Classification And Assessment Of Autism Based On Eye-movement Technology And Deep Learning

Autism spectrum disorders(ASD)is a neurodevelopmental disease with social communication disorders and repetitive stereotyped interests and behaviors as its core symptoms.A large number of studies have found the difference of eye movement behavior between ASD children and typical development children in social attention and restricted interest.The long-short term memory network(LSTM)model of deep learning technology can classify the eye movement fixation data of time series.In this study,two visual preference paradigms were used to research the differences of visual preference between ASD children and typical development children for social stimulation and object repetitive movement,and LSTM model was used to classify eye movement data to distinguish ASD children and typical development children.In this study,recruited 25 children with ASD and 42 typical development children,and the two groups were matched in gender and age,with an average age of about 50 months.They participated in the eye movement experiment of study 1 and study 2.Study 1 adopted a mixed design of 2(group: ASD,typical development)* 2(stimulus type: social video,geometric video)to explore the visual attention patterns of two groups of children on different types of video.The research results showed that the first gaze of normal children on social stimuli,while there was no significant difference in the perception speed and visual preference of ASD children on the two types of stimuli.The two indicators of visual acuity before the first fixation and the percentage of viewing time for social stimulation have good discrimination for children with ASD.The second study adopted a mixed design of 2(group: ASD,typical development)* 2(object motion type: repetitive motion,random motion)to explore the visual preferences of children in the two groups for object motion types.The research results showed that there was no significant difference between the visual preferences of children with ASD and normal children.In Study 3,LSTM model was used to classify the data of Study 1,with the accuracy of 64.02%,specificity of 64.95% and sensitivity of 62.42%.The classification accuracy,specificity and sensitivity of the data in study 2 were 68.72%,71.71% and 62.28%.Therefore,the following conclusions are reached in this study:(1)Children with ASD have social attention disorder,which is manifested as slower detection of social stimuli and less visual preference.Differences in visual preference for social information can distinguish children with ASD from children with TD.(2)There was no difference in visual preference between repetitive and random motion of objects in ASD children.(3)Eye movement technology combined with deep learning has certain value for the classification and assessment of ASD.