Towards an Automated Analysis of Social-Communication Behaviors in Autism Spectrum Disorder

Background: Clinical assessment of autism spectrum disorder (ASD) involves sophisticated and psychometrically sound assessment of characteristic behaviors such as emotional expressions, reciprocity, and atypical mannerisms. Available gold standard assessments require highly trained evaluators and an expert clinical eye – constraining crucial early detection and intervention needs in communities. Additionally, available tools have limited utility in research settings seeking objective assessments of subtle behaviors for mechanistic and treatment research.

Objectives: This collaborative project integrates artificial intelligence (AI) methods and clinical expertise to advance methods for detecting and monitoring social-communication behaviors associated with for ASD and related neuropsychiatric conditions.

Methods: Forty-seven individuals were enrolled in the pilot feasibility study to refine protocols and procedures for quantitative assessment of social-communication behaviors with approximately half (n=23) completing initial analysis and quality control checks. Participants included typically developing children (n=13), individuals with rare genetic disorders associated with ASD (n=7), and children with idiopathic ASD (n=28). System components included sensors such as off-the-shelf cameras and microphones, secure data analytic protocols to transfer data from clinics to “core lab”, and analytic software to turn sensor data into clinically relevant behavior codes. Data below are reported from images collected during administration of the Autism Diagnostic Observation Schedule-2 (ADOS-2, Lord, Rutter, DiLavore, Risi, Gotham, & Bishop, 2012). Evaluators wore Tobii Glass 2 in clinical assessment rooms equipped with 2-D and 3-D cameras. No additional modifications were made to the clinical environment or ADOS-2 protocol. Data was also collected during multi-modal clinical research protocols utilizing semi-structured behavioral assessments such as the Unstructured Imitation Assessment (UIA, McDuffie, et al, 2007), clinical assessments such as the Clinical Global Impressions Scale (CGI), and caregiver reports such as the Social Responsiveness Scale-2(SRS-2, Constantino, 2012).

Results: Analytic software measured several behavioral features of ASD including gaze contact, gaze distribution, responsive gaze, facial expressions of emotion, vocalizations, speech content, speech quality, and integration of vocalization and body language. Data from subset of verbally fluent, children with ASD (n=8) were compared to clinical scores from the ADOS2 Module 3 for initial clinical validation. Spearman’s rho was used to provide an estimate of the strength of the relationship between biometric data and qualitative clinical scores. These preliminary analyses suggest moderate to strong correlations between ADOS-2 summary codes (e.g. amount of reciprocal social communication) with patient behaviors (e.g. average duration patient looks at clinician), clinician behaviors (e.g. clinician initiates eye contact) as well as interactions (e.g. average duration of eye contact between patient and clinician). Analysis and system training is underway for the full scope of social-communication behaviors described above, as well as validation with clinical assessments and caregiver reports of social-communication behaviors.

Conclusions: An understanding of hallmark social-communication features of ASD has been crucial to advancing clinical care and research. Automated, objective assessments are exciting and promising developments but require standardization and clinical validation. Research applications may see more near-term benefits from use of digital biomarkers to elucidate relationships between observable behaviors and molecular & neurobiological variables; as well as measure target engagement & outcomes from clinical trials.