Machine learning using clinical data at baseline predicts the medium-term efficacy of ustekinumab in patients with ulcerative colitis

Abstract

Abstract Predicting the therapeutic response to biologics before administration is a key clinical challenge in ulcerative colitis (UC). We previously reported a model for predicting the efficacy of vedolizumab (VDZ) for UC using a machine-learning approach. Ustekinumab (UST) is now available for treating UC, but no model for predicting its efficacy has been developed. When applied to patients with UC treated with UST, our VDZ prediction model showed positive predictive value (PPV) of 56.3% and negative predictive value (NPV) of 62.5%. Given this limited predictive ability, we aimed to develop a UST-specific prediction model with clinical features at baseline including background factors, clinical and endoscopic activity, and blood test results, as we did for the VDZ prediction model. The top 10 features (Alb, monocytes, height, MCV, TP, Lichtiger index, white blood cell count, MCHC, partial Mayo score, and CRP) associated with steroid-free clinical remission at 6 months after starting UST were selected using random forest. The predictive ability of a model using these predictors was evaluated by fivefold cross-validation. Validation of the prediction model with an external cohort showed PPV of 68.8% and NPV of 71.4%. Our study suggested the importance of establishing a drug-specific prediction model.

Keywords

• Ulcerative colitis
• Machine learning
• Ustekinumab
• Steroid-free clinical remission
• Prediction model