Automated Detection of Stereotypical Motor Movements

Background: Stereotypical motor movements can be one of the most disruptive types of behavior seen in persons on the autism spectrum. Engagement in these behaviors can lead to social stigmatization and complicate social interaction. Moreover, if stereotypical motor movements become dominant in an individual’s behavioral repertoire, they can interfere with the performance of established skills and acquisition of new skills, and may lead to self-injurious behavior. Unfortunately, a lack of accurate and timely measures has made it difficult to determine what function(s) stereotypical motor movements serve and whether interventions aimed at reducing, replacing, or preventing them are effective.

Objectives: Utilize sensor technology to provide a measure of stereotypical motor movements that may be more objective, detailed, and precise than rating scales and direct observation, and more time-efficient than video-based methods. Use of wireless accelerometers and pattern recognition software to automatically detect stereotypical motor movements will be demonstrated and experimental data validating this innovative application of technology will be presented.

Methods: Six participants on the autism spectrum who engaged in stereotypical hand flapping and body rocking were observed repeatedly in both laboratory and classroom settings (for a total of 11 hours) while wearing three small, unobtrusive wireless and wearable movement sensors placed on the left wrist, right wrist, and torso. A digital camera recorded each session. The camera was connected to a computer that synchronized saved video with accelerometer data streams. Start time, end time, and type of stereotypical motor movement were coded both in real-time and offline by two independent raters using custom video coding software. Real-time human coding and computerized pattern recognition performance was compared to offline, “ground truth” video annotations (average inter-rater reliability across offline codes was .95).

Results: In addition to excellent compliance and performance with the sensors across participants and settings, findings revealed that, on average, real-time human coding was correct only 60% of the time across participants in the laboratory and classroom, while computerized pattern recognition algorithms (C4.5 classifier using five acceleration time and frequency domain features as input) correctly identified (verified with 10-fold cross validation) approximately 90% of stereotypical motor movements observed across participants in the laboratory and approximately 85% observed across participants in the classroom.

Conclusions: The technology-assisted assessment strategy developed in this work has significant clinical implications. First, reliable recording of stereotypical motor movements could enable researchers to study what functional relations may exist between stereotypical motor movements and specific antecedents and consequences. These relations may arise differentially in various environmental settings, in the presence of demand tasks, or in the presence of physiological influences. Second, wireless devices that record stereotypical motor movements could enable teachers, therapists, and caregivers to monitor movement behavior and gather data that can assist with treatment decisions. Finally, documentation of stereotypical motor movements before and after an experimental treatment could facilitate efficacy studies of behavioral and pharmacologic interventions intended to decrease the incidence or severity of stereotypical motor movements.