Prediction for children with autism spectrum disorder based on digital behavioral features during free play.

Background: Play is an indispensable and meaningful activity in children's daily life. Research has shown that autistic children often exhibit differences in play development. The core traits of autism, such as distinct patterns in social interaction and communication, focused interests, and repetitive behaviors, frequently manifest in their play. Therefore, play may serve as an insightful measure of these differences. Unlike previous studies focusing on play behaviors only, we explored other behaviors associated with autism during free play, and constructed a clinical prediction model for effectively screening autistic children.

Methods: Participants, including 123 autistic children and 123 neurotypical children aged 1-6 years, engaged in a 1.5-min free play with fixed toys, which was videotaped. A novel behavior-coding scheme was used to code these videos for 19 autistic behaviors, including play. The coding details of the 19 behaviors were then converted and expanded to 81 digital behavior indicators, including counts, duration, and proportion.

Results: The autistic children showed less functional play and imaginative play and reduced social communication and interactions, such as eye contact, facial expressions, and vocalizations, compared to the neurotypical children during free play. Furthermore, 5 behavioral indicators were selected for the prediction model through stepwise logistic regression, including 1 on socially oriented vocalizations and 4 on count and duration of functional play. The receiver operating characteristic (ROC) curve revealed a good prediction performance with an area under the curve (AUC) of 0.826, a sensitivity of 85.4%, and a specificity of 68.3%.

Conclusion: Our findings highlight differences in play performance and social communication and interactions during free play among autistic children. Based on these findings, we constructed a good clinical prediction model, which might be a potential digital tool used by clinicians to effectively screen autistic children.