Quantitative Identification of ADHD Tendency in Children With Immersive Fingertip Force Control Tasks

Attention-deficit/hyperactivity disorder (ADHD) is a prevalent neurodevelopmental disorder that affects children. However, the traditional scale-based diagnosis methods rely more on subjective experiences, leading to a demand of objective biomarkers and quantified diagnostic methods. This study proposes a quantitative approach for identifying ADHD tendency based on fingertip pressing force control paradigm with immersive visual feedback. By extracting nine behavioral features from reaction time and dynamic force fluctuation features with high temporal and amplitude resolution, the proposed method can effectively capture the continuous changes in attention levels for ADHD diagnosis. The extracted features were analyzed using independent sample t-test and Pearson correlation to determine their association with ADHD-RS scale scores. Results showed that 12 statistical indicators were effective for distinguishing ADHD children from typically developed children, and several features of force control ability were also associated with core ADHD symptoms. A support vector machine (SVM) based classifier is trained for ADHD diagnosis and achieved an accuracy of 78.5%. This work provides an objective and quantitative approach for identifying ADHD tendency within a short testing time, and reveals the inherent correlation between the attention levels and the extracted features of reaction time and force fluctuation dynamics.


I. INTRODUCTION
A TTENTION Deficit/Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental disorder affecting children globally, with an estimated childhood prevalence of 5% [1] and 6.4% in China [2].Attention deficit, impulse and hyperactivity are three core symptoms of this disorder, with multidimensional behavior and cognitive function defect which is atypical for their age.The condition can cause learning disabilities, mood disorders, and social difficulties, creating numerous obstacles for affected children that negatively impact their academic performance, career prospects, and overall quality of life.It also poses a significant burden to families and society at large [3], [4].The impairment of higher cognitive functions in ADHD children may be related to changes in basic sensory levels [5].Studies have found that children with ADHD have impaired sensory processing functions in various fields such as vision and haptic [6], [7].
ADHD diagnoses in recent years have relied largely on subjective descriptions from parents and/or teachers via clinical interviews, questionnaires, and scales.This approach has been associated with variable diagnosis results, potentially leading to misdiagnosis or missed diagnoses due to differences in the diagnostic abilities of the evaluators and their subjective judgments.There is an urgent need to identify relevant biological markers and objective diagnostic methods that can provide reliable and valid information for early ADHD diagnosis and rehabilitation.
Non-invasive functional neuroimaging techniques, such as functional magnetic resonance imaging (fMRI) and positron emission tomography (PET), have shown promise for objective assessment of ADHD [8], [9].However, these techniques demand large scanning instruments, restrict body position and head movement, which poses limitations in confined spaces [10].Functional near-infrared spectroscopy (fNIRS) and electroencephalography (EEG) offer advantages such as portability, low cost, and relative motion artifact tolerance.Nonetheless, their low spatial resolution and limited recording depth restrict their applicability in clinical settings [11].Furthermore, since the selected brain regions in fNIRS could differ from the involved activated regions during performing the tasks by patients, the findings were inconsistent and the clinical guidance value was limited [12].
Some researchers used take performance data and machine learning-based methods to objectively evaluate ADHD tendency.Slobodin et al. [13] utilized a dataset of the performance from the continuous performance test (CPT) task, and used random forest and neural network methods for classification, achieving an accuracy of 87%.Yasumura et al. [14] employed a dataset from the Stroop task paradigm and collected NIRS data, before using SVM algorithm for classification and achieved an accuracy of 86.25%.Mikolas et al. [15] utilized medical case data including demographic variables, symptom ratings, attention scores from audio-visual tasks and IQ.They used SVM algorithm and achieved an accuracy of 68.1%.Maniruzzaman et al. [16] used data from a national children's health survey, and used eight machine learning methods, achieving a highest accuracy of 85.5%.Haptic paradigm can capture and quantify fine finger force control ability with high temporal and magnitude resolution, and it can thus be used to investigate the relationship between fine finger force control ability and the attentional state of the participant.The human haptic system is a promising avenue for evaluating attention due to the strong haptic sensitivity, precise motor and force control capabilities of the hands, and the selective attention ability of the haptic channel.Fine finger force data has large data volume due to its high temporal sampling rate, and several time and force-based features can be extract from fine finger force