Associations Between Coordination and Wearable Sensor Variables Vary by Recording Context but Not Assessment Type

Abstract

Motor coordination is an important driver of development and improved coordination assessments could facilitate better screening, diagnosis, and intervention for children at risk of developmental disorders. Wearable sensors could provide data that enhance the characterization of coordination and the clinical utility of that data may vary depending on how sensor variables from different recording contexts relate to coordination. We used wearable sensors at the wrists to capture upper-limb movement in 85 children aged 6–12. Sensor variables were extracted from two recording contexts. Structured recordings occurred in the lab during a unilateral throwing task. Unstructured recordings occurred during free-living activity. The objective was to determine the influence of recording context (unstructured versus structured) and assessment type (direct vs. indirect) on the association between sensor variables and coordination. The greatest associations were between six sensor variables from the structured context and the direct measure of coordination. Worse coordination scores were associated with upper-limb movements that had higher peak magnitudes, greater variance, and less smoothness. The associations were consistent across both arms, even though the structured task was unilateral. This finding suggests that wearable sensors could be paired with a simple, structured task to yield clinically informative variables that relate to motor coordination.

Keywords:

• accelerometry
• wearable sensors
• motor coordination
• developmental coordination disorder

INTRODUCTION

Childhood development is driven by movement (Bhat et al., 2011; McQuillan et al., 2021; Sacrey et al., 2014; Soska et al., 2010). Every day, movement creates interactions that are the foundation for social, cognitive, and physical development (Adolph & Hoch, 2019; Cheung et al., 2021; Kuzik et al., 2020; Soska et al., 2010; Wilson et al., 2018). Impaired motor coordination leads to missed interaction opportunities, which may alter developmental trajectories (Iverson, 2021; Salaj & Masnjak, 2022). There is accumulating evidence that impairments in motor coordination relate to an array of developmental outcomes such as educational achievement, reading skills, social development, and emotional regulation (Lee et al., 2020; McQuillan et al., 2021; Sutera et al., 2007; Teitelbaum et al., 1998). Poor motor coordination has a profound impact on quality of life (Caçola & Killian, 2018; Zwicker et al., 2013) and often goes unrecognized and under-treated (Harrowell et al., 2018; Ketcheson et al., 2021). Broadly, motor coordination refers to controlling the kinematic parameters of body segments during goal-directed movements. Motor coordination is intimately related to motor competence which refers to proficiency in all goal-directed movements that require coordination and control of the body (Cattuzzo et al., 2016). Coordinated movement patterns emerge from the constraints imposed on the mover. Those constraints may be specific to the environment, the task, or the individual (Newell, 1986). Motor deficits are more reproducible and reliably quantifiable than other impairments, such as social-communication deficits, which makes coordination an ideal clinical diagnostic for neurodevelopmental disorders (Mostofsky et al., 2009). Improved methods of quantifying coordination are needed to facilitate improvements in screening, early diagnosis, and motor skill interventions.

Coordination assessments can be broadly divided into two categories, direct and indirect. In a direct assessment, a person or technology assigns a score to the child’s performance. Technology-based direct assessments can include video recording, motion capture, and accelerometry. Direct measures entail the application of an objective scoring standard across all participants, with some subjective elements present when rated by a person. The gold standard for assessing coordination in developmental populations is a direct, standardized assessment administered by a qualified professional (Staples & Reid, 2010; Wilson et al., 2018). While these standardized assessments provided valuable data, they are also time-consuming, require expert clinicians, and test administration most often occurs in a clinical or research environment which may not elicit the child’s natural motor behaviors and therefore may not reflect true functional ability (Hoyt et al., 2020). Indirect assessments are based on self-, teacher-, or caregiver-ratings and do not require training. They are cost-effective and easy to use but accuracy varies (Rice et al., 2013). Indirect assessments often have low to moderate correlation with direct measures of the same construct. For example, in a sample of 107 children, caregiver-reported physical activity correlated with direct accelerometer measurements at r = 0.30 (Rice et al., 2013). Other work has found a correlation of r = 0.23 between a caregiver-completed motor skills questionnaire and the skill-based Zurich Neuromotor Assessment (Zysset et al., 2018), and a correlation of r = 0.20 between a parent-reported hyperactivity score and a direct video-based assessment (Wehrmann & Müller, 2015).

