Research On Auxiliary Screening For Children With Autism Spectrum Disorder Based On Motor Features In Videos Naturalistic Social Interactions

Autism spectrum disorder(ASD)is a neurodevelopmental disorder characterized by deficits in social interaction and communication and restrictive,repetitive behaviors.Early special education and intervention improve the outcomes and life quality.Early identification of children with ASD is the premise of receiving early education and intervention.The process of identification depends on assessments by qualified experts.However,due to the shortage of experts in this field,children with ASD are often not detected in time.Screening is a key step of ASD identification,aiming at finding out children at risk for ASD quickly and referring them to diagnostic evaluations.Developing an effective auxiliary screening method with low cost would improve the efficiency of screening and optimize the process of identification,which promotes the early detection of children with ASD.Since videos of naturalistic social interactions are accessible and can capture the social deficits of children,which are core symptoms in ASD,they are important information sources of screening.Of the plenty features embedded in the videos,quantified motor features are associated diagnostic features and easy to be extracted automatically,thus having the potential for use in auxiliary automatic screening methods.In addition,machine learning techniques,which mined the latent pattern from the data,have been used in the automatic classification of children with ASD.To foster the implementation of efficient and low-cost auxiliary screening methods,it is necessary to investigate the automatic classification method for children with ASD based on machine learning techniques and motor features in videos of naturalistic social interactions.The development of this method needs to solve the following key problems:(1)data cleaning,(2)feature exploration,(3)feature selection and classification.The specific problems include:(1)controlling the effect of social engagement on interaction to reduce the influence of irrelevant factors on the classification performance;(2)discovering motor features suitable for natural interactive scenes to improve the discriminative power of the feature sets;(3)selecting discriminative,sparse,and stable feature sets to optimize classifier performance and generalization.To solve the above problems,this dissertation first designed a coding method for dyadic social engagement states,extracted video segments with the same dyadic social engagement state which improved the homogeneity of social interaction data;meanwhile,through literature research,we searched for motor features with context robustness and associated with core symptoms of ASD,and investigated whether preschool-aged children showed abnormality in such characteristics;at last,the best feature subset was selected from the motor features to optimize the classifier performance,and feature selection was also used to reduce the feature space dimension to build a classifier with good generalization ability and discover stable features in it.The main contributions of the dissertation are as follows:(1)Design,application,and validation of the coding scheme of social engagement states.Based on the coding method of social engagement in mother-infant interactions,early childhood social ability assessment tools,literature on social interactions,and our preliminary observations on children’s social interactions,a coding scheme of dyadic social engagement states suitable for teacher-preschooler naturalistic play interactions was proposed.We collected videos of the interactions and trained raters to code.Through the analysis of the time proportion of each code,it was verified that the design of the coding scheme was reasonable and the data cleaning based on the coding scheme was necessary.(2)Analysis of the characteristics of interpersonal motor synchrony in preschool-aged children with ASD in naturalistic social interactions.Through literature research on interpersonal motor synchrony in individuals with ASD,we proposed that interpersonal motor synchrony may be a feature set suitable for automatic classification of children with ASD,and conducted empirical research to explore the characteristics of interpersonal motor synchrony in preschool-aged children with ASD and its relationship with autistic traits.Specifically,computer vision techniques were used to automatically extract features of body translation and head rotation movement from videos of social interactions,then signal processing technology and synchrony calculation method were used to extract interpersonal motor synchrony features corresponding to the two forms of movement.Statistical methods were used to test whether there was a significant difference in motor synchrony between the ASD group and the typically developing group and analyzed the correlation between the interpersonal motor synchrony and autistic traits of children with ASD.The above empirical research provided the theoretical and scientific basis for introducing interpersonal motor synchrony features in the building of automatic classifiers for children with ASD.(3)Classification of children with ASD based on motor features from videos of naturalistic social interactions and feature analysis.First,feature selection methods were used to discover the optimal feature subset from the motor features to build classifiers with the best performance;second,the minimum feature subsets that can meet the criteria of clinical screening tools were selected to build classifiers with acceptable performance and strong generalization ability.Finally,stable features were extracted from the minimum feature subset to facilitate the rapid building of new classifiers when the dataset is expanded,which also provided information for the discovery of behavioral markers for ASD.Through the above studies,this dissertation finally implemented an auxiliary screening method for children with ASD based on motor features from videos of naturalistic social interactions,which solves the problem of insufficient automation of related methods.The method has good performance and could effectively reduce the labor cost of screening work,which helps to promote the early detection of children with ASD,and provides technical assistance for the popularization of special education.