Developmental Differences in Reaching-and-Placing Movement and Its Potential in Classifying Children with and without Autism Spectrum Disorder: Deep Learning Approach

Autism Spectrum Disorder (ASD) is among the most prevalent neurodevelopmental disorders, yet the current diagnostic procedures rely on behavioral analyses and interviews and lack objective screening methods. This study seeks to address this gap by integrating upper limb kinematics and deep learning methods to identify potential biomarkers that could be validated in younger age groups in the future to enhance the identification of ASD. Forty-one school-age children, with and without an ASD diagnosis (Mean age ± SE = 10.3 ± 0.4; 12 Females), participated in the study. A single Inertial Measurement Unit (IMU) was affixed to the child’s wrist as they engaged in a continuous reaching and placing task. Deep learning techniques were employed to classify children with and without ASD. Our findings suggest delays in motor planning and control in school-age children compared to healthy adults. Compared to TD children, children with ASD exhibited poor motor planning and control as seen by greater number of movement units, more movement overshooting, and prolonged time to peak velocity/acceleration. Compensatory movement strategies such as greater velocity and acceleration were also seen in the ASD group. More importantly, using Multilayer Perceptron (MLP) model, we demonstrated an accuracy of ~ 78.1% in classifying children with and without ASD. These findings underscore the potential use of studying upper limb movement kinematics during goal-directed arm movements and deep learning methods as valuable tools for classifying and, consequently, aiding in the diagnosis and early identification of ASD upon further validation in younger children.


