Machine Learning Model For Predicting The Response To Mesalamine Treatment In Crohn’s Disease

Objective:Mesalamine is a common treatment for Crohn’s disease in clinical practice,but its efficacy is controversial.This study aimed to develop a machine learning model incorporating body composition features to improve prediction of mesalamine treatment response in Crohn’s disease,and to explore the factors related to the efficacy of mesalamine.Methods:107 patients with Crohn’s disease treated with mesalamine were analyzed retrospectively,diagnosed in Xiangya Hospital of Central South University from March 2013 to August 2020.According to the therapeutic effect of mesalamine,these patients were divided into response group and non response group.The general demographic data,clinical characteristics data,laboratory data and body composition data of patients were collected.Analyze whether there is statistical difference between the two groups.Then,107 patients were separated randomly into a training and validation group.The prediction models were developed using machine learning methods(least absolute shrinkage and selection operator,random forest,and support vector machine)using just clinical/laboratory values(SVM-Clinic-Labtest),using computed tomography body composition features and clinical/laboratory values(SVM-Combined),or using multivariable logistic regression(LR).The performance of the model was evaluated by the Receiver operating characteristic curve,correction curve and decision curve of the subjects.Results:Onset age(OR=1.031,95%CI: 1.001 to 1.062),white blood cell count(OR=0.867,95%CI: 0.768 to 0.978),neutrophil count(OR=0.869,95%CI: 0.758 to 0.995),monocyte count(OR=0.118,95%CI: 0.029 to0.486),erythrocyte sedimentation rate(OR=0.987,95%CI: 0.977 to0.998),C-reactive protein(OR=0.989,95%CI: 0.979 to 1.000),globulin(OR=0.893,95%CI: 0.834 to 0.957),Subcutaneous fat area(OR= 1.009,95%CI: 1.001 to 1.017)and the ratio of visceral fat to skeletal muscle area(OR=0.022,95%CI: 0.002 to 0.318)were predictors of mesalazine treatment response.Globulin level and the ratio of visceral fat to skeletal muscle area were independent predictors.After incorporating body composition features,the SVM-Combined model showed good discrimination between the responder and non-responder groups,with an area under the curve of 0.957(95% CI: 0.910 to 1.000)in the training group and 0.953(95% CI: 0.883 to 1.000)in the validation group.This was significantly higher than for the SVM-Clinic-Labtest model(area:training group,0.910 [95% CI: 0.841 to 0.980];validation group,0.910[95% CI: 0.799 to 1.000]),and LR model(area: training group,0.788[95% CI: 0.686 to 0.890];validation group: 0.625 [95% CI: 0.413 to0.837]).Favorable calibration performance and clinical applicability of the machine learning model were observed using calibration and decision curve analysis.Conclusions:1.The model constructed by least absolute shrinkage and selection operator,random forest and support vector machine is better than multivariable logistic regression in predicting the response of patients with Crohn’s disease to mesalamine;2.Compared with using clinical and laboratory characteristics,incorporating body composition characteristics into the model further improves the prediction performance;3.Globulin level and the ratio of visceral fat to skeletal muscle area are independent predictors of mesalamine treatment response,which are negatively correlated with the efficacy.