Identify autism spectrum disorder via dynamic filter and deep spatiotemporal feature extraction

Abstract Early intervention and treatment are crucial for individuals with autism spectrum disorder (ASD). However, it is challenging to identify individuals with ASD at an early age, i.e. under 3 years old, due to the lack of an effective and objective identification method. The mainstream clinical diagnosis relies on long-term observation of children’s behaviors, which is time-consuming and expensive, and thus how to accurately and quickly distinguish children with ASD in early childhood has become a critical issue. In this paper, we propose an eye movement based model to identify children with ASD. Specifically, children are required to freely observe some images. At the same time, their eye movements are recorded to analyze. Both the observed image and eye movements are input into our model. The input data are processed by the embedding layer, dynamic filters and LSTM block, respectively. Eventually, the spatiotemporal features are extracted to identify the eye movements belonging to a child with ASD or a typically developed child. Experiments on the Saliency4ASD dataset demonstrate that the proposed model achieves state-of-the-art performance in identifying children with ASD.