View in EDS HTML Full Text PDF Full Text

Identifying Autism with Head Movement Features by Implementing Machine Learning Algorithms

Saved in:
Bibliographic Details
Title: Identifying Autism with Head Movement Features by Implementing Machine Learning Algorithms
Language: English
Authors: Zhao, Zhong, Zhu, Zhipeng, Zhang, Xiaobin, Tang, Haiming, Xing, Jiayi, Hu, Xinyao, Lu, Jianping, Qu, Xingda (ORCID 0000-0003-1764-0357)
Source: Journal of Autism and Developmental Disorders. Jul 2022 52(7):3038-3049.
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: 12
Publication Date: 2022
Document Type: Journal Articles
Reports - Research
Descriptors: Autism, Pervasive Developmental Disorders, Motion, Human Body, Children, Identification, Nonverbal Communication, Artificial Intelligence
DOI: 10.1007/s10803-021-05179-2
ISSN: 0162-3257
Abstract: Our study investigated the feasibility of using head movement features to identify individuals with autism spectrum disorder (ASD). Children with ASD and typical development (TD) were required to answer ten yes--no questions, and they were encouraged to nod/shake head while doing so. The head rotation range (RR) and the amount of rotation per minute (ARPM) in the pitch (head nodding direction), yaw (head shaking direction) and roll (lateral head inclination) directions were computed, and further fed into machine learning classifiers as the input features. The maximum classification accuracy of 92.11% was achieved with the decision tree classifier with two features (i.e., RR_Pitch and ARPM_Yaw). Our study suggests that head movement dynamics contain objective biomarkers that could identify ASD.
Abstractor: As Provided
Entry Date: 2022
Accession Number: EJ1339362
Database: ERIC
Full text is not displayed to guests.
FullText Links:
  – Type: pdflink
    Url: https://content.ebscohost.com/cds/retrieve?content=AQICAHj0k_4E0hTGH8RJwT4gCJyBsGNe_WN95AvKlDbXJGqwxwH5mH4qfDJTI5062ghiHwecAAAA4TCB3gYJKoZIhvcNAQcGoIHQMIHNAgEAMIHHBgkqhkiG9w0BBwEwHgYJYIZIAWUDBAEuMBEEDJenNFGYIzUE0HKlNAIBEICBmU9faW61pFwMdjKVBruEriTvqTLPsvy2eZ5LnGXyMbzA9uHxY8zxhXPO2-97fMp10Y9bsXYGRT9c0_T2Eq8GywVePYvMoMo_FcwYK-MeUpciP294exfua2zjf-mk8myzOb0rLy3P8yjKUDWAJZbj3Hfs8RxEb9fm-bI6RXl7BINIhEHj8H6pIkY1XZFKh0PqTAM75-2TeJf5KQ==
Text:
  Availability: 1
  Value: <anid>AN0157570732;aut01jul.22;2022Jun28.04:39;v2.2.500</anid> <title id="AN0157570732-1">Identifying Autism with Head Movement Features by Implementing Machine Learning Algorithms </title> <p>Our study investigated the feasibility of using head movement features to identify individuals with autism spectrum disorder (ASD). Children with ASD and typical development (TD) were required to answer ten yes–no questions, and they were encouraged to nod/shake head while doing so. The head rotation range (RR) and the amount of rotation per minute (ARPM) in the pitch (head nodding direction), yaw (head shaking direction) and roll (lateral head inclination) directions were computed, and further fed into machine learning classifiers as the input features. The maximum classification accuracy of 92.11% was achieved with the decision tree classifier with two features (i.e., RR_Pitch and ARPM_Yaw). Our study suggests that head movement dynamics contain objective biomarkers that could identify ASD.</p> <p>Keywords: Autism; Biomarkers; Diagnosis; Head movement; Machine learning</p> <hd id="AN0157570732-2">Introduction</hd> <p></p> <hd id="AN0157570732-3">Background and Significance</hd> <p>Autism spectrum disorder (ASD) is a complex neurodevelopmental condition that is characterized by compromised social communication (American Psychiatric Association, [<reflink idref="bib1" id="ref1">1</reflink>]). At present, the diagnosis of ASD heavily relies on qualitative behavioral evaluation. The diagnostic accuracy might be negatively affected by factors such as caregivers' report bias and clinicians' insufficient experiences of detecting ASD (Möricke et al., [<reflink idref="bib30" id="ref2">30</reflink>]; Tebartz van Elst et al., [<reflink idref="bib36" id="ref3">36</reflink>]). In addition, the shortage in specialists and the lengthy examinations make the diagnostic procedure a time-consuming process. Data from the United States revealed that more than a year on average is needed between the initial screening and the confirmed diagnosis (Wiggins et al., [<reflink idref="bib40" id="ref4">40</reflink>]). According to a recent meta-analysis examining data from 35 countries (Hof et al., [<reflink idref="bib20" id="ref5">20</reflink>]), the mean age at ASD diagnosis was 60.48 months, and it became 43.18 months when only taking into account children aged no greater than 10 years old. The delayed diagnosis directly translates to the postponed intervention, which subsequently affects the prognosis of the affected children (Dawson et al., [<reflink idref="bib14" id="ref6">14</reflink>]). Therefore, it becomes a critical issue to develop an efficient tool that could shorten the diagnostic procedure and to attain an objective diagnosis.</p> <p>Given that machine learning (ML) is particularly advantageous in reviewing large quantity of data and identifying patterns that could be barely observed with human eyes, recent years have witnessed an increasing interest in applying ML to either shorten standardized ASD diagnostic instruments or to detect ASD with objectively measured features. For example, a few ML studies reported that the diagnostic accuracy could be excellently preserved after removing redundant items from diagnostic instruments such as the Autism Diagnostic Observation Schedule (ADOS) and the Autism Diagnostic Interview-Revised (ADI-R) (Duda et al., [<reflink idref="bib15" id="ref7">15</reflink>]; Wall et al., [<reflink idref="bib37" id="ref8">37</reflink>], [<reflink idref="bib38" id="ref9">38</reflink>]). With respect to using objectively measured features to identify ASD with ML algorithms, previous studies found that features obtained from neuroimaging, eye tracking, EEG and kinematic data could well classify individuals with and without ASD (Crippa et al., [<reflink idref="bib12" id="ref10">12</reflink>]; Grossi et al. [<reflink idref="bib19" id="ref11">19</reflink>]; Kang et al., [<reflink idref="bib24" id="ref12">24</reflink>]; Li et al., [<reflink idref="bib26" id="ref13">26</reflink>]; Liu et al., [<reflink idref="bib28" id="ref14">28</reflink>]; Plitt et al., [<reflink idref="bib34" id="ref15">34</reflink>]; Wan et al., [<reflink idref="bib39" id="ref16">39</reflink>]; Yamagata et al., [<reflink idref="bib41" id="ref17">41</reflink>]; Zhao et al., [<reflink idref="bib42" id="ref18">42</reflink>]). For instance, Wan et al. ([<reflink idref="bib39" id="ref19">39</reflink>]) eye-tracked children with ASD and those with typical development (TD) in a video watching scenario. Their results showed that visual fixation measures could discriminate these two groups of participants with a classification accuracy of 85.1% (Wan et al., [<reflink idref="bib39" id="ref20">39</reflink>]). Presumably, expanding the list of objective biomarkers has the potential to promote an objective and automatic identification of ASD.</p> <p>Prior research has demonstrated a high prevalence of motor functioning abnormalities in individuals with ASD (Licari et al., [<reflink idref="bib27" id="ref21">27</reflink>]), who have been observed to exhibit an atypical pattern of motor behavior in postural control, gait stability, arm movements, gross and fine motor control (Bojanek et al., [<reflink idref="bib7" id="ref22">7</reflink>]; Fournier et al., [<reflink idref="bib16" id="ref23">16</reflink>]). Thus, it is feasible to objectively identify ASD by using kinematic markers. Further, individuals with ASD might have delayed motor development during their infancy (Landa & Garrett-Mayer, [<reflink idref="bib25" id="ref24">25</reflink>]; Ozonoff et al., [<reflink idref="bib32" id="ref25">32</reflink>]), which could subsequently contribute to the social impairments (Bhat et al., [<reflink idref="bib6" id="ref26">6</reflink>]; Perochon et al., [<reflink idref="bib33" id="ref27">33</reflink>]). All these evidences highlight that kinematic analysis would not only help discover atypical motor behavior in ASD individuals, but also help obtain novel behavioral markers of ASD.