18 - A DEEP LEARNING APPROACH TO PREDICT CORONARY ARTERY ANEURYSM IN LARGE NUMBER OF PEDIATRIC PATIENTS WITH KAWASAKI DISEASE

Aim: To predict coronary artery aneurysm using an artificial intelligence model developed based on laboratory and demographic variables before an initial IVIG administration for the treatment of acute Kawasaki disease.

Methods: Data collected through the nationwide surveys (2012-2014 and 2015-2017) in Korea were reviewed. Data from 16,889 patients who had all the demographic and laboratory variables and coronary artery outcome data required to construct the model, were analyzed. Ten percent of the subjects were randomly selected and classified into the test group, and the remaining 90% of the subjects were categorized into the training group. Among the subjects in the training group, 20% of them were randomly selected for tuning the hyper-parameters. We used Scikit-learn package in Python 3.7 environment for construction of neural network model. The developed model underwent receiver operating characteristic (ROC) curve analysis.

Results: A deep neural network model was constructed based on the multi-layer perceptron classifier algorithm. SHAP (SHapley Additive exPlanations) feature learning was conducted concurrently. The results of ROC curve analysis of the model were 0.8419 of area under the curve, 0.8385 of accuracy, 0.4598 of sensitivity, and 0.9499 of specificity. Ten input variables with relatively large mean absolute value of SHAP were (in descending order): serum protein level, total bilirubin level, complete presentation of disease, N terminal-pro BNP level, sodium level, platelet number, pyuria, age, sex, and serum glutamic-oxaloacetic transaminase level.

Conclusion: A deep neural network model with a high level of specificity for predicting coronary artery aneurysms was developed. It could be helpful in identifying an acute Kawasaki disease patient needing an intensified initial treatment.