Differences in eye movement data extraction methods and machine learning models-Applied to the detection of children with autism

Abstract

The detection of children with autism spectrum disorders (ASD) have always been a difficult problem For children with ASD it is hard to pay attention to others as well as to interact with people in society (Myles Brenda & Simpson 2001) If we do not concern those children with ASD they will have difficulties making social connection throughout their lives In order to reduce the behavioral difference in children with autism we need some ways to detect ASD more easily first To distinguish between children with ASD and typically developed children we used an eye-tracker to record the eye-movement of the children participants An eye-tracker is a device that tracks the center of visual field of a person By using this device the researchers can know where a participant is focusing on This device also provides some analysis details like the average of fixation durations etc Also many of the previous studies about ASD used an eye-tracker for measurement since it is an easier tool to record and analyze children with ASD So eye-tracker is widely used for experiments concerning ASD If you want to analyze the attention of people it is a good way to use an eye-tracker as the measuring device The eye tracking data can help us detect ASD because the visual attentional pattern of children with ASD is different from that of normal children Most of the children with ASD exhibit abnormal eye-tracking patterns when looking at various people and objects (Sasson & Elison 2012) especially when looking at strangers This is because while recognizing a less familiar human children with ASD tend to avoid the important parts of the face such as eyes nose mouth etc Some research also used the unfamiliar faces as the stimuli for experiment and obtained several good findings Based on those findings we used unfamiliar faces as stimuli in our experiment as well We aim to develop a tool which can detect children with ASD more easily and quickly so that maybe by assigning homework for them to improve their behavior we can help them earlier In order to further improve the detection of ASD we used pictures of both native and foreign faces Some research pointed out that while recognizing native faces and foreign faces the eye-movement of children with ASD is different from that of normal children (Wilson Palermo Burton & Brock 2011) This implies that using foreign faces as stimuli might increase the difference of eye-movement between children with ASD and normal children and therefore reduce the ambiguity and improve the accuracy of the detection Hence in this study we compared the results of the two conditions using either native faces or foreign faces as the stimuli For the detection of autism previous research has proposed several prediction models (Liu Li & Yi 2016; Tyagi Mishra & Bajpai 2018) with different characteristics Among those models most studies used SVM or DNN models to predict ASD However there is no benchmark that allows us to have a fare comparison with those models Although Tyagi Mishra & Bajpai (2018) made a comparison between different predict models to predict adult autism they did not separate the eye-tracker variables and questionnaire variables so it was hard to tell which of the variables is the most important Moreover they only used adult participants in the research so the results cannot be utilized in early detection and early treatment Thus in this research we also compared the methods and structures of different machine learning models Before analyzing the eye-movement recorded by the eye-tracker the researchers need to select their areas of interest (AOI) on the stimuli The eye-tracker research in the past used human area of interest (human AOI) to select the important areas in the visual stimuli Liu Li & Yi (2016) have used k-means algorithm to select the important parts of the interest (Auto AOI) and obtained more than 80% of accuracy However they did not have a comparison between human AOI & Auto AOI so it is unclear which AOI selection method is the best way for detecting children with ASD We hope our research can be used to develop a powerful system that can automatically distinguish between children with ASD and normal children based on the pattern of eye-movement To achieve our goals we divided our research into three sections First we compared the results of ASD detection of different machine learning models namely support vector machine (SVM) and deep neural network (DNN) Second we examined whether the stimuli of faces from different countries help our model to detect autism more accurately or not Third we compared the difference between Human AOI and Auto AOI which was generated by k-means algorithm The results of the research can provide not only a system to detect ASD easier and more accurately but also an advance of mental disorder diagnosis For future research we hope that our research can be used as a basis for diagnosis of any clinical disease Furthermore we hope our study can connect machine learning technology with clinical disease in order to achieve early detection and early treatment

Date of Award	2020
Original language	English
Supervisor	Fu-Zen Shaw (Supervisor)