Introduction
Autism Spectrum Disorder (ASD) ranks among the most prevalent neurodevelopmental disorders in the U.S., affecting approximately 1 in 36 children [1].As de ned by the Diagnostic and Statistical Manual of Mental Disorders, 5th Edition, children with ASD exhibit distinct characteristics, including social communication di culties and the presence of repetitive behaviors [2].However, due to the disorder's high heterogeneity, diagnosing ASD is still very challenging [3].The current diagnostic process for ASD relies on behavioral analyses and interviews, without well-established objective biomarkers [3].
Additionally, the core diagnostic features, such as social communication di culties, often do not become apparent until later stages of development, limiting their utility in early ASD identi cation [4].While not constituting the primary diagnostic criteria, motor impairments are highly prevalent and signi cantly related to other social communication di culties of children with ASD [5][6][7].More importantly, many motor signs, such as atypical reach-to-grasp movements, manifest early in life and hold promise as diagnostic biomarkers [4,8].Given the potential value of studying objective motor measures, our current study explored the developmental differences in arm movement kinematics during goal-directed reaching and the application of deep learning algorithms for the classi cation of Typically Developing (TD) children and children with ASD.Our ndings provide deeper insights into motor development and hold signi cant clinical potential for early ASD identi cation upon further validation in younger children.
During early development, early spontaneous arm movements change into more re ned, object-oriented actions, with object contact and grasping emerging between 4 to 6 months of age [9][10][11].Subsequent to the onset of grasping, reaching movements undergo further re nement, adopting a smoother and more uent trajectory through feedback corrections to initial reaches [12].A longitudinal analysis of hand and joint kinematics during spontaneous and purposeful reaching from 8 to 20 weeks indicated gradual increase in movement length and more purposeful reaching in presence of a toy [9][10][11].In addition, early on infants controlled their shoulder motions and later close to reach onset controlled both shoulder and elbow motions to make toy contact [13,14].This re nement in reaching movements will continue to persist into childhood.A study targeting school-age children found higher peak velocity and straighter movement trajectories in children aged 11 to 15 years old compared to 5 to 7 years old children when reaching toward different directions [15].In the current study, our focus extends to typically developing school-age children, as well as those diagnosed with ASD.We aim to compare their reaching and placing movements with those of healthy adults to examine differences in reaching trajectories between the two groups.
Although not included in the primary diagnostic criteria which mainly include social communication di culties and presence of repetitive behaviors, children with ASD exhibit challenges in both gross and ne motor performance, including motor incoordination, de cient posture control, as well as poor upper limb motor control/dexterity [16][17][18][19].Recent insights gleaned from parent screening questionnaires in the SPARK database reveal that over 85% of children with ASD are at risk for motor di culties, underscoring the critical importance of recognizing and addressing these motor issues within this population [20].
Documentation of motor challenges in children with ASD also goes beyond screening tools and behavioral assessments; and extends to quantitative measures such as movement kinematics [21,22].
Using a motion tracking system, Yang et al. (2014) revealed that children with ASD, when reaching to grasp a cylinder-shaped target, displayed prolonged movement times, increased jerks, and a greater number of movement units compared to their TD peers [21].In a more complex motor task involving ball interceptions, children with ASD exhibited larger movement amplitudes, greater velocity, and more movement units compared to TD children [22].These ndings collectively suggest that children with ASD have impaired motor planning and control, which is quanti able through motion tracking techniques.In the current study, we used a child-friendly, lightweight motion-tracking system a xed to the children's wrists to capture their upper movement kinematics during a reaching and placing task.Moreover, we extend the task from single to continuous, repetitive reaching and placing actions, making it a prolonged goal-directed action to detect ASD-related differences.
Motor di culties observed in children with ASD have a profound impact on their daily living skills and are linked to various developmental domains [5][6][7].Motor skills are known to serve as the foundation for the emergence of other social and cognitive skills during development [23,24].For instance, locomotor skills facilitate the exploration of one's surroundings, while manual dexterity fosters object manipulation and creates opportunities for joint attention and social interactions [23][24][25][26].Recognizing the important role of motor skill in broader development, a comprehensive analysis of the SPARK database revealed that motor performance in children with ASD predicts developmental outcomes in social communication, repetitive behavior, language, and functional domains [6].More importantly, motor signs in children with ASD often precede the emergence of social communication symptoms, making them an ideal biomarker for early identi cation [4].Even in infancy, children later diagnosed with ASD exhibit less variability in general movements compared to their non-diagnosed counterparts [27].In the current study, we seek to understand the relationships between the potential movement biomarkers and the adaptive functioning of children with and without ASD and explore the use of these biomarkers in classifying children with and without ASD.
Given  1).There were no signi cant differences in age, sex, race, and ethnicity between the two groups (Table 1).
Recruitment of participants was conducted through online postings, phone calls, and distribution of iers in local schools, community centers, ASD advocacy groups, and through the Simons Powering Autism Research (SPARK) participant research match service (https://www.sfari.org/resource/spark/).To ensure eligibility and gather essential demographic information such as age, sex, race, and ethnicity, we conducted screening interviews with parents via phone calls.For the TD group, children between 6 to 17 years were included and were excluded if they had any neurological or developmental diagnoses/delays, a history of preterm birth, signi cant birth complications, or a family history of ASD.In contrast, children in the ASD group were included if they met the following criteria: 1) their ages ranged from 6 to 17 years, 2) had a professionally con rmed ASD diagnosis, supported by either a school record (e.g., Individualized Education Plan) or medical/neuropsychological records from a quali ed psychiatrist or clinical psychologist (using the Autism Diagnostic Observation Schedule (ADOS) or Autism Diagnostic Interview-Revised (ADI-R) [30,31].On the other hand, children with ASD were excluded if they were unable to follow one-step instructions (e.g., "please pick up the blocks"), or if they exhibited signi cant sensory and behavioral challenges that hindered their ability to wear the fNIRS cap or complete the reaching-andplacing task.Additionally, parents were requested to complete the Vineland Adaptive Behavioral Scales-2nd edition (VABS) and Social Responsive Scale (SRS) to assess their child's adaptive functioning and social responsiveness (Table 1) [32,33].Children with ASD were consistently found to have poor adaptive functions (indicated by lower VABS scores) and social responsiveness (indicated by higher SRS T score) compared to the TD group (ps < 0.05).All study procedures were carried out in accordance with the Declaration of Helsinki.All inform consent and assent forms as well as all study procedures were approved by the University of Delaware Institutional Review Board (UD IRB).Before participating in the study, all children and their parents signed the inform consent or assent forms approved by the University of Delaware Institutional Review Board.Written parental and experimenter permission/consent to use their pictures for this publication has also been taken.[Insert Table 1 here]

