Leveraging routinely collected health data to improve understanding of language development in children identified as late talkers

ABSTRACT – Overall
The overall goal of the Duke Autism Center of Excellence is to use an innovative, translational digital health and
computational approach to address the critical need for more effective autism screening tools, objective outcome
measures, and brain-based biomarkers that can be used in clinical trials with young autistic children. An
Administrative Core, Dissemination and Outreach Core, and Data Management and Analysis Core will support
three Projects. Project 1 will recruit a large population of 16- to 30-month-old toddlers through primary care clinics
to evaluate the accuracy of a remotely administered novel digital phenotyping application (app) for detecting
early signs of autism. The app automatically quantifies observations of children’s behavior using computer vision
analysis and is deployed on widely available devices. The usability of the app for longitudinal outcome monitoring
will be assessed at 16-30, 36, and 48 months of age. The feasibility of using computer vision analysis to measure
patterns of caregiver-child interactions from videos recorded at home will be explored. Project 2 will develop a
complementary autism screening approach by using North Carolina Medicaid and Blue Cross Blue Shield claims
data (N ~ 230,000, autism cases ~6,000) to create a generalizable autism prediction model based on routine
health data collected from birth to 18 months. Then, using Duke University Health System electronic health
records (EHR; N ~ 64,000, autism cases ~ 800), this Project will use natural language processing to assess the
added predictive value of EHR elements not captured in claims data (e.g., clinician notes). Both data sets will be
leveraged to gain insight into the nature and prevalence of medical conditions in infants and toddlers who are
later diagnosed with autism. Projects 1 and 2 will collaboratively engage primary care providers and other
stakeholders to design an automated clinical decision support tool for autism screening that, in the future, could
be integrated into the primary care provider’s clinical workflow. Project 3 will use an innovative machine learning
computational method to develop a multimodal biomarker that combines features of electroencephalographic
(EEG) activity and synchronized measures of children’s behavior (e.g., social attention) automatically coded via
computer vision analysis, with a focus on neural connectivity measured via traditional methods (coherence,
phase-lag index) and novel neural network analysis methods (discriminative cross-spectral factor analysis)
developed by our team. This multimodal approach will be evaluated in 3–6-year-old autistic children without
intellectual disability (ID), age- and sex-matched neurotypical children, and autistic children with ID (IQ