Novel Computational Analysis of Prosody in ASD and the Broad Autism Phenotype

点击翻译按钮获取中文摘要

Abstract Pragmatic (i.e., social) language impairments are a core feature of Autism Spectrum Disorder (ASD), and differences have also been observed more subtly in the Broad Autism Phenotype (BAP; a cluster of subclinical features related to ASD, which are believed to reflect genetic liability in clinically unaffected relatives). Prosody is a central component of pragmatics, which includes aspects of speech and language that modulate and enhance meaning at grammatical, pragmatic, and affective levels including stress and intonation, which are conveyed by changes in fundamental frequency, intensity, and rate. Atypical prosody in ASD has been identified as the most prominent feature that immediately identifies an individual with ASD as “odd” compared to typically developing peers, causing significant obstacles to social interaction and integration1-3. More subtle differences in prosody have also been observed among parents of individuals with ASD, and may serve as a key marker of ASD genetic risk, measurable in affected and unaffected individuals. Because speech samples may be easily obtained, identification of key prosodic profiles in ASD and profiles reflecting genetic liability to ASD, could have widespread implication as a biomarker for detection of ASD risk, and as a tool for assessment and interventions focused on this clinically significant feature of ASD. Limitations in current work pose significant barriers to the objective and efficient characterization of prosody, where current methods typically rely on subjective perceptual ratings which, though clinically valid, are difficult to obtain and apply objectively in treatment contexts, and are unfeasible for application with large samples. Moreover, without extensive training, perceptual rating methods are not adequately sensitive for capturing important variation and heterogeneity in ASD prosodic profiles, or identifying the often quite subtle yet biologically meaningful prosodic differences that may be observed in clinically unaffected relatives. These factors together impose substantial barriers to reproducibility, limit scalability, and render prosodic characterization unfeasible for use by clinicians. This project attempts to address these challenges by applying sophisticated computational modeling of extensive existing speech and language samples collected through a larger companion project (R01DC010191, PI: Losh), to characterize prosodic profiles in ASD and in parents. In Preliminary Data, we demonstrate evidence of distinct prosodic profiles of individuals with ASD and parents, along with relationships to broader pragmatic language abilities and neural processing of speech sounds in ASD and parents, supporting the goals of this project to apply sophisticated computational tools to speech data obtained across multiple contexts in order to 1) identify prosodic profiles that characterize ASD and the BAP in parents, and 2) examine how prosodic profiles relate to broader clinical-behavioral phenotypes, and aggregate within families, potentially contributing phenotypic signatures of clinically and etiologically more homogeneous subgroups.