P1085 Machine learning to predict the efficacy of ustekinumab for ulcerative colitis

H Morikubo,R Tojima,T Maeda,K Matsuoka,M Matsuura,J Miyoshi,S Tamura,T Hisamatsu

DOI: https://doi.org/10.1093/ecco-jcc/jjad212.1215

2024-01-01

Journal of Crohn's and Colitis

Abstract:Abstract Background Ustekinumab (UST), an anti-IL12/23p40 monoclonal antibody, is now widely used for ulcerative colitis (UC). Meanwhile, there are no established models to predict the efficacy of UST for UC. We previously demonstrated that the machine-leaning approach was useful for finding predictors and developing a prediction tool in vedolizumab (VDZ) treatment for UC (Miyoshi et al. Sci. Rep. 2021). In this study, we aimed to develop a new prediction model for UST in patients with UC. Methods UC patients who started UST as a remission induction therapy in Kyorin University Hospital were enrolled as the training cohort (Cohort 1), and at Toho University Sakura Medical Center were enrolled as the test cohort (Cohort 2). Clinical information of 44 items including age, sex, past/concomitant therapy for UC, clinical and endoscopic activity for UC, blood sample, and so on, in week 0, and clinical activity for UC in week 22 were collected. The outcome was defined as steroid-free clinical remission (SFCR) in week 22. Clinical remission was defined as the Lichtiger index (LI) of four or lower. Patients who terminated UST or needed surgery before week 22 were regarded as not achieving SFCR. Random Forest was used for the clinical data in Cohort 1 to identify the promising predictors to achieve SFCR in week 22. Prediction models were developed with these predictors and their prediction ability was validated in Cohort 2. Results 25 patients (SFCR/non-SFCR=13/12 patients) and 46 patients (SFCR/non-SFCR=27/19 patients) were enrolled as Cohort 1 and Cohort 2, respectively, in this study. First, we examined if a VDZ prediction model that we previously reported could work for predicting the UST efficacy. In Cohort 1, the model showed a positive predictive value (PPV) of 56.3% and a negative predictive value (NPV) of 62.5%. This supports the notion that a prediction model for UST should be different from that for VDZ. From the 44 clinical information, 10 clinical features with the highest contribution to SFCR with UST were detected in Cohort 1. These clinical features included serum albumin, monocyte fraction, height, MCV, total protein, LI, white blood cell count, MCHC, partial Mayo score, and CRP (Figure 1). With these 10 clinical features, the seven prediction models for UST were developed using different algorithms. These models were applied to Cohort 2 to validate the prediction ability. The model that showed the best prediction ability was a support-vector machine polynomial model with PPV 68.8% and NPV 71.4%, respectively. Conclusion We developed a new prediction model for the middle-term efficacy of UST in patients with UC. Further studies are needed to develop a prediction model for each molecular-targeted drug.

gastroenterology & hepatology

What problem does this paper attempt to address?

The problem this paper attempts to address is: developing a new predictive model to evaluate the mid-term efficacy of Ustekinumab (UST) in treating patients with ulcerative colitis (UC). The background information indicates that although UST, as an anti-IL-12/23p40 monoclonal antibody, has been widely used in the treatment of ulcerative colitis, there is currently no reliable model to predict its efficacy in UC patients. Researchers have previously demonstrated that machine learning methods can be used to identify predictive factors and develop predictive tools, particularly in the context of Vedolizumab (VDZ) treatment for UC. Therefore, in this study, they attempt to use machine learning techniques to identify key clinical features that can predict the efficacy of UST and to develop a predictive model based on these features.