data, whereas machine learning-based algorithms are suitable to be applied due to its strong feature learning and representation ability.The haptic interaction task involves perceiving body mechanoreceptor information and actively exerting body motion and force.Crucially, it requires accurate force control, which is a unique aspect of haptic interaction [17].Researches on the default mode network in healthy adults have shown that sustained attention is directly involved in the process of haptic perception and force control, and may modulate the related behavioral performance [18], [19].Several studies have clarified the important relationship between attention, cognitive ability, and haptic performance.Wang et al. found that haptic tasks can temporarily enhance attention levels on typically developed adults [20].Puts et al. examined haptic-related task performance differences in children with ADHD, finding that they exhibited poorer results on simple and choice reaction time tasks, static detection threshold, simultaneous frequency discrimination, and temporal order judgment with carrier stimulus tasks.The authors speculated that lower attention levels partially account for such dysfunction in haptic tasks among children with ADHD [21].Additionally, researchers also found that children with ADHD have poor ability in maintaining force stability [22].Haptic processing disorder may thus be linked to core symptoms in ADHD children, leading to more interpersonal, emotional, and behavioral problems than experienced by typically developed children [23], [24], [25].
Considering the substantial involvement of human attention and cognition ability in the haptic task, we conducted haptic experiments to investigate the method of identifying ADHD tendency.By using electromechanical sensors with high accuracy and sampling rate to acquire force information, haptic tasks can capture fine changes in time-varying human attention state.This allows for obtaining pressing force data with high temporal and amplitude resolution, providing useful features for assessing attention and cognitive ability [26].In the meantime, the assistance of virtual reality (VR) technology helps to create a diverse and immersive environment with three-dimensional visual stimulation, enabling diagnostics to be performed in an environment similar to the real world with greater effectiveness [27].Several previous studies have demonstrated the effectiveness of VR environment in the assessment of attention in children with ADHD, including VR classroom [28], [29] and VR test [30].
Based on the advantages of the haptic tasks, this study proposes a visual-haptic experimental paradigm in VR environment for identifying ADHD tendency.Experiments with a larger number of participants is needed to improve the reliability for the result.Participants were required to keep a visual target within a specific range through fine regulation of fingertip force (Fig. 1).Nine behavioral features based on reaction time and force fluctuation performances were extracted from the behavioral experiment data collected from experiments of both ADHD and typically developed children.A comparative test was conducted to investigate the correlation between force control ability and ADHD core symptoms.An SVM-based classification model was formulated to diagnose ADHD.This study hypothesized that (1) ADHD children have abnormal performance in force control ability during haptic tasks compared to typically developed children; (2) Force control ability is correlated with ADHD core symptoms; (3) The proposed quantitative ADHD tendency identifying method based on haptic tasks can improve diagnostic accuracy.(4) Delay in information processing in the brain cortex is the primary reason for poor haptic task performance in ADHD children.
In conclusion, this study proposes a novel and objective approach for identifying ADHD tendency, with advantages of high temporal resolution, high force amplitude resolution of designed paradigm, and high sensitivity to the attention level and fine force regulation ability of children.From clinical perspectives, early detection and timely treatment of ADHD during childhood development are critical, underscoring the importance of accurate diagnosis.Firstly, our method attempts to mitigate subjectivity associated with current methods relying solely on subjective descriptions from parents and/or teachers through clinical interviews, questionnaires, and scales.Secondly, the proposed paradigm is effective to arouse the interest of children since it was designed as an interesting and interactive game.From an engineering perspective, the proposed approach aims to explore the feasibility of using pure fingertip force related experiments to assist clinical diagnosis, with potentially lower time cost and lower equipment cost compared to neuroimaging-based methods such as electroencephalography and near-infrared spectroscopy.Additionally, we proposed the SVM algorithm to analyze the force data, which is deemed suitable for the relatively small-scale dataset.We validated the hypothesis of quantifying attention fluctuation using fine force control with rapid regulation behavior.The relationship between pressing force data and internal attention state regulation was obtained through simple behavioral modeling, partially revealing the underlying mechanism for the behavioral differences in the haptic task.

