Using Tablet-Based Gameplay for the Identification of Autism-Related Movement Patterns
We would like to demonstrate a solution which automatically assesses the risk that a child might have ASD, by analyzing movements using smart gameplay. Disruption of normal movement patterns is a cardinal feature of ASD, however, it has recently been proposed that abnormalities in the development of intentional movements, including prospective planning and execution of movements, can be considered as one of the early markers of ASD. Crucially, such deficiencies can be observed before manifestations of syndromes typically associated with autism, i.e., deficiencies in social interaction or reading emotions (Trevarthen and Delafield-Butt, 2013). Modern technology provides unprecedented access to motor information about the user. Inertial motion sensors coupled with gyroscopes and magnetometers have been miniaturised and integrated into consumer microelectronics such as smart phones, tablets, and wearable devices, opening new possibilities for their application in autism research and diagnosis.
To assess whether children with ASD can be distinguished from typically developing ones, and those with other developmental disorders on the basis of movement analysis conducted during a smart device gameplay
Participants: 46 children with ASD, 20 with other developmental disorders (i.e.,. Down Syndrome, intellectual impairment, aphasia) and 369 typically developing children. The sample size of the control group was significantly larger than the experimental group to reflect the population structure, thus allowing precise calibration of the computer learning algorithms used for the data analysis.
Two mobile game-like applications for children aged 2-5 were used. In the first, a child's goal was to share a piece of food and distribute it evenly among four children depicted on the screen. In the second game, a child's task was to outline a shape (e.g. a squirrel or a snail) and fill it with colour. The materials were designed in collaboration with developmental psychologists. There was a 2-minute training and a 5-minute test session for each application. During the gameplay, touch data and data from tablet’s sensors (gyroscope and accelerometer) were collected.
Data were analyzed by means of computer learning algorithms. 10-fold cross validation technique was used to ensure the models would not be prone to overfitting. Movement patterns of children with ASD and typically developing children were investigated by means of Random Greedy Forest (RGF) algorithm. Area under the receiver operating characteristic curve (AUC) was selected to be the metric, as it does not require a fixed classification threshold. The results demonstrated that AUC was relatively high in case of both applications. The algorithms performed with 72% sensitivity and 88% specificity for differentiation of ASD from typically developing children. To assess whether the models successfully differentiate children with ASD from those with other developmental disorders simple models, namely, SVM and logistic regression were used because the latter group was relatively small (N=20). The models achieved 65% specificity and 79% sensitivity, suggesting that it is possible to differentiate ASD from other disorders.
We provide evidence that ASD has a prominent motor component that can be identified by smart device gameplay, which should be taken into account in early assessment.