Early detection of autism leads to earlier treatment, which is proven to have a major impact on outcomes. In spite of recent advances in early Autism Spectrum Disorders (ASD) detection, the average age of diagnosis in the US is still around five. ASD diagnosis is currently performed via behavioral assessment, which requires highly specialized training, is not widely available in rural areas, and may be applied inconsistently. The need for specialized training for the administration of behavioral assessment and the effort involved in individual assessments preclude large scale deployment of these diagnostic methods in clinics and pediatricians' offices as well as large scale population studies. The Infant Brain Imaging Study (IBIS) is an early detection study at the University of Washington Autism Center which assesses behavioral and brain development in infants at high familial risk for ASD. Behavioral assessments include specialized observations of gross motor function, an area of development that is uniquely highlighted in the first year of life. This study along with others, highlight atypical motor development as the first step in the emergence of autism-related symptoms. Analyzing behavioral video data in order to assess/score individual subjects is a process that is time-intensive, subjective, and requires extensive training to attain reliability. We will build a Human Action Recognition Engine (HARE) that leverages computer vision tools to automatically extract, quantify and classify known motor actions - from video datasets - adding a significantly more efficient and standardized method to augment the current diagnostic standard of care. In this Phase I proposal, we will: 1. Develop the HARE prototype: automatic segmentation of subject of interest; determination of 3D orientation; extraction of features that are used in classification of actions from a predefined set defined in the IBIS behavioral assessment battery; 2. Leverage the intermediate outputs of the AR engine in establishing techniques to detect and de-identify faces of multiple, closely-interacting human subjects in video toward further processing and data sharing; 3. Explore early markers to classify subjects, based on actions detected, into ASD and non-ASD groups and evaluate the sensitivity and specificity of the classification engine. This Phase I effort will pave the way forthe creation of an action-annotated video repository from HARE's action recognition output. The repository will provide a rich source of highly-accessible data toward training and further research discoveries. Finally, the HARE system can systematically identify new, previously unidentified motor actions that may relate to increased risk for later developmental difficulties, particularly ASD. These novel early risk markers - in combination with existing assessments - would allow reliable, earlier identification of ASD.
We propose to build an action-annotated video repository that will aid objective behavioral assessment of infants at risk for Autism Spectrum Disorders (ASD) and serve as a valuable training tool. We offer a methodology that can systematically reveal salient early motor markers that can result in efficacious screening of high risk infants.
