Integrating Deep Learning with Behavior Imaging to Accelerate Industry’s ‘Learning’ of Autism Core Deficits

Friday, May 11, 2018: 10:00 AM-1:30 PM
Hall Grote Zaal (de Doelen ICC Rotterdam)
R. Oberleitner1, J. Schwartz2, C. J. Smith3, M. J. Morrier4, U. Reischl5, J. Fodor6, B. Martin6 and C. E. Rice7, (1)Behavior Imaging Solutions, Boise, ID, (2)Behavior Imaging, Boise, ID, (3)Southwest Autism Research and Resource Center, Phoenix, AZ, (4)Psychiatry and Behavioral Sciences, Emory University School of Medicine, Decatur, GA, (5)Boise State University, Boise, ID, (6)University of Idaho, Moscow, ID, (7)Emory Autism Center, Decatur, GA
Background: During the past several decades, the worldwide prevalence of children diagnosed with autism spectrum disorder (ASD) has increased dramatically. Since ASD does not have clear biomarkers and is a behaviorally defined disorder, a key element in the diagnostic process is direct observation of behavior and expert diagnosticians. This approach can lead to significant variability in results . However, new approaches using Artificial Intelligence (AI) are being developed that have the potential to improve the diagnostic accuracy of this process.

Objectives: 1. How Deep Learning Method Can Be Deployed to 'Learn' from 1 to 1000 Clinicians' Assessments of Patient Video Data 2. How Clinicians Evaluating Children via Archived Video can document Atypical Behaviors related to Autism

Methods: A Deep Learning (DL) computational method called a ‘DL Classifier’ is being introduced to help clinicians analyze video information collected by a telehealth assessment called NODATM (Naturalistic Observation Diagnostic Assessment). NODA results in a diagnostic category (ASD v. non ASD) based on clinician-annotated videos performed by expert diagnosticians. Clinicians identify both atypical and typical behaviors in children at specific moments on video samples parents collect and share from their home. Given a large enough dataset, the DL Classifier delineates the behavioral information into recognized patterns that can be automatically ‘flagged’ on video data as typical or atypical - for review by an expert diagnostician. The more clinicians train the DL Classifier, the better it gets in recognizing atypical and typical behaviors. A DL Classifier was developed to compute patterns in 244 videos captured and previously shared by families as part of the NODA diagnostic assessment process; these videos cumulatively had 6,243 atypical / typical tags previously applied at select time points by clinicians. The Classifier was applied in several iterative computational methods to analyze 1. still images within NODA video, then 2. short video segments associated with associated tags applied by clinicians.

Results: All ‘Deep Learning’ tags suggested by the Classifier were compared to tags used to generate the classifier. Classifier resulted in 80% accuracy based on analysis of still images, and 86% accuracy based on analysis of video segments.

Conclusions: A more commercial-grade 'DL Classifier' is being integrated into the existing Behavior Imaging / NODA platform, and we’re predicting accuracy will improve with more Classifier development, and more training of the NODA 'data set' by clinicians. When further validated, the proposed NODA DL classifier represents a significant advancement in the diagnosis and care of children with ASD, especially by building an industry-wide consensus on autism symptoms, and specifically -- through a reduction of time that will be required to obtain an accurate diagnosis.