Wearable sensors could be an alternative method of directly quantifying motor coordination. They allow for continuous, passive monitoring of the participants’ motion in the natural environment. Wearable sensors have been used in developmental research to identify delays (Abrishami et al., 2019; Heinze et al., 2010), monitor sleep (Benson et al., 2019), and characterize physical activity levels (Pate et al., 2006; Wood et al., 2009). The sensors are small, lightweight, and no more obtrusive than a wristwatch. Bilateral wearable sensors worn at the wrists are a reliable and valid way to capture features of real-world upper-limb use (Bailey et al., 2014; Hoyt et al., 2019; Lang et al., 2013; Urbin et al., 2015). Accelerometers within the sensors provide time-series data from which sensor variables can be calculated. Explorations into the use of wearable sensors have used various clinical populations and have had various goals (Bailey et al., 2015; Benson et al., 2019; Heathers et al., 2019; Patros et al., 2017; Wanigatunga et al., 2019), including some work that has sought to replace clinical assessments with wearable sensor metrics (Ernesto et al., 2022; Lambercy et al., 2016). The goal of this investigation, however, was to explore the capability of sensors to capture motor behavior in development in order to augment, not replace, existing clinical assessment tools. Previous work in this area is very limited. One study examined a small sample of infants using sensor variables from the lower extremity to examine neurodevelopmental risk (Wilson et al., 2021). Other work has established that sensor variables can detect asymmetrical motor impairments across childhood. This was done by comparing key sensor variables across a group with cerebral palsy and a reference group (Hoyt et al., 2019). This team did find correlations between the sensor variables and a clinical, video-based assessment of upper-limb function (Hoyt et al., 2020) but they did not measure motor coordination globally. As far as we know, no other studies have explored the use of wearable sensors to quantify overall coordination level in children.

In a recent pilot study, we recruited 48 children (22 with ASD) to explore how indirect measures of coordination, hyperactivity, and autistic symptoms related to sensor variables (Konrad et al., 2022). We found that the indirect coordination measure correlated modestly with select sensor variables. The sensor variables were obtained from unstructured recording time, when the participant was in their free-living, natural environment and their specific activities were uncontrolled and unknown. Sensor variables can also be obtained from structured recording time, when the participant performs a known activity in a controlled environment (e.g. clinic, research lab). In structured recordings, the environment, equipment, and task parameters are common to all participants but moment-to-moment individual movements are not constrained. Our pilot study examined only the associations between an indirect measure of coordination and sensor variables from unstructured recordings; however, structured contexts may create different relations between coordination scores and sensor variables compared to the general, unconstrained movements of an unstructured recording. Therefore, it remains an important, unanswered question as to how the associations would differ if the sensor variables were obtained from a structured context and/or if a direct measure of motor coordination was used, since achieving the strongest relationship possible between sensor variables and coordination score is important to gauge the capability of wearable sensors to capture this construct. The current study builds upon the pilot by examining sensor variables from structured and unstructured recordings and adding a direct, criterion-based measure of motor coordination. We know of no previous work that examines the differential relations between wearable sensor variables and motor competence across structured and unstructured recording contexts. Differing relations would have ramifications for further research and the clinical deployment of wearable sensors.

The purpose of the current study was to examine the ability of wearable sensors to index motor coordination in a cohort of children aged 6–12. We had two objectives. The first objective was to determine the influence of recording context (unstructured vs. structured) on the association between sensor variables and coordination. Our first hypothesis predicted greater magnitudes (more negative or more positive) in associations between sensor variables and coordination score in the unstructured context compared to the structured context because we expected the natural, unconstrained movements and longer recording duration would provide more data and yield sensor variables that better quantify individual variation. The second objective was to determine the influence of assessment type (direct vs. indirect) on the association between sensor variables and coordination. Our second hypothesis predicted greater magnitudes (more negative or more positive) in associations between sensor variables and coordination when coordination was assessed with a direct measure compared to the indirect caregiver-report which is subject to bias and lacks an objective frame of reference (Wolraich et al., 2004). We expected greater magnitudes of associations between the sensor variables and the direct coordination assessment because both are objective scores based on common criteria. Knowledge about how recording context and coordination assessment type influence the overall ability of sensor variables to index coordination has important implications for the future use of wearable sensors as clinical assessment tools.

