Estimating Social Responsiveness of Adults with ASD in Group Conversations

Abstract

ASD (autism spectrum disorder) is a developmental disorder with problems in social communication. The diagnosis of ASD is usually made in childhood, but there are some individuals who have communication problems but are not diagnosed until later in life. Early detection of such individuals and appropriate treatment and support are important issues. The Social Responsiveness Scale was developed to objectively measure symptoms associated with ASD, making it suitable for ASD screening. In this study, we propose machine learning models that estimate SRS-2 scores using multimodal information from group communication videos. First, as an analysis of the communication characteristics of adults with ASD, we examined the features that significantly correlate with the SRS-2 scores. Next, based on these results, we created estimation models using audio, facial, and language features. Ablation studies revealed that combining features from multiple modalities improved estimation performance.