A personalised algorithm predicting the risk of intravenous corticosteroid failure in acute ulcerative colitis

An episode of acute ulcerative colitis (UC) represents an important watershed moment in a patient's disease course.

To derive a personalised algorithm for identifying patients at high risk of corticosteroid non-response from variables available at hospital presentation using a large prospectively collected acute UC patient database and machine learning-based techniques.

We analysed data from 682 consecutive presentations of acute UC. We used an Akaike information criterion-based elastic net model to select variables based on the 419 earliest presentations of acute UC (1996-2017). We constructed two risk-scoring algorithms, with and without utilising additional endoscopic variables, using logistic regression models. We validated these risk scores on separate cohorts of 181 (2018-2022) and 82 (2015-2022) acute UC presentations.

The partial risk of rescue (ROR) score included the admission indices of oral corticosteroid treatment, bowel frequency ≥6/24 h, albumin, CRP ≥12 mg/mL and log₁₀CRP. The full ROR score incorporates the same variables with the addition of the Mayo endoscopic subscore and disease extent. The AUCs in the main validation cohort were 0.76 (95% CI: 0.69-0.83) and 0.78 (95% CI: 0.71-0.85) for the partial and full ROR scores, respectively.

These pragmatic personalised risk scores (available at www.severecolitis.com) have comparably strong performance characteristics and usability enabling the identification of individuals at high risk of corticosteroid non-response before or after endoscopic assessment. The ROR scores have the potential to challenge conventional acute UC treatment paradigms by identifying patients who may benefit from early rescue therapy or participation in relevant clinical trials.