A. Participants
This study involved 20 children with ADHD and 20 typically developed children who served as a control group.The Ethical Committee of Peking University Sixth Hospital approved this study, and all participants voluntarily agreed to take part after their parents or legal guardians signed informed consent forms.The experiment consisted of two groups: the ADHD group and the control group.A detailed calculation of sample size can be found in Note 2 in supplementary material.
(1)ADHD group: The inclusion criteria were as follows: ① Children diagnosed with ADHD by a psychiatrist at or above the attending physician level in the outpatient department of Peking University Sixth Hospital, based on the American Diagnostic and Statistical Manual of Mental Disorders, Fifth edition (DSM-V), and Kiddie Schedule for Affective Disorders and Schizophrenia for School-Age Children-Present and Lifetime Version (KSADS-PL);② Aged between 7 to 16; ③ Right-handed; ④ No previous medication for ADHD.The exclusion criteria were as follows: ① Patients suffering from schizophrenia, mood disorders, mental retardation, or other primary mental disorders other than oppositional defiant disorder and anxiety disorder based on KSDS-PL assessment; ② Wechsler children's intelligence test score < 80; ③ Suffering from nervous system disorders or other serious physical disorders.
(2) Control group: The inclusion criteria were as follows: ① Typically developed children recruited from Beijing who were screened by psychiatrists at or above the attending level, and did not meet the previous or current diagnosis of ADHD, as well as other schizophrenia, emotional disorders, mental retardation, or other primary mental disorders; ② Aged from 7 to 16; ③ Right-handed.We excluded children with neurological disorders or other serious physical illnesses.

