Risk-Stratified Biologic Efficacy in Ulcerative Colitis: A Multicenter Machine Learning Study.

BACKGROUND

Ulcerative colitis (UC) exhibits a heterogeneous clinical course, complicating prognostication and therapeutic decision making. Current tools inadequately predict progression or identify patients most likely to benefit from biologic therapies. We aimed to develop a machine learning model for risk stratification and evaluate its utility in optimizing biologic therapy outcomes.

METHODS

In this multicenter retrospective study, we analyzed 481 UC patients as the training cohort and 131 external validation patients. Disease progression-defined as treatment escalation, UC-related hospitalization, or surgery-served as the primary endpoint. Four models (Cox regression, logistic regression, random forest, XGBoost) were developed to predict progression risk. Biologic-treated patients (n = 235) were stratified into risk groups using the optimal model, with outcomes including mucosal healing, relapse, and acute severe UC assessed.

RESULTS

The random forest model demonstrated superior performance, achieving an area under the curve of 0.959 in training set and 0.759 in validation set. High-risk biologic-treated patients (n = 172) exhibited lower mucosal healing rates (33.7% vs 55.2%; P = .049) and higher hazards of clinical relapse (hazard ratio [HR], 3.35; P = .003), hospitalization (HR, 2.03; P = .014), and acute severe UC (HR, 2.70; P = .030) compared with low-risk patients (n = 63). No differences in serological remission, surgery, or biologic switching were observed.

CONCLUSIONS

Our random forest model enables precise risk stratification in UC, distinguishing patients with divergent responses to biologics. Low-risk patients derive significant benefit from timely biologics, while high-risk subgroups may require intensified strategies. This framework advances personalized UC management, though prospective validation is warranted.