</p> <p>In terms of using kinematic features to identify ASD, existing studies implemented different motor tasks. For instance, Zhao et al. ([<reflink idref="bib42" id="ref28">42</reflink>]) extracted kinematic features from a motor task in which participants were required to perform the utmost complex oscillatory arm movements. Entropy analyses were conducted to obtain a variety of kinematic features that captured the level of restrictedness of movement. These features were fed into five ML classifiers, and a maximum classification accuracy of 88.37% was reached with the <emph>k</emph>-nearest neighbor (KNN, <emph>k</emph> = 1) classifier (Zhao et al., [<reflink idref="bib42" id="ref29">42</reflink>]). Interestingly, kinematic features extracted from other motor tasks, including touch-sensitive tablet interaction, reach-to-drop task, and imitation of hand movements (Anzulewicz et al., [<reflink idref="bib2" id="ref30">2</reflink>]; Crippa et al., [<reflink idref="bib12" id="ref31">12</reflink>]; Li et al., [<reflink idref="bib26" id="ref32">26</reflink>]), have also been used in ASD identification, and desirable identification accuracy has been reported in these studies.</p> <p>To date, no study has been found to investigate whether head movement dynamics could be utilized to identify ASD. In reality, there exists ample research reporting an atypical patterns of head movement in individuals with ASD. For instance, ASD individuals have been found to exhibit stereotyped head movement such as head rolling, shaking and banging (Goldman et al., [<reflink idref="bib18" id="ref33">18</reflink>]; Hutt & Hutt, [<reflink idref="bib21" id="ref34">21</reflink>]; Sorosky et al., [<reflink idref="bib35" id="ref35">35</reflink>]). Recently, Zhao et al. ([<reflink idref="bib43" id="ref36">43</reflink>]) utilized a computer vision algorithm—OpenFace 2.0—to extract the head movement time series when participants were engaged in a face-to-face conversation with an interviewer. This research demonstrated a significantly higher level of stereotypy and excessive head movement in children with ASD (Zhao et al., [<reflink idref="bib43" id="ref37">43</reflink>]). Similarly, Martin et al. ([<reflink idref="bib29" id="ref38">29</reflink>]) implemented Zface (<ulink href="http://zface.org/">http://zface.org/</ulink>) (Jeni et al., [<reflink idref="bib23" id="ref39">23</reflink>]) to track head movement in children with ASD while watching videos. It was reported that children with ASD exhibited greater head turning displacement and higher velocity of head turning and lateral inclinations when watching the social stimuli (Martin et al., [<reflink idref="bib29" id="ref40">29</reflink>]). Furthermore, two recent studies developed computer vision algorithms to detect head turning events and to calculate time latency in the 'response to name' experiment. Their results showed that toddlers with ASD less frequently oriented to name and it took them longer time to do so as compared to the TD group (Campbell et al., [<reflink idref="bib8" id="ref41">8</reflink>]; Perochon et al., [<reflink idref="bib33" id="ref42">33</reflink>]). In sum, all these findings provide concrete evidence for atypical head movement dynamics in ASD individuals, which strongly suggested that head movement might contain valuable information that could be leveraged to identify ASD.</p> <p>In the present study, ML was implemented to explore the feasibility of classifying children with ASD from those with TD using kinematic features extracted from head movement. Two groups of participants (ASD and TD) were engaged in a conversation with an interviewer. They were verbally required to use head nodding/shaking while answering ten yes–no questions. We aimed to determine: (<reflink idref="bib1" id="ref43">1</reflink>) whether children with ASD differed from those with TD with respect to the frequency of using head nodding/shaking while answering these questions; (<reflink idref="bib2" id="ref44">2</reflink>) whether and what specific kinematic feature (or feature combination) extracted from head movement could accurately classify children with ASD and TD by implementing ML algorithms.</p> <hd id="AN0157570732-4">Methods</hd> <p></p> <hd id="AN0157570732-5">Participants</hd> <p>Participants were the same as in our previous study (Zhao et al., [<reflink idref="bib43" id="ref45">43</reflink>]), but data of this study came from a different conversational session. Specifically, 20 children with ASD and twenty-three children with TD were recruited as participants. ASD children were enrolled from a local psychiatric hospital, based on the following criteria: (a) between 6 and 13 years old; (b) confirmed diagnosis of ASD by a licensed psychiatrist with no less than 5 years' clinical experience by strictly following the DSM-IV criteria; (c) at least average nonverbal intelligence (initially screened by the psychiatrist and subsequently confirmed with the Raven's Advanced Progressive Matrices as <emph>IQ</emph> ≥ <emph>70</emph>); (d) absence of pathological conditions including ADHD and schizophrenia. The TD children were included from nearby communities if they were between 6 and 13 years old, and had no physical or mental disorders. No ASD/ADHD was reported in the first-degree relatives of the TD group. The experimental protocol followed the Declaration of Helsinki, and participants' caregivers signed the written informed consent form. Participants were paid 200 CNY for the participation of the experiment.</p> <hd id="AN0157570732-6">Experimental Procedure</hd> <p>Participants were engaged in a conversation with a 33-year old female interviewer. Four conversational sessions were chronologically arranged: generic question, hobby sharing, yes–no question, and question raising. Data of this study were obtained from the yes–no question session, in which participants were asked to answer ten yes–no questions (please refer to Appendix for detailed information), and they were verbally required by the interviewer to use head nodding/shaking while doing so. The group membership of the participants was not informed to the interviewer, and she was specifically required to behave and speak in the same fashion across all participants.</p> <p>Participants were seated 80 cm away from the interviewer's chair (Fig. 1), and two cameras were used to record the conversation from different angles and for different purposes. One camera (Samsung HMX-F90, sampling frequency 25 Hz) recorded both persons' behavior from the side view, with each person separated equally on the left and right side of the image. The other camera (Logitech C270, sampling frequency 30 Hz, Fig. 1) was placed beside the interviewer to capture participants' behavior from the front view, and only the data recorded by this camera was used to study head movement in both groups of participants.