Exploratory Risk Patterns of Steroid Dependency in Ulcerative Colitis: A Multicenter Pilot Study Combining Logistic Regression and Self-Organizing Map Analysis

Introduction: Systemic corticosteroids remain widely used in ulcerative colitis (UC), yet approximately half of initial responders relapse during tapering or soon after discontinuation (steroid dependency). Understanding patients at risk of steroid dependency may inform individualized treatment strategies. Here, we explore factors associated with steroid dependency using conventional regression analysis and an exploratory artificial neural network approach. Methods: Consecutive patients with UC who received their first course of systemic corticosteroids from May 2012 to March 2020 were retrospectively screened, and those who responded to corticosteroids at day 30 were enrolled from four institutions. Factors associated with steroid dependency were assessed by multivariable logistic regression and a self-organizing map (SOM), an unsupervised neural network. Steroid dependency was defined as clinical relapse during tapering or within 3 months after completion of corticosteroids. Results: A total of 107 patients who showed a clinical response at day 30 were analyzed. Thirty-three (30.8%) patients developed steroid dependency. In the multivariable logistic analysis, extensive colitis, initial dose of prednisolone, and two-item patient-reported outcome score at day 30 were independently associated with steroid dependency. In the SOM analysis, patients were categorized into 8 clusters with variable rates of steroid dependency, indicating heterogeneous multidimensional clinical patterns beyond those captured by conventional regression analysis. Conclusion: By combining an unsupervised SOM with conventional logistic regression, our study provided an exploratory visualization of multidimensional clinical patterns associated with steroid dependency. These findings are hypothesis-generating and require prospective validation with a larger sample size before clinical application.