An Automatic Framework for Detecting Autism Spectrum Disorder From EEG Signals Using TFD

Autism spectrum disorder (ASD) is an intricate neurodevelopmental disorder with many neurological problems. Social interaction and communication issues, repetitive behaviors, and limited interests are its main symptoms. Manual ASD diagnosis testing is prone to human error, time-consuming, and difficult owing to contamination from a number of factors. Electroencephalogram (EEG) signals are extensively utilized to identify ASD as they represent brain abnormalities. This study employed a novel method that included pre-processing, segmentation, and time-frequency distribution (TFD) of various algorithms such as short-time Fourier transformation (STFT), continuous wavelet transformation (CWT), and smoothed pseudo-Wigner-Ville distribution (SPWVD), which produced corresponding spectrograms, scalograms, and SPWVD-TFD. These TFDs are introduced into the DenseNet-121 and ResNet-101 pre-trained (ImageNet dataset) models, and then subsequently fed into the proposed ASD-Net. Deep learning networks (DLMs) models were utilized to identify ASD and Normal subject using these TFD images. We acquired a 97.35% mean accuracy utilizing the SPWVD-based TFD and ASD-Net model. When compared to the benchmark DenseNet-121 and ResNet-101, the developed convolution neural network (CNN) model with five convolution layers (CLs) not only needs less learnable parameters but is also computationally efficient and quick.