An Echocardiogram‐Based Ultrasound Radiomic Model Predicts Recovery of Coronary Artery Diameter in Kawasaki Disease

OBJECTIVE: To develop a predictive model based on echocardiogram-based ultrasound radiomic features for the recovery of coronary artery diameter in Kawasaki disease (KD) with coronary artery lesions (CALs). METHODS: A total of 206 children with acute KD and left CALs who underwent coronary ultrasound before and 8 weeks after treatment were retrospectively enrolled. The patients were randomly divided into training and testing sets at a 7:3 ratio. Ultrasound radiomic features were extracted from regions of interest delineated along the coronary artery wall on pretreatment ultrasound images. Machine learning models based on five different algorithms, including logistic regression (LR), support vector machine (SVM), random forest, extra trees, and light gradient boosting machine (LightGBM), were established to predict the recovery of coronary artery diameter. Receiver operating characteristic (ROC) curve analysis was performed to evaluate the diagnostic efficiency of the models, and the area under the curve (AUC), sensitivity, specificity, and accuracy were calculated. Decision curve analysis (DCA) was used to evaluate the clinical benefits of the use of the models. RESULTS: Based on Z score normalization of coronary artery diameters, diameters returned to normal in 77 patients in the training set and 37 in the testing set, whereas 67 and 25 patients, respectively, showed no such recovery after treatment. A total of 1562 ultrasound radiomic features were extracted; following the use of feature selection processes, 9 optimally performing features were employed for machine learning model development. The LightGBM model showed the best diagnostic performance, with an AUC, sensitivity, and specificity of 0.867 (95% CI: 0.811-0.923), 59.7%, and 97.0%, respectively, in the training set and of 0.857 (95% CI: 0.756-0.958), 86.5%, and 76.0%, respectively, in the testing set. DCA revealed that over a threshold range of 0.4 ~ 0.8, the model offers good clinical benefits. CONCLUSION: Machine learning models can predict the recovery of coronary artery diameter in KD and thus may offer important value for clinical decision-making.