View in EDS HTML Full Text PDF Full Text

Use of Oculomotor Behavior to Classify Children with Autism and Typical Development: A Novel Implementation of the Machine Learning Approach

Saved in:
Bibliographic Details
Title: Use of Oculomotor Behavior to Classify Children with Autism and Typical Development: A Novel Implementation of the Machine Learning Approach
Language: English
Authors: Zhao, Zhong, Wei, Jiwei, Xing, Jiayi, Zhang, Xiaobin, Qu, Xingda, Hu, Xinyao, Lu, Jianping
Source: Journal of Autism and Developmental Disorders. Mar 2023 53(3):934-946.
Availability: Springer. Available from: Springer Nature. One New York Plaza, Suite 4600, New York, NY 10004. Tel: 800-777-4643; Tel: 212-460-1500; Fax: 212-460-1700; e-mail: customerservice@springernature.com; Web site: https://link.springer.com/
Peer Reviewed: Y
Page Count: 13
Publication Date: 2023
Document Type: Journal Articles
Reports - Research
Descriptors: Children, Autism Spectrum Disorders, Symptoms (Individual Disorders), Eye Movements, Interpersonal Communication, Classification, Accuracy, Disability Identification
DOI: 10.1007/s10803-022-05685-x
ISSN: 0162-3257
1573-3432
Abstract: This study segmented the time series of gaze behavior from nineteen children with autism spectrum disorder (ASD) and 20 children with typical development in a face-to-face conversation. A machine learning approach showed that behavior segments produced by these two groups of participants could be classified with the highest accuracy of 74.15%. These results were further used to classify children using a threshold classifier. A maximum classification accuracy of 87.18% was achieved, under the condition that a participant was considered as 'ASD' if over 46% of the child's 7-s behavior segments were classified as ASD-like behaviors. The idea of combining the behavior segmentation technique and the threshold classifier could maximally preserve participants' data, and promote the automatic screening of ASD.
Abstractor: As Provided
Entry Date: 2023
Accession Number: EJ1368763
Database: ERIC
Full text is not displayed to guests.
FullText Links:
  – Type: pdflink
    Url: https://content.ebscohost.com/cds/retrieve?content=AQICAHj0k_4E0hTGH8RJwT4gCJyBsGNe_WN95AvKlDbXJGqwxwF6jkeb937Iw_cPRyDZeDIBAAAA4jCB3wYJKoZIhvcNAQcGoIHRMIHOAgEAMIHIBgkqhkiG9w0BBwEwHgYJYIZIAWUDBAEuMBEEDGR1OyiCPbypFXqHMAIBEICBmnRD6ZMwH3lpvROE-A8TS2yWSS0BLagqo4yPoUC1BfDmKP5hC26Bk-nMgBVFH7e7HVYVTrY-7eRtnqeiDG0CDyEVsl9mvaVYdXzUNIxA16HCGpXwbyomtNLzRdVVX0otkkqQz0o1aZyxSvRgFD94vfID4aaNMWctfSvmSd8cetj5czLE2BR19hT-Q0NjYBHQax2cHX6AnC2i2SQ=
Text:
  Availability: 1
  Value: <anid>AN0162233062;aut01mar.23;2023Mar08.01:30;v2.2.500</anid> <title id="AN0162233062-1">Use of Oculomotor Behavior to Classify Children with Autism and Typical Development: A Novel Implementation of the Machine Learning Approach </title> <p>This study segmented the time series of gaze behavior from nineteen children with autism spectrum disorder (ASD) and 20 children with typical development in a face-to-face conversation. A machine learning approach showed that behavior segments produced by these two groups of participants could be classified with the highest accuracy of 74.15%. These results were further used to classify children using a threshold classifier. A maximum classification accuracy of 87.18% was achieved, under the condition that a participant was considered as 'ASD' if over 46% of the child's 7-s behavior segments were classified as ASD-like behaviors. The idea of combining the behavior segmentation technique and the threshold classifier could maximally preserve participants' data, and promote the automatic screening of ASD.</p> <p>Keywords: Autism; Behavior segmentation; Entropy; Eye-tracking; Machine learning; Oculomotor</p> <hd id="AN0162233062-2">Introduction</hd> <p>Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by persistent impairments in social communication and the presence of restricted and repetitive behavior (American Psychiatric Association, [<reflink idref="bib1" id="ref1">1</reflink>]). Data from the United States reported that the prevalence of ASD has increased up to 1/44 now (Maenner et al., [<reflink idref="bib24" id="ref2">24</reflink>]). However, the current diagnosis of ASD heavily relies on clinicians' experience and caregivers' report, which is highly time-consuming and might be biased by subjective factors (Möricke et al., [<reflink idref="bib27" id="ref3">27</reflink>]; Tebartz van Elst et al., [<reflink idref="bib38" id="ref4">38</reflink>]). Therefore, developing an objective and efficient tool to identify individuals with ASD becomes a critical issue.</p> <p>Owing to the advantage of machine learning (ML) in identifying patterns that could be hardly recognized by human eyes, recent years have witnessed a mounting interest in applying ML algorithms to make objective and efficient detection of ASD (Crippa et al., [<reflink idref="bib6" id="ref5">6</reflink>]; Duda et al., [<reflink idref="bib7" id="ref6">7</reflink>]; Li et al., [<reflink idref="bib21" id="ref7">21</reflink>]; Liu et al., [<reflink idref="bib23" id="ref8">23</reflink>]; Wan et al., [<reflink idref="bib42" id="ref9">42</reflink>]). For instance, in order to improve the diagnostic efficiency, several studies have applied ML to develop effective simplified versions of the Autism Diagnostic Observation Schedule (ADOS) and the Autism Diagnostic Interview-Revised (ADI-R) (Duda et al., [<reflink idref="bib7" id="ref10">7</reflink>]; Wall et al., [<reflink idref="bib40" id="ref11">40</reflink>], [<reflink idref="bib41" id="ref12">41</reflink>]). In terms of improving the diagnostic objectivity, a sizable body of literature showed that features extracted from neuroimaging, eye-tracking, electroencephalography (EEG) and kinematic data could accurately identify ASD by using ML (Crippa et al., [<reflink idref="bib6" id="ref13">6</reflink>]; Grossi et al., [<reflink idref="bib13" id="ref14">13</reflink>]; Li et al., [<reflink idref="bib21" id="ref15">21</reflink>]; Liu et al., [<reflink idref="bib23" id="ref16">23</reflink>]; Osterling & Dawson, [<reflink idref="bib28" id="ref17">28</reflink>]; Plitt et al., [<reflink idref="bib30" id="ref18">30</reflink>]; Wan et al., [<reflink idref="bib42" id="ref19">42</reflink>]; Yamagata et al., [<reflink idref="bib44" id="ref20">44</reflink>]; Zhao et al., [<reflink idref="bib49" id="ref21">49</reflink>]).</p> <p>With the aid of the eye-tracking technology, both atypical visual fixation and oculomotor behavior has been frequently reported in individuals with ASD (Miller et al., [<reflink idref="bib25" id="ref22">25</reflink>]; Sumner et al., [<reflink idref="bib37" id="ref23">37</reflink>]; Wan et al., [<reflink idref="bib42" id="ref24">42</reflink>]). So far, several studies have shown that eye-tracking technology might be a promising tool to facilitate the objective detection of ASD by implementing the ML approach (Eraslan et al., [<reflink idref="bib9" id="ref25">9</reflink>]; Liu et al., [<reflink idref="bib23" id="ref26">23</reflink>]; Wan et al., [<reflink idref="bib42" id="ref27">42</reflink>]; Yaneva et al., [<reflink idref="bib45" id="ref28">45</reflink>]). For instance, Wan et al. ([<reflink idref="bib42" id="ref29">42</reflink>]) eye-tracked children with ASD and typical development (TD) while they were watching a 10-s video displaying a female speaking. Visual fixations at seven areas of interest were computed and fed into the ML classifier. Their results showed that a classification accuracy of 85.1% could be achieved with two features when classifying these two groups of participants (Wan et al., [<reflink idref="bib42" id="ref30">42</reflink>]). Similarly, other studies reported that eye-tracking data obtained from both image viewing tasks and face-to-face conversations could help discriminate individuals with ASD from those with TD by using the ML approach (Liu et al., [<reflink idref="bib23" id="ref31">23</reflink>]; Yaneva et al., [<reflink idref="bib45" id="ref32">45</reflink>]; Zhao et al., [<reflink idref="bib46" id="ref33">46</reflink>]). Interestingly, some studies combined eye-tracking features with those obtained from other modalities (e.g., EEG, kinematics and conversational length) (Baranek, [<reflink idref="bib3" id="ref34">3</reflink>]; Osterling & Dawson, [<reflink idref="bib28" id="ref35">28</reflink>]; Zhao et al., [<reflink idref="bib46" id="ref36">46</reflink>]), and their results supported the idea that combined features produced better classification performance than a single modality. For example, Zhao et al. ([<reflink idref="bib46" id="ref37">46</reflink>]) investigated whether combining features on visual fixation and conversational length would yield higher classification accuracy than using features from either of these two modalities alone. Their results demonstrated a maximum classification accuracy of 92.31% with the combined features and only 84.62% with visual fixation features alone (Zhao et al., [<reflink idref="bib46" id="ref38">46</reflink>]).</p> <p>Noticeably, the above-mentioned eye-tracking based ML studies extracted features from visual fixation patterns (Eraslan et al., [<reflink idref="bib9" id="ref39">9</reflink>]; Liu et al., [<reflink idref="bib23" id="ref40">23</reflink>]; Wan et al., [<reflink idref="bib42" id="ref41">42</reflink>]; Yaneva et al., [<reflink idref="bib45" id="ref42">45</reflink>]; Zhao et al., [<reflink idref="bib46" id="ref43">46</reflink>]). No study has been found to identify ASD with oculomotor behavior. In reality, visual fixation patterns reflect the situation of visual attention, and oculomotor behavior refers to the eye movement in the eye sockets. Although visual fixation patterns and oculomotor behavior measures could be highly correlated simply owing to the fact that alteration of visual attention is always accompanied by oculomotor movement (Johnson et al., [<reflink idref="bib17" id="ref44">17</reflink>]), they are two different concepts. For example, in the case of starring at a target while smoothly shaking one's head, visual attention is not altered but eyes move in the sockets. Despite that a few studies demonstrated similar oculomotor functions between individuals with ASD and their TD counterparts (Avni et al., [<reflink idref="bib2" id="ref45">2</reflink>]; van der Geest et al., [<reflink idref="bib39" id="ref46">39</reflink>]), atypical oculomotor performance has also been reported in people with ASD by a variety of studies (Miller et al., [<reflink idref="bib25" id="ref47">25</reflink>]; Schmitt et al., [<reflink idref="bib35" id="ref48">35</reflink>]; Sumner et al., [<reflink idref="bib37" id="ref49">37</reflink>]). For example, prior studies implemented visually-guided tasks, in which participants were requested to fixate on a peripheral target as fast as possible after focusing on a central fixation target. It has been reported that compared to the TD group, participants with ASD exhibited reduced saccade accuracy (Miller et al., [<reflink idref="bib25" id="ref50">25</reflink>]; Schmitt et al., [<reflink idref="bib35" id="ref51">35</reflink>]), longer saccade latency (Miller et al., [<reflink idref="bib25" id="ref52">25</reflink>]), and increased instability of fixation (Johnson et al., [<reflink idref="bib17" id="ref53">17</reflink>]; Sumner et al., [<reflink idref="bib37" id="ref54">37</reflink>]) in these visually-guided tasks.