Objective To explore the value of machine learning(ML)models based on ultrasonic texture features(TF)of coronary artery for predicting incomplete Kawasaki disease(IKD)in children.Methods Forty-eight children with IKD and 48 children without KD(non-KD)were enrolled with propensity score matching and divided into training set(n=67,34 cases of IKD and 33 cases of non-KD)and test set(n=29,14 of IKD and 15 of non-KD)at the ratio of 7∶3.Based on clinic-laboratory indicators(C-L)in training set and TF obtained with texture analysis of coronary artery ultrasound images,the optimal C-L-related features and TF were selected.Based on the optimal C-L correlated features,TF and their combinations,6 ML models,including random forest(RF),support vector machine(SVM),logistic regression(LR),gradient boosting decision tree(GBDT),decision tree(DT)and eXtreme gradient boosting(XGBoost)were respectively constructed for predicting IKD in children.The models were then trained in training set and validated in test set,and the best C-L ML,TF ML and C-L-TF ML models were selected.The area under the curve(AUC)of the best ML models were compared,and the clinical value of the best TF ML model was observed with decision curve analysis(DCA).Results Totally 3 optimal C-L related features and 8 optimal TF were selected.Among the constructed C-L ML,TF ML and C-L-TF ML models,C-L-LR model,TF-LR model and C-L-TF-SVM model were the optimal ones,with AUC in training set of 0.891,0.985 and 0.965,while in test set of 0.676,0.971 and 0.948,respectively.No significant difference of AUC was found between TF-LR model and C-L-TF-SVM model in both training set and test set(both P＞0.05),which were both greater than those of C-L-LR model(all P＜0.05).TF-LR model achieved higher clinical benefits in both training set and test set.Conclusion Ultrasound TF-LR model of coronary artery could be used to effectively predict IKD in children.

Objective To explore the value of machine learning(ML)models based on ultrasonic texture features(TF)of coronary artery for predicting incomplete Kawasaki disease(IKD)in children.Methods Forty-eight children with IKD and 48 children without KD(non-KD)were enrolled with propensity score matching and divided into training set(n=67,34 cases of IKD and 33 cases of non-KD)and test set(n=29,14 of IKD and 15 of non-KD)at the ratio of 7∶3.Based on clinic-laboratory indicators(C-L)in training set and TF obtained with texture analysis of coronary artery ultrasound images,the optimal C-L-related features and TF were selected.Based on the optimal C-L correlated features,TF and their combinations,6 ML models,including random forest(RF),support vector machine(SVM),logistic regression(LR),gradient boosting decision tree(GBDT),decision tree(DT)and eXtreme gradient boosting(XGBoost)were respectively constructed for predicting IKD in children.The models were then trained in training set and validated in test set,and the best C-L ML,TF ML and C-L-TF ML models were selected.The area under the curve(AUC)of the best ML models were compared,and the clinical value of the best TF ML model was observed with decision curve analysis(DCA).Results Totally 3 optimal C-L related features and 8 optimal TF were selected.Among the constructed C-L ML,TF ML and C-L-TF ML models,C-L-LR model,TF-LR model and C-L-TF-SVM model were the optimal ones,with AUC in training set of 0.891,0.985 and 0.965,while in test set of 0.676,0.971 and 0.948,respectively.No significant difference of AUC was found between TF-LR model and C-L-TF-SVM model in both training set and test set(both P＞0.05),which were both greater than those of C-L-LR model(all P＜0.05).TF-LR model achieved higher clinical benefits in both training set and test set.Conclusion Ultrasound TF-LR model of coronary artery could be used to effectively predict IKD in children.