</p> <p>Graph: Fig. 1 Experimental setup</p> <hd id="AN0157570732-7">Head Movement Data Analysis</hd> <p></p> <hd id="AN0157570732-8">Head Nodding and Shaking Frequencies</hd> <p>To examine whether ASD participants differed from their TD counterparts with respect to head movement, two female college students, who were blind to both the purpose of the experiment and the participant's group membership, watched the videos and coded whether participants used head nodding/shaking while answering the questions. For each participant, the frequency of head nodding and shaking were coded independently by both judges. Specifically, a head shaking/nodding behavior was noted if the participant nodded/shake head while answering a question. Reliabilities of the two judges' ratings on the head nodding and shaking frequencies were computed using Spearman's rho correlation. Results showed that the Spearman's rho correlation coefficients for head nodding and shaking were 0.787 and 0.895 respectively, suggesting a strong agreement on ratings between these two judges (Chow et al., [<reflink idref="bib11" id="ref46">11</reflink>]).</p> <p>The two judges' coded frequencies were then averaged to obtain the head nodding and shaking frequencies for each participant. In addition, we also computed the frequency of no nodding/shaking as the total number of questions (ten) subtracting the sum of head nodding/shaking frequencies. It was hypothesized that the ASD children would be less likely to use head nodding/shaking when answering these questions, and thus a greater frequency of no head nodding/shaking would be observed in the ASD group.</p> <hd id="AN0157570732-9">Head Movement Dynamics Extraction</hd> <p>The OpenFace 2.0 (<ulink href="http://cmusatyalab.github.io/openface/">http://cmusatyalab.github.io/openface/</ulink>) was utilized to extract the time series of the participant's head movement from the video clip. It is an automatic computer-vision algorithm that performs facial landmark detection, head pose tracking, facial action unit recognition and eye trajectory tracking (Baltrusaitis et al., [<reflink idref="bib4" id="ref47">4</reflink>]). The present study used the head pose tracking function to compute the angular displacement in the pitch (head nodding direction), yaw (head shaking direction) and roll (lateral head tilting direction) directions of the head movement (Fig. 2). The accuracy of OpenFace 2.0 was assessed on two publicly available datasets (Baltrušaitis et al., [<reflink idref="bib3" id="ref48">3</reflink>]; Cascia et al., [<reflink idref="bib10" id="ref49">10</reflink>]). It was reported that the mean absolute degree error for the pitch direction was between 2.4 and 3.1 degrees, the yaw direction 3.2–3.5 degrees, and the roll direction 2.4–3.1 degrees, which demonstrates the state-of-the-art performance as compared to other similar head pose tracking frameworks (Baltrusaitis et al., [<reflink idref="bib4" id="ref50">4</reflink>]). Apart from the high accuracy at head pose tracking, OpenFace 2.0 enables head tracking in the case of face occlusion or insufficient light. This is particularly important for the present study since the videotape was not recorded from the direct frontal direction, and it happened that part of the participant's face was self-occluded when the head rotated to the far left side. On average, the OpenFace 2.0 failed to track 1.76% of the video frames (ASD: 3.46% and TD 0.23%) in this study. With the aid of OpenFace 2.0, the time series of the angular displacement in the pitch, yaw and roll direction could be obtained.</p> <p>Graph: Fig. 2 Illustration of head pose tracking and the three directions of head movement. The red dots around the face are the facial landmarks detected by OpenFace 2.0, and the blue cuboid represents the head direction.</p> <hd id="AN0157570732-10">Feature Computation</hd> <p>Matlab (2017b, MathWorks, Natick, MA, USA) was used in the feature computation procedure. The rotation range (RR) and the amount of rotation per minute (ARPM) in the pitch, yaw and roll directions were computed as features. RR refers to the maximum angular displacement participants made in a specific direction. ARPM denotes the amount of head movement participants made in a particular direction per minute. To be noted is that we computed ARPM instead of the whole amount of movement to address the issue that participants spent different time to finish the conversation. To obtain these two variables, the raw data of the time series of the angular displacement in a specific direction were interpolated with the Spline interpolation method, which helped smooth the raw data by excluding outliers and reducing noise.</p> <p>The RR was computed as the subtraction between the maximum positive value and the minimum negative value (Fig. 3a). To calculate the ARPM, the angular velocity was first estimated as the first derivative of the angular displacement. Subsequently, the amount of head movement in a specific direction was obtained as the integral of the angular velocity time series, which refers to area under the velocity curve (Fig. 3b). Finally, the ARPM could be achieved as:</p> <p> <ephtml> <math display="block" xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mtext>ARPM</mtext><mspace width="0.166667em" /><mo>=</mo><mspace width="0.166667em" /><mfrac><mrow><mtext>The</mtext><mspace width="0.166667em" /><mtext>total</mtext><mspace width="0.166667em" /><mtext>amount</mtext><mspace width="0.166667em" /><mtext>of</mtext><mspace width="0.166667em" /><mtext>rotation</mtext><mspace width="0.166667em" /><mtext>of</mtext><mspace width="0.166667em" /><mtext>the</mtext><mspace width="0.166667em" /><mtext>trial</mtext><mspace width="0.166667em" /><mo>×</mo><mspace width="0.166667em" /><mn>60</mn><mspace width="0.166667em" /><mfenced close=")" open="("><mtext>seconds</mtext></mfenced></mrow><mrow><mtext>The</mtext><mspace width="0.166667em" /><mtext>length</mtext><mspace width="0.166667em" /><mtext>of</mtext><mspace width="0.166667em" /><mtext>the</mtext><mspace width="0.166667em" /><mtext>trial</mtext><mspace width="0.166667em" /><mtext>measured</mtext><mspace width="0.166667em" /><mtext>in</mtext><mspace width="0.166667em" /><mtext>seconds</mtext></mrow></mfrac></mrow></math> </ephtml> </p> <p>Graph</p> <p>Graph: Fig. 3 a Exemplary time series of angular displacement after interpolation. The graph illustrates that rotation range (RR) refers to the distance between the maximum positive value and the minimum negative value. b Exemplary time series of angular velocity. The dark area under the velocity curve was calculated as the amount of rotation of the trial.</p> <p>In the end, six features were obtained, namely RR and ARPM in the pitch, yaw and roll directions.</p> <hd id="AN0157570732-11">Machine Learning Process</hd> <p></p> <hd id="AN0157570732-12">Description of Dataset</hd> <p>Data of five participants (2 ASDs and 3 TDs) were lost due to technical problems caused by the camera. Therefore, the present study included 18 ASD and 20 TD participant samples. Given that six variables were computed as features to characterize head movement, a 38 (participants) * 6 (features) matrix was used as the ML dataset.</p> <hd id="AN0157570732-13">Classifiers</hd> <p>Five classifiers were used in this ML procedure, including support vector machine (SVM), linear Discriminant analysis (LDA), decision tree (DT), random forest (RF), and ensembles learning with boosting (ENS). The introduction to these classifiers is presented as follows.</p> <p>SVM is a supervised learning approach that typically solves the binary classification problem through the creation of an optimal hyperplane in the multidimensional space with a set of labeled training samples. Testing samples are classified to either category based on the sign of the distance vector to the hyperplane. SVM has been widely implemented in prior ML studies to classify individuals with and without ASD (Crippa et al., [<reflink idref="bib12" id="ref51">12</reflink>]; Zhao et al., [<reflink idref="bib42" id="ref52">42</reflink>]).