Experimental Procedure
Upon con rming eligibility, the participating children were seated at a table across from an adult experimenter.Both the child and adult reached for their own sets of blocks that were arranged in a circular manner, with the containers placed at the child's right and on the adult's left-hand side (as depicted in Fig. 1A).The child was instructed to use their right hand throughout the experiment, while the adults moved their left hand in order to mirror the child's actions.The inertial measurement units (IMUs, Xsens, Inc.) were tted to the wrists of both the child's and the adult's moving hands (i.e., the child's right and the adult's left hand).During the experiment, the adult experimenter and the child took turns reaching for the blocks and placing them into the container one by one, according to pictorial instructions (Fig. 1B).E-prime software was used to trigger the start of each trial, and an experimenter marked the end of each trial when the child and experimenter nished putting all blocks into the containers.All children and adults completed a total of 6 reaching-placing trials.Note that any one out of 4 adult experimenters were randomly paired with the child participants during the reach-place task.This task has been described in our previous publications, wherein we reported cortical activation ndings in children with ASD [34,35].
The current study focuses on the kinematics of goal-directed, reach-place actions based on pictorial instructions that explained the sequence of blocks to be picked up.

Movement Kinematics
We computed twelve kinematic parameters from the IMU data collected during the reaching and placing task performed by both children and adults.Speci cally, we calculated reaction time as the time interval between the initiation signal and the onset of movement.Movement initiation was determined when both the relative distance (in relation to the starting point) and acceleration spiked to 20% of their maximum values.Once movement initiations were identi ed, we computed additional statistics, including total distance, average velocity, maximum velocity, time to peak velocity, average acceleration, maximum acceleration, and time to peak acceleration.Furthermore, we assessed the number of movement units along the three motion vectors (X, Y, and Z).An individual movement unit was de ned as the moment when a kinematic movement vector crosses zero, representing a signi cant change in direction.We calculated three types of movement units (Type 1 to Type 3), by detecting instances where velocity (Type 1), acceleration (Type 2), and jerk (Type 3) crossed zero in the x, y, and z vectors.To consolidate the zerocrossing incidence across these three axes, we calculated the Root Mean Squares (RMS) for each of the three movement unit types.

Statistical Analyses
One-way ANOVAs were performed to assess disparities in kinematic variables among adults, TD children, and children with ASD.For variables exhibiting signi cant differences, post-hoc t-tests were employed to further explore group differences.In order to explore the associations between movement kinematics variables and the adaptive functioning of the children, Pearson correlations were conducted.All statistical analyses were conducted using IBM SPSS (Version 29, SPSS, Inc.).

Deep Learning Model
Due to the limited number of training data, instead of using an end-to-end deep learning model to predict the ASD classi cation with the raw kinematic data, we used kinematic parameters as the input to the model.Nine out of the twelve kinematic parameters exhibited signi cant (or borderline signi cant) differences between children with and without ASD, as detailed in Result 3.1 and visualized in Fig. 3. 3. Results

