Recognition of autism spectrum disorder based on face processing abnormality

Abstract

Autism spectrum disorder (ASD), a highly complex neurodevelopmental disorder with complicated causes and processes, has core symptoms including impairment of social interaction, narrow interests, and repetitive stereotypical behaviors. Many methods have been applied to the early diagnosis of ASD, one of which is eye-tracking technology. Faces are important social stimuli, and face processing is very important for social interaction. Eye movements during face processing are a source of information which can reveal characteristics of neural development. We here study eye movements during the processing of different types of faces in autistic children. We use machine learning on eye movement information to automatically identify ASD. We recruited a total of 81 children aged 3 to 6 years (40 ASD, 41 age- and gender-matched typically-developing controls) to look at faces of strangers from foreign races, faces of strangers from the children’s own race, and faces of familiar people from the children’s own race. We quantized eye-fixation coordinates using K-means, partitioning face images into K =64 different cell-like regions (areas of interest). We then used the fixation coordinate frequency distribution as features, along with the minimal redundancy and maximal relevance (mRMR) method for feature selection and support vector machines (SVM) for classification of ASD versus control. Results showed that maximum classification accuracies based on foreign face, own-race stranger face, and familiar face respectively reached 78.89% (sensitivity 81.67%, specificity 74.00%), 73.89% (sensitivity 70.88%, specificity 76.38%), and 79.44% (sensitivity 82.67%, specificity 79.71%). The areas under the receiver operating characteristic curves (AUC) were 0.8065, 0.8217, and 0.8387, respectively. Then, we combined features from the three types of faces to obtain a total of 192 features and used the mRMR method, selecting 27 features. With these features, the maximum classification accuracy reached 90.28% (sensitivity 91.33%, specificity 86.83%), and the AUC was 0.9317. Independent sample t -tests showed that, compared with controls, ASD children had atypical facial scanning. In particular, ASD children paid less attention to some areas around the eyes of foreign faces, while gazes on unfamiliar faces focused more on the body and background. In summary, our results show that face scanning characteristics are feasible for classifying ASD versus typically developing children, implying that eye-tracking technology combined with machine learning can provide auxiliary evaluation indices for clinical diagnosis of ASD.