</p> <p>LDA is a dimension reduction technique that creates a linear combination of features to address a classification question. In the present study (a binary classification task), LDA projects all data points in the high-dimensional space onto one straight line. A threshold would be obtained to optimally differentiate these two groups of samples.</p> <p>DT is a flowchart with a tree-like structure. Each node in the DT model represents a test on an attribute, each branch denotes an outcome of the test, and each leaf node corresponds to a class label. The construction of a decision tree model is to find attributes that return the highest information gain. DT holds the advantage of strong interpretability and computational efficiency, but it is prone to overfitting.</p> <p>Different from the DT classifier, RF builds a model with multiple simple tress by means of training with a dataset of random features on a random portion of observations. Test sample are classified into either category by the majority of votes from these trees. The RF classifier could well address the overfitting problem of the DT algorithm.</p> <p>ENS is a process that combines multiple learning algorithms to solve a classification or prediction problem. The present study used the AdaBoost ensemble learning algorithm, in which the week classifier is a DT. The ML is the process of upgrading the "weak learning algorithm" to the "strong learning algorithm". Specifically, a weighted majority voting method was used to increase the weight of weak classifiers with a small classification error rate, and to reduce the weight of weak classifiers with a large classification error rate.</p> <hd id="AN0157570732-14">Feature Selection</hd> <p>It is reasonable to argue that only variables significantly different between the ASD group and the TD group may contribute to ASD identification. At the same time, more features for ML approaches require higher computational resources. Thus, after obtaining six features (RR and ARPM in the pitch, yaw and roll directions), measures were taken to select only discriminant features into the ML process for the purpose of improving computation efficiency, and indiscriminate ones were discarded. Discriminant features were defined as variables that were statistically significant between the ASD group and the TD group. Specifically, linear mixed-effect models (LMEMs) were performed to examine which feature could significantly discriminate the ASD children from the TD counterparts by controlling potential confounders. In all LMEMs, each of the six features were entered as the dependent variable, with group (ASD vs. TD) as the fixed factor. The participant's age, IQ, head nodding frequency, head shaking frequency, and frequency of no nodding/shaking were entered as random factors. Likelihood ratio tests (LR-test) were conducted to attain the <emph>p</emph>-values, Chi-Square values and the associated degrees of freedom. The logic of LR-test was to compare two models—one with the examined factor (i.e., group) and the other one without. A <emph>p</emph>-value ≤ 0.05 indicated that the dependent variable was a discriminate feature, and a <emph>p</emph>-value > 0.05 an indiscriminate feature. Chi-Square values were obtained according to the Wilk's Theorem, which states that negative two times the log likelihood ratio of two models approaches a Chi-Square distribution.</p> <p>In the second step, a random feature combination (RFC) approach was applied in the procedure of selecting features into classifiers for model training and testing. RFC was implemented since only six features were computed in the present study. As compared to other feature selection solutions such as the forward feature selection or the backward feature selection, the RFC suffers from the limitation of high computation demanding. However, it typically yields equal or higher classification accuracy simply owing to the fact that all features and feature combinations could be tested. With the help of the RFC approach, the classification accuracy of all the one-feature model, two-feature model, ..., <emph>n</emph>-feature (<emph>n</emph> equals the number of discriminant features) model were tested.</p> <hd id="AN0157570732-15">Classification Task</hd> <p>Matlab was used to perform the classification task. The goal was to examine which feature (or feature combination) would yield the best classification performance. All possible feature combinations were fed into the five classifiers to evaluate accuracy, specificity, and sensitivity. Accuracy refers to the proportion of samples correctly classified in both groups of participants. Specificity represents the proportion of TD samples that are correctly identified, and sensitivity the percentage of ASD samples that are correctly labeled. To prevent overfitting, leave-one-out cross validation (LOOCV) was applied to all classifiers. Specifically, LOOCV entailed that one participant sample was used as the testing set and all the other participant samples were used as the training set. LOOCV was repeated until all participant samples have been used once as the test set. The whole ML process is presented in Fig. 4.</p> <p>Graph: Fig. 4 The ML procedure</p> <hd id="AN0157570732-16">Results</hd> <p></p> <hd id="AN0157570732-17">Comparisons Between ASD and TD</hd> <p>With respect to the frequencies of head nodding, shaking and no nodding/shaking, Shapiro–Wilk tests showed significant departures from normality in data regarding ASD head nodding (<emph>w</emph>(<reflink idref="bib18" id="ref53">18</reflink>) = 0.88, <emph>p</emph> = 0.02), ASD head shaking (<emph>w</emph>(<reflink idref="bib18" id="ref54">18</reflink>) = 0.83, <emph>p</emph> < 0.01), ASD no nodding/shaking (<emph>w</emph>(<reflink idref="bib18" id="ref55">18</reflink>) = 0.53, <emph>p</emph> < 0.01), and TD no nodding/shaking (<emph>w</emph>(<reflink idref="bib20" id="ref56">20</reflink>) = 0.73, <emph>p</emph> < 0.01). Based on these results, independent-samples Mann–Whitney U tests were performed. Results revealed that the ASD children were not significantly different from the TD children on head nodding (Median<subs>_ASD</subs> = 8, Median<subs>_TD</subs> = 5.75, <emph>p</emph> = 0.276, <emph>r</emph> = − 0.18), head shaking (Median<subs>_ASD</subs> = 2, Median<subs>_TD</subs> = 3.25, <emph>p</emph> = 0.063, <emph>r</emph> = − 0.30), or no nodding/shaking (Median<subs>_ASD</subs> = 0, Median<subs>_TD</subs> = 0, <emph>p</emph> = 0.290, <emph>r</emph> = − 20). Only two participants (1 ASD and 1 TD) did not use head nodding/shaking at all while answering questions.</p> <p>Results on the comparisons between ASD and TD with respect to the demographic information and head movement are presented in Table 1. Concerning the differences in head movement dynamics, LMEMs showed that ASD children had significantly greater RR and ARPM in all the pitch, yaw, and roll directions as compared to the TD group, with all <emph>p</emph>-values lower than 0.05. Therefore, all of these six features were fed into the ML classifiers.