Movement Kinematics
In Fig. 2, we presented the movement trajectories of adults, TD children, and children with ASD.Notably, children with ASD have wider and jerkier movement trajectories compared to both TD adults and children.
Statistical analyses through a one way-ANOVA revealed signi cant differences between Adults, TD children, and children with ASD for all but two kinematic parameters (i.e., type 1 and 3 movement units; Supplementary Table S1).Post-hoc t-tests of the remaining 10 variables revealed distinct features of these differences.Speci cally, longer movement time, greater total displacement, averaged velocity, maximum velocity, and averaged acceleration were found in adults compared to the TD children (ps < 0.05; Fig. 3B, 3C, 3D, 3E, 3G).On the contrary, adults showed faster reaction time, time to peak velocity, time to peak acceleration, as well as fewer Type 2 movement units compared to the TD children (ps < 0.05; Fig. 3A, 3F, 3I, 3K), suggesting developmental differences in temporal motor control parameters.For the ASD-related differences, children with ASD demonstrated increased values in nearly all kinematic parameters, including total displacement, averaged velocity, maximum velocity, time to peak velocity, averaged acceleration, maximum acceleration, time to peak acceleration, and Type 2 movement units (ps < 0.05; Fig. 3C, 3D, 3E, 3F, 3G, 3H, 3I, 3K).Additionally, a notable trend toward signi cance was observed in reaction time, suggesting that children with ASD required more time to initiate their movements (p = 0.09; Fig. 3A).

Correlations between Motor Performance and Adaptive Functions
TD children showed signi cant correlations between the movement kinematics with the VABS Daily living and socialization scores, whereas children with ASD showed signi cant correlations between the VABC communication and socialization scores (ps < 0.001; Table 2).Speci cally, in TD children, averaged velocity and acceleration were positively correlated with VABC social scores (r = 0.336 ~ 0.341, p < 0.001, Table 2), while the movement time, and movement units (Type 1 to 3) were negatively correlated with the VABS daily living and socialization scores (r = -0.341~-0.525,p < 0.001; Table 2).For children with ASD, movement time, displacement, averaged velocity, averaged acceleration, and maximum acceleration were positively correlated with their VABS communication and/or socialization scores (r = 0.237 ~ 0.452, p < 0.001; Table 2).

Classi cation of Typically Developing Adults and Children
The deep learning model achieved a 10-fold cross-validation accuracy of 98.00% in classifying between adults and TD children.Feature importance analyses showed that the averaged velocity is the most important factor for the developmental classi cation (Fig. 4).

Classi cation of Children with and without ASD
The deep learning model achieved a 10-fold cross-validation accuracy of 78.1% in classifying between children with and without ASD.Feature importance analyses showed that the Averaged Acceleration, followed by Movement Units, Displacement, and Maximum Acceleration are the important factors for the TD vs ASD classi cation (Fig. 5).

Discussion
Autism Spectrum Disorder (ASD) is one of the most prevalent neurodevelopmental disorders, necessitating objective diagnostic approaches.Through the integration of reaching kinematic measures and deep learning methods, the current study aimed to understand the relationship between motor performance and adaptive functioning in children with and without ASD by obtaining action-based, objective biomarkers of ASD.Our ndings showed many differences in reaching kinematics of schoolaged children with ASD compared to TD children/adults, with children having longer reaction times and time to peak velocity/acceleration, shorter movement times and total displacements, lower average velocity/acceleration and maximum acceleration, as well as fewer movement units compared to healthy adults.When comparing children with and without ASD, children with ASD showed greater total displacement, averaged velocity/acceleration, maximum velocity/acceleration, time to peak velocity/acceleration, as well as movement units compared to the TD children.Reaching kinematics also correlated to the daily living and socialization scores of TD children and to the communication and socialization scores of children with ASD.More importantly, our deep learning models achieved ~ 98% accuracy in classifying TD children and adults and ~ 78.1% in classifying children with and without ASD.
Our ndings support the use of reaching kinematics and deep learning methods in classifying children with and without ASD and have the potential to assist in diagnosis/early identi cation process upon further validation in younger children.

