DWTN-Net: Autism Evaluation Through P300 Detection Using a Discrete Wavelet Transform Neural Network

People with Autism Spectrum Disorder (ASD) suffer from disability in joint attention and also lack the ability of social interaction. One of the most common event-related potential is p300, which can be used in the treatment of ASD patients. Identification of p300 peak from the Electroencephalogram (EEG) signal of the ASD patients is a challenging task since in this case, the p300 wave, unlike the neuro-typical (healthy) subjects, may not be clearly distinguishable from the non-p300 waves. The EEG signals available in this case are obtained from different locations of the brain and thus contain spatio-temporal variation that needs to be effectively captured. Considering the potential of deep neural network based models in handling the complex spatio-temporal pattern, in this paper, a two-branch convolutional neural network (CNN) based architecture is proposed, which can efficiently identify the presence of p300 wave for the ASD patients. The proposed architecture "DWTN-Net" (Discrete Wavelet Transform Neural Network), is a two branch CNN model consisting of a spatial CNN branch and a temporal CNN branch. Depthwise convolutional layers and separable convolutional layers are also used in order to keep the number of training parameters small. A publicly available dataset "BCIAUT-P300", composed of the brain signals of ASD patients, is used for the evaluation of the proposed method. The signals were Discrete Wavelet Transformed (DWT) into detailed and approximation coefficients corresponding to high and low frequency bands before feeding them to train the model. This resulted in a better performance for the model which is demonstrated in the result section.