Objective Measurements of Children’s Gaze and Vocal Behavior Associated with Clinician-Rated Autism Symptoms

• In Event: Objective (Automated) Measurement of Infant and Child Behavior: Reliability, Validity, and Discovery

Abstract

Autism spectrum disorder (ASD) is characterized by social communication difficulties, such as eye contact, and restricted and repetitive patterns of behavior, including atypical vocal qualities. Current best practice for assessing ASD symptoms is based on expert, but subjective, clinician observation during the Autism Diagnostic Observation Schedule-2 (ADOS-2). Automated objective detection of children’s social communication and restricted/repetitive behaviors from video has the potential to provide quantifiable measures of key behavioral features of ASD. The current study examined associations between ASD symptom severity during the ADOS-2 and objectively quantified gaze behaviors (i.e., social gaze and gaze involving smile) and vocalization features (i.e., vocalization pitch, number of consonants and vowels) measured with a low-cost adult-worn audio/video recorder (see Figure 1).

Sixty-six children (49 boys, Mage=39.97mo, SD=10.58) with suspected ASD (61 confirmed ASD) were administered the ADOS-2. A hypothesis-blind examiner provided social affect (SA), restricted and repetitive behavior (RRB), and total calibrated severity scores (CSS). Children’s gaze and gaze involving smile directed toward adults during the assessment were recorded with a camera (Pivothead/Orca) contained in eye-glasses, which were processed with a pipeline of algorithms to detect gaze (Chong et al., 2020) and smiling. Measures included gaze duration expressed as a proportion of ADOS-2 duration and proportion of gaze duration involving smile. Child vocalizations and cries were detected from the camera audio using LENA and processed by PRAAT to determine the fundamental frequency (F0) and by Sphinx-4 to detect the average count of consonants and vowels per utterance (ACPU).

In the model predicting RRB CSS, Bayesian model averaging (BMA) indicated both child ACPU (.983) and child cry F0 (.968) had high marginal inclusion probabilities. The model with the highest posterior probability (.369) includes both child cry mean F0, b=.71, t(65)=3.83, p<.001 and child ACPU, b=-1.37, t(65)=-3.56, p<.001, F(2,63)=15.50, p<.001, adjusted R2=.31 (see Figure 2). In the model predicting SA CSS, BMA indicated that proportion of time gazing at parent (.530) and proportion of gaze at parent involving smile (.581) had high marginal inclusion probabilities, but the substantive model with the highest posterior probability included only social gaze at the parent (.200), b=-185.44, t(58)=-2.66, p=.01, F(1,57)=7.08, p=.01, adjusted R2=.10. Finally, in the model predicting Total CSS, BMA indicated that proportion of time gazing at parent (.678) had the highest marginal inclusion probability and was the only predictor in the model with the highest posterior probability (.282), b=-171.10, t(58)=-2.86, p=.006, F(1,57)=8.16, p=.006, adjusted R2=.11.

Children who produced higher pitch of cries and used fewer consonants and vowels per utterance received higher RRB symptom severity scores. Children who demonstrated more gaze toward parent received higher SA symptom severity scores. Social gaze at the parent emerged as the strongest overall predictor of autism severity. The objective measurement of key behavioral features of ASD appear to have the potential to produce quantitative indices of ASD symptoms.

Authors

• Jacquelyn Michele Moffitt, University of Miami
  Presenting Author

• Lynn K Perry, University of Miami

• Yeojin Amy Ahn, University of Miami

• Jennifer Durocher, University of Miami

• Daniel Messinger, University of Miami