Abstract
OBJECTIVES: To develop a model incorporating computed tomography (CT) radiomic features and clinical parameters for predicting bronchiolitis obliterans (BO) with adenovirus pneumonia in children. METHODS: A total of 165 children with adenovirus pneumonia between October 2013 and February 2020 were enrolled retrospectively. Among them, BO occurred in 70 patients, and the remaining 95 patients did not have BO. These children were stratified into training and testing groups at a ratio of 7:3. Manual segmentation of lesions in baseline CT images during acute pneumonia was performed to extract radiomic features. Multiple statistical methods were used to determine the best radiomic features. Combined models based on radiomic and clinical features were established via logistic regression (LR), random forest (RF), and support vector machine (SVM) algorithms. Model performance was evaluated via the area under the receiver operating characteristic curve (AUC). RESULTS: A total of 2,264 radiomic features were extracted from the lesions, from which 10 optimal radiomic features were ultimately selected. The length of hospitalization, number of pneumonia lobes, and optimal radiomic features were incorporated into the combined models. In the training group, the AUCs of the combined LR, RF and SVM models were 0.946, 0.977, and 0.971, respectively; while in the testing group, they yielded AUCs of 0.890, 0.859, and 0.885, respectively. The predictive performance of these combined models surpassed that of the radiomic and clinical models. CONCLUSION: Combining CT-based radiomic features with clinical parameters can offer an effective noninvasive model to predict BO in children with adenovirus pneumonia.