Machine Learning Predicts First Elective Bowel Resection in Newly Diagnosed Pediatric Crohn Disease Patients.

BACKGROUND: Pediatric onset Crohn disease (pCD) requires surgery more frequently than in adults. While risk scores and relative survival analysis are available, they do not accurately calculate the probability of surgery for pCD at an individual level. In this study, we propose a novel method to predict the time to perform a first elective bowel resection on individual pCD patients. Therefore, we machine learned models for predicting Individual Survival Distributions (ISDs) using clinical and laboratory data from a prospective, national pCD cohort (n=934). ISDs provide the probability of the decision to perform first bowel resection at each future point on individual pCD patients. METHODS: We developed two SuperLearners (SLs) that learned from baseline alone and baseline with week-10 longitudinal data, respectively. The SLs considered different base-survival-prediction models and used k-fold (internal) cross-validation to select the best model to estimate predictions. We used k-fold (external) cross-validation to estimate the quality of the learned model, seeking the model with the best (lowest) Truncated Mean Absolute Error (tMAEpo). RESULTS: Our models showed tMAEpo of 755 and 840 days for baseline and longitudinal models, respectively, which is the average number of days our prediction deviated from the true bowel resection times. The clinical features selected by our SLs aligned with previous studies and include penetrating and stricturing disease behaviors and perianal modifier. CONCLUSIONS: This study demonstrates the ability to produce an ISD predictor capable of forecasting elective bowel resection over future times. This novel approach can build helpful tools in planning future treatment for novel pCD patients.