</p> <hd id="AN0162233062-3">Brief Introduction to This Study</hd> <p>Data of this study came from our previous research (Zhao et al., [<reflink idref="bib48" id="ref55">48</reflink>]), in which both children with ASD and TD were eye-tracked in a face-to-face conversation. A head-mounted eye tracker (Tobii Pro Glasses 2) was tightly attached to the participant's head (without causing self-reported discomfort) to avoid relative movement between the head and the eye tracker. Since the scene camera is fixed on the eye tracker, gaze behavior mapped onto the video taken by the scene camera represents how eyes move in the sockets, and thus it could be used to estimate the oculomotor performance. Note that gaze behavior refers to the raw time series of gaze data exported by the Tobii Pro Lab, and it directly reflects oculomotor behavior in the present study. Our previous study examined both the regularity and amount of oculomotor behavior at different length in both children with ASD and TD. The regularity of oculomotor behavior was computed as the entropy of gaze allocation, with the higher entropy indicating greater level of randomness or lower level of regularity. It was found that gaze allocation was significantly more random at all lengths of gaze behavior, and the amount of oculomotor behavior was only significantly greater within 3 s in children with ASD (Zhao et al., [<reflink idref="bib48" id="ref56">48</reflink>]). These results suggested that gaze allocation might contain more powerful information to classify these two groups of participants than the amount of oculomotor behavior. Therefore, we only extracted features from gaze allocation in the present study. Note that gaze refers to the raw signal collected by the eye tracker, and fixation represents maintaining the visual gaze on a single location. Previous studies typically used visual fixation metrics to understand the situation of visual attention in individuals with ASD (Jones & Klin, [<reflink idref="bib19" id="ref57">19</reflink>]; Santos et al., [<reflink idref="bib33" id="ref58">33</reflink>]). Since gaze includes both fixations and saccades, gaze allocation is different from fixation. In addition, gaze allocation is different from gaze behavior. More specifically, gaze allocation captured how gaze behavior was mapped spatially without taking into account the temporal information of the gaze behavior.</p> <p>Specifically, the present study computed the entropy and range of gaze allocation as features of oculomotor performance. The time series of gaze behavior in each participant was segmented into shorter behaviors at different lengths. The purpose of this procedure was to generate gaze behavior segments for both groups of participants. Gaze behavior segments produced by the ASD children were defined as ASD behaviors, and those by the TD children as TD behaviors. The objective of the present study was two-fold. First, we investigated whether these two types of gaze behavior could be precisely classified by implementing a ML approach. Second, gaze behavior segments were labeled by the ML model as ASD-like and TD-like behaviors. Based on the proportion of ASD-like behavior each participant had, it was explored whether ASD and TD participants could be accurately classified.</p> <hd id="AN0162233062-4">Method</hd> <p></p> <hd id="AN0162233062-5">Experimental Setup and Procedure</hd> <p>In this study, twenty children with ASD and twenty-three children with TD were enrolled, with both groups of participants aged between 6 and 13 years old, and with at least average non-verbal intelligence. Specifically, Children with ASD were recruited from a first-class mental health center in China. Due to a limited access to established ASD diagnostic instruments such as ADOS and ADI-R, the accuracy of the ASD diagnosis in the present study was assured through a variety of other rigorous procedures. Specifically, by strictly following the DSM-IV criteria, the diagnosis of ASD was first made by a licensed psychiatrist with no less than five years' clinical experience, and then the diagnosis was further evaluated by a senior psychiatrist. Additional consultation with at least two other senior psychiatrists would be conducted if disagreement between psychiatrists took place. All these procedures ensured the correctness of ASD diagnosis in our study. In addition, the inclusion criteria for the ASD group included: (a) aged between 6 and 13 years old; (b) having at least average non-verbal intellectual ability (IQ was first screened by the psychiatrist and subsequently assessed as <emph>IQ</emph> ≥ <emph>70</emph> with the Raven's Advanced Progressive Matrices); (c) absence of other clinical conditions such as schizophrenia and attention-deficit hyperactivity disorder (ADHD), and not on medication at the time of experiment; (d) being capable of maintaining average verbal communication (assessed by a speech-language psychologist in preliminary screening). Children with TD were healthy participants recruited from nearby communities and they also received the Raven's Advanced Progressive Matrices to confirm that their IQ was higher than 70. None of the children with TD reported any physical or mental disorders, and no diagnosis of ASD/ADHD in the first-degree relatives was reported. Data of 4 participants (1 ASD and 3 TDs) were excluded due to technical problems in the eye-tracking process. Data of the participants' demographics are presented in Table 1.</p> <p>Table 1 Participants' demographics</p> <p> <ephtml> <table frame="hsides" rules="groups"><thead><tr><th align="left" /><th align="left"><p>ASD</p></th><th align="left"><p>TD</p></th><th align="left"><p>Group comparison</p></th><th align="left"><p><italic>p</italic> value</p></th></tr></thead><tbody><tr><td align="left"><p>Sex (M:F)<sup>†</sup></p></td><td align="left"><p>17:2</p></td><td align="left"><p>17:3</p></td><td align="left"><p>χ<sup>2</sup>(1) =.174</p></td><td char="." align="char"><p>.676</p></td></tr><tr><td align="left"><p>Age in months (Mean ± SD)<sup>‡</sup></p></td><td align="left"><p>99.6 ± 25.1</p></td><td align="left"><p>108.8 ± 27.0</p></td><td align="left"><p>t(37) = 1.09</p></td><td char="." align="char"><p>.283</p></td></tr><tr><td align="left"><p>IQ (Mean ± SD)<sup>‡</sup></p></td><td align="left"><p>100.8 ± 22.7</p></td><td align="left"><p>116.1 ± 22.7</p></td><td align="left"><p>t(37) = 2.45</p></td><td char="." align="char"><p>.019<sup>*</sup></p></td></tr></tbody></table> </ephtml> </p> <p>*.01 < <emph>p</emph> <.05 <sups>†</sups>Chi-square test was performed <sups>‡</sups>Independent samples t-test was performed</p> <p>Participants were engaged in a structured conversation with an interviewer, who was seated 80 cm away from the participants. Four sessions/topics were arranged for all participants: Generic question (also referred to the 1st session), Hobby sharing (2nd session), Yes–no question (3rd session), and Question raising (4th session). Only data of the first session (Generic question) were analyzed in the present study. Data of one session instead of all the four sessions were used because gaze behavior might be sensitive to conversational topic (Jones et al., [<reflink idref="bib18" id="ref59">18</reflink>]; Zhao et al., [<reflink idref="bib47" id="ref60">47</reflink>]). Characteristics of gaze behavior segments would be affected by the conversational topic. Thus, it would be less effective for the ML model to classify gaze behavior segments produced by the two groups of participants when using combined data from different sessions. Further, the first session was chosen as it was the longest session among the four so that more gaze behavior segments could be generated. In this particular session, the interviewer posed six questions to the participants, and the purpose of this session was to help both people get familiar with each other (please refer to Appendix for questions asked in this session). The average lengths for the ASD and TD children were 191 s and 110 s, respectively. Mann–Whitney <emph>U</emph> test demonstrated that the ASD children spent significantly longer time in this session than the TD group (<emph>U</emph> = 48, <emph>p</emph> < 0.01). It was assumed that longer session lengths for the ASD group might be attributed to their social deficits. Children with ASD might have experienced greater difficulty in processing the social information (e.g., motivation, mental state, and emotion) conveyed by the interviewer, and thus resulting in longer session lengths.</p> <p>During the conversation, the participants were asked to wear an eye tracker (Tobii Pro Glasses 2, Tobii Technology, Stockholm, Sweden; sampling frequency: 50 Hz), which served the purpose of tracking the participants' natural gaze behavior without constraining head movement. Before the start of each conversation, the one-point calibration procedure was conducted, in which participants were asked to gaze at the center point of the calibration card placed around the position of the interlocutor's body (within Tobii's recommended calibration distance of 0.5–1.5 m). Once the calibration was complete, participants were asked to look at a few targets (e.g., the interviewer's eyes, nose and different parts of body) to verify the success of calibration in the live view video. An acceptable calibration was considered if the red circle indicating the gaze point overlapped the targets. Conversations were not launched until the eye tracker had been successfully calibrated. Participants were neither told the genuine function of the eye tracker, nor suggested to move the glasses or to make abrupt or intense head movements. Post-experiment questions asked whether they knew the genuine function of the glasses, and none of them were aware that the eye tracker was used for recording gaze behavior. The eye tracker had a scene camera, which captured the front scene of the wearer (Fig. 1a). The size of the scene camera was 1920 pixels (x-axis) × 1080 pixels (y-axis) (Fig. 1b). After recording, the x and y coordinates of each gaze point were exported by the Tobii Pro Lab in units of pixels.