Predicting surgical intervention in pediatric intussusception using machine learning model

Abstract

Aim: To develop and validate a model predicting surgical treatment of intussusception in children.

Methods: Design: Retrospective study of charts and development of a model for predicting surgical treatment using logistic regression and machine learning using''Knime'' platform.

Setting: Data collection occurred in the Department of Pediatric Surgery between January 2013 and December 2022.

Patients: Children aged less than 3 years old with the diagnosis of ileocolic intussusception.

Results: One hundred and nine children were assigned to the training set, and 47 were assigned to the validation set. There were no significant differences between the two sets in clinical characteristics and surgical reduction. Surgical reduction was performed in 64 patients in the training set and 23 patients in the validation set (p=0.259). The univariate analysis showed that the duration of symptoms, mental state, palpable abdominal mass, bloody stools, elevated white blood cells, intraperitoneal effusion on ultrasound, and mass length were significantly associated with surgical treatment. After Logistic regression, bloody stools (p=0.033; OR=2.61), the duration of symptoms (p=0.028; OR=1.02), and the length of the intussusception (p=0.014; OR=1.265) were identified as independent risk factors for surgical treatment. The clinic-pathologic risk factors incorporated in the machine learning model were bloody stools, the duration of symptoms, and the length of the intussusception. This model was highly predictive, with a sensitivity and specificity of 95% for the SVM-derived model.

Conclusions: This model may be applied to facilitate pre-surgery decisions for children with intussusception. Larger prospective multicenter studies are needed to validate the model.

Keywords: Intussusception; Prediction; Surgery.