</p> <p>Table 1 Subject demographics and group comparisons</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>a</sup></p></td><td align="left"><p>16:2</p></td><td align="left"><p>16:4</p></td><td align="left"><p>χ<sup>2</sup>(1) =.563</p></td><td char="." align="char"><p>.453</p></td></tr><tr><td align="left"><p>Age in months (Mean ± SD)<sup>b</sup></p></td><td align="left"><p>101.1 ± 24.9</p></td><td align="left"><p>116.1 ± 22.8</p></td><td align="left"><p>t(36) = 1.92</p></td><td char="." align="char"><p>.062</p></td></tr><tr><td align="left"><p>IQ (Mean ± SD)<sup>b</sup></p></td><td align="left"><p>100.6 ± 23.3</p></td><td align="left"><p>119.0 ± 16.0</p></td><td align="left"><p>t(29.7) = 2.80</p></td><td char="." align="char"><p>.009*</p></td></tr><tr><td align="left"><p>RR_Pitch (Mean ± SD)<sup>c</sup></p></td><td align="left"><p>45.6 ± 32.6</p></td><td align="left"><p>11.3 ± 8.4</p></td><td align="left"><p>χ<sup>2</sup>(1) = 12.29</p></td><td char="." align="char"><p> <.001**</p></td></tr><tr><td align="left"><p>RR_Yaw (Mean ± SD)<sup>c</sup></p></td><td align="left"><p>45.7 ± 26.7</p></td><td align="left"><p>17.4 ± 16.4</p></td><td align="left"><p>χ<sup>2</sup>(1) = 11.34</p></td><td char="." align="char"><p> <.001**</p></td></tr><tr><td align="left"><p>RR_Roll (Mean ± SD)<sup>c</sup></p></td><td align="left"><p>37.8 ± 28.7</p></td><td align="left"><p>10.2 ± 5.5</p></td><td align="left"><p>χ<sup>2</sup>(1) = 12.61</p></td><td char="." align="char"><p> <.001**</p></td></tr><tr><td align="left"><p>ARPM_ Pitch (Mean ± SD)<sup>c</sup></p></td><td align="left"><p>83.8 ± 67.5</p></td><td align="left"><p>28.1 ± 26.4</p></td><td align="left"><p>χ<sup>2</sup>(1) = 5.89</p></td><td char="." align="char"><p>.015*</p></td></tr><tr><td align="left"><p>ARPM_ Yaw (Mean ± SD)<sup>c</sup></p></td><td align="left"><p>91.3 ± 67.9</p></td><td align="left"><p>33.4 ± 34.0</p></td><td align="left"><p>χ<sup>2</sup>(1) = 7.42</p></td><td char="." align="char"><p>.006**</p></td></tr><tr><td align="left"><p>ARPM_ Roll (Mean ± SD)<sup>c</sup></p></td><td align="left"><p>72.7 ± 58.6</p></td><td align="left"><p>24.9 ± 22.8</p></td><td align="left"><p>χ<sup>2</sup>(1) = 7.88</p></td><td char="." align="char"><p>.005**</p></td></tr></tbody></table> </ephtml> </p> <p>*.01 < p <.05 ; ** p <.01 <sups>a</sups>Chi-square test was performed <sups>b</sups>Independent samples t-test was performed <sups>c</sups>Likelihood ratio test was performed to examine whether group was a significant predictor of the feature in the LMEMs</p> <hd id="AN0157570732-18">ML Classification Performances</hd> <p>The maximum accuracy of each classifier was achieved with different numbers of features (Fig. 5). It is shown that the SVM classifier yielded the highest accuracy with only one feature, the DT classifier with two features, and the LDA, RF and ENS classifiers with three features. The results also showed that all classifiers achieved the highest accuracy over 81% with no more than three features, indicating that all the selected ML classifiers were able to successfully identify ASD children from the TD group (Table 2). Among all classifiers, the maximum classification accuracy was fulfilled with the DT classifier (accuracy = 92.11%) with two features: RR_Roll and ARPM_Yaw. The sensitivity and specificity of the corresponding model was 88.9% and 95% respectively (area under curve: AUC = 0.9167).</p> <p>Graph: Fig. 5 The maximum classification accuracy of the five classifiers as a function of the number of features</p> <p>Table 2 Maximum performance of the ML classifiers</p> <p> <ephtml> <table frame="hsides" rules="groups"><thead><tr><th align="left"><p>Classifier</p></th><th align="left"><p>No. of features that achieved highest accuracy</p></th><th align="left"><p>Accuracy</p></th><th align="left"><p>Specificity</p></th><th align="left"><p>Sensitivity</p></th><th align="left"><p>AUC</p></th><th align="left"><p>Feature/Feature combination</p></th></tr></thead><tbody><tr><td align="left"><p>SVM</p></td><td char="." align="char"><p>1</p></td><td char="." align="char"><p>0.8158</p></td><td char="." align="char"><p>0.8500</p></td><td char="." align="char"><p>0.7778</p></td><td char="." align="char"><p>0.8167</p></td><td align="left"><p>RR_Roll</p></td></tr><tr><td align="left"><p>LDA</p></td><td char="." align="char"><p>3</p></td><td char="." align="char"><p>0.8421</p></td><td char="." align="char"><p>0.8500</p></td><td char="." align="char"><p>0.8333</p></td><td char="." align="char"><p>0.8222</p></td><td align="left"><p>RR_Yaw, ARPM_Pitch, and ARPM_Yaw</p></td></tr><tr><td align="left"><p>DT</p></td><td char="." align="char"><p>2</p></td><td char="." align="char"><p>0.9211</p></td><td char="." align="char"><p>0.9500</p></td><td char="." align="char"><p>0.8889</p></td><td char="." align="char"><p>0.9167</p></td><td align="left"><p>RR_Roll and ARPM_Roll</p></td></tr><tr><td align="left"><p>RF</p></td><td char="." align="char"><p>3</p></td><td char="." align="char"><p>0.8421</p></td><td char="." align="char"><p>0.9000'</p></td><td char="." align="char"><p>0.7778</p></td><td char="." align="char"><p>0.8500</p></td><td align="left"><p>RR_Roll, ARPM_Yaw, and ARPM_Roll</p></td></tr><tr><td align="left"><p>ENS</p></td><td char="." align="char"><p>3</p></td><td char="." align="char"><p>0.8684</p></td><td char="." align="char"><p>0.9500</p></td><td char="." align="char"><p>0.7778</p></td><td char="." align="char"><p>0.8556</p></td><td align="left"><p>RR_Pitch, RR_Yaw and ARPM_Yaw</p></td></tr></tbody></table> </ephtml> </p> <p> <emph>RR</emph> Rotation range, <emph>ARPM</emph> Amount of rotation per minute</p> <p>To evaluate the performance of all the ML classifiers, we plotted the receiver operating characteristic (ROC) curves for each classifier respectively (Fig. 6). The maximum AUC reached 0.9167 with the DT classifier. All the other classifiers produced an AUC value greater than 0.81, indicating a good classification performance for all the selected classifiers.</p> <p>Graph: Fig. 6 ROC curves for all classifiers. The true positive rate and the false positive rate refer to the accuracy of the model to correctly and incorrectly predict the ASD individuals respectively</p> <hd id="AN0157570732-19">Discussion</hd> <p>The present study established ML frameworks to classify children with ASD and TD. With the help of a head pose tracking algorithm—OpenFace 2.0, six features were computed and further fed into five ML classifiers. Results demonstrated that the maximum classification accuracy of 92.11% was reached by the DT classifier with two features (i.e., RR_Roll and ARPM_Yaw). Our finding suggested that social head movement contains kinematic information that could be leveraged to identify ASD. To the best of the authors' knowledge, our study is the first to use head movement related features to detect ASD.</p> <p>By having participants answer yes–no questions, it was initially hypothesized that children with ASD would exhibit less head nodding/shaking behavior. Contrary to this expectation, the ASD children in the present study did not show a diminished frequency of head nodding/shaking while answering yes–no questions. This result was inconsistent with several previous findings. For example, Oi showed that ASD children had significantly less non-verbal (head nodding and pointing) responses to yes–no questions (Oi, [<reflink idref="bib31" id="ref57">31</reflink>]). In conversations other than answering yes–no questions, previous studies were supportive of the idea that children with ASD exhibit less head-nodding/shaking behavior when listening to interviewer (Capps et al., [<reflink idref="bib9" id="ref58">9</reflink>]; García-Pérez et al., [<reflink idref="bib17" id="ref59">17</reflink>]). The inconsistency between this study and previous findings might be attributed to the fact that our participants were deliberately required to use head nodding/shaking while answering yes–no questions. Previous studies suggest that the reduced use of head nodding/shaking in ASD individuals might only be observed in unintentional scenarios (Capps et al., [<reflink idref="bib9" id="ref60">9</reflink>]; García-Pérez et al., [<reflink idref="bib17" id="ref61">17</reflink>]; Oi, [<reflink idref="bib31" id="ref62">31</reflink>]). In this vein, our task may not a good one to reveal an ASD vs. TD difference in terms of head nodding/shaking frequency. Future studies might consider examining the unintentional use of head nodding/shaking when answering yes–no questions when investigating whether individuals with ASD are less likely to perform these behaviors.