</p> <p>Graph: Fig. 1 a Scene camera and the coordinate system. x and y coordinates correspond to the horizontal and vertical directions of gaze behavior respectively. b Illustration of the coordinate system of the scene video camera and the position of gaze allocation</p> <hd id="AN0162233062-6">Data Processing</hd> <p>As the first step, the whole time series of gaze data in each participant was segmented into shorter, non-overlapping behavior segments at different lengths, ranging from 1 to 10 s, with the interval of 1 s (Fig. 2a). Specifically, for behavior segments at <emph>k</emph>-second, the first segment included all the gaze points between the 1st and the 50 × <emph>k-th</emph> points in the time series, the second segment incorporated the gaze points between the (50 × <emph>k</emph> + 1)<emph>-th</emph> and 2 × 50 × <emph>k-th</emph> points, and the <emph>i</emph>-th segment incorporated the gaze points between the [(<emph>i</emph>-1) × 50 × <emph>k</emph> + 1]<emph>-th</emph> and <emph>i</emph> × 50 × <emph>k-th</emph> points. By repeating the same approach, the segmentation window slid along the time series to obtain the rest of the behavior segments in a non-overlapping way. Segments with substantial data loss (> 30%) were discarded (Wang et al., [<reflink idref="bib43" id="ref61">43</reflink>]). The number of segments at different lengths for both groups of participants is presented in Fig. 2b, which shows that the longer the segment the fewer the quantity.</p> <p>Graph: Fig. 2 a Creation of gaze behavior segments with the length of k-second(s). b Illustration of the number of gaze behavior segments at different lengths in both groups of participants</p> <p>Second, six features accounting for oculomotor behavior were computed for each gaze behavior segment. Based on our previous finding that gaze was more randomly allocated at all examined lengths in children with ASD as compared to the TD peers (Zhao et al., [<reflink idref="bib48" id="ref62">48</reflink>]), these six features included the ranges and entropy of gaze allocation on the x- and y-axis, and in the x–y coordinate plane. The range of gaze allocation was computed as the distance between two gaze points that were most remotely allocated on the x- or y-axis, or in the x–y coordinate plane (Fig. 3a), and it reflected the maximal distance that oculomotor behavior covered spatially. Entropy analysis has been commonly used in various fields to compute the complexity/regularity of behavior (Fournier et al., [<reflink idref="bib10" id="ref63">10</reflink>]; Kolmogorov, [<reflink idref="bib20" id="ref64">20</reflink>]; Li et al., [<reflink idref="bib22" id="ref65">22</reflink>]; Shannon, [<reflink idref="bib36" id="ref66">36</reflink>]; Zhao et al., [<reflink idref="bib50" id="ref67">50</reflink>]). In this study, the entropy of gaze allocation was used to reflect the regularity of oculomotor behavior. It was computed with Shannon entropy as:</p> <p> <ephtml> <math display="block" xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mtext>Entropy</mtext><mspace width="0.166667em" /><mo>=</mo><mspace width="0.166667em" /><mo>-</mo><mspace width="0.166667em" /><munderover><mo movablelimits="false">∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>n</mi></munderover><mi>p</mi><mfenced close=")" open="("><msub><mi>x</mi><mi>m</mi></msub></mfenced><mrow /><mo>∗</mo><mo>log</mo><mn>2</mn><mi>p</mi><mfenced close=")" open="("><msub><mi>x</mi><mi>m</mi></msub></mfenced></mrow></math> </ephtml> </p> <p>Graph</p> <p>Graph: Fig. 3 a Range of gaze allocation on the x- and y-axis, and in the x–y coordinate plane. b Exemplary illustration of gaze allocation for an ASD participant with a 5-s segment</p> <p>To compute the entropy on the x- and y-axis, the first step was to create equal-distance bins, with 10 pixels per bin. Given 1920 pixels on the x-axis and 1080 pixels on the y-axis, 192 bins were created for the x-axis and 108 bins for the y-axis. In these two cases, <emph>n</emph> equaled 192 and 108 in the entropy calculation on the x- and y-axis, respectively. <emph>p(x</emph><subs><emph>m</emph></subs><emph>)</emph> represented the probability of gaze allocated in the <emph>m-</emph>th bin. To calculate the entropy of gaze allocation in the x–y coordinate plane, given that the size of the scene camera video was 1920 × 1080 pixels, the image was divided into 20,736 (192 × 108) equally sized blocks (Fig. 3b). This meant that the size of each block was 10 × 10 pixels. Therefore, <emph>n</emph> equaled 20,736, and <emph>p</emph>(<emph>x</emph><subs><emph>m</emph></subs>) referred to the probability of gaze allocation in the <emph>m-</emph>th block.</p> <p>Since the number of gaze behavior segments varied with the length of the segment, the obtained feature dataset was different for different segment lengths. Specifically, for the <emph>k</emph>-second gaze behavior segment, a matrix of <emph>N</emph> (= the corresponding number of gaze behavior segment at <emph>k</emph>-second, illustrated in Fig. 2b) × 6 (number of features) was obtained. Gaze behavior segments generated by children with ASD were labeled as ASD behaviors, and those by children with TD as TD behaviors.</p> <hd id="AN0162233062-7">Classification of ASD and TD Gaze Behavior Segments</hd> <p>This study implemented support vector machine (SVM) to classify gaze behavior segments for each examined length, separately. In previous literature, SVM has been often used in ML studies in the field of ASD (Liu et al., [<reflink idref="bib23" id="ref68">23</reflink>]; Zhao et al., [<reflink idref="bib49" id="ref69">49</reflink>]). The aim of this classifier is to establish an optimal hyperplane in a multidimensional space with labelled training samples. Testing samples are classified based on the sign of the distance vector to the hyperplane. The distance to the hyperplane determines the probability they belong to a specific category.</p> <p>Forward feature selection (FFS) was used to select features for model training and testing. Specifically, FFS is an iterative process starting with the examination of each individual feature by testing their classification performance. The feature with the highest classification accuracy would be preserved, and then be combined with each of the other features to form two-feature models, whose classification performances were further evaluated. The two features with the optimal classification accuracy were retained and used to establish three-feature models by combining them with each of the remaining features. By repeating these procedures, the one-feature, two-feature, ..., <emph>q</emph>-feature models with the highest classification accuracy would be obtained (<emph>q</emph> equals the total number of examined features intended to be fed into the ML model). In sum, FFS helped achieve not only the model with the highest classification accuracy, but also the corresponding feature or feature combination.</p> <p>To minimize the potential overfitting problem, tenfold cross validation (tenfold-CV) was applied in the ML model training and testing. Specifically, all behavior samples were randomly partitioned into 10 subsamples. For each iteration, nine of these ten subsamples were used to train the ML model, and the remaining subsample was used as the test set. Following the process of FFS, each cross-validation iteration was trained separately using the same feature(s) as the other nine iterations. Note that feature(s) fed into the ML model were determined by the FFS process. The trained ML model labeled all the behavior segments in the test set as ASD-like or TD-like behaviors. This procedure was repeated 10 times until all the ten subsamples were labeled. Finally, all behavior segments in all participants were labeled as ASD-like or TD-like behaviors. Therefore, each behavior sample would be both labeled as an ASD/TD behavior segment and an ASD-like/TD-like behavior segment. Note that ASD/TD behavior segments are defined as those produced by the children with ASD/TD, whereas ASD-like/TD-like behavior segments refer to those labeled by the ML model. A child with ASD/TD could only produce ASD/TD behavior segments, but he/she could have both TD-like and ASD-like behavior segments. Based on these results, the sensitivity, specificity, and the accuracy of the ML models in classifying behavior samples were calculated. Specifically, sensitivity was computed as the percentage of ASD behavior segments that were correctly labeled as ASD-like behavior by the ML model, and specificity to the percentage of TD behavior segments that were labeled as TD-like behavior. Accuracy was computed as the percentage of behavior segments that were correctly labeled as ASD-like and TD-like behaviors across all behavior samples.</p> <hd id="AN0162233062-8">Classification of ASD and TD Children</hd> <p>Based on the assumption that children with ASD would produce higher percentage of ASD-like behaviors than children with TD, threshold classifier was used to classify children with ASD and TD. Specifically, once all the gaze behavior segments from a child were classified as ASD-like or TD-like behaviors, the percentages of classified ASD-like behaviors and TD-like behaviors were calculated. If the percentage of the classified ASD-like behaviors exceeded a predetermined threshold, the participant would be classified as an ASD child. Otherwise, he or she would be classified as a TD child. Different percentage thresholds were tested for each examined length of gaze behavior segment, and the optimal one that was associated with the highest classification accuracy was determined.</p> <hd id="AN0162233062-9">Results</hd> <p></p> <hd id="AN0162233062-10">Classification Performance on Gaze Behavior Segments</hd> <p>Information concerning the group differences on the computed six features is listed in Table 2, and Fig. 4 presents the optimal gaze behavior segment classification accuracy achieved at different segment lengths. To illustrate that the classification accuracy was higher than the chance level, results on majority class prediction are also plotted as a comparison. In majority class prediction, every behavior was classified as the category with higher percentage in quantity. For example, more ASD behavior segments were generated for the 1-s behavior in these two groups of participants, and thus the majority class prediction labeled all the 1-s gaze behaviors as ASD-like behaviors. Figure 4 demonstrates that the classification accuracy at different lengths of gaze behavior segments is higher than the one with majority classification prediction.