Detection of Autism Spectrum Disorder by a Fast Deep Neural Network

Autism spectrum disorder is a psychiatric illness that refers to a wide range of conditions caused by a biologically determined developmental disorder with onset of symptoms within the first three years of life. Autism can be diagnosed at any stage of life with problems beginning in childhood and continuing into adolescence and adulthood. Given the immense attraction gained by deep learning as one of the most successful paradigms for a plethora of real-world medical applications, in this paper we explore the possibility of using fast deep learning models for the detection of autism in children. To this end, random deep neural networks are one of the most important alternatives, in particular because they strike a good balance in the trade-off between accuracy and efficiency. We propose a deep neural architecture that employs the randomization of some parameters in a complex structure for the detection of autism spectrum disorder. The proposed approach is validated by using a threedimensional dataset consisting of body joint positions taken from videos of both suffering and sane children. To evaluate the classification performance of the proposed network, the latter is compared with a fully trainable, non-randomized version of the same model and with stateof-the-art binary classifiers applied to the same data. Numerical results show that the proposed method outperforms reference benchmarks in terms of accuracy and speed, demonstrating the inherent capabilities of the implemented system that makes use of such specific features.