</p> <p>Although no significant difference was found in head nodding/shaking frequency, our results showed that the ASD children exhibited significantly higher level of RR and ARPM in all the pitch, yaw and roll directions of head movement as compared to the TD children. On one hand, these results were consistent with our previous finding by having participants answer generic questions (Zhao et al., [<reflink idref="bib43" id="ref63">43</reflink>]), and with those reported in Martin et al.'s study, which presented that the ASD children had higher amount of head turning displacements and velocity in a video watching scenario (Martin et al., [<reflink idref="bib29" id="ref64">29</reflink>]). On the other hand, this finding suggested that children with ASD nodded/shake head in a different fashion from those with TD, with ASD children presenting excessive amount of head movements. In fact, human perceptual system is good at detecting specific patterns of movement (e.g., head nodding, smiling), but it is not good at giving precise metric measures of movement. These results demonstrated an advantage head pose tracking technique has over behavioral coding techniques—finding motor patterns that might be difficult to spot through human observations. This idea is consistent with Perochon et al.'s finding which showed that their computer vision algorithms were able to find the difference in latency to respond to their names, which was hardly detectable by human eyes (Perochon et al., [<reflink idref="bib33" id="ref65">33</reflink>]).</p> <hd id="AN0157570732-20">Limitations and Future Directions</hd> <p>Although both the sample size and classification performance of our ML framework are comparable to prior studies using kinematic features to identify ASD (Anzulewicz et al., [<reflink idref="bib2" id="ref66">2</reflink>]; Crippa et al., [<reflink idref="bib12" id="ref67">12</reflink>]; Li et al., [<reflink idref="bib26" id="ref68">26</reflink>]), we admit that our sample size was small. This is mainly due to the great difficulty of recruiting a large sample of ASD participants and having them complete the motor tasks. Small sample size is always associated with the overfitting problem. To address this concern, this study implemented LOOCV in model training and testing, which is a widely used technique to prevent overfitting (Berrar, [<reflink idref="bib5" id="ref69">5</reflink>]). In addition, the number of features that yielded the highest classification accuracy in all our examined classifiers did not exceed three, suggesting that the models would not be too complex to overfit. Given the significant heterogeneity in ASD (Jacob et al., [<reflink idref="bib22" id="ref70">22</reflink>]), however, future explorations should include larger samples of ASD individuals with different behavioral presentations to help better understand what behavioral features could accurately identify ASD.</p> <p>The age of our participants was between 6 and 13 years, which was much older than the mean age at diagnosis (Hof et al., [<reflink idref="bib20" id="ref71">20</reflink>]). Thus, our study could only be viewed as an initial examination of social head behaviors in an older sample to establish proof of concept. However, it is worth noting that head pose tracking technique could well serve ASD early screening purposes owing to its constraint-free characteristics, which determines that behavioral data could be collected in natural settings without wearing sophisticated devices or following experimental instructions. Therefore, the use of these computer-vision based techniques as in the present study might facilitate the detection of ASD in everyday life, which might lead to earlier confirmed diagnosis of ASD (Daniels & Mandell, [<reflink idref="bib13" id="ref72">13</reflink>]; Hof et al., [<reflink idref="bib20" id="ref73">20</reflink>]).</p> <hd id="AN0157570732-21">Conclusion</hd> <p>The present study utilized a computer vision algorithm to extract kinematic features from head movement dynamics, and established ML frameworks to examine the classification performance of these features. Our results demonstrated that only two features could be able to classify children with ASD and TD with a high accuracy, suggesting a considerable value head movement dynamics may pose on ASD identification. With the advancement of technology that enables the acquirement of objective features, additional novel biomarkers of ASD could be identified. Indeed, features obtained with neuroimaging, eye tracking, EEG and motion capture approaches have exhibited promising results in detecting ASD (Crippa et al., [<reflink idref="bib12" id="ref74">12</reflink>]; Grossi et al. [<reflink idref="bib19" id="ref75">19</reflink>]; Liu et al., [<reflink idref="bib28" id="ref76">28</reflink>]; Plitt et al., [<reflink idref="bib34" id="ref77">34</reflink>]; Zhao et al., [<reflink idref="bib42" id="ref78">42</reflink>]). It is assumed that the current diagnostic system of ASD, which are established on behavioral evaluation of social communication and restricted and repetitive behavior, would be greatly challenged. Future ASD diagnosis might be made by combining features obtained from all these modalities of data collection.</p> <hd id="AN0157570732-22">Acknowledgments</hd> <p>The authors thank Xianpeng Zhang, Zeming Huang, Chuang Luo, and Jian Cai for data collection.</p> <hd id="AN0157570732-23">Author Contributions</hd> <p>Z Z, XZ, X.Q and J.L designed the experiment and recruited the participants. Z.Z, Z.Z, H.T, J.X and X.H analyzed data. Z.Z, X.Z and X.Q drafted and revised the manuscript.</p> <hd id="AN0157570732-24">Funding</hd> <p>The study was financially supported by 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 (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="AN0157570732-25">Declarations</hd> <p></p> <hd id="AN0157570732-26">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="AN0157570732-27">Appendix 1</hd> <p>See Table 3.</p> <p>Table 3 Detail of the yes–no question answering conversation</p> <p> <ephtml> <table frame="hsides" rules="groups"><tbody><tr><td align="left"><p>1. Do you like apples?</p></td></tr><tr><td align="left"><p>2. Do you like to go to the zoo?</p></td></tr><tr><td align="left"><p>3. Do you like to go to school?</p></td></tr><tr><td align="left"><p>4. Do you like reading?</p></td></tr><tr><td align="left"><p>5. Do you like painting?</p></td></tr><tr><td align="left"><p>6. Do you like watching cartoons?</p></td></tr><tr><td align="left"><p>7. Do you like sports?</p></td></tr><tr><td align="left"><p>8. Do you like watching movies?</p></td></tr><tr><td align="left"><p>9. Do you like traveling?</p></td></tr><tr><td align="left"><p>10. Do you like shopping?</p></td></tr></tbody></table> </ephtml> </p> <hd id="AN0157570732-28">Publisher's Note</hd> <p>Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.</p> <ref id="AN0157570732-29"> <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="ref30" type="bt">2</bibl> <bibtext> Anzulewicz A, Sobota K, Delafield-Butt JT. Toward the Autism Motor Signature: Gesture patterns during smart tablet gameplay identify children with autism. Scientific Reports. 2016. 10.1038/srep31107. 27553971. 4995518</bibtext> </blist> <blist> <bibl id="bib3" idref="ref48" type="bt">3</bibl> <bibtext> Baltrušaitis T, Robinson PL. Morency 3D Constrained Local Model for rigid and non-rigid facial tracking. 2012 IEEE Conference on computer vision and pattern recognition. 2012; IEEE: 2610-2617. 