Parameter Selection for Analyzing Conversations with Autism Spectrum Disorder

Tahiya Chowdhury,Veronica Romero,Amanda Stent

DOI: https://doi.org/10.21437/Interspeech.2023-1885

2024-01-18

Abstract:The diagnosis of autism spectrum disorder (ASD) is a complex, challenging task as it depends on the analysis of interactional behaviors by psychologists rather than the use of biochemical diagnostics. In this paper, we present a modeling approach to ASD diagnosis by analyzing acoustic/prosodic and linguistic features extracted from diagnostic conversations between a psychologist and children who either are typically developing (TD) or have ASD. We compare the contributions of different features across a range of conversation tasks. We focus on finding a minimal set of parameters that characterize conversational behaviors of children with ASD. Because ASD is diagnosed through conversational interaction, in addition to analyzing the behavior of the children, we also investigate whether the psychologist's conversational behaviors vary across diagnostic groups. Our results can facilitate fine-grained analysis of conversation data for children with ASD to support diagnosis and intervention.

Audio and Speech Processing,Artificial Intelligence,Sound

What problem does this paper attempt to address?

This paper discusses how to assist in the diagnosis of Autism Spectrum Disorder (ASD) children by analyzing their voice and language features in diagnostic conversations. The focus of the study is to identify the most informative acoustic-prosodic and linguistic indicators in different conversational tasks, as well as the behavioral differences of psychologists in conversations, to help identify ASD children and support diagnosis and intervention.

The study used assessment data from the Autism Diagnostic Observation Schedule (ADOS-2), which included 14 different subtasks involving 29 children aged 10 to 15 (14 with ASD and 15 typically developing children). They analyzed acoustic features (such as pitch and rhythm) and linguistic features (such as word frequency, social-cognitive vocabulary count, and semantic similarity).

The study found significant correlations between certain acoustic-prosodic indicators, linguistic indicators, and ASD diagnosis results, which were helpful in differentiating ASD from typically developing children in various conversational contexts. Additionally, differences were found in the language use of psychologists when conversing with ASD children.

Through correlation analysis and a random forest classifier, the study identified a minimal feature set that could effectively classify ASD and typically developing children. The classification results indicated that linguistic features were more discriminative of ASD than acoustic features, and combining both could further improve accuracy.

The limitations of the study include imperfect automatic preprocessing, inclusion of only two diagnostic groups (ASD and typically developing), and the use of English conversations only. Future work will expand to multimodal conversational data, examine other neurodevelopmental disorders and languages, and consider more conversational contexts.