Machine learning-based prediction of response to Ustekinumab with Crohn's disease.

Background: A reliable approach to predict the response to Ustekinumab (UST) in patients with Crohn's disease (CD) is lacking. Objectives: This study aims to develop and validate machine learning (ML) models to predict the response to UST and further achieve personalized therapy. Design: Retrospective multi-center study. Methods: This study included 162 CD patients treated with UST between May 2022 and May 2024. Four ML algorithms (extreme gradient boosting, random forest, logistic regression, and support vector machine) were integrated to identify the optimal model, and Shapley Additive exPlanations (SHAP) interpretation was used for visual explainability. Two models were established to forecast the response to UST, with the outcomes of the response situation at week 26 and secondary loss of response (sLOR) status at week 52, respectively. Eighty-two CD patients from the other five centers were applied for the week-26 model's external validation. Results: XGBoost performed excellently among the four ML algorithms. The week-26 model exhibited good performances of 0.88 area under the receiver operating characteristic curve (AUC), 0.92 area under the precision-recall curve, and 0.86 F1 score. The sLOR model demonstrated acceptable predictive performance with 0.74 AUC. Conclusion: We developed and validated models to predict UST response for CD patients and interpreted related factors by the SHAP method. We hope that the models can assist physicians in identifying patients who are suitable for UST at baseline and further explore who are at high risk for sLOR.