10.1109/CVPR.2012.6247980</bibtext> </blist> <blist> <bibl id="bib4" idref="ref47" type="bt">4</bibl> <bibtext> Baltrusaitis T, Zadeh A, Lim YC. Morency OpenFace 2.0: Facial behavior analysis toolkit. 2018 13th IEEE International conference on automatic face & gesture recognition (FG 2018). 2018; IEEE: 59-66. 10.1109/FG.2018.00019</bibtext> </blist> <blist> <bibl id="bib5" idref="ref69" type="bt">5</bibl> <bibtext> Berrar D. Cross-validation. Encyclopedia of bioinformatics and computational biology. 2019; Elsevier: 1542-1545</bibtext> </blist> <blist> <bibl id="bib6" idref="ref26" type="bt">6</bibl> <bibtext> Bhat AN, Galloway JC, Landa RJ. Relation between early motor delay and later communication delay in infants at risk for autism. Infant Behavior and Development. 2012; 35; 4: 838-846. 10.1016/j.infbeh.2012.07.019. 22982285</bibtext> </blist> <blist> <bibl id="bib7" idref="ref22" type="bt">7</bibl> <bibtext> Bojanek EK, Wang Z, White SP, Mosconi MW. Postural control processes during standing and step initiation in autism spectrum disorder. Journal of Neurodevelopmental Disorders. 2020; 12; 1: 1. 10.1186/s11689-019-9305-x. 31906846. 6945692</bibtext> </blist> <blist> <bibl id="bib8" idref="ref41" type="bt">8</bibl> <bibtext> Campbell K, Carpenter KL, Hashemi J, Espinosa S, Marsan S, Borg JS. Computer vision analysis captures atypical attention in toddlers with autism. Autism. 2018; 23: 613-628</bibtext> </blist> <blist> <bibl id="bib9" idref="ref58" type="bt">9</bibl> <bibtext> Capps L, Kehres J, Sigman M. Conversational abilities among children with autism and children with developmental delays. Autism. 1998; 2; 4: 325-344. 10.1177/1362361398024002</bibtext> </blist> <blist> <bibtext> Cascia ML, Sclaroff S, Athitsos V. Fast, reliable head tracking under varying illumination: An approach based on registration of texture-mapped 3D models. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2000; 22; 4: 322-336. 10.1109/34.845375</bibtext> </blist> <blist> <bibtext> Chow R, Chiu N, Bruera E, Krishnan M, Chiu L, Lam H. Inter-rater reliability in performance status assessment among health care professionals: A systematic review. Annals of Palliative Medicine. 2016; 5; 2: 83-92. 10.21037/apm.2016.03.02. 27121736</bibtext> </blist> <blist> <bibtext> Crippa A, Salvatore C, Perego P, Forti S, Nobile M, Molteni M. 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> <bibtext> Daniels AM, Mandell DS. Explaining differences in age at autism spectrum disorder diagnosis: A critical review. Autism. 2014; 18; 5: 583. 10.1177/1362361313480277</bibtext> </blist> <blist> <bibtext> Dawson G, Jones EJH, Merkle K, Venema K, Lowy R, Faja S. Early behavioral intervention is associated with normalized brain activity in young children with autism. Journal of the American Academy of Child & Adolescent Psychiatry. 2012; 51; 11: 1150-1159. 10.1016/j.jaac.2012.08.018</bibtext> </blist> <blist> <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> <bibtext> Fournier KA, Hass CJ, Naik SK, Lodha N, Cauraugh JH. Motor coordination in autism spectrum disorders: A synthesis and meta-analysis. Journal of Autism and Developmental Disorders. 2010; 40; 10: 1227-1240. 10.1007/s10803-010-0981-3</bibtext> </blist> <blist> <bibtext> García-Pérez RM, Lee A, Hobson RP. On intersubjective engagement in autism: A controlled study of nonverbal aspects of conversation. Journal of Autism and Developmental Disorders. 2007; 37; 7: 1310-1322. 10.1007/s10803-006-0276-x</bibtext> </blist> <blist> <bibtext> Goldman S, Wang C, Salgado MW, Greene PE, Kim M, Rapin I. Motor stereotypies in children with autism and other developmental disorders. Developmental Medicine & Child Neurology. 2009; 51; 1: 30-38. 10.1111/j.1469-8749.2008.03178.x</bibtext> </blist> <blist> <bibtext> Grossi E, Valbusa GM. Buscema 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> 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; 2: 136236132097110</bibtext> </blist> <blist> <bibtext> Hutt C, Hutt SJ. Effects of environmental complexity on stereotyped behaviours of children. Animal Behaviour. 1965; 13; 1: 1-4. 10.1016/0003-3472(65)90064-3</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> Jeni LA, Cohn JF, Kanade T. Dense 3D face alignment from 2D videos in real-time. 2015 11th IEEE International conference and workshops on automatic face and gesture recognition (FG). 2015; IEEE: 1-8</bibtext> </blist> <blist> <bibtext> Kang J, Han X, Song J, Niu Z, Li X. The identification of children with autism spectrum disorder by SVM approach on EEG and eye-tracking data. Computers in Biology and Medicine. 2020. 10.1016/j.compbiomed.2020.103722. 32892114</bibtext> </blist> <blist> <bibtext> Landa R, Garrett-Mayer E. Development in infants with autism spectrum disorders: A prospective study. Journal of Child Psychology & Psychiatry. 2010; 47; 6: 629-638. 10.1111/j.1469-7610.2006.01531.x</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> Licari MK, Alvares GA, Varcin K, Evans KL, Cleary D, Reid SL. Prevalence of motor difficulties in autism spectrum disorder: Analysis of a population-based cohort. Autism Research. 2020; 13; 2: 298-306. 10.1002/aur.2230. 31625694</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> Martin KB, Hammal Z, Ren G, Cohn JF, Cassell J, Ogihara M. Objective measurement of head movement differences in children with and without autism spectrum disorder. Molecular Autism. 2018; 9; 1: 14. 10.1186/s13229-018-0198-4</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. 10.1007/s10803-015-2562-y. 26334871</bibtext> </blist> <blist> <bibtext> Oi M. Do Japanese children with high-functioning autism spectrum disorder respond differently to Wh-questions and Yes/No-questions?. Clinical Linguistics & Phonetics. 2010; 24; 9: 691-705. 10.3109/02699206.2010.488313</bibtext> </blist> <blist> <bibtext> Ozonoff S, Young GS, Goldring S, Greiss-Hess L, Herrera AM, Steele J. Gross motor development, movement abnormalities, and early identification of autism. Journal of Autism and Developmental Disorders. 2008; 38; 4: 644. 10.1007/s10803-007-0430-0</bibtext> </blist> <blist> <bibtext> Perochon S, Di Martino M, Aiello R, Baker J, Carpenter K, Chang Z. A scalable computational approach to assessing response to name in toddlers with autism. Journal of Child Psychology and Psychiatry. 2021. 10.1111/jcpp.13381. 33641216</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. 10.1016/j.nicl.2014.12.013</bibtext> </blist> <blist> <bibtext> Sorosky AD, Ornitz EM, Brown MB, Ritvo ER. Systematic observations of autistic behavior. Archives of General Psychiatry. 1968; 18; 4: 439-449. 10.1001/archpsyc.1968.01740040055007</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> 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. Applying eye tracking to identify autism spectrum disorder in children. Journal of Autism and Developmental Disorders. 2019; 49; 1: 209-215. 10.1007/s10803-018-3690-y. 30097760</bibtext> </blist> <blist> <bibtext> Wiggins LD, Baio J, Rice C. Examination of the time between first evaluation and first autism spectrum diagnosis in a population-based sample. Journal of Developmental and Behavioral Pediatrics. 2006; 27; 2 Suppl: S79-87. 10.1097/00004703-200604002-00005</bibtext> </blist> <blist> <bibtext> Yamagata B, Itahashi T, Fujino J, Ohta H, Nakamura M, Kato N. Machine learning approach to identify a resting-state functional connectivity pattern serving as an endophenotype of autism spectrum disorder. Brain Imaging and Behavior. 2018. 10.1007/s11682-018-9973-2</bibtext> </blist> <blist> <bibtext> Zhao Z, Zhang X, Li W, Hu X, Qu X, Cao X. Applying machine learning to identify autism with restricted kinematic features. IEEE Access. 2019. 10.1109/ACCESS.2019.2950030. 33777590. 