METHODS

Participants

We recruited 6–12-year-old children through targeted social media advertising, word-of-mouth, and from a community occupational therapy clinic. This age range was chosen because: (1) motor coordination is expected to progress as more advanced, adult-like motor control, such as motor imagery (Smits-Engelsman & Wilson, 2013) and feedforward control (Wilson & Hyde, 2013), emerges during this period, (2) any major motor or non-motor developmental deficits are likely known by this time, and (3) we expected that participants 6 years old and older could follow the directions associated with a motor learning throwing task and tolerate a ∼2 hour lab visit. This study aimed to recruit participants across the continuum of atypical and typical development to facilitate a dimensional approach rather than a categorical, case-control approach (see Analysis). Participants could be typically developing or have diagnoses of Developmental Coordination Disorder (DCD), as well as Autism Spectrum Disorder (ASD) or Attention-Deficit/Hyperactivity Disorder (ADHD). These developmental diagnoses were included because of their high incidence and close association with impaired coordination (Ketcheson et al., 2021; Mokobane et al., 2019; Murray, 2010). Participants whose primary language was not English or those with other activity limiting conditions or injuries were excluded. The participant’s medical history was based on caregiver-report. One parent or legal guardian signed an informed consent form prior to their child’s participation and participants provided their verbal assent. Families received $50 per participant and participants were allowed to choose a small toy at the conclusion of the lab visit. The research protocol was approved by the Washington University Human Research Protection Office (IRB #:202201076), and conformed to the principles of the Declaration of Helsinki.

Study Protocol

Participants and a caregiver came to the lab for a single study visit. The study procedure is depicted in Figure 1. On arrival, two wearable sensors were applied to the participant’s wrists just proximal to the ulnar styloids. The caregiver and participant were educated on the wearing protocol. The participant wore the sensors during the lab visit and continued to wear them for the next two days outside of the lab. Our recent pilot study using the same sensor protocol found that the variables extracted were stable across two recorded days (Konrad et al., 2022). Based on that result, we expected two recording days would provide adequate data while also minimizing participant burden. We encouraged continuous wearing during waking hours, but parents and participants were instructed to remove the sensors if the participant was swimming or experiencing skin irritation. Wearing the sensors while sleeping overnight was not required. We used either Actigraph GT3X-BT or GT9X Link Activity Monitors (Actigraph, Pensacola, FL). Both these devices resemble a wristwatch and have a 3-axis, solid-state accelerometer with a dynamic range of ±8 gravitational units (g). Data were recorded at 30 Hz. After the 2-day wearing period, the sensors were returned to the lab with shipping materials provided to the caregiver or picked up at the participant’s home. During the lab visit, participants completed a motor learning task associated with another study. For this task, the participant performed approximately 200 underhand bean bag throws to a target 2 m away with the dominant limb only. The throws were completed in blocks of 25 and short breaks (1–2 minutes) were given between blocks. The beanbags were either handed to the participant or placed on a chair immediately next to the participant. There was a short delay between each throw while the previous throw was scored and the target was reset for the next throw. The throwing task took approximately 45 minutes with the accelerometers continuously recording for the whole duration. The recording began just before the first throw and ended immediately after the last throw. This means that the structured recording consisted of the bean bag throws, the inter-throw periods, and the breaks between blocks, that is, all upper-limb movements made during these periods were recorded. The other laboratory assessments were not included in the structured recording.

Figure 1. Study protocol. DCDQ = developmental coordination disorder questionnaire. Conners = conners-3 parent report. SRS = social responsiveness survey. KBIT-2 = Kaufman Brief Intelligence Test. MABC = movement assessment battery for children. Note that the Bean bag throw and MABC-2/KBIT-2 were reversed for some participants when multiple participants were being tested at once and alternated through the tasks.

In addition to assessing motor coordination, we assessed hyperactivity, autistic traits, and intelligence in order to characterize the sample. The questionnaires were delivered to caregivers using REDCap electronic data capture tools (Harris et al., 2019). Direct assessments were administered during the study visit, either preceding or following the throwing task (Figure 1).

Assessments

Motor Coordination

Caregivers completed the Developmental Coordination Disorder Questionnaire (DCDQ) (Cairney et al., 2008). The DCDQ is widely used in clinical care and research and was the best indirect assessment of motor coordination and competence available for our purpose and age range (Cancer et al., 2020). Scores on the DCDQ range from 0 to 75, with higher scores indicating greater motor coordination. The raw total scores were used in our analyses, as there are no standardized scores. The DCDQ has a Cronbach’s alpha of 0.94 and correlates with the Movement Assessment Battery for Children at r = 0.55 (Wilson et al., 2009). To provide a direct measure of motor coordination and competence, participants were assessed with the Movement Assessment Battery for Children (MABC-2). The MABC-2 has eight items in three domains: manual dexterity, aiming and catching, and balance. The MABC-2 has demonstrated good reliability and validity. It has a Cronbach’s alpha of 0.90 and correlates with the Bruininks–Oseretsky Test of Motor Proficiency (BOT-2) at ρ = 0.90 (Wuang et al., 2012). The total standard score (mean = 10, sd = 3) was used in our analysis. Higher standard scores indicate greater motor coordination.

