Deep Learning Diagnostic Analysis of the kinematic movement data recorded from individuals with Neurodevelopmental Disorders

We have used Deep Learning ML techniques to analyze the kinematic movement patterns of individuals with Neurodevelopmental disorders (NDD). We studied individuals with Autism Spectrum Disorder (ASD), Attention Deficit Hyperactive Disorder (ADHD), comorbid ASD and ADHD (ASD+ADHD) vs Typically Developing (TD). Motor deficits have previously been identified in subjects with NDDs. We used XSENS high definition motor sensors, (www.xsens.com), to measure the angular velocity, angular acceleration and jerk, as well as the linear acceleration and linear jerk. The subjects carried out a reaching task. The reaching task consisted in moving the subjects’ arms towards touch screen targets appearing intermittently on the monitor. We implemented a supervised deep neural network to predict NDD conditions based on the raw kinematic sensors data, without any preconceived assumptions about the data. We assumed that the sensors raw kinematic data would contain such NDD classification information that is unravelled when using the AI Deep Learning techniques. We generated training sets with, typically, 75% of the full NDD and TD data sets measured with the sensors to train a neural network. Each of the output neurons were trained to represent a different NDD condition. We showed that the network was not overfitting through the accuracy on the held out set that was never used to train the network. Given the small size of the datasets, we applied K-Fold Cross Validation to train the network. The final accuracy on the held out set was 71.4%. This shows that DL is indeed able to correctly identify the diagnostics of most of the remaining NDD subjects without any a priori information about their cognitive abilities. It also clearly separates the NDDs from the TDs. This remarkable DL motor predictive power could be used as an early step in diagnosing the NDD conditions for potential treatments.