Artificial intelligence-powered early identification of refractory constipation in children.

Background: Children experiencing refractory constipation, resistant to conventional pharmacological approaches, develop severe symptoms that persist into adulthood, leading to a substantial decline in their quality of life. Early identification of refractory constipation may improve their management. We aimed to describe the characteristics of colonic anatomy in children with different types of constipation and develop a supervised machine-learning model for early identification. Methods: In this retrospective study, patient characteristics and standardized colon size (SCS) ratios by barium enema (BE) were studied in patients with functional constipation (n=77), refractory constipation (n=63), and non-constipation (n=65). Statistical analyses were performed and a supervised machine learning (ML) model was developed based on these data for the classification of the three groups. Results: Significant differences in rectum diameter, sigmoid diameter, descending diameter, transverse diameter, and rectosigmoid length were found in the three groups. A linear support vector machine was utilized to build the early detection model. Using five features (SCS ratios of sigmoid colon, descending colon, transverse colon, rectum, and rectosigmoid), the model demonstrated an accuracy of 81% [95% confidence interval (CI): 79.17% to 83.19%]. Conclusions: The application of using a supervised ML strategy obtained an accuracy of 81% in distinguishing children with refractory constipation. The combination of BE and ML model can be used for practical implications, which is important for guiding management in children with refractory constipation.