Hyperactivity

We used a subtest from The Conners-3 Parents’ Rating Scale (Cabral et al., 2020), the hyperactivity/impulsivity content scale raw score, to quantify caregiver-reported levels of hyperactivity. The Conners-3 is a commonly used behavioral rating scale for identifying ADHD in children and adolescents and includes items relating to cognitive, behavioral, and emotional symptoms. The hyperactivity/impulsivity content scale score range from 0 to 42 with mean scores ranging from 6 to 9 and scores greater than ∼20 (depending on age and sex) indicating hyperactivity. This content scale has a test–retest reliability of 0.77–0.83 (Thorell et al., 2018).

Autistic Traits

We assessed autistic traits with the Social Responsiveness Scale (SRS-2) School Age, a metric of quantitative autistic traits indexed as deficits in reciprocal social behavior, which is disrupted in ASD. The SRS indexes quantitative autistic traits across the entire population, including individuals with and without ASD (Constantino & Todd, 2000), and differentiates individuals with ASD at a level of 3 standard deviations above the mean. The SRS-2 has a Cronbach’s alpha of ≥0.92 at all ages (Constantino & Gruber, 2012). We extracted the SRS T-score for the analysis. T-scores ≤59 are indicative of typical development while T-scores >59 are consistent with ASD.

Intelligence

Intelligence was assessed with the Kaufman Brief Intelligence Test (KBIT-2) (Carlozzi, 2011). The KBIT-2 is easy to administer and does not require any reading. It provides verbal, non-verbal, and composite intelligence scores. We used the IQ composite score (mean = 100, sd = 10). Scores below 75 are indicative of intellectual disability. The KBIT-2 has a test–retest correlation of r = 0.90 and correlates with the Wechsler Intelligence Scale for Children-4th edition at r = 0.77 (Kaufman & Kaufman, 2004).

Accelerometry Data

We downloaded accelerometry data from the wearable devices using ActiLife 6 software (Actigraph, Pensacola, FL). Data were plotted and visually inspected to confirm that the sensors were worn for the specified wearing period. Both the original 30 Hz data and a copy of the down-sampled 1 Hz data were extracted. Most sensor variables were calculated using a time resolution of 1 s. The frequency-based complexity measure requires greater time resolution to capture changes in limb acceleration and the 30 Hz data were used. As part of Actigraph’s proprietary process, a 7th order Infinite Impulse Response filter was used to band-pass filter 1 Hz data between 0.25 and 2.5 Hz in order to remove constant linear accelerations, such as gravity and non-human accelerations due to external motion (such as riding in a car or elevator). For the 30 Hz variables, data were band-pass filtered from 0.2 to 12 Hz, also to remove constant linear acceleration components. The periods during which the participant did not wear the sensors (i.e. swimming) or slept were manually trimmed. These periods were determined using the caregiver-completed log and confirmed by visual inspection of the data. The accelerations from each axis were combined into a single vector magnitude value using the equation √(x² + y² + z²). Using that vector magnitude, the sensor variables were calculated for each of the three separate time periods (Lab visit throwing activity, Day 1, Day 2). The sensor recordings from the throwing activity provided the structured data for this analysis, and the recordings from after the lab visit until the end of the two days provided the unstructured data.

