Identification of autism spectrum disorder gait patterns based on three-dimensional gait analysis / Che Zawiyah Che Hasan

Autism spectrum disorder (ASD) is a complex and lifelong neurodevelopmental disorder that affects the brain growth, functional capabilities, and quality of life of an individual. The existence of movement and gait abnormalities particularly in children with ASD are presently regarded as additional evidence that supports the diagnosis of ASD. In clinical practice, most of the assessment of gait abnormalities are primarily grounded on subjective judgements made by experienced clinicians which are usually manually interpreted, time-consuming, burdensome, and often include subjective and inaccurate evaluations. Hence, automated identification of abnormalities in ASD gait patterns is important for early intervention and post-treatment monitoring. So far, however, there has been little discussion dealing specifically with automated identification of ASD gait patterns. Thus, this study endeavours to propose an automated machine learning-based approach for accurate identification of ASD and normal gait patterns on the basis of dominant gait features acquired from three dimensional (3D) gait analysis. The proposed approach consisted of five sequential stages of data acquisition, data processing, features extraction, features selection, and model classification. The gait data of 30 children with ASD and 30 healthy typically developing (TD) children were acquired using a state-of-the-art 3D motion capture system and two force plates during the self-selected speed of barefoot walking. Time-series parameterisation techniques were applied to the kinematic and kinetic waveforms to extract useful gait features. Two statistical feature selection techniques, namely the statistical hypothesis tests and the stepwise method of discriminant analysis were utilised to select dominant gait features that would best differentiate between ASD and TD gait patterns. Four different machine learning classifiers which include linear discriminant analysis (LDA), k-nearest neighbour (KNN), kernel-based support vector machines (SVMs), and artificial neural networks (ANN) were employed to perform the classification tasks. The superior classification performance was achieved using the ANN classifier with six dominant gait features. The 10-fold cross-validation test results showed that the proposed classification model was able to produce the optimum classification performance with 98.3% accuracy, 96.7% sensitivity, and 100.0% specificity. These findings suggest the potential use of the proposed methods as an aided tool that may be beneficial for clinicians to perform an automated and accurate diagnosis of ASD gait patterns as well as for evaluation purposes of the treatment programmes.