Feasibility of a Mobile Phone-Delivered Study of Social and Emotional Behaviors in Young Children at Risk for Autism

Background: Effective and accessible screening of young children for autism requires reliable, valid, and efficient tools for evaluation of a child’s social-emotional behaviors. Coupling parent-report measures with evidence based observational assessments significantly improve the accuracy of screening and diagnosis. Current validated tools for structured assessment of children’s behavior are expensive, training intensive, and time-consuming to administer. Development of tools for low-cost, automatic, objective, and quantifiable measurements of young children’s observable behaviors has the potential to increase access to screening and early intervention for children around the globe.

Objectives: To test the feasibility of an iPhone application (“app”) to engage parents directly in the collection of survey and video data about their children within their homes and use automatic computer vision algorithms to quantify children’s emotions and attention in the uploaded videos.

Methods: Autism & Beyond, a study of young children ages 12 to 72 months, is an iOS app built on Apple’s ResearchKit framework. The iPhone app includes self-guided econsent, parent-report surveys, and video activities with the child. The camera on the device recorded the child’s behaviors and emotions as s/he watches short film clips while sitting on the parent’s lap. Autism risk status of children in the study was based on parent-reported autism diagnosis and/or a positive M-CHAT-R/F (children 16-30 months). Behavioral variables automatically extracted from the videos include positive emotions, attention, and social referencing. Here we present results from first 6 months of the study.

Results: In the first sixth months of the study, 878 families met inclusion criteria. 295 children were in the high risk autism group (parent report of autism diagnosis: 218, high MCHAT: 51, Both: 26). Mean age of high risk children was 43.2 months (SD 14.9); mean age in the low risk group was 40.6 months (SD 16.4). The majority of parents were non-Hispanic/Latino Caucasians (65%), college educated (60%), and employed (63%). 862 participants completed additional surveys about child behaviors. 403 parents agreed to upload full videos of their child and completed at least one video activity with their child. 300 of 878 uploaded only facial landmarks extracted from their child’s video. The child’s face was identified in 84-92% of the video frames. Children in the high risk autism group showed decreased mean positive affect (-7%, SD 4%, p=0.04 in one stimulus) and lower predicted probability of social referencing (9% vs 24% p=0.02 at oldest ages in one stimulus).

Conclusions: Over 6 months, nearly 900 parents downloaded the study app, consented to participate, and completed study tasks in their homes. The quality of the video data is excellent. Automatic computer vision coding of emotions and head position in the videos enabled us to identify differences between children with and without a high risk for autism. These results are a first step toward the development of globally accessible, affordable, and easy to use app-based tools to improve the science and practice of screening for autism. Collaborations in South Africa and Argentina are enabling us to expand the reach of Autism & Beyond beyond the US.