Wearable sensors generate large volumes of data and a large range of candidate variables can be calculated from the data (Barth et al., 2021; Konrad et al., 2022; Lang et al., 2020). Sensor variables from bilateral wrist accelerometry can be categorized according to four characteristics of human movement: (a) duration, (b) symmetry, (c) intensity, and (d) complexity (Barth et al., 2021; Konrad et al., 2022). Duration variables quantify the amount of time that one or both limbs were moving. Symmetry variables quantify the relative contribution of the two limbs during movement. Intensity variables quantify the magnitude of the movement accelerations for one or both limbs and are reported in activity counts where faster and/or more frequent movements of the upper limb produce higher values. One Actigraph activity count is equivalent to 0.001664 gravitational units (g). Complexity variables form a heterogeneous category that includes the variance of the magnitude, time-series variability, smoothness, and descriptive characteristics (weighted mean and SD) of the frequency spectrum of the vector magnitude time series. Our selection of meaningful variables from those available was guided by theory and the results of the pilot study. All of the duration variables were excluded as these do not appear to index any aspects of coordination and showed trivial to small, and non-significant, associations in the pilot. Likewise, all symmetry variables were excluded as this population was not expected to have asymmetrical motor impairments. Those sensor variables that significantly associated with the DCDQ score in the pilot were included. Finally, we included variables that were not calculated at the time of the pilot study. See Table 1 for the final set of 11 intensity and complexity variables. Intensity and complexity variables may be able to capture the variation among the participants in the sequence, timing, and scaling of muscle force that is required for coordinated movement patterns. That variation may then match with individual variation in coordination score. Our limb specific sensor variables were calculated as dominant (D) and non-dominant (ND) rather than right and left, since children in this age range have established hand dominance. Hand dominance was assessed by asking the participant and confirming with a writing sample.

Table 1. Accelerometer sensor variables.

Movement characteristic	Variable	UL	Hz	Definition
Intensity	Peak magnitude	ND & D	1	The maximum activity count that occurred during the recording period. Peak magnitude may scale to the demands of the movement goal where extremes of the distribution indicate neuromuscular control deficits (Law et al., 2011; Pereira et al., 2001)
	Magnitude	ND & D	1	The median activity count. A measure of how intense each participant’s upper-limb movements tend to be (Bailey et al., 2015)
	Bilateral magnitude	B	1	The sum of the median activity counts from each limb. An aggregate measure of how intense the participant’s overall upper-limb activity is over the recording period (Barth et al., 2021)
Complexity	Variance	ND & D	1	The SD of the activity counts. Quantifies the overall degree to which the participant’s upper limb movement fluctuates in intensity. This value may scale to the motor task where individuals at the extremes of the distribution have difficulty meeting their movement goals (Stergiou & Decker, 2011)
	Average Jerk	ND & D	30	Rate of change of acceleration in gravitational units averaged across the recording period. Quantifies the smoothness of the limb movements where lower values indicate smoother movement (De Lucena et al., 2017; Smith & Lang, 2019)
	Mean of frequency spectrum	D	30	The weighted mean of the component frequencies from the dominant limb acceleration time series. An aggregate measure of the frequency at which the participant tends to move their upper limb which may relate to disordered movement (Hssayeni et al., 2021)
	Variance of frequency spectrum	D	30	The weighted SD of the component frequencies from the dominant-limb acceleration time series. Quantifies the degree to which the frequency of the participant’s upper limb movement fluctuates. Higher values are hypothesized to indicate more complex movement patterns (Wilson et al., 2021)

UL, upper limb; D, calculated for the dominant limb; ND, calculate for the non-dominant limb; B, calculated using both limbs.

Statistical Analysis

All statistical analyses were performed in the R environment version 4.2.2 (R Core Team, 2022). Motor coordination was treated as a quantitative trait (Faraone & Larsson, 2019; Lichtenstein et al., 2010). A quantitative trait is a measurable biological characteristic (Abiola et al., 2003; Milner & Buck, 2010) that varies among individuals to produce a continuous population distribution (Kadesjö & Gillberg, 1998; Romano et al., 2006). Cutoff scores are often applied to the distribution for the purpose of categorization. Categorization is clinically important and often governs access to healthcare services (i.e. school services), but categories based on cutoff scores collapse much of the variance in the distribution of a trait. Thus, coordination was treated as a continuous variable without group assignment. Comparison of sensor variables across the recording days was done with paired t-tests. Spearman correlations were performed between the MABC-2 and the indirect behavioral assessments.

To test both hypotheses, random intercepts hierarchical linear models (HLM) were built with the LME4 package (Bates et al., 2015). First, we examined the influence of recording context (unstructured or structured) on the association with a direct coordination measure, the MABC-2 (hypothesis 1). Eleven models were built, one for each sensor variable. The sensor variable was the dependent variable while the MABC-2 standard score, the recording context (coded 0 = unstructured, 1 = structured), and their interaction were the independent variables. All sensor variables and the MABC-2 score were z-transformed (mean = 0, sd = 1). This type of model was possible because the two recording contexts gave each participant two repeated measures for each sensor variable. The interaction term was the first statistic of interest for evaluating the first hypothesis because it identified the presence of differences in the association between the sensor variables and the MABC-2 standard score as a function of context (i.e. since the variables were standardized, it is equivalent to the difference between the correlations of each recording context and the MABC-2 standard score). For those sensor variables with significant interactions, the hypothesis was confirmed if the simple effects in the unstructured context were greater: that the sensor variables collected in an unstructured context may serve as a better index of motor coordination. The simple effects were used because they describe the effect of one variable at the different levels of another variable (Judd et al., 2017; Schabenberger et al., 2000). We did not use the main effects, which describe the effect of one variable on average across all levels of another variable. To further characterize the differences between the recording contexts, we also compared the variances of the sensor variables using Levene’s test with p-values adjusted to control for the False Discovery Rate (Benjamini & Yekutieli, 2001).

