Applying Machine Learning to Identify Autism With Restricted Kinematic Features

Existing diagnosis of the autism spectrum disorder (ASD) heavily depends on the informant’s evaluation of the patient’s behavior, which is both time consuming and labor demanding. In order to develop a rapid diagnostic tool with high accuracy, machine learning (ML) approaches have been proposed to explore the feasibility of identifying ASD with a limited number of features extracted from behavioral evaluation, neuroimaging and kinematic data. Though restricted and repetitive behavior (RRB) is one of the cardinal symptoms of ASD, no study has been conducted to investigate whether restricted kinematic features (RKF) could be used to identify ASD. The present study aimed to address this question. Twenty children with high functioning autism and twenty-three children with typical development (TD) were recruited. They were instructed to perform a motor task that required the execution of the utmost variant movement. Entropy and 95% range of the movement amplitude, velocity and acceleration were computed as indices of RKF. Five ML classifiers were trained and tested including support vector machine (SVM), Linear Discriminant Analysis (LDA), Decision tree (DT), Random forest (RF), and K nearest neighbor (KNN). Results showed that the KNN algorithm ( $k = 1$ ) yielded the highest classification accuracy with four kinematic features (accuracy: 88.37%, specificity: 91.3%, sensitivity: 85%, AUC: 0.8815). Our study demonstrated that RKF could help robustly identify ASD. It is inferred that the application of ML on genetic, neuroimaging, psychological and kinematic features might pose a considerable challenge to the current diagnostic criteria of ASD, and might potentially lead to an automated and objective diagnosis of ASD.