Automated Detection of Stereotypical Motor Movements in Individuals with Autism Spectrum Disorder Using Wireless 3-Axis Accelerometers and Computerized Pattern Recognition Algorithms

Background: A primary barrier to understanding and successfully intervening upon Stereotypical Motor Movements (SMM) in individuals with autism spectrum disorder (ASD) is a lack of tools for researchers, clinicians, and caregivers to easily, accurately, and consistently measure them. Traditional measures of SMM rely primarily on paper-and-pencil rating scales, direct behavioral observation, and video-based coding, all of which have limitations.

Objectives: To further explore whether wireless accelerometer sensing technology and pattern recognition algorithms can provide an automatic measure of SMM that is more objective, detailed, and precise than rating scales and direct observation, and more time-efficient than video-based methods. 

Methods: This investigation is a direct replication of Goodwin, Albinali, & Intille (JADD, 2011), wherein the same six individuals with ASD were observed again three years later in their classrooms while wearing three, three-axis accelerometers and engaging in body rocking, hand flapping, and/or simultaneous body rocking and hand flapping. We evaluated automated recognition results compared to manually coded video records at both study time points that yielded an overall average of 0.90 inter-rater reliability, and thus served as ground truth, for two different classifiers – Support Vector Machine and Decision Tree – in combination with different feature sets based on time-frequency characteristics of accelerometer data.

Results: Average automated recognition accuracy across all participants over time ranged from 81.2% (TPR: 0.91; FPR: 0.21) to 99.1% (TPR: 0.99; FPR: 0.01) for all combinations of classifiers and feature sets. We also conducted analyses of kinematic parameters associated with observed SMM over time using raw acceleration data. These included observed SMM within a bout (i.e., contiguous time range in which an individual is engaged in SMM) and across bouts (i.e., pooled SMM bouts) per participant. Intensity (i.e., how vigorously a participant engaged in SMM), duration (i.e., how long a participant engaged in SMM), latency (i.e., time delay between bouts), and estimated movement frequency (i.e., number of moves within a SMM) for rocking, flapping, and "flaprock" were also derived. Finally, we calculated engagement proportion as the percentage of time a participant engaged in SMM while being observed.

Conclusions: Person-dependent algorithms can accurately and consistently measure and describe SMM automatically in individuals with ASD over time in real-word settings. An algorithm that consistently achieves good automated recognition performance across settings and over time could advance autism research and enable new intervention tools that help researchers, clinicians, and caregivers monitor, understand, and cope with these behaviors. For research, automating SMM detection could free a human observer to concentrate on and note environmental antecedents and consequences necessary to determine what functional relations exist for this perplexing and often disruptive class of behavior. For intervention, mobile classifiers could be integrated into a real-time intervention system where just-in-time feedback is provided when SMM are detected to better manage or reduce the occurrence, duration, and/or intensity of these episodes. Finally, such a system could facilitate efficacy studies of behavioral and pharmacological interventions intended to decrease the incidence or severity of SMM by evaluating change in kinematic parameters over time.