</p> <p>Table 2 Comparisons of six features between ASD and TD at different lengths</p> <p> <ephtml> <table frame="hsides" rules="groups"><thead><tr><th align="left" /><th align="left" colspan="3"><p>Entropy_xy-plane (bits)</p></th><th align="left" colspan="3"><p>Entropy_x-axis (bits)</p></th><th align="left" colspan="3"><p>Entropy_y-axis (bits)</p></th></tr><tr><th align="left" /><th align="left"><p>ASD(M ± SD)</p></th><th align="left"><p>TD(M ± SD)</p></th><th align="left"><p>Group comparison</p></th><th align="left"><p>ASD(M ± SD)</p></th><th align="left"><p>TD(M ± SD)</p></th><th align="left"><p>Group comparison</p></th><th align="left"><p>ASD(M ± SD)</p></th><th align="left"><p>TD(M ± SD)</p></th><th align="left"><p>Group Comparison</p></th></tr></thead><tbody><tr><td align="left"><p>1 s</p></td><td char="±" align="char"><p>4.2 ± 0.4</p></td><td char="±" align="char"><p>3.6 ± 0.3</p></td><td align="left"><p>t(37) = 4.3</p><p>p < 0.001</p></td><td char="±" align="char"><p>3.0 ± 0.5</p></td><td char="±" align="char"><p>2.5 ± 0.3</p></td><td align="left"><p>t(23.8) = 3.6</p><p>p = 0.001</p></td><td char="±" align="char"><p>3.1 ± 0.4</p></td><td char="±" align="char"><p>2.7 ± 0.3</p></td><td align="left"><p>t(37) = 4.0</p><p>p < 0.001</p></td></tr><tr><td align="left"><p>2 s</p></td><td char="±" align="char"><p>5.0 ± 0.4</p></td><td char="±" align="char"><p>4.5 ± 0.3</p></td><td align="left"><p>t(37) = 4.4</p><p>p < 0.001</p></td><td char="±" align="char"><p>3.6 ± 0.5</p></td><td char="±" align="char"><p>3.1 ± 0.3</p></td><td align="left"><p>t(27.1) = 3.6</p><p>p = 0.001</p></td><td char="±" align="char"><p>3.7 ± 0.4</p></td><td char="±" align="char"><p>3.3 ± 0.3</p></td><td align="left"><p>t(37) = 3.8</p><p>p < 0.001</p></td></tr><tr><td align="left"><p>3 s</p></td><td char="±" align="char"><p>5.5 ± 0.5</p></td><td char="±" align="char"><p>5.0 ± 0.4</p></td><td align="left"><p>t(37) = 3.7</p><p>p = 0.001</p></td><td char="±" align="char"><p>4.0 ± 0.6</p></td><td char="±" align="char"><p>3.5 ± 0.3</p></td><td align="left"><p>t(28.4) = 3.0</p><p>p = 0.006</p></td><td char="±" align="char"><p>4.0 ± 0.4</p></td><td char="±" align="char"><p>3.6 ± 0.3</p></td><td align="left"><p>t(37) = 3.6</p><p>p = 0.001</p></td></tr><tr><td align="left"><p>4 s</p></td><td char="±" align="char"><p>5.9 ± 0.5</p></td><td char="±" align="char"><p>5.4 ± 0.4</p></td><td align="left"><p>t(37) = 3.9</p><p>p < 0.001</p></td><td char="±" align="char"><p>4.2 ± 0.6</p></td><td char="±" align="char"><p>3.8 ± 0.3</p></td><td align="left"><p>t(28.8) = 3.1</p><p>p = 0.004</p></td><td char="±" align="char"><p>4.3 ± 0.4</p></td><td char="±" align="char"><p>3.9 ± 0.3</p></td><td align="left"><p>t(37) = 3.3</p><p>p = 0.002</p></td></tr><tr><td align="left"><p>5 s</p></td><td char="±" align="char"><p>6.2 ± 0.5</p></td><td char="±" align="char"><p>5.7 ± 0.4</p></td><td align="left"><p>t(37) = 4.0</p><p>p < 0.001</p></td><td char="±" align="char"><p>4.4 ± 0.6</p></td><td char="±" align="char"><p>3.9 ± 0.4</p></td><td align="left"><p>t(28.9) = 2.7</p><p>p = 0.011</p></td><td char="±" align="char"><p>4.4 ± 0.4</p></td><td char="±" align="char"><p>4.0 ± 0.3</p></td><td align="left"><p>t(37) = 3.5</p><p>p = 0.001</p></td></tr><tr><td align="left"><p>6 s</p></td><td char="±" align="char"><p>6.4 ± 0.5</p></td><td char="±" align="char"><p>5.9 ± 0.4</p></td><td align="left"><p>t(37) = 3.6</p><p>p = 0.001</p></td><td char="±" align="char"><p>4.5 ± 0.6</p></td><td char="±" align="char"><p>4.1 ± 0.4</p></td><td align="left"><p>t(28.6) = 2.5</p><p>p = 0.017</p></td><td char="±" align="char"><p>4.6 ± 0.4</p></td><td char="±" align="char"><p>4.2 ± 0.3</p></td><td align="left"><p>t(37) = 3.3</p><p>p = 0.002</p></td></tr><tr><td align="left"><p>7 s</p></td><td char="±" align="char"><p>6.6 ± 0.5</p></td><td char="±" align="char"><p>6.1 ± 0.4</p></td><td align="left"><p>t(31.8) = 3.4</p><p>p = 0.002</p></td><td char="±" align="char"><p>4.6 ± 0.7</p></td><td char="±" align="char"><p>4.2 ± 0.4</p></td><td align="left"><p>t(28.0) = 2.4</p><p>p = 0.025</p></td><td char="±" align="char"><p>4.6 ± 0.4</p></td><td char="±" align="char"><p>4.3 ± 0.3</p></td><td align="left"><p>t(37) = 3.2</p><p>p = 0.003</p></td></tr><tr><td align="left"><p>8 s</p></td><td char="±" align="char"><p>6.8 ± 0.5</p></td><td char="±" align="char"><p>6.3 ± 0.4</p></td><td align="left"><p>t(37) = 3.6</p><p>p = 0.001</p></td><td char="±" align="char"><p>4.7 ± 0.7</p></td><td char="±" align="char"><p>4.3 ± 0.4</p></td><td align="left"><p>t(28.8) = 2.2</p><p>p = 0.033</p></td><td char="±" align="char"><p>4.7 ± 0.4</p></td><td char="±" align="char"><p>4.4 ± 0.3</p></td><td align="left"><p>t(37) = 3.1</p><p>p = 0.003</p></td></tr><tr><td align="left"><p>9 s</p></td><td char="±" align="char"><p>6.9 ± 0.5</p></td><td char="±" align="char"><p>6.5 ± 0.4</p></td><td align="left"><p>t(37) = 3.4</p><p>p = 0.001</p></td><td char="±" align="char"><p>4.7 ± 0.7</p></td><td char="±" align="char"><p>4.4 ± 0.4</p></td><td align="left"><p>t(29.8) = 2.1</p><p>p = 0.042</p></td><td char="±" align="char"><p>4.8 ± 0.4</p></td><td char="±" align="char"><p>4.5 ± 0.3</p></td><td align="left"><p>t(37) = 3.3</p><p>p = 0.002</p></td></tr><tr><td align="left"><p>10 s</p></td><td char="±" align="char"><p>7.1 ± 05</p></td><td char="±" align="char"><p>6.6 ± 0.4</p></td><td align="left"><p>t(37) = 3.2</p><p>p = 0.003</p></td><td char="±" align="char"><p>4.8 ± 0.7</p></td><td char="±" align="char"><p>4.4 ± 0.4</p></td><td align="left"><p>t(29.2) = 2.1</p><p>p = 0.048</p></td><td char="±" align="char"><p>4.9 ± 0.4</p></td><td char="±" align="char"><p>4.5 ± 0.3</p></td><td align="left"><p>t(37) = 3.0</p><p>p = 0.006</p></td></tr></tbody></table> </ephtml> </p> <p>Table 2 Comparisons of six features between ASD and TD at different lengths</p> <p> <ephtml> <table frame="hsides" rules="groups"><thead><tr><th align="left" /><th align="left" colspan="3"><p>Range_xy-plane (pixels)</p></th><th align="left" colspan="3"><p>Range_x-axis (bits)</p></th><th align="left" colspan="3"><p>Range_y-axis (bits)</p></th></tr><tr><th align="left" /><th align="left"><p>ASD(M ± SD)</p></th><th align="left"><p>TD(M ± SD)</p></th><th align="left"><p>Group Comparison</p></th><th align="left"><p>ASD(M ± SD)</p></th><th align="left"><p>TD(M ± SD)</p></th><th align="left"><p>Group Comparison</p></th><th align="left"><p>ASD(M ± SD)</p></th><th align="left"><p>TD(M ± SD)</p></th><th align="left"><p>Group Comparison</p></th></tr></thead><tbody><tr><td align="left"><p>1 s</p></td><td char="±" align="char"><p>321.5 ± 105.4</p></td><td char="±" align="char"><p>254.8 ± 79.8</p></td><td align="left"><p>t(37) = 2.2</p><p>p = 0.031</p></td><td char="±" align="char"><p>250.6 ± 95.4</p></td><td char="±" align="char"><p>184.9 ± 64.1</p></td><td align="left"><p>t(31.3) = 2.5</p><p>p = 0.017</p></td><td char="±" align="char"><p>198.9 ± 60.4</p></td><td char="±" align="char"><p>166.3 ± 56.9</p></td><td align="left"><p>t(37) = 1.7</p><p>p = 0.090</p></td></tr><tr><td align="left"><p>2 s</p></td><td char="±" align="char"><p>471.4 ± 150.4</p></td><td char="±" align="char"><p>378.3 ± 110.5</p></td><td align="left"><p>t(37) = 2.2</p><p>p = 0.033</p></td><td char="±" align="char"><p>385.5 ± 147.5</p></td><td char="±" align="char"><p>286.5 ± 93.6</p></td><td align="left"><p>t(30.2) = 2.5</p><p>p = 0.019</p></td><td char="±" align="char"><p>296.4 ± 81.0</p></td><td char="±" align="char"><p>251.8 ± 74.7</p></td><td align="left"><p>t(37) = 1.8</p><p>p = 0.082</p></td></tr><tr><td align="left"><p>3 s</p></td><td char="±" align="char"><p>557.7 ± 182.9</p></td><td char="±" align="char"><p>460.1 ± 139.1</p></td><td align="left"><p>t(37) = 1.9</p><p>p = 0.068</p></td><td char="±" align="char"><p>458.0 ± 186.6</p></td><td char="±" align="char"><p>367.7 ± 126.7</p></td><td align="left"><p>t(37) = 1.8</p><p>p = 0.084</p></td><td char="±" align="char"><p>365.6 ± 102.9</p></td><td char="±" align="char"><p>311.7 ± 98.0</p></td><td align="left"><p>t(37) = 1.7</p><p>p = 0.102</p></td></tr><tr><td align="left"><p>4 s</p></td><td char="±" align="char"><p>644.1 ± 197.1</p></td><td char="±" align="char"><p>526.3 ± 154.3</p></td><td align="left"><p>t(37) = 2.1</p><p>p = 0.044</p></td><td char="±" align="char"><p>550.8 ± 202.4</p></td><td char="±" align="char"><p>437.0 ± 149.5</p></td><td align="left"><p>t(37) = 2.0</p><p>p = 0.052</p></td><td char="±" align="char"><p>411.6 ± 112.6</p></td><td char="±" align="char"><p>360.6 ± 106.3</p></td><td align="left"><p>t(37) = 1.5</p><p>p = 0.154</p></td></tr><tr><td align="left"><p>5 s</p></td><td char="±" align="char"><p>677.8 ± 218.5</p></td><td char="±" align="char"><p>584.0 ± 167.9</p></td><td align="left"><p>t(37) = 1.5</p><p>p = 0.140</p></td><td char="±" align="char"><p>580.3 ± 230.6</p></td><td char="±" align="char"><p>492.2 ± 170.1</p></td><td align="left"><p>t(37) = 1.4</p><p>p = 0.181</p></td><td char="±" align="char"><p>446.4 ± 123.5</p></td><td char="±" align="char"><p>393.8 ± 110.2</p></td><td align="left"><p>t(37) = 1.4</p><p>p = 0.168</p></td></tr><tr><td align="left"><p>6 s</p></td><td char="±" align="char"><p>725.3 ± 232.5</p></td><td char="±" align="char"><p>615.3 ± 186.1</p></td><td align="left"><p>t(37) = 1.6</p><p>p = 0.111</p></td><td char="±" align="char"><p>630.4 ± 240.5</p></td><td char="±" align="char"><p>533.7 ± 188.7</p></td><td align="left"><p>t(37) = 1.4</p><p>p = 0.170</p></td><td char="±" align="char"><p>470.6 ± 124.9</p></td><td char="±" align="char"><p>415.9 ± 118.7</p></td><td align="left"><p>t(37) = 1.4</p><p>p = 0.169</p></td></tr><tr><td align="left"><p>7 s</p></td><td char="±" align="char"><p>768.5 ± 231.6</p></td><td char="±" align="char"><p>666.9 ± 177.2</p></td><td align="left"><p>t(37) = 1.5</p><p>p = 0.131</p></td><td char="±" align="char"><p>671.2 ± 245.2</p></td><td char="±" align="char"><p>579.1 ± 185.4</p></td><td align="left"><p>t(37) = 1.3</p><p>p = 0.192</p></td><td char="±" align="char"><p>503.4 ± 140.9</p></td><td char="±" align="char"><p>447.0 ± 120.0</p></td><td align="left"><p>t(37) = 1.4</p><p>p = 0.185</p></td></tr><tr><td align="left"><p>8 s</p></td><td char="±" align="char"><p>791.2 ± 248.5</p></td><td char="±" align="char"><p>689.4 ± 192.0</p></td><td align="left"><p>t(37) = 1.4</p><p>p = 0.159</p></td><td char="±" align="char"><p>696.0 ± 265.0</p></td><td char="±" align="char"><p>609.6 ± 197.4</p></td><td align="left"><p>t(37) = 1.2</p><p>p = 0.253</p></td><td char="±" align="char"><p>518.7 ± 138.7</p></td><td char="±" align="char"><p>462.5 ± 124.9</p></td><td align="left"><p>t(37) = 1.3</p><p>p = 0.191</p></td></tr><tr><td align="left"><p>9 s</p></td><td char="±" align="char"><p>817.4 ± 271.6</p></td><td char="±" align="char"><p>710.3 ± 188.3</p></td><td align="left"><p>t(37) = 1.4</p><p>p = 0.159</p></td><td char="±" align="char"><p>721.0 ± 284.3</p></td><td char="±" align="char"><p>624.1 ± 200.5</p></td><td align="left"><p>t(37) = 1.2</p><p>p = 0.225</p></td><td char="±" align="char"><p>548.3 ± 150.9</p></td><td char="±" align="char"><p>487.7 ± 130.9</p></td><td align="left"><p>t(37) = 1.3</p><p>p = 0.188</p></td></tr><tr><td align="left"><p>10 s</p></td><td char="±" align="char"><p>872.5 ± 302.0</p></td><td char="±" align="char"><p>741.4 ± 186.4</p></td><td align="left"><p>t(37) = 1.6</p><p>p = 0.109</p></td><td char="±" align="char"><p>769.1 ± 317.3</p></td><td char="±" align="char"><p>661.9 ± 201.7</p></td><td align="left"><p>t(37) = 1.3</p><p>p = 0.213</p></td><td char="±" align="char"><p>574.0 ± 164.0</p></td><td char="±" align="char"><p>494.6 ± 121.5</p></td><td align="left"><p>t(37) = 1.7</p><p>p = 0.093</p></td></tr></tbody></table> </ephtml> </p> <p>Graph: Fig. 4 Classification accuracy for gaze behavior segments at different lengths using the SVM classifier (in orange) and major class predication approach (in blue), respectively</p> <p>As is shown, the optimal classification accuracy increases with the length of the gaze behavior segment, and the accuracy maintained above 70% with segments longer than 5 s. A maximum accuracy of 74.15% was obtained with the 9-s gaze behavior segments.</p> <p>The classification performance with different number of feature(s) for the 9-s behavior segments is presented in Table 3, which shows that the maximum classification accuracy was yielded with only one feature—entropy in the x–y coordinate plane.