Seed Grant Profile
Toward new assessments of problem behavior to increase treatment effectivenessMany treatment programs aim to reduce the extent to which individuals with developmental disabilities engage in problem behaviors. The ability to collect accurate data on the occurrence of these behaviors is key to determining response to treatment, yet the parent- or teacher-report checklists typically used to gather such data do not capture precise frequencies of behavior. The main goal of the proposed research is to combine body-worn sensors (accelerometers) with computational tools (machine learning techniques, namely semi-supervised statistical modeling approaches for sequential sensor data) to obtain objective, accurate measures of the frequency of specific problem behaviors. We hypothesize that activity recognition (AR) algorithms, applied to accelerometry data gathered through on-body sensing, can automatically detect and differentiate among three relevant classes of problem behaviors (self-injury, aggression to others, disruptive behaviors involving objects) with comparable fidelity to human expert judgments. The specific aims are: (1) To collect a comprehensive set of problem behavior accelerometry data from children and adolescents with ASD, along with human expert annotations of the data; (2) To develop AR algorithms to detect instances of problem behaviors, and evaluate the accuracy of automatic segmentations with respect to human expert coding; (3) To develop AR algorithms to distinguish among different classes of problem behavior and evaluate the accuracy of classification with respect to human expert coding; (4) To evaluate how classification accuracy is affected by the number and placement of sensors. Accelerometry data will be collected during usual intake procedures at a local behavior treatment clinic. Sessions will be recorded to facilitate subsequent human coding of problem behavior frequency and type. We will develop a fully functional prototype of an automatic analysis system for problem behavior. The long-term goal is to develop an automated problem-behavior assessment system that can be deployed in naturalistic settings over longer time scales.
Investigators: Agata Rozga (Georgia Institute of Technology), Nathan Call (Marcus Autism Center/Children's Healthcare of Atlanta),