Measuring Imitation Ability in Autism Using Dynamic Time Warping Applied to a Dance Videogame Task

Poster Presentation
Friday, May 3, 2019: 11:30 AM-1:30 PM
Room: 710 (Palais des congres de Montreal)
R. N. Rochowiak1, C. Pacheco2, R. Nicholas3, E. Mavroudi2, S. Rengarajan4, G. Miller4, B. Messenger4, A. S. Pillai4, R. Vidal2, S. H. Mostofsky5 and B. Tuncgenc6, (1)Center for Neurodevelopmental and Imaging Research, Kennedy Krieger Institute, Baltimore, MD, (2)Johns Hopkins University, Baltimore, MD, (3)University of Nottingham, Nottingham, United Kingdom, (4)Kennedy Krieger Institute, Baltimore, MD, (5)Department of Neurology, Johns Hopkins University School of Medicine, Baltimore, MD, (6)Neurology, Johns Hopkins University, Baltimore, MD

Impaired motor imitation is commonly reported in individuals with autism spectrum disorder (ASD) and is thought to contribute to impaired development of reciprocal-social skills. To date, most imitation studies have relied on human observer coding (HOC) to detect presence/absence of certain elements based on behavioral coding schemes. The reliance on HOC is time-consuming, introduces coder bias and reliability issues, and limits assessment to predetermined elements delineated in coding schemes. Using a computer-based technique would eliminate coder bias, reduce time spent, allow for reliable comparisons across sites, and provide more complete assessment of movement dynamics. Further, most studies have examined imitation of single, discreet movements performed in unrealistic settings. Using a videogame task to assess motor imitation ability would allow for assessment across a range of realistic contexts.


This study aims to develop and validate the novel use of a machine learning algorithm, dynamic time warping (DTW), to investigate imitation ability in typically-developing (TD) children and children with ASD during performance of a dance videogame task.


Thirty-six children aged 8-12 years (22 ASD; 3 females per group) participated. Children imitated a video avatar performing a 1-minute dance sequence comprised of 18 novel whole-body movements. Children’s movements were recorded using two Kinect Xbox depth cameras at 30 frames-per-second, placed in front of and behind the child (Figure 1a). The x-y-z coordinates of 20 joints were extracted from the depth recordings using iPi Motion Capture Software.

HOC was used to assess imitation performance and to validate the accuracy of DTW. For each move, elements describing key changes in limb locations were defined (Figure 1a). Children received a score of 1 for each element performed; as such, higher HOC scores indicate better imitation accuracy (Figure 1b).

DTW temporally aligns the child’s time-course to the avatar’s time-course by minimizing the Euclidean distance between them. DTW distances of the 18 moves were averaged to make up the child’s total DTW distance (Figure 1c). Higher DTW distances indicate greater difference between the child and the avatar, and hence poorer imitation. Preprocessing performed included positioning the hip as point of reference, normalizing the children’s limb length to the avatar’s skeleton, and adjusting the children’s rotation to match the avatar’s at the beginning of the activity.


HOC scores were reliability-coded by two hypothesis-blind coders (κ= .915, p< .001). HOC scores and DTW distances were significantly correlated (r(35)= -.77, p< .001) and this correlation held equally strong within the ASD (r(22)= -.77, p< .001) and TD (r(14)= -.74, p< .001; Figure 2) groups. Furthermore, DTW distances revealed significantly better imitation in TD than in ASD group (p= .02), and HOC scores confirmed this trend with near-significant values (p= .09; Figure 2).


The present study provides evidence for the use of a computer-based algorithm (DTW) to investigate autism-associated differences in motor imitation during a naturalistic dance videogame task. This approach for assessing motor imitation could prove useful in establishing biomarkers for assessing diagnosis and response to intervention.