Machine Learning Based Prediction Model for Closed-Loop Small Bowel Obstruction Using Computed Tomography and...

The aim of the study was to develop a prediction model for closed-loop small bowel obstruction integrating computed tomography (CT) and clinical findings.

The radiology database and surgical reports from 2 suburban teaching hospitals were retrospectively reviewed for patients undergoing surgery for suspected closed-loop small bowel obstruction (CLSBO). Two observers independently reviewed the CT scans for the presence of imaging features of CLSBO, blinded to the surgically confirmed diagnosis and clinical parameters. Random forest analysis was used to train and validate a prediction model for CLSBO, by combining CT and clinical findings, after randomly splitting the sample into 80% training and 20% test subsets.

Surgery confirmed CLSBO in 185 of 223 patients with clinically suspected CLSBO. Age greater than 52 years showed 2.82 (95% confidence interval = 1.13-4.77) times higher risk for CLSBO (P = 0.021). Sensitivity/specificity of CT findings included proximal dilatation (97/5%), distal collapse (96/2%), mesenteric edema (94/5%), pneumatosis (1/100%), free air (1/98%), and portal venous gas (0/100%). The random forest model combining imaging/clinical findings yielded an area under receiver operating curve of 0.73 (95% confidence interval = 0.58-0.94), sensitivity of 0.72 (0.55-0.85), specificity of 0.8 (0.28-0.99), and accuracy of 0.73 (0.57-0.85). Prior surgery, age, lactate, whirl sign, U/C-shaped bowel configuration, and fecalization were the most important variables in predicting CLSBO.

A random forest model found clinical factors including prior surgery, age, lactate, and imaging factors including whirl sign, fecalization, and U/C-shaped bowel configuration are helpful in improving the prediction of CLSBO. Individual CT findings in CLSBO had either high sensitivity or specificity, suggesting that accurate diagnosis requires systematic assessment of all CT signs.