B. Measure
The KSADS-PL is a semi-diagnostic tool that assesses current and past mental disorders in children and adolescents according to DSM-V criteria.KSADS-PL classified ADHD into three clinical subtypes, namely attention deficit dominant (ADHD-I), hyperactive impulsivity dominant (ADHD-HI) and mixed (ADHD-C).In addition, the tool can diagnose serious comorbidities such as childhood affective disorder and psychotic disorder, as well as behavioral problems such as oppositional defiant disorder, conduct disorder, and tic disorder.With an interrater agreement of 93-100%, test-retest reliability κ coefficients ranging from 0.63 to 1.00, KSADS-PL is considered highly reliable.
The ADHD rating scale (ADHD-RS) consists of 18 items and uses a four-point scoring method (1)-4 points) based on the DSM-IV diagnostic criteria for ADHD.Among these items, the first 9 items assess attention deficit symptoms, while the last 9 items assess hyperactivity/impulsivity symptoms.Total scores can be calculated for overall ADHD symptoms (DQ_TO), as well as attention deficit (DQ_IA) and hyperactive impulse (DQ_HI) dimensions to reflect symptom severity.Higher scores indicate greater symptom severity in the corresponding dimension.
To minimize the impact of low IQ on cognitive function, all participants with ADHD in this study were required to have an IQ higher than 80. Chinese Wechsler Intelligence Scale for Children, C-WISC: Wechsler Intelligence Scale for Children (Version 4), introduced and revised by Houchan Zhang, is an important tool for evaluating children's intelligence levels.The fourth edition of Intelligence test can be used to obtain the fullscale IQ of an individual according to the norm, including four dimensions of verbal comprehension, perceptual reasoning, working memory and processing speed in specific subtests.
The experimental paradigm is designed as a virtual isolated assessment chamber using Unity 3D (Unity Technologies, UK) and helmet-mounted display (HMD, Oculus Inc. USA) to provide visual information.During experiments, subjects wear HMD and start with two circular rings symmetrically displaced in the middle of the frame.Then one cylinder appears in the VR frames above the left or right circular ring randomly, and the subject presses the physical button at the corresponding side on the table with their finger while actively controlling the exerted force to keep within the required range.This range is represented in the VR frame as the thickness of the appeared cylinder.The magnitude of the varying finger force is indicated in the VR frame as the height of a thin disk concentric with the cylinder, allowing the subject actively adjust the exerted finger force based on the visual feedback of the relative vertical position between the disk and the static cylinder as shown in Fig. 2. A detailed explanation of the experiment procedure can be found in the supplementary video.
In this experiment, each subject is required to complete 3 sessions.The first two sessions follow the same procedure, but differ in difficulty level.The purpose of these sessions is to compare the force control abilities of two groups of subjects and to assess the impact of experimental difficulty on ADHD tendency identifying results.In session one and two, the target cylinder remains visually for a maximum of 1.7s and 2.5s, respectively.If the subject maintains finger forces within the allowed range for a continuous 200ms, the target cylinder will disappear immediately at the end of this 200ms period, and the trial is considered a success.Conversely, if the subject fails to maintain finger forces within the allowed range for 200ms during the 1.7s or 2.5s interval, the target cylinder will vanish automatically at the end of that interval, and the trial is considered a failure.If the subject maintains finger forces within the allowed range for less than 200ms before exceeding the range, they must sustain it for an additional continuous 200ms to achieve success.Upon completion, the subject will receive audio and visual feedback, the cylinder will disappear, and the finger will be lifted to prepare for the next trial.Different from the first two sessions, the third session focuses on the fine force fluctuation during force regulation.Therefore, the trial time is fixed at 2s and cannot be completed in advance, and no incentive visual or sound feedback is provided for success or failure.Each session consists of 100 trials and takes 5 to 8 minutes to complete, with a 2-minute rest period between each session.A detailed explanation of the experiment time can be found in Note 1 in the supplementary material.The force sensors (Honeywell FSG15N1A, Honeywell International, USA) record data with a sampling rate of 20Hz and they are assembled and encapsulated in a metal housing with only keys connected to the sensors stretching out to be pressed [26].Only two keys (pressed by the left and right index fingers respectively) are used in the experiments.The difficulty was different between the three sessions, and it remained the same within each session.The difficulty of the trials is related to the experiment parameters, including target pressing force F T , target pressing force tolerance F TOL , time tolerance required to force T TOL , and maximum time duration for one trial T TRIAL , while the former three parameters remained the same during experiment, and only the maximum allowed time duration for one trial was different between the three sessions.A detailed parameters introduction is provided in Table S1.Besides, the difficulty was not adjusted between participants due to the consideration of fairness.The experimental paradigm designed in our manuscript has no relationship with handedness.The orders for the right and left extremities were assigned randomly.Both hands of the subjects need to press during experiment, and each trial randomly assign one hand by giving a visual cue of a target cylinder appearing on the left or right in the VR environment.Therefore, the involvement of both hands in the task is symmetrical, and there is no deliberate arrangement of handedness preference in the task allocation.Fig. 2 illustrates the experiment paradigm in one trial.

C. Statistical Analysis
The evaluation ability of the extracted reaction time and force fluctuation-based features was assessed through statistical analyses.An independent sample t-test was conducted for each feature and the clinical scale results.The difference was considered statistically significant when the P value was less than 0.05.Pearson correlation analysis was performed between the extracted features and the clinical scale results to evaluate the effectiveness of the features in the identification of ADHD symptoms tendency.

A. Demographic Data of Participants
In this study, twenty children diagnosed with ADHD (4 females, aged 7 to 10 years old, mean ± SD: 9.26±1.116years old) and twenty typically developed children (7 females, aged 7 to 10 years old, 9.18±0.816years old) were recruited.Two boys with ADHD had tic disorder, while one boy with ADHD had anxiety disorder.The groups did not differ significantly with regard to age or gender (p>0.05).However, there were significant differences in ADHD-RS scale scores (DQ-IA, DQ-HI, DQ-TO) between the two groups (p < 0.01).Detailed information is provided in Table I.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