Developmental Differences in Goal-directed Reaching Movements
In line with prior research [15,39,40], the current study highlights the ongoing re nement of upper limb movements in school age children.Speci cally, we observed a longer reaction time in TD children compared to healthy adults, indicating delayed anticipatory control of reaching.Anticipatory control of movements also involves motor planning /executive functioning processes such as motor planning, inhibitory control, working memory, and cognitive shifting, that are known to stabilize in early adulthood [41].Therefore, consistent with other studies, slower reaching reaction times in children relative to adults [42], aligns with the expected developmental trajectory of anticipatory control and executive functioning.
We also found shorter movement time and total displacement in children compared to adults, ndings that are counterintuitive and not consistent with previous ndings [43].This discrepancy may stem from the unique experimental setup of the study, wherein adult participants were paired with children to complete the reaching and placing task.It is conceivable that the adults exaggerated their movements for the bene t of the child partner, thus executing movements with larger amplitude and taking more time to complete the task.Despite the shorter movement times in children, they moved slower and jerkier and showed more movement units compared to healthy adults, a characteristic consistently observed in previous studies, suggesting insu cient online movement control [15,44].Moreover, children demonstrated distinct temporal parameters compared to adults during the reaching and placing task, featuring lower averaged velocity/acceleration, maximum velocity, and longer time to peak velocity/acceleration.These ndings align with previous kinematic observations, indicating insu cient feedforward/anticipatory estimates of movement trajectories in the service of online movement control [39,40].extended time to peak velocity/acceleration and more substantial online movement corrections [22,47].For instance, during interceptive tasks, children with ASD did not rely on visual cues for movement planning and had greater movement units, as was observed in the study by Chen et al. (2019) [22].
Children with ASD demonstrated a unique movement pattern of increased movement speed/acceleration (i.e., greater averaged and maximum velocity/acceleration), and greater movement overshooting (i.e., greater total displacement) compared to their TD peers.These ndings, consistent with previous studies using horizontal sinusoidal arm movements and interceptive tasks [22,48], may signify a compensatory strategy employed by children with ASD to overcome the delays in movement initiation.In summary, children with ASD's delays in movement initiation, coupled with faster yet less uid movements, may result in overshooting as well as lower accuracy.

Relationships between Movement Kinematics Adaptive Functioning
Consistent with the past ndings [5,6], our study found signi cant correlations between children's motor skills and adaptive functioning.These connections highlight the fundamental role of motor skills in overall functioning as well as its inter-dependence with other developmental domains [24].For example, the acquisition of new motor skills, such as manual dexterity and locomotor skills, not only fosters exploration and learning but also exerts a cascading in uence on social communication, cognitive abilities, and overall functioning [23].While a general correlation exists between motor skills and adaptive functioning in both TD and ASD groups, the nature of these relationships differed for children with and without ASD.Speci cally, TD children exhibiting better adaptive functioning and socialization skills displayed more advanced movement kinematics such as shorter movement times, greater velocity/acceleration, and reduced movement units.In contrast, among children with ASD, those with better social communication performance exhibited prolonged movement times, greater displacement, velocity, and acceleration.This suggests that children with higher functioning ASD may employ compensatory mechanisms during the execution of the reach-place task.

Deep Learning Classi cation and Future Implications
the growing body of literature on motor di culties in children with ASD, the translation of these ndings to effective tools for early identi cation remains limited.Some researchers have advocated for the inclusion of motor di culties as diagnostic criteria or speci ers; however, the evaluation of motor di culties still heavily relies on parent reports and subjective behavioral assessments [5][6][7].To our knowledge, this is the rst study to leverage continuous upper limb movement kinematics and deep learning methods for the classi cation of individuals with and without ASD.Our ndings underscore the feasibility of employing a comprehensive examination of movement kinematics and deep learning methods to classify children with ASD which might facilitate ASD diagnosis upon validation in younger children.
While the current study included a reasonable sample size for an investigation centered on movement kinematics (N = 41), it may be considered relatively small for training an end-to-end deep learning model or using other network designs for time-series data such as Long Short-Term Memory network (LSTM).
Future research endeavors should include larger sample sizes to enhance the optimization of deep learning models.Beyond dataset expansion, several improvements could be made to the deep learning model.The current model requires pre-processing and identi cation of important kinematic parameters before putting them into the model.While this approach is bene cial for determining differences in movement strategies among adults, TD children, and children with ASD, it may also limit the maximum accuracy of the model.Future deep learning models could be developed to train on the raw kinematic data, allowing the model to autonomously calculate internal parameters optimized for classi cation.For example, an ideal architecture for such a model might incorporate recurrent layers, con gured into an autoencoder with an attention mechanism to mitigate the risk of vanishing gradients.These re nements stand as potential avenues for advancing the e cacy and e ciency of using kinematic parameters and deep learning models in classifying children with and without ASD and facilitating ASD diagnosis upon validation in younger children.

