Spatial structure based deep feature fusion network for autism spectrum disorder classification

Autism spectrum disorder (ASD) is a neuro-developmental disorder that has serious impact on person’s social interaction, communication, and behavior. Non intrusive detection through analysis of electroencephalogram (EEG) signals with deep learning techniques has been the prominent area of research in ASD diagnosis. The current deep learning techniques processed EEG signals without considering their spatial structure. They treat EEG features as independent or grid-like sequences, ignoring the non-Euclidean structure of brain connectivity. As the result, the features do not capture the intricate relationship between signals obtained from multiple electrodes considering both spatial and long term temporal dependencies. To solve this problem, this work proposes a spatial structure based feature fusion network (SSFFN) which considers two attributes of spatial proximity and functional connectivity to capture local and long range interactions between brain regions as fused features. These fused features are used for ASD classification using long short term (LSTM) predictor considering the long term temporal dependencies. Experimental analysis of proposed solution against Sheffield dataset and KAU dataset showed a 1% higher accuracy and 4.4% higher recall compared to most recent time frequency synergy networks.