Exploiting visual behaviour for autism spectrum disorder identification

In this contribution, a model for revealing the presence of autism spectrum disorder by exploiting visual information is developed. This condition is characterized by a deficit in social behaviour and nonverbal interactions such as specific facial expressions, reduced eye contact, and body gestures. Advancements in multimedia technologies can help in understanding symptoms for early detection of the disorder. In the proposed model, both the image content and the viewing behaviour are used for defining relevant features to be used in a machine learning-based classifier. A training phase is realized by taking multiple images and scanpaths representing the viewing behaviour of persons affected and not by the disorder. The influence of specific objects in the scene is considered. Finally, the number of fixations towards centre of the scene and duration for which a subject looked at the central area is also considered. A decision tree based classifier is used for training the model. The achieved results show that by taking into account the semantic and image features extracted from content, fixation, and center-bias, it is possible to estimate the presence of autism spectrum disorder. The results obtained in the performed experiments are promising even if they show room for improvement.