Gross Motor Features during Balance Training Distinguish Youth with Autism Spectrum Disorder Compared to Youth with Typical Development

Background: Motor challenges are commonly reported in individuals with ASD (Fournier et al., 2013), and fine motor skills during tablet play were shown to reliably distinguish between children with autism spectrum disorder (ASD) and children with typical development (Anzulewicz et al., 2016). However, it is unclear whether gross motor movements would similarly distinguish youth with ASD and youth with typical development. Understanding markers of gross motor function that are more common in ASD would help clarify the nature of motor challenges in this population.

Objectives: (1) Use machine learning to determine whether whole body movement and postural stability during balance tasks can reliably distinguish between youth with ASD and youth with typical development. (2) Investigate whether the results of our machine learning classification corresponded to performance on a standardized motor task. (3) Explore which features of whole-body movement are most informative in the classification algorithm.

Methods: Kinematic and postural sway data were collected in 46 youth with ASD and 18 age-matched youth with TD (ages 7.0-17.9 years), as part of a biofeedback-based videogame training to enhance balance in youth with ASD (Travers et al., 2018). Kinematic data from one-hour sessions were recorded with a Microsoft Kinect Camera, and postural sway data were recorded with a Wii Balance Board. Given the heterogeneity within the ASD profile, a 3:1 ratio of ASD to TD was selected, to allow the motor data to be representative of the diversity within the autism spectrum. Training sessions 2-4 (of 18) were used in these analyses to characterize motor skills before the majority of the training (sessions 5-18) but after the initial intake assessment (session 1). These data were cleaned, informative features were extracted, and an ensemble of random forests were trained. The intake session included confirmation of an ASD diagnosis (providing training labels) and a standardized measure of overall motor function (Bruininks & Bruininks, 2005).

Results: The machine learning algorithm demonstrated average sensitivity of .90, average precision of .84, and average F1 score of .86, using 5-fold cross-validation. The machine learning results were highly correlated with the standardized motor scores, r = -.50, p < .001, suggesting that the machine learning results aligned with an outside measure of motor ability. Postural sway movement was the feature that most robustly distinguished between the two groups, followed by movement in the right shoulder, left foot, right hand, right foot, left hand, right wrist, neck, and right elbow.

Conclusions: Gross motor features during a balance-training task were able to distinguish between the youth with ASD and the youth with typical development with a high degree of sensitivity and precision. Further, the metrics rendered from the machine learning were highly related to overall motor skill, suggesting that our data and machine learning results were representative of participant motor function (and not other potential sources of noise in the data). The results further suggest postural sway data (both left-to-right and forward-to-back movement) may be a distinctive feature for individuals with ASD and a key skill to target in therapy.