B. Behavioral Features Extracted in Experiments and Correlation Analysis
During the three experimental sessions, finger force data from both groups are collected.Four force fluctuation-based features were extracted through directly analyzing the force curve in each trial, along with four reaction time-based features that represent different stages of the process in one trial (Fig. 3, Fig. 4).
Each trial was divided into three stages: the first covers the period from trial onset to when the participant presses the key reaching the threshold, after that the second stage starts and lasts until the participant reach the required force range and keep the force within it.The third stage spans the remainder of the trial.For the third session, a feature calculating the percentage of time that the participant manages to keep the force within the required range was extracted.Regarding force fluctuation-based features, the peak force and overshooting value beyond the target force were extracted in each trial, but only positive overshooting values were included.The force fluctuation range in the third time stage, represented by the difference value between the maximum force and the minimum force exerted in the third stage, was extracted for the first two session.For the third session, the force fluctuation range during the last one second in the trial was extracted.Besides, the mean success rate was also calculated for the first two session.A detailed explanation of the experiment parameters and behavior features can be found in Table S1-3.These features were used to compare force modulate ability between the two groups using statistical analysis, including mean value (AVG), standard deviation (STD), median value, and coefficient of variation (COV), which is calculated by dividing the standard deviation with the mean value using the following equation: where σ stands for standard deviation and µ stands for average value.Fig. 4 presents a comprehensive list of the nine behavior features, along with their respective statistical indicators and the level of significance differences between the two groups.

C. Correlation Analysis Between ADHD-RS Scores and Statistical Indicators
Fig. 5 and Fig. 6 demonstrate the statistical indicators calculated based on the behavioral features of both groups that correspond to ADHD-RS scores.
1) ADHD Group: Fig. 5 displays the statistical indicators of the ADHD group.In session one, the standard deviation of reaction time positively correlated with DQ-HI (r=0.566,p<0.05) and DQ-TO (r=0.456,p<0.05).Conversely, no other statistical indicators showed any correlation with ADHD-RS scores in session one.In session two, the average success rate was negatively correlated with DQ-IA (r=−0.497,p<0.05),DQ-HI (r=−0.535,p<0.05), and DQ-TO (r=−0.592,p<0.01).Additionally, the standard deviation of reaction time negatively correlated with DQ-HI (r=0.523,p <0.05) and DQ-TO (r=0.514,p<0.05).Reaction time in stage two also showed a positive correlation with DQ-HI (r=0.447,p<0.05).No force fluctuation-based statistical indicators showed any significant correlation with ADHD-RS scores in session two.In the third session, the force fluctuation value during the last one-second interval positively correlated with DQ-HI (r=0.476,p<0.05) and DQ-TO (r=0.489,p<0.05), while the average value of time duration of force within the required range negatively correlated with DQ-HI (r=−0.450,p<0.05).
2) Typically Developed Group: Fig. 6 illustrates the statistical indicators of the control group.In session one, the standard deviation of reaction time positively correlated with DQ-IA (r=0.476,p<0.05) and DQ-TO (r=0.506,p<0.05).The coefficient of variation of reaction time also positively correlated with DQ-TO (r=0.465,p<0.05).Fisher's z transformations showed no statistically significant difference in the correlation coefficients between the ADHD group and the control group for DQ-TO scores with the coefficient of variation of reaction time in the first session (z=−0.1898,p=0.8495).

D. Classification Based on SVM
1) SVM Model: Twelve statistical indicators exhibiting significant differences (p<0.001) between the two groups were chosen as input features to develop a data-driven classification algorithm.Due to limited sample size and feature dimensions, we utilized an SVM-based classification algorithm.The formula for the algorithm is as follows: where w is the parameter matrix for the hyperplane, x is the feature value, and b is the bias.

2) Model Training and Performance:
The values of each indicator were normalized using the Z-score normalization method to ensure identical range.The training and test sets were randomly divided into ten folds using cross-validation, with each data sample used as a test data to ensure result validity.The cross-validation process was conducted ten times, and the final results were obtained by averaging them.We used linear kernel function to obtain the support vector.The linear kernel   the optimal parameter for the support vector was obtained after training.The accuracy was calculated based on this support vector.The total data set included 20 ADHD children and 20 typically developed children.The training set and test set was partitioned based on the ten-fold cross validation method, which employed 90% of the total data set as training set each time and trains ten times to make sure each data sample have equal chance.Therefore, the magnitude of the training set for each training process was 36.With an accuracy of 78.5% and an F1 score of 0.844, the confusion matrix was computed and visualized in Fig. 7.
3) Visualization: The classification result is illustrated in a two-dimensional feature space after dimensionality reduction using Principal Component Analysis (PCA) algorithm (Fig. 8).The formula of PCA algorithm is the following: Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.where P is the eigenvector of covariance matrix of the feature.matrix X.The two most-significant feature dimensions are extracted to form this two-dimensional feature space, and the separating line is calculated as the intersecting line between this space and the separating hyperplane.

