SCH: Enhancing Automated Prediction of Challenging Behavior in Individuals with Autism Using Biosensor Data and Machine Learning - PROJECT SUMMARY (See instructions): Autism Spectrum Disorder (ASD) is one of the most common childhood disorders (1 in 44 ). Individuals with ASD have a higher prevalence of challenging behavior (e.g., aggression, self-injury, emotion dysregulation) that interferes with adaptive development, ranks among the most common causes for referral to behavioral healthcare services, and incurs high healthcare costs. Over the past four years, the team at Northeastern University has made significant research progress developing machine learning procedures that automate the detection of challenging behavior onsets in individuals with ASD using wearable biosensor data (cardiovascular, electrodermal, and physical activity). Despite our promising results, issues still need to be addressed to enable practical daily use in real-world contexts. This includes the need for extensively labeled data, individual calibration from population models to specific individuals, and handling the non-stationary nature of challenging behavior and physiological data. This project aims to advance fundamental machine learning theory and techniques that facilitate rapid model individualization and continuous online model adaptation with little or no labeled data. To this end, we will contribute to areas including domain adaptation, transfer learning, lifelong learning, and robust modeling and inference. Three Specific Aims guide the project: (1) Rapid physiological and behavioral data model individualization; (2) Continuous lifelong physiological and behavioral data model adaptation; and (3) Validation of model individualization and adaptation techniques with prospective data collected in a clinical setting from our partners at the Marcus Autism Center at Emory University who specialize in functional analysis of challenging behavior in individuals with ASD. Across these Aims, we will explore applications of semi-supervised learning theory, data importance weighting, Support Vector Machines, neural network models, Hierarchical Markov-Modulated Point Process Models, and Bayesian evidence fusion. The modeling and inference techniques we develop will be of general applicability to other health application contexts involving event prediction (e.g., seizure detection) and human action/decision-making (e.g., intensive care unit triage).
