Currently submitted to: JMIR Pediatrics and Parenting
Date Submitted: Jan 11, 2020
(closed for review but you can still tweet)
Prediction Model for children's risk of autism spectrum disorder: a method based on deep learning artificial neural network
Over the past two decades, the prevalence of autism spectrum disorder (ASD) among children worldwide has been rising rapidly. While Lab-based test were commonly used in the clinical diagnosis of autism, it is not feasible to test all new-borns for autism.
This study focused on the risk factors of children with autism in the pre-pregnancy, early gestational (first month to six month) and perinatal stages as proposed by clinical epidemiological studies, and applied a deep learning artificial neural network to establish an early warning model of children with autism.
A multivariate questionnaire on risk factors for autism in children was developed in this study. Parents of children with (n = 137) and without autism (n = 186) in five Chinese cities were investigated. The data were split into two completely independent datasets: training set (80%) and test set (20%). The sensitivity, specificity, and accuracy of four risk factor set (RFS) models were compared. The AUCs of four prediction models were also compared.
The sensitivity and accuracy values of the RFS-B model were superior to those of the other three models. The specificity of the RFS-C was superior to that of the other three models. The AUCs of the four RFS models were computed to be 0.876, 0.905, 0.850 and 0.870.
The results of the present study indicate that the deep learning artificial neural network has potential value in early risk prediction for children with autism.