IV. DISCUSSION
This study develops a VR-based visual-haptic system for attention level evaluation with high temporal resolution and amplitude resolution.We collected fingertip pressing force data in an immersive virtual environment and analyzed it using SVM algorithm.Only children with ADHD who did not take any medication were included in our experiments, and we designed strict inclusion and exclusion criteria to avoid interferences from medication or other disorders.
This study defined and extracted various behavioral features based on the reaction time and force fluctuation during the experiment.Twelve statistical indicators calculated based on the value of the behavioral features had significant differences (P <0.001) between the groups (Fig. 4).Children with ADHD performed worse in force control than typically developed children.The VR-based visual-haptic force control task has great potential in the quantitative identifying of ADHD tendency, as the extracted features based on force control ability may serve as a biological marker.The finger force regulation ability has benefits from the high sampling rate and resolution of the force sensors, as well as the neural mechanism of the fine finger force regulation process.The success of each trial requires both a high degree of concentration over time and precision and accuracy of muscle force regulation.Attention level plays a crucial role in monitoring force errors from visual feedback, while muscle force regulation is vital for outputting fine finger force within the required range.
Clinical scales are commonly used in ADHD diagnoses.To evaluate the correlation between the force control ability in visual-haptic paradigm and the ADHD-RS scale results, this study conducted several correlation analyses.The results in Fig. 5 and Fig. 6 demonstrate that several behavioral features correlate with ADHD scale scores.Considering that these scale scores can partially represent the core symptoms of ADHD, including inattention and impulsive hyperactivity, these behavioral features can serve as objective biomarkers for ADHD assessment in children.However, some other features were found to be not statistically significant correlated with ADHD scale scores.In future researches, the experiment paradigm should be improved to enrich the task types and feature categories.Besides, the preprocessing algorithm should be optimized to better remove invalid trials.In conclusion, our analysis reveals a notable association between force control performance features and the core symptoms of ADHD, with a particular emphasis on the average reaction time and force fluctuation.These findings suggest that the extracted features within the visual-haptic paradigm may serve as sensitive indicators for assessing the level of attention deficit in children with ADHD.By identifying these relevant statistical indicators, we aim to highlight the potential differences between children with ADHD and typically developed children, thereby enhancing the reliability of the SVM-based classification model.
Based on correlation analysis results between statistical indicators and scale scores, we discuss physiological reasons as follows.At the physiological level, the force control process hypothesis involves three steps in the first stage of the trial time: visual perception, processing in visual cortex, and decision making in the central neural system.In the second stage of the trial time, the motor cortex and muscle execution process are involved.As demonstrated in Fig. 1(c), the brain processes both visual input data and haptic input and output data.The decision-making process is closely related to the attention network of human brain.Studies have shown that in children with ADHD, the activation level of this network is lower compared to that of typically developing children [31].This network involves the bilateral ventral attention system, including the right inferior prefrontal cortex and basal ganglia, as well as the fronto-temporo-parietal cognitive control network.The reduced activation of these regions in ADHD children contributes to delays in the decision-making process, which could further lead to a longed reaction time in the experiment.Regarding force generation and regulation, prior research has emphasized the significance of the basal Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
ganglia and prefrontal cortex [32].Children with ADHD often exhibit poorer performance in tasks requiring force generation, possibly due to alterations in the functioning of these brain regions.The cerebellum also performs a vital function in controlling movements, with cerebellar vermis believed to affect the downstream control system involved in regulating body posture and movement [33].The volume of cerebellar vermis in children with ADHD is significantly reduced [34].Therefore, these may be the basis for the poor performance of children with ADHD on attention-demanding force control tasks, providing an effective framework for understanding the mechanism of force control ability on ADHD neuropathology.Delay in the decision-making process can be reflected in the value of four reaction time-based features, and force control ability deterioration in the value of four force fluctuation-based features.In addition to the average value of these features, we also calculated their standard deviation and coefficient of variation to represent the influence of individual differences in each group, providing a data input with higher discrimination for the SVM-based classification model.Although attention deficits or hyperactive/impulsive behavior are not distinct for ADHD and are also involved in autism spectrum disorder (ASD) or cerebral palsy (CP), these impairments are not their typical features.ASD is a neurodevelopmental disorder that features social communication disorders, narrow interests or activities, and repetitive stereotyped behaviors.CP is one of the primary disabling disorders during infant and child development, which manifests mainly as central motor disorders and abnormal posture, often accompanied by various impairments such as intellectual, language, visual, and auditory disorders.Both ASD and CP involve attention deficits or hyperactive/impulsive behavior to varying degrees, but these impairments are not their typical features.This study excluded participants with neurological diseases or ASD; therefore, our statistical indicators cannot be used to predict ASD or CP.This study has following limitations.Firstly, more neurobehavioral measurements, such as eye tracking or electroencephalogram, remains to be combined with the behavioral features extracted in this study to further improve the classification accuracy.Secondly, future studies need to optimize the machine learning-based ADHD tendency identifying model by increasing sample sizes and incorporating validation groups.Thirdly, future research should broaden the dataset by incorporating children below 7 years old or above 10 years old.Children under the age of 7 tend to exhibit lower cognitive ability and experimental cooperation, thus limiting the applicability of the proposed method among this age group.To improve the applicability for identifying ADHD in children over the age of 10, the sample size may be expanded for the corresponding age group.Finally, the physiological correlation between the brain function and haptic perception and control process remains to be further explored.

