AI-assisted Initiation to Joint Attention Evaluation for Autism Spectrum Disorder Detection

Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by social deficits that affects a considerable percentage of the global population. Early detection of ASD allows for early intervention and ensures a better prognosis; however, there is a lack of screening tools that can objectively measure social deficiency and calculate the probability of ASD. Joint attention is a social skill that develops from an early age, and poor joint attention skills are thought to be a characteristic feature of ASD. We have devised a protocol for the collection of video data for joint attention from toddlers and preschool-aged children. Ten-second video clips of instances where the participant is expected to initiate joint attention are used as input data for the convolutional neural network-long short-term memory-attention classification model. The model distinguished children with ASD and those demonstrating with typical developmental patterns with high accuracy and precision. We validated our model using an attention mechanism to ascertain that the classifications were made based on certain social gestures rather than on random chance. To that end, this work could be further developed into an automatic computational system for autism screening in the general population, as well as an assistive diagnostic tool for mental health specialists to objectively measure the severity of social deficits.