Wavelet Scattering Transform and Deep Learning Networks based Autism Spectrum Disorder Identification using EEG Signals

Abstract

Autism Spectrum Disorder (ASD), a neurological abnormality that influences how an individual perceives and interacts with others, which leads to issues with social interaction and communication. In accordance with the Centers for Disease Control and Prevention, 1 in every 44 children in USA is affected by ASD. The identification of ASD is based on behavioural characteristics and it generally takes a long time from the initial observation of behavioural signs to the final diagnosis, due to the complexity and diversity of ASD symptoms. The application of Electroencephalography (EEG) signals, recorded from 14 ASD affected children and 14 healthy controls, as a potential biomarker for ASD categorisation, was analysed in this study. After pre-processing, second-order Wavelet Scattering Transform (WST) coefficients were extracted from the EEG signals and Deep Learning (DL) based ASD detection networks (WST-ASDNets) were used for categorisation of ASD and control subjects. Long Short Term Memory Network (LSTM) based WST-ASDNet and Convolution Neural Network (CNN) based WST-ASDNet achieved accuracy of 94% and 92% respectively, in ASD subject identification. The results demonstrate that the proposed WST-ASDNets can efficiently classify ASD and the usage of WST coefficients extracted from EEG signals can be used as potential biomarker for ASD categorisation.