Structure equation model and neural network analyses to predict coronary artery lesions in Kawasaki disease: a single-centre retrospective study

Abstract

A new method to predict coronary artery lesions (CALs) in Kawasaki disease (KD) was developed using a mean structure equation model (SEM) and neural networks (Nnet). There were 314 admitted children with KD who met at least four of the six diagnostic criteria for KD. We defined CALs as the presence of a maximum z score of ≥ 3.0. The SEM using age, sex, intravenous immunoglobulin resistance, number of steroid pulse therapy sessions, C-reactive protein level, and urinary β2-microglobulin (u-β2MG/Cr) values revealed a perfect fit based on the root mean square error of approximation with an R2 value of 1.000 and the excellent discrimination of CALs with a sample score (SS) of 2.0 for a latent variable. The Nnet analysis enabled us to predict CALs with a sensitivity, specificity and c-index of 73%, 99% and 0.86, respectively. This good and simple statistical model that uses common parameters in clinical medicine is useful in deciding the appropriate therapy to prevent CALs in Japanese KD patients.