An active learning framework for adaptive autism healthcare

项目总结

Project Summary Neuropsychiatric disorders are the single greatest cause of disability due to non-communicable disease worldwide, accounting for 14% of the global burden of disease. The current standards of care suffer from subjectivity, inconsistent delivery, and limited access with growing waitlists. Data science solutions, in particular artificial intelligence (AI) that can port to more ubiquitous mobile tools and that are not restricted for use in clinical settings, have great potential to complement or even replace aspects of the standards of care. We propose to develop a novel data science solution for one of the most pressing mental health burdens, autism, which is up in prevalence by more than 200% since 1990, among the fastest growing pediatric concerns today, and highly representative of many other mental health conditions. We have invented a prototype mobile system called Guess What (GW) that noninvasively turns the focus of the camera on the child through a fluid social engagement with his/her social partner in a way that reinforces prosocial learning while simultaneously measuring the child’s developmental learning progress. At its simplest level, the GW app engages and challenges the child to imitate social and emotion-centric prompts shown on the screen of a smartphone held just above the eyes of the individual with whom the child is playing. Preliminary work to-date resulted in positive user feedback, evidence of high engagement for both the parents and children, and meaningful gains in socialization in the child. A single session produces 90 seconds of enriched social video and sensor data, opening up an exciting opportunity for the game play to passively generate labeled training libraries that enable the development of novel models that are extremely difficult to build without sufficient amounts of domain- relevant training data. Our grant plan will explore this opportunity by designing and optimizing game modes, creating a reusable active learning framework for growth of domain-relevant training libraries, and by creating at least 3 “autism-feature-aware” neural networks that detect child emotion, eye gaze, and hand gestures. Our project will show that GW can not only gamify crowdsourced construction of novel AI models that automatically classify important features of child development – providing a way to address many challenges with AI in medicine today -- but that it can also serve as a mobile therapy for repeat use to target core autism deficits while also tracking progress at the same time.