Constructing data-adaptive dictionaries for robust sparse feature selection in classification of noisy electro-dermal activity data

Hypersensitivity to sensory stimuli causes overstimulation, inducing overwhelming emotional distress in individuals with an autism spectrum disorder (ASD). Reveal is a wearable device designed by Awake Labs that monitors anxiety levels in ASD children and interfaces with parents and caregivers. It predicts behavioural meltdowns by tracking and classifying key physiological markers of anxiety using machine learning technology. However, the features between which this model is trained to differentiate were developed ad hoc, and built from data that was collected from adults without ASD. Hence, it is unclear if the current technology is optimal for detecting anxiety in ASD children, since markers of anxiety in ASD children have much greater variability. Moreover, in the current technology, there is no way of ensuring that features remain reliable over long periods of time do the best markers of anxiety in an ASD child change as that child ages? We will solve each of these problems using a variant of a popular tool in signal and image processing. We construct a framework that uses this variant to select features in a mathematically rigorous way. TO BE CONT’D

Faculty Supervisor:

Ozgur Yilmaz

Student:

Aaron Berk

Partner:

Awake Labs Inc

Discipline:

Mathematics

Sector:

Medical devices

University:

University of British Columbia