Comparison of Different Interaction Formats for Automatized Analysis of Symptoms in Children with Autism Spectrum Disorder

Abstract

Observation and assessment of interactional difficulties in children with suspected Autism Spectrum Disorder (ASD) are part of the gold standard in the ASD diagnostic process. The risk for a diagnosis of ASD can be assessed involving the three typical symptom categories speech/language, facial expression, and interaction. However, identifying ASD turns out to be staff-intensive and time consuming and hereby often delaying the start of therapy and support for the child and the family. Automatized analyses, for example of eye contact or mimic response, could facilitate early screening and hereby contribute to more efficient diagnostic routines. Such automatized screening can build on advanced language and mimic processing technology already available. However, the validity of automated screening depends on the elicitability of behavior that signifies ASD. Therefore, a proof of principle is required to demonstrate that a mediated approach does indeed elicit ASD symptoms that are usually observed during natural face-to-face interaction. Research aim: The goal of this paper is to show preliminary results on the validation of diagnostic comparability of real and simulated interactions between a child with ASD and an examiner. Specifically, the simulated interactions would be most useful for an automated screening if they could be based on pre-recorded stimuli and not only be mediated in real time. Method: For such a proof of principle approach we apply a within-design with five conditions in which the emerging symptoms in children diagnosed with ASD are contrasted with regard to speech, facial expressions, and communication behavior. Both, the authenticity of the communication situation (face-to-face vs. video call vs. pre-recorded video) as well as the child's interlocutor (real person vs. video recording vs. digital avatar) are varied and conditions fully counterbalanced. After each condition, the child is prompted to reflect about the perceived authenticity of the test situation and their interactional involvement. Inclusion criteria include boys with an established diagnosis of ASD at elementary school age who are verbally fluent. Results and implications: Although data collection is still ongoing we present preliminary results based on two children. The observations suggest that digitally delivered content is highly appealing and perceived as appropriate for children with ASD. Further data collection and analyses will inform whether typical and relevant ASD symptoms in the participants will be observable in mediated and even fully automatized conditions. However, our first impressions already demonstrate a great potential for automated measurements with low personnel effort.