Machine Learning Predicts the Need for Surgical Intervention in Adhesive Small Bowel Obstruction

Abstract

Objectives: To explore the predictive performance on the need for surgical intervention in patients with adhesive small bowel obstruction (ASBO) using machine-learning (ML) algorithms and investigate the optimal timing for transition to surgery. Methods: One hundred and six patients with ASBO who initially underwent long transnasal intestinal tube (LT) decompression were enrolled in this retrospective study. Traditional logistic regression analysis and ML algorithms were used to evaluate the risk of need for surgical intervention. Results: Non-operative management (NOM) by LT decompression failed in 28 patients (26%). Multivariate logistic regression analysis identified a drainage volume 665 ml via LT on day 1, interval between ASBO diagnosis and LT intubation, and small bowel dilatation at 48 h after LT intubation to be independent predictors of transition to surgery (odds ratios 7.10, 1.42, and 19.81, respectively; 95% confidence intervals 1.63-30.94, 1.00-2.02, and 3.04-129.10; P-values 0.009, 0.047, and 0.002). The random forest algorithm showed the best predictive performance of five ML algorithms tested, with an area under the curve of 0.889, accuracy of 0.864, and precision of 0.667 in the test set. 97.4% of patients without transition to surgery (n=78) had passes of first flatus until three days. Conclusions: This is the first study to demonstrate that ML algorithm can predict the need for surgery in patients with ASBO. The guideline recommended period for initial NOM of 72 h seems to be reasonable. These findings can be used to develop a framework for earlier clinical decision-making in these patients.

Keywords

• adhesive small bowel obstruction
• long transnasal intestinal tube
• machine-learning
• non-operative management
• surgery