Conclusions
Using a continuous reach and place paradigm, the current study found differences in movement kinematics among TD children compared to healthy adults, revealing the ongoing re nement of anticipatory control and motor planning throughout school age.More importantly, children with ASD exhibited different pro les indicating less re ned anticipatory control and motor planning, as seen by movement overshooting, prolonged time to peak velocity/acceleration, and a higher frequency of movement units.Interestingly, children with ASD also demonstrated compensatory movement strategies, employing increased velocity and acceleration to offset their reaching di culties.More importantly, the implementation of a deep learning model yielded a noteworthy 78% accuracy in effectively classifying children with ASD from TD children, thereby substantiating the potential utility of movement kinematics and deep learning as an objective screening tool for ASD diagnosis, upon further validation in younger children.These ndings hold promise for improving early identi cation methods and underscore the signi cance of using objective measures to quantify motor di culties in children with ASD.
These nine parameters were computed for all trials of reaching and placing tasks performed by the children, resulting in a dataset of approximately 221 samples.Approximately 65% of the dataset represented children with ASD, with the remaining portion comprising typically developing children.Prior to training, each statistic was standardized across the entire dataset using min-max normalization.This normalized dataset served as the foundation for training a straightforward Multilayer Perceptron (MLP) model.The model consisted of four fully connected layers, incorporating three batch normalizations interspersed among the layers, as well as three leaky ReLU activation functions.A sigmoid function was employed for the output layer.The ADAM optimizer was utilized, with the initial learning rate set at 1e-5, subsequently adjusting to 1e-6 once the training accuracy reached 95%.Each model underwent 200 epochs of training, and all accuracy evaluations were conducted through ten-fold cross-validation.To determine the signi cance of each feature, permutation feature importance was computed[36][37][38].The model was fully trained for 200 epochs, and the lowest validation loss was recorded.Subsequently, each kinematic parameter was individually shu ed to introduce randomization, and the trained model was reevaluated for classifying the validation set.The difference between the initial best loss and the new loss resulting from the permutation of a feature represented its permutation feature importance.

4. 2
ASD-related Differences in Goal-directed Reaching MovementsCompared to TD children, children with ASD differed in their kinematics during the reach-place task as seen by the prolonged reaction times, increased time to peak velocity/acceleration, and higher frequency of movement units.Delayed executive functioning/motor planning in children with ASD could also contribute to the longer reaction times, as highlighted in studies byDemetriou et al. (2018) andWilloughby et al. (2020) [45, 46].Recent ndings indicate children with ASD have more di culties with anticipatory movement control compared to TD children.Challenges in utilizing visual information for anticipatory/feedforward control of movements are evident in children with ASD may have led to was involved in the data collection, kinematic data analyses, and manuscript writing.J.M. and W.H. were involved in the development of the deep learning model.A.B. supervised multiple aspects of the project including recruitment, screening, student training, data collections, kinematic data analysis, and writing and revisions of the manuscript.A.G. supervised the development of deep learning models, data analyses, and writing and revisions of the manuscript.

Figures
Figures

Figure 1 Experimental setups and examples for picture instruction Figure 2 Figure 3
Figure 1

Table 1
Demographic Information for Participated Children.

Table 2
Correlations Between the Movement Kinematics and Children's Adaptive Functioning [Insert Table2 here]