ABSTRACT
Objective To analyze the correlation between echocardiography report narratives and the risk level of congenital heart disease in children,and to validate the feasibility and value of employing text mining technique in such task.Methods Echocardiography reports were retrospectively analyzed for 1 042 children with congenital heart disease.We adopted natural language processing (NLP) technique to generate features from the clinical narratives for machine learning algorithms.Decision trees were trained to predict the risk level of patients.Model performance was evaluated by means of classification accuracy and normalized mean absolute error (NMAE),which were averaged among 50 rounds of stratified 10-fold cross validation.By analyzing branches of the decision tree,we formulated the possible decision path of a clinician and identifyied the key information in the clinical narratives.Results Compared with the auto-generated 3-grams,the selected features yielded a better performance.After feature selection,the predict accuracy was improved from 32.82％ to 48.57％,while the NMAE reduced from 0.33 to 0.25.Conclusions Based on echocardiography report narratives,the risk levels of congenital heart disease in children can be evaluated by our model with an accuracy level of 75 ％.Echocardiographic terms that describe the lesion provide significant information to support the clinical decision making.