</p> <p>Table 3 Classification performance of different number of feature(s) for the 9-s behavior segments</p> <p> <ephtml> <table frame="hsides" rules="groups"><thead><tr><th align="left"><p>Number of features</p></th><th align="left"><p>Added feature</p></th><th align="left"><p>Accuracy (%)</p></th><th align="left"><p>Sensitivity (%)</p></th><th align="left"><p>Specificity (%)</p></th></tr></thead><tbody><tr><td align="left"><p>1</p></td><td align="left"><p>Entropy_x-y</p></td><td char="." align="char"><p>74.15</p></td><td char="." align="char"><p>75.45</p></td><td char="." align="char"><p>72.51</p></td></tr><tr><td align="left"><p>2</p></td><td align="left"><p> ~ + Range_y-axis</p></td><td char="." align="char"><p>73.90</p></td><td char="." align="char"><p>74.62</p></td><td char="." align="char"><p>73.10</p></td></tr><tr><td align="left"><p>3</p></td><td align="left"><p> ~ + Entropy_x-axis</p></td><td char="." align="char"><p>73.13</p></td><td char="." align="char"><p>75.57</p></td><td char="." align="char"><p>70.32</p></td></tr><tr><td align="left"><p>4</p></td><td align="left"><p> ~ + Range_x-y</p></td><td char="." align="char"><p>73.66</p></td><td char="." align="char"><p>75.60</p></td><td char="." align="char"><p>71.32</p></td></tr><tr><td align="left"><p>5</p></td><td align="left"><p> ~ + Entropy_y-axis</p></td><td char="." align="char"><p>73.39</p></td><td char="." align="char"><p>75.62</p></td><td char="." align="char"><p>71.05</p></td></tr><tr><td align="left"><p>6</p></td><td align="left"><p> ~ + Range_x-axis</p></td><td char="." align="char"><p>73.41</p></td><td char="." align="char"><p>75.10</p></td><td char="." align="char"><p>71.61</p></td></tr></tbody></table> </ephtml> </p> <p>Entropy_x-y: entropy in the x–y plane In forward feature selection (FFS), ~ represents all features in the previous iteration; for example, ~ represents all 3 previously selected features in the 4<sups>th</sups> iteration</p> <hd id="AN0162233062-11">Classification Performance on Children</hd> <p>Table 4 shows the performance of children classification using threshold classifiers, which demonstrates that the highest classification accuracy of 87.18% (sensitivity: 89.47%, specificity: 85.00%) was achieved with the 7-s behavior segments when the percentage threshold (optimal threshold) was 46%. The ROC (Receiver Operator Characteristic) of the threshold classifier is plotted in Fig. 5. Note that in the case of the highest children classification accuracy, the maximum accuracy of classifying 7-s behavior segments was 72.65% (Sensitivity: 78.00%, Specificity: 66.52%) by using three features: entropy in the x–y plane, entropy on y-axis, and range on y-axis.</p> <p>Table 4 Classification performance on children using gaze behavior at different lengths</p> <p> <ephtml> <table frame="hsides" rules="groups"><thead><tr><th align="left"><p>Behavior length</p></th><th align="left"><p>Highest accuracy</p></th><th align="left"><p>Sensitivity</p></th><th align="left"><p>Specificity</p></th><th align="left"><p>Threshold</p></th></tr></thead><tbody><tr><td align="left"><p>1 s</p></td><td char="." align="char"><p>0.79</p></td><td char="." align="char"><p>0.79</p></td><td char="." align="char"><p>0.80</p></td><td char="." align="char"><p>0.54</p></td></tr><tr><td align="left"><p>2 s</p></td><td char="." align="char"><p>0.77</p></td><td char="." align="char"><p>0.68</p></td><td char="." align="char"><p>0.85</p></td><td char="." align="char"><p>0.52</p></td></tr><tr><td align="left"><p>3 s</p></td><td char="." align="char"><p>0.82</p></td><td char="." align="char"><p>0.74</p></td><td char="." align="char"><p>0.90</p></td><td char="." align="char"><p>0.56</p></td></tr><tr><td align="left"><p>4 s</p></td><td char="." align="char"><p>0.77</p></td><td char="." align="char"><p>0.68</p></td><td char="." align="char"><p>0.85</p></td><td char="." align="char"><p>0.51</p></td></tr><tr><td align="left"><p>5 s</p></td><td char="." align="char"><p>0.77</p></td><td char="." align="char"><p>0.79</p></td><td char="." align="char"><p>0.75</p></td><td char="." align="char"><p>0.40</p></td></tr><tr><td align="left"><p>6 s</p></td><td char="." align="char"><p>0.77</p></td><td char="." align="char"><p>0.74</p></td><td char="." align="char"><p>0.80</p></td><td char="." align="char"><p>0.53</p></td></tr><tr><td align="left"><p>7 s</p></td><td char="." align="char"><p>0.87</p></td><td char="." align="char"><p>0.89</p></td><td char="." align="char"><p>0.85</p></td><td char="." align="char"><p>0.46</p></td></tr><tr><td align="left"><p>8 s</p></td><td char="." align="char"><p>0.77</p></td><td char="." align="char"><p>0.63</p></td><td char="." align="char"><p>0.90</p></td><td char="." align="char"><p>0.54</p></td></tr><tr><td align="left"><p>9 s</p></td><td char="." align="char"><p>0.79</p></td><td char="." align="char"><p>0.68</p></td><td char="." align="char"><p>0.90</p></td><td char="." align="char"><p>0.55</p></td></tr><tr><td align="left"><p>10 s</p></td><td char="." align="char"><p>0.85</p></td><td char="." align="char"><p>0.84</p></td><td char="." align="char"><p>0.85</p></td><td char="." align="char"><p>0.56</p></td></tr></tbody></table> </ephtml> </p> <p>Graph: Fig. 5 Receiver operator characteristic (ROC) of the threshold classifier when classifying children based on the proportion of the 7-s ASD-like behavior segments</p> <hd id="AN0162233062-12">Discussion</hd> <p>The present study partitioned the time series of gaze behavior into shorter behavior segments, and used ML to classify these segments produced by children with ASD and TD. Results demonstrated that oculomotor behavior in children with ASD could be discriminated from that in children with TD with an above-chance accuracy. Further, our results showed that segmenting the whole time series of gaze behavior into 7-s behavior segments could help classify children with ASD and TD with a maximum accuracy of 87.18%. These results demonstrated that oculomotor behavior contained valuable information that could be used to identify ASD behavior as well as individuals with ASD.</p> <p>The present study was innovative in two respects. First, we investigated whether oculomotor behavior could be accurately classified between ASD and TD, and whether children with ASD and TD could be discriminated from each other using oculomotor behavior. In contrast to the present study, previous studies extracted features from visual fixation patterns, and it was only explored whether individuals with ASD and TD could be discriminated (Eraslan et al., [<reflink idref="bib9" id="ref70">9</reflink>]; Liu et al., [<reflink idref="bib23" id="ref71">23</reflink>]; Wan et al., [<reflink idref="bib42" id="ref72">42</reflink>]; Yaneva et al., [<reflink idref="bib45" id="ref73">45</reflink>]). Our findings of the accurate classifications on both oculomotor behavior segments and children suggest the feasibility of creating the ASD oculomotor behavior template (i.e., representative ASD oculomotor behavior). This specific idea is similar to Eraslan et al.'s, which proposed performing a scan-path trend analysis to identify representative visual fixation sequences for both individuals with TD and ASD ([<reflink idref="bib9" id="ref74">9</reflink>]). Individuals could be appropriately classified based on the similarity of their visual scan path to the representative sequences. A major difference between our approach and Eraslan et al.'s is that we used oculomotor behavior and they utilized visual fixation behavior. Future studies are encouraged to combine these two approaches to expect a higher classification accuracy.</p> <p>The second innovation of this study lies in that we segmented gaze behavior into shorter lengths, and classified children with ASD and TD based on the proportion of ASD-like behavior. Unlike the present study, previous studies typically computed features from the whole time series of data (Liu et al., [<reflink idref="bib23" id="ref75">23</reflink>]; Wan et al., [<reflink idref="bib42" id="ref76">42</reflink>]; Zhao et al., [<reflink idref="bib46" id="ref77">46</reflink>]). One of the major advantages of the behavior segmentation technique is that data of participants could be maximumly preserved. More specifically, data loss is a common but inevitable phenomenon when collecting time series data with instruments (Raven, [<reflink idref="bib32" id="ref78">32</reflink>]). Participants with greater data loss would have to be removed from further analysis, which may lead to biased conclusions. With the aid of the behavior segmentation technique, however, participants with great data loss could be preserved so long as part of their gaze behavior is intactly recorded. The specific advantage may well serve the automatic screening of ASD since future screening of ASD might be realized by integrating time series data from multiple modalities (e.g., eye-tracking, EEG, neuroimaging, kinematic data) (Crippa et al., [<reflink idref="bib6" id="ref79">6</reflink>]; Grossi et al., [<reflink idref="bib13" id="ref80">13</reflink>]; Li et al., [<reflink idref="bib21" id="ref81">21</reflink>]; Liu et al., [<reflink idref="bib23" id="ref82">23</reflink>]; Osterling & Dawson, [<reflink idref="bib28" id="ref83">28</reflink>]; Plitt et al., [<reflink idref="bib30" id="ref84">30</reflink>]; Wan et al., [<reflink idref="bib42" id="ref85">42</reflink>]; Yamagata et al., [<reflink idref="bib44" id="ref86">44</reflink>]; Zhao et al., [<reflink idref="bib49" id="ref87">49</reflink>]). Once the behavior segmentation technique is implemented to these data, participants could be maximally preserved for classification tasks.</p> <p>Entropy in the x–y coordinate plane emerged as a prominent feature in the models that produced the highest accuracy when classifying both the 9-s and 7-s gaze behavior segments. These results were consistent with Zhao et al.'s finding that eyes moved in a less regular fashion in children with ASD ([<reflink idref="bib48" id="ref88">48</reflink>]). Given the close relation between oculomotor behavior and the central neural system that controls it (Guyon & Elisseeff, [<reflink idref="bib14" id="ref89">14</reflink>]; Pritschet et al., [<reflink idref="bib31" id="ref90">31</reflink>]), future scientific endeavor is encouraged to investigate the neural underpinnings of the irregularity of oculomotor behavior in individuals with ASD.