7996404</bibtext> </blist> <blist> <bibtext> Zhao Z, Zhu Z, Zhang X, Tang H, Xing J, Hu X. Atypical head movement during face-to-face interaction in children with autism spectrum disorder. Autism Research: Official Journal of the International Society for Autism Research. 2021. 10.1002/aur.2478</bibtext> </blist> </ref> <aug> <p>By Zhong Zhao; Zhipeng Zhu; Xiaobin Zhang; Haiming Tang; Jiayi Xing; Xinyao Hu; Jianping Lu and Xingda Qu</p> <p>Reported by Author; Author; Author; Author; Author; Author; Author; Author</p> </aug> <nolink nlid="nl1" bibid="bib30" firstref="ref2"></nolink> <nolink nlid="nl2" bibid="bib36" firstref="ref3"></nolink> <nolink nlid="nl3" bibid="bib40" firstref="ref4"></nolink> <nolink nlid="nl4" bibid="bib20" firstref="ref5"></nolink> <nolink nlid="nl5" bibid="bib14" firstref="ref6"></nolink> <nolink nlid="nl6" bibid="bib15" firstref="ref7"></nolink> <nolink nlid="nl7" bibid="bib37" firstref="ref8"></nolink> <nolink nlid="nl8" bibid="bib38" firstref="ref9"></nolink> <nolink nlid="nl9" bibid="bib12" firstref="ref10"></nolink> <nolink nlid="nl10" bibid="bib19" firstref="ref11"></nolink> <nolink nlid="nl11" bibid="bib24" firstref="ref12"></nolink> <nolink nlid="nl12" bibid="bib26" firstref="ref13"></nolink> <nolink nlid="nl13" bibid="bib28" firstref="ref14"></nolink> <nolink nlid="nl14" bibid="bib34" firstref="ref15"></nolink> <nolink nlid="nl15" bibid="bib39" firstref="ref16"></nolink> <nolink nlid="nl16" bibid="bib41" firstref="ref17"></nolink> <nolink nlid="nl17" bibid="bib42" firstref="ref18"></nolink> <nolink nlid="nl18" bibid="bib27" firstref="ref21"></nolink> <nolink nlid="nl19" bibid="bib16" firstref="ref23"></nolink> <nolink nlid="nl20" bibid="bib25" firstref="ref24"></nolink> <nolink nlid="nl21" bibid="bib32" firstref="ref25"></nolink> <nolink nlid="nl22" bibid="bib33" firstref="ref27"></nolink> <nolink nlid="nl23" bibid="bib18" firstref="ref33"></nolink> <nolink nlid="nl24" bibid="bib21" firstref="ref34"></nolink> <nolink nlid="nl25" bibid="bib35" firstref="ref35"></nolink> <nolink nlid="nl26" bibid="bib43" firstref="ref36"></nolink> <nolink nlid="nl27" bibid="bib29" firstref="ref38"></nolink> <nolink nlid="nl28" bibid="bib23" firstref="ref39"></nolink> <nolink nlid="nl29" bibid="bib11" firstref="ref46"></nolink> <nolink nlid="nl30" bibid="bib10" firstref="ref49"></nolink> <nolink nlid="nl31" bibid="bib31" firstref="ref57"></nolink> <nolink nlid="nl32" bibid="bib17" firstref="ref59"></nolink> <nolink nlid="nl33" bibid="bib22" firstref="ref70"></nolink> <nolink nlid="nl34" bibid="bib13" firstref="ref72"></nolink>
Header DbId: eric
DbLabel: ERIC
An: EJ1339362
AccessLevel: 3
PubType: Academic Journal
PubTypeId: academicJournal
PreciseRelevancyScore: 0
IllustrationInfo
Items – Name: Title
  Label: Title
  Group: Ti
  Data: Identifying Autism with Head Movement Features by Implementing Machine Learning Algorithms
– 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="%22Zhu%2C+Zhipeng%22">Zhu, Zhipeng</searchLink><br /><searchLink fieldCode="AR" term="%22Zhang%2C+Xiaobin%22">Zhang, Xiaobin</searchLink><br /><searchLink fieldCode="AR" term="%22Tang%2C+Haiming%22">Tang, Haiming</searchLink><br /><searchLink fieldCode="AR" term="%22Xing%2C+Jiayi%22">Xing, Jiayi</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><br /><searchLink fieldCode="AR" term="%22Qu%2C+Xingda%22">Qu, Xingda</searchLink> (ORCID <externalLink term="http://orcid.org/0000-0003-1764-0357">0000-0003-1764-0357</externalLink>)
– 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>. Jul 2022 52(7):3038-3049.
– 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: 12
– Name: DatePubCY
  Label: Publication Date
  Group: Date
  Data: 2022
– Name: TypeDocument
  Label: Document Type
  Group: TypDoc
  Data: Journal Articles<br />Reports - Research
– Name: Subject
  Label: Descriptors
  Group: Su
  Data: <searchLink fieldCode="DE" term="%22Autism%22">Autism</searchLink><br /><searchLink fieldCode="DE" term="%22Pervasive+Developmental+Disorders%22">Pervasive Developmental Disorders</searchLink><br /><searchLink fieldCode="DE" term="%22Motion%22">Motion</searchLink><br /><searchLink fieldCode="DE" term="%22Human+Body%22">Human Body</searchLink><br /><searchLink fieldCode="DE" term="%22Children%22">Children</searchLink><br /><searchLink fieldCode="DE" term="%22Identification%22">Identification</searchLink><br /><searchLink fieldCode="DE" term="%22Nonverbal+Communication%22">Nonverbal Communication</searchLink><br /><searchLink fieldCode="DE" term="%22Artificial+Intelligence%22">Artificial Intelligence</searchLink>
– Name: DOI
  Label: DOI
  Group: ID
  Data: 10.1007/s10803-021-05179-2
– Name: ISSN
  Label: ISSN
  Group: ISSN
  Data: 0162-3257
– Name: Abstract
  Label: Abstract
  Group: Ab
  Data: Our study investigated the feasibility of using head movement features to identify individuals with autism spectrum disorder (ASD). Children with ASD and typical development (TD) were required to answer ten yes--no questions, and they were encouraged to nod/shake head while doing so. The head rotation range (RR) and the amount of rotation per minute (ARPM) in the pitch (head nodding direction), yaw (head shaking direction) and roll (lateral head inclination) directions were computed, and further fed into machine learning classifiers as the input features. The maximum classification accuracy of 92.11% was achieved with the decision tree classifier with two features (i.e., RR_Pitch and ARPM_Yaw). Our study suggests that head movement dynamics contain objective biomarkers that could identify ASD.
– Name: AbstractInfo
  Label: Abstractor
  Group: Ab
  Data: As Provided
– Name: DateEntry
  Label: Entry Date
  Group: Date
  Data: 2022
– Name: AN
  Label: Accession Number
  Group: ID
  Data: EJ1339362
PLink https://search.ebscohost.com/login.aspx?direct=true&site=eds-live&db=eric&AN=EJ1339362
RecordInfo BibRecord:
  BibEntity:
    Identifiers:
      – Type: doi
        Value: 10.1007/s10803-021-05179-2
    Languages:
      – Text: English
    PhysicalDescription:
      Pagination:
        PageCount: 12
        StartPage: 3038
    Subjects:
      – SubjectFull: Autism
        Type: general
      – SubjectFull: Pervasive Developmental Disorders
        Type: general
      – SubjectFull: Motion
        Type: general
      – SubjectFull: Human Body
        Type: general
      – SubjectFull: Children
        Type: general
      – SubjectFull: Identification
        Type: general
      – SubjectFull: Nonverbal Communication
        Type: general
      – SubjectFull: Artificial Intelligence
        Type: general
    Titles:
      – TitleFull: Identifying Autism with Head Movement Features by Implementing Machine Learning Algorithms
        Type: main
  BibRelationships:
    HasContributorRelationships:
      – PersonEntity:
          Name:
            NameFull: Zhao, Zhong
      – PersonEntity:
          Name:
            NameFull: Zhu, Zhipeng
      – PersonEntity:
          Name:
            NameFull: Zhang, Xiaobin
      – PersonEntity:
          Name:
            NameFull: Tang, Haiming
      – PersonEntity:
          Name:
            NameFull: Xing, Jiayi
      – PersonEntity:
          Name:
            NameFull: Hu, Xinyao
      – PersonEntity:
          Name:
            NameFull: Lu, Jianping
      – PersonEntity:
          Name:
            NameFull: Qu, Xingda
    IsPartOfRelationships:
      – BibEntity:
          Dates:
            – D: 01
              M: 07
              Type: published
              Y: 2022
          Identifiers:
            – Type: issn-print
              Value: 0162-3257
          Numbering:
            – Type: volume
              Value: 52
            – Type: issue
              Value: 7
          Titles:
            – TitleFull: Journal of Autism and Developmental Disorders
              Type: main
ResultId 1