V. CONCLUSION
This study explored the feasibility of using VR-based visual-haptic system to diagnose ADHD in children.We found significant differences in extracted behavior features between groups during the force control task, including differences in features and hyperactive and impulsive symptoms of ADHD-RS scale results, with fewer differences found regarding inattention symptoms.The SVM-based classification model achieved the accuracy of 78.5% in distinguishing ADHD children from typically developed children.In conclusion, the VR-based visual-haptic system has potential in the clinical application of ADHD tendency assessment.While this study represents a preliminary exploration of force control ability as a biomarker in identifying ADHD tendency, future studies must incorporate more neurobehavioral measures, such as eye tracking or electroencephalography, to continue developing the system in clinical settings.

Fig. 1 .
Fig. 1.Brain function in haptic perception and control process.(a) The force output is demonstrated in blue curve while the visual perception is demonstrated in red straight line.(b) The amplitude and temporal resolution can reach 0.05N and 20ms respectively.(c) The brain integrates the input information and controls the output force.

Fig. 2 .
Fig. 2. Experimental paradigm.(a) The critical hardware and software arrangements are highlighted in yellow frames.(b) The subject presses one key during one trial.(c-e) The VR frames shown in one successful trial.

Fig. 3 .
Fig. 3. Force curve in one trial and typical features extracted.The time in one trial is divided into three stages, and the real-time fingertip force is compared with the required force range.

Fig. 4 .
Fig. 4. Boxplot of statistical indicators of all participant.(a) Four reaction time-based features: whole reaction time, reaction time in stage one, reaction time in stage two, time duration for force in required range.(b) Task-based feature: succeed within required time.(c) Four force fluctuationbased features: peak value, overshooting value, fluctuation range in stage three, fluctuation during last one second; no statistical indicator of fluctuation range in stage three was found to have significant difference between ADHD and control group.'S1': Employed in session 1; 'S2': Employed in session 2; 'S3': Employed in session 3; * p<0.05; * * p<0.001.

Fig. 8 .
Fig. 8. Classification result in two-dimensional feature space.The green line represents the hyperplane projected to this two-dimensional feature plane.CG: control group.