</p> <p>The performance of classifying individuals with ASD and TD was comparable to that reported by other ML eye-tracking studies (Eraslan et al., [<reflink idref="bib9" id="ref91">9</reflink>]; Liu et al., [<reflink idref="bib23" id="ref92">23</reflink>]; Wan et al., [<reflink idref="bib42" id="ref93">42</reflink>]; Yaneva et al., [<reflink idref="bib45" id="ref94">45</reflink>]; Zhao et al., [<reflink idref="bib46" id="ref95">46</reflink>]). Interestingly, our previous work achieved a maximum classification accuracy of 84.62% by using solely visual fixation features (Zhao et al., [<reflink idref="bib46" id="ref96">46</reflink>]) when classifying the same sample of participants as in the present study. The present study was different from Zhao et al. ([<reflink idref="bib46" id="ref97">46</reflink>]) in three main respects. First, features used to establish ML models and their physical significances were different. Zhao et al. ([<reflink idref="bib46" id="ref98">46</reflink>]) extracted features from visual fixation patterns, and they reflected the characteristics of visual attention to different areas of interest (Zhao et al., [<reflink idref="bib46" id="ref99">46</reflink>]). As mentioned earlier, the present study computed features out of gaze allocation patterns, which captured the regularity and range of the oculomotor behavior. Second, Zhao et al. ([<reflink idref="bib46" id="ref100">46</reflink>]) extracted features from the whole time series (Zhao et al., [<reflink idref="bib46" id="ref101">46</reflink>]), but the present study used the behavior segmentation technique and the threshold classifier to classify both behavior segments and children. Given that eye-tracking data is vulnerable to data loss, the ML approach presented in this study is superior to Zhao et al.'s ([<reflink idref="bib46" id="ref102">46</reflink>]) in terms of preserving participants' data. Third, the classification accuracy of the present study was slightly higher than Zhao et al. ([<reflink idref="bib46" id="ref103">46</reflink>]) when using visual fixation features alone (i.e., 84.62%) (Zhao et al., [<reflink idref="bib46" id="ref104">46</reflink>]), suggesting that oculomotor performance might be more powerful than visual fixation features. It is also recommended that future ML studies combine features from both visual fixation patterns and oculomotor performance to examine whether the classification accuracy could be further improved.</p> <p>The classification accuracy of our study was comparable to that of the ADOS-2 and slightly higher than that of the ADI-R, as reported in a recent meta-analysis (Gabrielsen et al., [<reflink idref="bib12" id="ref105">12</reflink>]). This may partly be due to the relatively small sample size as well as lower heterogeneity and complexity of participants in the present study. However, as a comparison, eye-tracking implementation is more objective and less time- and labor- demanding than these two gold-standard diagnostic instruments. Given that existing ML studies using eye-tracking data have reported high accuracy of classifying children with ASD and TD (Eraslan et al., [<reflink idref="bib9" id="ref106">9</reflink>]; Wan et al., [<reflink idref="bib42" id="ref107">42</reflink>]; Yaneva et al., [<reflink idref="bib45" id="ref108">45</reflink>]; Zhao et al., 2021a), it is proposed that eye-tracking technology might be incorporated in clinical practice to complement existing diagnostic procedure. The inclusion of eye-tracking tests may improve the objectivity and efficiency of the current diagnostic methods. Specifically, a variety of eye-tracking tasks might be completed to see whether the individual's gaze behavior matches the characteristics of ASD. However, a major issue needs to be noted is that the current eye-tracking based ML studies focused on classifying individuals with ASD from those with TD only (Eraslan et al., [<reflink idref="bib9" id="ref109">9</reflink>]; Wan et al., [<reflink idref="bib42" id="ref110">42</reflink>]; Yaneva et al., [<reflink idref="bib45" id="ref111">45</reflink>]; Zhao et al., 2021a). It remains unexplored whether ML is still an efficient tool in differentiating people with ASD from those with other clinical phenotypes (e.g., developmental delay and ADHD). Future studies are highly encouraged to address this question.</p> <p>The present study computed the entropy and range of gaze allocation as features capturing the oculomotor performance, and our ML results showed that entropy plays a prominent role in classification tasks. It needs to be noted that higher level of entropy in the ASD group might be largely attributed to their attentional deficit of selectively attending to social information out of irrelevant information (Chita-Tegmark, [<reflink idref="bib5" id="ref112">5</reflink>]; Frazier et al., [<reflink idref="bib11" id="ref113">11</reflink>]). Specifically, children with TD fixated more on important social information (i.e., face, and body), and less on background in face-to-face conversations, whereas visual fixations in children with ASD were more equally assigned between the social and irrelevant information (Zhao et al., [<reflink idref="bib47" id="ref114">47</reflink>]). The specific viewing pattern in the ASD group would lead to more sparsely allocated gaze points, and thus to a higher entropy value. Therefore, the atypical oculomotor performance in the children with ASD was inseparable from visual attention in our study. Future investigations are encouraged to separate the inherent oculomotor function from visual attention deficits in unconstrained viewing settings (e.g., face-to-face interactions).</p> <hd id="AN0162233062-13">Limitations and Future Directions</hd> <p>To ensure that participants would be able to converse with the interviewer, this study only included children aged between 6 and 13 years with at least average intellectual ability. Children with severe symptoms were not included. In addition, our participant sample was mainly composed of boys, and previous studies showed significant gender difference in individuals with ASD in terms of behavioral presentation and cognitive domains (Bölte et al., [<reflink idref="bib4" id="ref115">4</reflink>]; Sasson et al., [<reflink idref="bib34" id="ref116">34</reflink>]). Given the significant heterogeneity in ASD (Jacob et al., [<reflink idref="bib16" id="ref117">16</reflink>]), the present study should only be considered as proof-of-concept research which showed the feasibility of using oculomotor behavior to identify ASD. Indeed, further investigation is required to explore whether oculomotor behavior could be utilized to accurately classify children with ASD with various presentations (e.g., different severity and balanced sex ratio).</p> <p>Another limitation was pertaining to the age of the participants. In our study, participants aged between 6 and 13 years old. A sizable body of literature has demonstrated the benefits of early intervention in ASD prognosis (Meghan Miller et al., [<reflink idref="bib26" id="ref118">26</reflink>]; Ozonoff et al., [<reflink idref="bib29" id="ref119">29</reflink>]). Therefore, the classification tasks performed on younger children is of greater significance for the sake of ASD screening. A recent meta-analysis examining children aged no greater than 10 years old recruited from thirty-five countries reported that the mean age at the diagnosis of ASD was 43.18 months (Hof et al., [<reflink idref="bib15" id="ref120">15</reflink>]). Future studies are encouraged to examine whether oculomotor behavior could discriminate individuals with ASD from other populations at an earlier age. Since atypical oculomotor performance has been reported in infants later diagnosed as ASD (Elison et al., [<reflink idref="bib8" id="ref121">8</reflink>]), it is assumed that oculomotor behavior could help early ASD screening.</p> <hd id="AN0162233062-14">Conclusion</hd> <p>By implementing a ML approach, the present study showed that oculomotor behavior in children with ASD could be well discriminated from that in children with TD with an accuracy above the chance level. In addition, oculomotor behavior could help classify children with ASD and TD. Different from previous ML studies which extracted features directly from the whole time series, our study provides a new perspective by showing that combining the behavior segmentation technique and the threshold classifier could well classify children with ASD and TD.</p> <hd id="AN0162233062-15">Author Contributions</hd> <p>ZZ, XZ, XQ, and JL designed the experiment and recruited participants. ZZ, JW, JX, and XH performed data analysis. ZZ, JW, XZ, and XQ drafted and revised the manuscript.</p> <hd id="AN0162233062-16">Funding</hd> <p>The study was financially supported by the National Natural Science Foundation of China (Grant No. 82171539), the SZU funding project (Grant No. 860-000002110259), the Science and Technology Innovation Committee of Shenzhen (Grant No. JCYJ20190808115205498), the Key medical discipline of GuangMing Shenzhen (Grant No. 12 Epidemiology), Sanming project of medicine in Shenzhen (Grant No. SZSM201612079), Key Realm R&D Program of Guangdong Province (Grant No. 2019B030335001), Shenzhen Key Medical Discipline Construction Fund (Grant No. SZXK042), and Shenzhen Double Chain Grant [2018]256.</p> <hd id="AN0162233062-17">Declarations</hd> <p></p> <hd id="AN0162233062-18">Conflict of interest</hd> <p>None of the authors declare a financial interest in any of the products or devices mentioned in the manuscript.</p> <hd id="AN0162233062-19">Appendix</hd> <p>See Table 5.</p> <p>Table 5 Details of the structured conversation: generic question</p> <p> <ephtml> <table frame="hsides" rules="groups"><tbody><tr><td align="left"><p>1. What is your name?</p></td></tr><tr><td align="left"><p>2. How is your name written?</p></td></tr><tr><td align="left"><p>3. What is the name of your school and what grade are you in?</p></td></tr><tr><td align="left"><p>4. Who is your best friend? What is your favorite thing to do together?</p></td></tr><tr><td align="left"><p>5. Could you please share with me the most interesting thing happened last week? Let me know the time, place, people and the whole process of the event</p></td></tr><tr><td align="left"><p>6. What is the plan for your summer vacation?</p></td></tr></tbody></table> </ephtml> </p> <hd id="AN0162233062-20">Abbreviations</hd> <p></p> <p>• ADHD</p> <p></p> <ulist> <item> Attention-deficit hyperactivity disorder</item> <p></p> </ulist> <p>• ADI-R</p> <p></p> <ulist> <item> Autism diagnostic interview-revised</item> <p></p> </ulist> <p>• ADOS</p> <p></p> <ulist> <item> Autism diagnostic observation schedule</item> <p></p> </ulist> <p>• ASD</p> <p></p> <ulist> <item> Autism spectrum disorder</item> <p></p> </ulist> <p>• DSM-V</p> <p></p> <ulist> <item> The diagnostic and statistical manual of mental disorders—5th Edition</item> <p></p> </ulist> <p>• EEG</p> <p></p> <ulist> <item> Electroencephalography</item> <p></p> </ulist> <p>• ML</p> <p></p> <ulist> <item> Machine learning</item> <p></p> </ulist> <p>• TD</p> <p></p> <ulist> <item> Typical development</item> </ulist> <hd id="AN0162233062-21">Publisher's Note</hd> <p>Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.</p> <ref id="AN0162233062-22"> <title> References </title> <blist> <bibl id="bib1" idref="ref1" type="bt">1</bibl> <bibtext> American Psychiatric Association. Diagnostic and statistical manual of mental disorders. 20135; American Psychiatric Publishing. 10.1176/appi.books.9780890425596</bibtext> </blist> <blist> <bibl id="bib2" idref="ref45" type="bt">2</bibl> <bibtext> Avni I, Meiri G, Michaelovski A, Menashe I, Shmuelof L, Dinstein I. Basic oculomotor function is similar in young children with ASD and typically developing controls. Autism Research. 2021; 14; 12: 2580-2591. 10.1002/aur.2592. 