to 70.2% in 2016-2021. Compared to the 1997-2005 period, the adjusted hazard ratios (aHRs) for 90-day colectomy were 1.1 (95% CI 0.4-2.8) in 2006-2015 and 1.5 (95% CI 0.6-3.9) in 2016-2021 (P = 0.225). Similarly, we did not observe any differences in rates of 90-day rehospitalizations (P = 0.624).INTERPRETATION: Here, we present an NLP algorithm for extracting features of endoscopic severity in patients with UC. Using this algorithm, we showed that the outcomes, including colectomy and rehospitalization, of patients hospitalized for severe UC in a large health system in the US have not significantly changed over the past 24 years.