34405961</bibtext> </blist> <blist> <bibl id="bib3" idref="ref34" type="bt">3</bibl> <bibtext> Baranek GT. Autism during infancy: A retrospective video analysis of sensory-motor and social behaviors at 9–12 months of age. Journal of Autism and Developmental Disorders. 1999; 29; 3: 213-224. 10.1023/A:1023080005650. 10425584</bibtext> </blist> <blist> <bibl id="bib4" idref="ref115" type="bt">4</bibl> <bibtext> Bölte S, Duketis E, Poustka F, Holtmann M. Sex differences in cognitive domains and their clinical correlates in higher-functioning autism spectrum disorders. Autism. 2011; 15; 4: 497-511. 10.1177/1362361310391116. 21454389</bibtext> </blist> <blist> <bibl id="bib5" idref="ref112" type="bt">5</bibl> <bibtext> Chita-Tegmark M. Social attention in ASD: A review and meta-analysis of eye-tracking studies. Research in Developmental Disabilities. 2016; 48: 79-93. 10.1016/j.ridd.2015.10.011. 26547134</bibtext> </blist> <blist> <bibl id="bib6" idref="ref5" type="bt">6</bibl> <bibtext> Crippa A, Salvatore C, Perego P, Forti S, Nobile M, Molteni M, Castiglioni I. Use of machine learning to identify children with autism and their motor abnormalities. Journal of Autism and Developmental Disorders. 2015; 45; 7: 2146-2156. 10.1007/s10803-015-2379-8. 25652603</bibtext> </blist> <blist> <bibl id="bib7" idref="ref6" type="bt">7</bibl> <bibtext> Duda M, Kosmicki JA, Wall DP. Testing the accuracy of an observation-based classifier for rapid detection of autism risk. Translational Psychiatry. 2014; 4: e424. 10.1038/tp.2014.65. 25116834. 4150240</bibtext> </blist> <blist> <bibl id="bib8" idref="ref121" type="bt">8</bibl> <bibtext> Elison, J. T, Paterson, S. J, Wolff, J. J, Reznick, J. S, Sasson, N. J, Gu, H, ... Piven, J. (2013). White matter microstructure and atypical visual orienting in 7-month-olds at risk for autism. American Journal of Psychiatry, 170(8), 899–908. https://doi.org/10.1176/appi.ajp.2012.12091150</bibtext> </blist> <blist> <bibl id="bib9" idref="ref25" type="bt">9</bibl> <bibtext> Eraslan, S, Yesilada, Y, Yaneva, V, & Harper, S. (2020). Autism detection based on eye movement sequences on the web: A scanpath trend analysis approach. Paper presented at the Proceedings of the 17th International Web for All Conference, Taipei, Taiwan. https://doi.org/10.1145/3371300.3383340</bibtext> </blist> <blist> <bibtext> Fournier KA, Amano S, Radonovich KJ, Bleser TM, Hass CJ. Decreased dynamical complexity during quiet stance in children with autism spectrum disorders. Gait & Posture. 2014; 39; 1: 420-423. 10.1016/j.gaitpost.2013.08.016</bibtext> </blist> <blist> <bibtext> Frazier TW, Strauss M, Klingemier EW, Zetzer EE, Hardan AY, Eng C, Youngstrom EA. A meta-analysis of gaze differences to social and nonsocial information between individuals with and without autism. Journal of the American Academy of Child & Adolescent Psychiatry. 2017; 56; 7: 546-555. 10.1016/j.jaac.2017.05.005</bibtext> </blist> <blist> <bibtext> Gabrielsen TP, Farley M, Speer L, Villalobos M, Baker CN, Miller J. Identifying autism in a brief observation. Pediatrics. 2015; 135; 2: e330-e338. 10.1542/peds.2014-1428. 25583913</bibtext> </blist> <blist> <bibtext> Grossi E, Valbusa G, Buscema M. Detection of an autism EEG signature from only two EEG channels through features extraction and advanced machine learning analysis. Clinical EEG and Neuroscience. 2020. 10.1177/1550059420982424. 33349054</bibtext> </blist> <blist> <bibtext> Guyon I, Elisseeff A. An introduction to variable and feature selection. Journal of Machine Learning Research. 2003; 3: 1157-1182</bibtext> </blist> <blist> <bibtext> Hof M, Tisseur C, Berckelear-Onnes IV, Nieuwenhuyzen AV, Ester WA. Age at autism spectrum disorder diagnosis: A systematic review and meta-analysis from 2012 to 2019. Autism. 2020; 25; 4: 862-873</bibtext> </blist> <blist> <bibtext> Jacob S, Wolff JJ, Steinbach MS, Doyle CB, Kumar V, Elison JT. Neurodevelopmental heterogeneity and computational approaches for understanding autism. Translational Psychiatry. 2019; 9; 1: 63. 10.1038/s41398-019-0390-0. 30718453. 6362076</bibtext> </blist> <blist> <bibtext> Johnson BP, Lum JA, Rinehart NJ, Fielding J. Ocular motor disturbances in autism spectrum disorders: Systematic review and comprehensive meta-analysis. Neuroscience and Biobehavioral Reviews. 2016; 69: 260-279. 10.1016/j.neubiorev.2016.08.007. 27527824</bibtext> </blist> <blist> <bibtext> Jones, R. M, Southerland, A, Hamo, A, Carberry, C, Bridges, C, Nay, S, ... Rozga, A. (2017). Increased eye contact during conversation compared to play in children with autism. Journal of Autism and Developmental Disorders, 47(3), 607–614. https://doi.org/10.1007/s10803-016-2981-4</bibtext> </blist> <blist> <bibtext> Jones W, Klin A. Attention to eyes is present but in decline in 2-6-month-old infants later diagnosed with autism. Nature. 2013; 504; 7480: 427-431. 10.1038/nature12715. 24196715. 4035120</bibtext> </blist> <blist> <bibtext> Kolmogorov AN. Entropy per unit time as a metric invariant of automorphism. Doklady Akademii Nauk Sssr. 1959; 124; 4: 754-755</bibtext> </blist> <blist> <bibtext> Li B, Sharma A, Meng J, Purushwalkam S, Gowen E. Applying machine learning to identify autistic adults using imitation: An exploratory study. PLoS ONE. 2017; 12; 8: e0182652. 10.1371/journal.pone.0182652. 28813454. 5558936</bibtext> </blist> <blist> <bibtext> Li Y, Mache MA, Todd TA. Complexity of center of pressure in postural control for children with autism spectrum disorders was partially compromised. Journal of Applied Biomechanics. 2019; 35; 3: 190-195. 10.1123/jab.2018-0042. 30676165</bibtext> </blist> <blist> <bibtext> Liu W, Li M, Yi L. Identifying children with autism spectrum disorder based on their face processing abnormality: A machine learning framework. Autism Research. 2016; 9; 8: 888-898. 10.1002/aur.1615. 27037971</bibtext> </blist> <blist> <bibtext> Maenner, M. J, Shaw, K. A, Bakian, A. V, Bilder, D. A, Durkin, M. S, Esler, A, ... Cogswell, M. E. (2021). Prevalence and characteristics of autism spectrum disorder among children aged 8 years - autism and developmental disabilities monitoring network, 11 sites, United States, 2018. MMWR Surveillance Summaries, 70(11), 1–16. https://doi.org/10.15585/mmwr.ss7011a1</bibtext> </blist> <blist> <bibtext> Miller M, Chukoskie L, Zinni M, Townsend J, Trauner D. Dyspraxia, motor function and visual-motor integration in autism. Behavioural Brain Research. 2014; 269: 95-102. 10.1016/j.bbr.2014.04.011. 24742861. 4072207</bibtext> </blist> <blist> <bibtext> Miller M, Iosif A-M, Hill M, Young GS, Schwichtenberg AJ, Ozonoff S. Response to name in infants developing autism spectrum disorder: A prospective study. The Journal of Pediatrics. 2017; 183: 141-146.e141. 10.1016/j.jpeds.2016.12.071. 28162768. 5370582</bibtext> </blist> <blist> <bibtext> Möricke E, Buitelaar JK, Rommelse NNJ. Do we need multiple informants when assessing autistic traits? The degree of report bias on offspring, self, and spouse ratings. Journal of Autism and Developmental Disorders. 2016; 46: 164-175. 10.1007/s10803-015-2562-y. 26334871</bibtext> </blist> <blist> <bibtext> Osterling J, Dawson G. Early recognition of children with autism: A study of first birthday home videotapes. Journal of Autism and Developmental Disorders. 1994; 24; 3: 247-257. 10.1007/BF02172225. 8050980</bibtext> </blist> <blist> <bibtext> Ozonoff, S, Iosif, A.-M, Baguio, F, Cook, I. C, Hill, M. M, Hutman, T, ... Young, G. S. (2010). A prospective study of the emergence of early behavioral signs of autism. Journal of the American Academy of Child & Adolescent Psychiatry, 49(3), 256–266.e252. https://doi.org/10.1016/j.jaac.2009.11.009</bibtext> </blist> <blist> <bibtext> Plitt M, Barnes KA, Martin A. Functional connectivity classification of autism identifies highly predictive brain features but falls short of biomarker standards. NeuroImage: Clinical. 2015; 7: 359-366. 10.1016/j.nicl.2014.12.013. 25685703</bibtext> </blist> <blist> <bibtext> Pritschet L, Powell D, Horne Z. Marginally significant effects as evidence for hypotheses: Changing attitudes over four decades. Psychological Science. 2016; 27; 7: 1036-1042. 10.1177/0956797616645672. 27207874</bibtext> </blist> <blist> <bibtext> Raven J. The Raven's progressive matrices: Change and stability over culture and time. Cognitive Psychology. 2000; 41; 1: 1-48. 10.1006/cogp.1999.0735. 10945921</bibtext> </blist> <blist> <bibtext> Santos A, Chaminade T, Da Fonseca D, Silva C, Rosset D, Deruelle C. Just another social scene: Evidence for decreased attention to negative social scenes in high-functioning autism. Journal of Autism and Developmental Disorders. 2012; 42; 9: 1790-1798. 10.1007/s10803-011-1415-6. 22160371</bibtext> </blist> <blist> <bibtext> Sasson NJ, Pinkham AE, Weittenhiller LP, Faso DJ, Simpson C. Context effects on facial affect recognition in schizophrenia and autism: Behavioral and eye-tracking evidence. Schizophrenia Bulletin. 2016; 42; 3: 675-683. 10.1093/schbul/sbv176. 26645375</bibtext> </blist> <blist> <bibtext> Schmitt LM, Cook EH, Sweeney JA, Mosconi MW. Saccadic eye movement abnormalities in autism spectrum disorder indicate dysfunctions in cerebellum and brainstem. Molecular Autism. 2014; 5; 1: 47. 10.1186/2040-2392-5-47. 25400899. 4233053</bibtext> </blist> <blist> <bibtext> Shannon CE. A mathematical theory of communication. Bell System Technical Journal. 1948; 5; 3: 3-55</bibtext> </blist> <blist> <bibtext> Sumner E, Hutton SB, Hill EL. Subtle oculomotor difficulties and their relation to motor skill in children with autism spectrum disorder. Advances in Neurodevelopmental Disorders. 2020; 5; 2: 144-155. 10.1007/s41252-020-00188-1</bibtext> </blist> <blist> <bibtext> Tebartz van Elst L, Pick M, Biscaldi M, Fangmeier T, Riedel A. High-functioning autism spectrum disorder as a basic disorder in adult psychiatry and psychotherapy: Psychopathological presentation, clinical relevance and therapeutic concepts. European Archives of Psychiatry and Clinical Neuroscience. 2013; 263; 2: 189-196. 10.1007/s00406-013-0459-3</bibtext> </blist> <blist> <bibtext> van der Geest JN, Kemner C, Camfferman G, Verbaten MN, van Engeland H. Eye movements, visual attention, and autism: A saccadic reaction time study using the gap and overlap paradigm. Biological Psychiatry. 2001; 50; 8: 614-619. 10.1016/s0006-3223(01)01070-8. 11690597</bibtext> </blist> <blist> <bibtext> Wall DP, Dally R, Luyster R, Jung JY, Deluca TF. Use of artificial intelligence to shorten the behavioral diagnosis of autism. PLoS ONE. 2012; 7; 8: e43855. 10.1371/journal.pone.0043855. 22952789. 3428277</bibtext> </blist> <blist> <bibtext> Wall DP, Kosmicki J, DeLuca TF, Harstad E, Fusaro VA. Use of machine learning to shorten observation-based screening and diagnosis of autism. Translational Psychiatry. 2012; 2: e100. 10.1038/tp.2012.10. 22832900. 3337074</bibtext> </blist> <blist> <bibtext> Wan, G, Kong, X, Sun, B, Yu, S, Tu, Y, Park, J, ... Kong, J. (2019). Applying eye tracking to identify autism spectrum disorder in children. Journal of Autism and Developmental Disorders, 49(1), 209-215. https://doi.org/10.1007/s10803-018-3690-y</bibtext> </blist> <blist> <bibtext> Wang Q, Hoi SP, Wang Y, Song C, Yi L. Out of mind, out of sight? Investigating abnormal face scanning in autism spectrum disorder using gaze-contingent paradigm. Developmental Science. 2020; 23; 1: e12856. 10.1111/desc.12856. 31081980</bibtext> </blist> <blist> <bibtext> Yamagata, B, Itahashi, T, Fujino, J, Ohta, H, Nakamura, M, Kato, N, ... Aoki, Y. (2019). Machine learning approach to identify a resting-state functional connectivity pattern serving as an endophenotype of autism spectrum disorder. Brain Imaging and Behavior, 13(6), 1689–1698. https://doi.org/10.1007/s11682-018-9973-2</bibtext> </blist> <blist> <bibtext> Yaneva, V, Ha, L. A, Eraslan, S, Yesilada, Y, & Mitkov, R. (2018). Detecting autism based on eye-tracking data from web searching tasks. In: Paper presented at the Proceedings of the Internet of Accessible Things. https://doi.org/10.1145/3192714.3192819</bibtext> </blist> <blist> <bibtext> Zhao Z, Tang H, Zhang X, Qu X, Hu X, Lu J. Classification of children with autism and typical development using eye-tracking data from face-to-face conversations: machine learning model development and performance evaluation. Journal of Medical Internet Research. 2021; 23; 8: e29328. 10.2196/29328. 34435957. 8440949</bibtext> </blist> <blist> <bibtext> Zhao, Z, Tang, H, Zhang, X, Zhu, Z, Xing, J, Li, W, ... Lu, J. (2021b). Characteristics of visual fixation in chinese children with autism during face-to-face conversations. Journal of Autism and Developmental Disordershttps://doi.org/10.1007/s10803-021-04985-y</bibtext> </blist> <blist> <bibtext> Zhao Z, Xing J, Zhang X, Qu X, Lu J. Random and short-term excessive eye movement in children with autism during face-to-face conversation. Journal of Autism and Developmental Disorders. 2021. 10.1007/s10803-021-05255-7. 34455513. 8854168</bibtext> </blist> <blist> <bibtext> Zhao, Z, Zhang, X, Li, W, Hu, X, Qu, X, Cao, X, ... Lu, J. (2019). Applying machine learning to identify autism with restricted kinematic features. IEEE Access, 7, 157614–157622. https://doi.org/10.1109/ACCESS.2019.2950030</bibtext> </blist> <blist> <bibtext> Zhao, Z, Zhu, Z, Zhang, X, Tang, H, Xing, J, Hu, X, ... Qu, X. (2021d). Atypical head movement during face-to-face interaction in children with autism spectrum disorder. Autism Research, 14(6), 1197–1208 https://doi.org/10.1002/aur.2478</bibtext> </blist> </ref> <aug> <p>By Zhong Zhao; Jiwei Wei; Jiayi Xing; Xiaobin Zhang; Xingda Qu; Xinyao Hu and Jianping Lu</p> <p>Reported by Author; Author; Author; Author; Author; Author; Author</p> </aug> <nolink nlid="nl1" bibid="bib24" firstref="ref2"></nolink> <nolink nlid="nl2" bibid="bib27" firstref="ref3"></nolink> <nolink nlid="nl3" bibid="bib38" firstref="ref4"></nolink> <nolink nlid="nl4" bibid="bib21" firstref="ref7"></nolink> <nolink nlid="nl5" bibid="bib23" firstref="ref8"></nolink> <nolink nlid="nl6" bibid="bib42" firstref="ref9"></nolink> <nolink nlid="nl7" bibid="bib40" firstref="ref11"></nolink> <nolink nlid="nl8" bibid="bib41" firstref="ref12"></nolink> <nolink nlid="nl9" bibid="bib13" firstref="ref14"></nolink> <nolink nlid="nl10" bibid="bib28" firstref="ref17"></nolink> <nolink nlid="nl11" bibid="bib30" firstref="ref18"></nolink> <nolink nlid="nl12" bibid="bib44" firstref="ref20"></nolink> <nolink nlid="nl13" bibid="bib49" firstref="ref21"></nolink> <nolink nlid="nl14" bibid="bib25" firstref="ref22"></nolink> <nolink nlid="nl15" bibid="bib37" firstref="ref23"></nolink> <nolink nlid="nl16" bibid="bib45" firstref="ref28"></nolink> <nolink nlid="nl17" bibid="bib46" firstref="ref33"></nolink> <nolink nlid="nl18" bibid="bib17" firstref="ref44"></nolink> <nolink nlid="nl19" bibid="bib39" firstref="ref46"></nolink> <nolink nlid="nl20" bibid="bib35" firstref="ref48"></nolink> <nolink nlid="nl21" bibid="bib48" firstref="ref55"></nolink> <nolink nlid="nl22" bibid="bib19" firstref="ref57"></nolink> <nolink nlid="nl23" bibid="bib33" firstref="ref58"></nolink> <nolink nlid="nl24" bibid="bib18" firstref="ref59"></nolink> <nolink nlid="nl25" bibid="bib47" firstref="ref60"></nolink> <nolink nlid="nl26" bibid="bib43" firstref="ref61"></nolink> <nolink nlid="nl27" bibid="bib10" firstref="ref63"></nolink> <nolink nlid="nl28" bibid="bib20" firstref="ref64"></nolink> <nolink nlid="nl29" bibid="bib22" firstref="ref65"></nolink> <nolink nlid="nl30" bibid="bib36" firstref="ref66"></nolink> <nolink nlid="nl31" bibid="bib50" firstref="ref67"></nolink> <nolink nlid="nl32" bibid="bib32" firstref="ref78"></nolink> <nolink nlid="nl33" bibid="bib14" firstref="ref89"></nolink> <nolink nlid="nl34" bibid="bib31" firstref="ref90"></nolink> <nolink nlid="nl35" bibid="bib12" firstref="ref105"></nolink> <nolink nlid="nl36" bibid="bib11" firstref="ref113"></nolink> <nolink nlid="nl37" bibid="bib34" firstref="ref116"></nolink> <nolink nlid="nl38" bibid="bib16" firstref="ref117"></nolink> <nolink nlid="nl39" bibid="bib26" firstref="ref118"></nolink> <nolink nlid="nl40" bibid="bib29" firstref="ref119"></nolink> <nolink nlid="nl41" bibid="bib15" firstref="ref120"></nolink>
Header DbId: eric
DbLabel: ERIC
An: EJ1368763
AccessLevel: 3
PubType: Academic Journal
PubTypeId: academicJournal
PreciseRelevancyScore: 0
IllustrationInfo
Items – Name: Title
  Label: Title
  Group: Ti
  Data: Use of Oculomotor Behavior to Classify Children with Autism and Typical Development: A Novel Implementation of the Machine Learning Approach
– Name: Language
  Label: Language
  Group: Lang
  Data: English
– Name: Author
  Label: Authors
  Group: Au
  Data: <searchLink fieldCode="AR" term="%22Zhao%2C+Zhong%22">Zhao, Zhong</searchLink><br /><searchLink fieldCode="AR" term="%22Wei%2C+Jiwei%22">Wei, Jiwei</searchLink><br /><searchLink fieldCode="AR" term="%22Xing%2C+Jiayi%22">Xing, Jiayi</searchLink><br /><searchLink fieldCode="AR" term="%22Zhang%2C+Xiaobin%22">Zhang, Xiaobin</searchLink><br /><searchLink fieldCode="AR" term="%22Qu%2C+Xingda%22">Qu, Xingda</searchLink><br /><searchLink fieldCode="AR" term="%22Hu%2C+Xinyao%22">Hu, Xinyao</searchLink><br /><searchLink fieldCode="AR" term="%22Lu%2C+Jianping%22">Lu, Jianping</searchLink>
– Name: TitleSource
  Label: Source
  Group: Src
  Data: <searchLink fieldCode="SO" term="%22Journal+of+Autism+and+Developmental+Disorders%22"><i>Journal of Autism and Developmental Disorders</i></searchLink>. Mar 2023 53(3):934-946.
– Name: Avail
  Label: Availability
  Group: Avail
  Data: Springer. Available from: Springer Nature. One New York Plaza, Suite 4600, New York, NY 10004. Tel: 800-777-4643; Tel: 212-460-1500; Fax: 212-460-1700; e-mail: customerservice@springernature.com; Web site: https://link.springer.com/
– Name: PeerReviewed
  Label: Peer Reviewed
  Group: SrcInfo
  Data: Y
– Name: Pages
  Label: Page Count
  Group: Src
  Data: 13
– Name: DatePubCY
  Label: Publication Date
  Group: Date
  Data: 2023
– Name: TypeDocument
  Label: Document Type
  Group: TypDoc
  Data: Journal Articles<br />Reports - Research
– Name: Subject
  Label: Descriptors
  Group: Su
  Data: <searchLink fieldCode="DE" term="%22Children%22">Children</searchLink><br /><searchLink fieldCode="DE" term="%22Autism+Spectrum+Disorders%22">Autism Spectrum Disorders</searchLink><br /><searchLink fieldCode="DE" term="%22Symptoms+%28Individual+Disorders%29%22">Symptoms (Individual Disorders)</searchLink><br /><searchLink fieldCode="DE" term="%22Eye+Movements%22">Eye Movements</searchLink><br /><searchLink fieldCode="DE" term="%22Interpersonal+Communication%22">Interpersonal Communication</searchLink><br /><searchLink fieldCode="DE" term="%22Classification%22">Classification</searchLink><br /><searchLink fieldCode="DE" term="%22Accuracy%22">Accuracy</searchLink><br /><searchLink fieldCode="DE" term="%22Disability+Identification%22">Disability Identification</searchLink>
– Name: DOI
  Label: DOI
  Group: ID
  Data: 10.1007/s10803-022-05685-x
– Name: ISSN
  Label: ISSN
  Group: ISSN
  Data: 0162-3257<br />1573-3432
– Name: Abstract
  Label: Abstract
  Group: Ab
  Data: This study segmented the time series of gaze behavior from nineteen children with autism spectrum disorder (ASD) and 20 children with typical development in a face-to-face conversation. A machine learning approach showed that behavior segments produced by these two groups of participants could be classified with the highest accuracy of 74.15%. These results were further used to classify children using a threshold classifier. A maximum classification accuracy of 87.18% was achieved, under the condition that a participant was considered as 'ASD' if over 46% of the child's 7-s behavior segments were classified as ASD-like behaviors. The idea of combining the behavior segmentation technique and the threshold classifier could maximally preserve participants' data, and promote the automatic screening of ASD.
– Name: AbstractInfo
  Label: Abstractor
  Group: Ab
  Data: As Provided
– Name: DateEntry
  Label: Entry Date
  Group: Date
  Data: 2023
– Name: AN
  Label: Accession Number
  Group: ID
  Data: EJ1368763
PLink https://search.ebscohost.com/login.aspx?direct=true&site=eds-live&db=eric&AN=EJ1368763
RecordInfo BibRecord:
  BibEntity:
    Identifiers:
      – Type: doi
        Value: 10.1007/s10803-022-05685-x
    Languages:
      – Text: English
    PhysicalDescription:
      Pagination:
        PageCount: 13
        StartPage: 934
    Subjects:
      – SubjectFull: Children
        Type: general
      – SubjectFull: Autism Spectrum Disorders
        Type: general
      – SubjectFull: Symptoms (Individual Disorders)
        Type: general
      – SubjectFull: Eye Movements
        Type: general
      – SubjectFull: Interpersonal Communication
        Type: general
      – SubjectFull: Classification
        Type: general
      – SubjectFull: Accuracy
        Type: general
      – SubjectFull: Disability Identification
        Type: general
    Titles:
      – TitleFull: Use of Oculomotor Behavior to Classify Children with Autism and Typical Development: A Novel Implementation of the Machine Learning Approach
        Type: main
  BibRelationships:
    HasContributorRelationships:
      – PersonEntity:
          Name:
            NameFull: Zhao, Zhong
      – PersonEntity:
          Name:
            NameFull: Wei, Jiwei
      – PersonEntity:
          Name:
            NameFull: Xing, Jiayi
      – PersonEntity:
          Name:
            NameFull: Zhang, Xiaobin
      – PersonEntity:
          Name:
            NameFull: Qu, Xingda
      – PersonEntity:
          Name:
            NameFull: Hu, Xinyao
      – PersonEntity:
          Name:
            NameFull: Lu, Jianping
    IsPartOfRelationships:
      – BibEntity:
          Dates:
            – D: 01
              M: 03
              Type: published
              Y: 2023
          Identifiers:
            – Type: issn-print
              Value: 0162-3257
            – Type: issn-electronic
              Value: 1573-3432
          Numbering:
            – Type: volume
              Value: 53
            – Type: issue
              Value: 3
          Titles:
            – TitleFull: Journal of Autism and Developmental Disorders
              Type: main
ResultId 1