Second, we examined the influence of the type of coordination assessment (direct assessment versus indirect assessment) on the association with sensor variables. This analysis used the structured sensor variables due to the lack of associations with the unstructured sensor variables (see Results). Eleven models, one for each structured sensor variable, were built to evaluate if the direct assessment of coordination has a stronger association with sensor variables than the indirect assessment (hypothesis 2). The coordination score was the dependent variable while the sensor variable, the coordination score type (coded 0 = indirect, 1 = direct), and their interaction were the independent variables. This model was possible because the two coordination assessments provided two repeated measures for each participant. Neither the MABC-2 nor the DCDQ was normally distributed. The DCDQ total score was log-transformed and the MABC-2 total standard score was square root-transformed to obtain normal distributions. Then both were z-transformed. Thus, differences between the simple effects for each assessment reflect differences in their association with coordination, not an artifact of their scaling. All of the sensor variables were z-transformed. The interaction term was the first statistic of interest because it identified the presence of differences in the associations of the sensor variable and each coordination assessment type (i.e. since all variables were standardized, it is equivalent to the difference between the correlations of each coordination score and the sensor variable). For those variables with significant interactions, our hypothesis was confirmed if the magnitude of the simple effects with the direct assessment were greater: that sensor variables (perhaps due to being a more objective tool) have stronger relationships with the direct versus indirect assessment. For both of these hypotheses, the p-values of the coefficients were adjusted to control for the False Discovery Rate due to multiple comparisons with the alpha level set at 0.05. The adjustment was made for 33 comparisons (3 regression coefficients * 11 models).

RESULTS

We enrolled a total of 88 participants. Eighty-five participants had complete accelerometry data and are included in this analysis; data were lost from two participants due to hardware issues and one participant did not perform the throwing task. Of the 85 participants, six were reported to have ASD and 10 were reported to have ADHD, with two having overlapping diagnoses. Demographic information and descriptive assessment scores are displayed in Table 2. The sample was enriched with participants with lower coordination, as evidenced by 32 participants with MABC-2 total percentile scores below 15 (indicates the presence or risk of DCD). The mean hyperactivity/impulsivity content scale score was 8.42 which is below the hyperactive range. Consistent with previous reports, the MABC-2 and DCDQ correlated at ρ = 0.47 (Wilson et al., 2009). The MABC-2 correlated with the SRS-2 at ρ = −0.35 and the Conners hyperactive/impulsivity score at ρ = −0.27.

Table 2. Participant characteristics.

	N = 85
Demographics
Age	9.44 (1.96)
	6.0–12.99
Sex
Male	58.8% (50)
Female	41.2% (35)
Race
Black or African-American	8.2% (7)
White	88.2% (75)
Asian	3.5% (3)
Ethnicity
Hispanic, Latino	9.4% (8)
Area deprivation index	35.9 (17.6)
	7.0–94.0
Motor coordination
DCDQ	61.8 (12.6)
	26.0–75.0
MABC total percentile	28.0 (25.5)
	0.1–99.0
Hyperactivity
Conners: Hyperactivity/Impulsivity	8.42 (7.87)
	0.00–34.00
Autistic traits
SRS T-score	44.0 (7.84)
	34.2–79.6
Intelligence
KBIT composite score	112 (14.6)
	62.0–141.0

Values are mean (SD) and range, or percentage (N).

There were no statistical differences between the unstructured sensor variables when day 1 was compared to day 2 with a paired t-test (all p-values >0.05). Each participant’s longer recording day was chosen for the analysis. The average unstructured time during waking hours for the longer day was 13.16 h (sd = 1.68). The average structured time during the throwing task was 40.6 min (sd = 5.6).

Structured Sensor Variables Are Better Able to Index Motor Coordination

Our first hypothesis was not supported. Six of the 11 sensor variables had significant interaction terms, with the correlation magnitudes in the structured context greater than those in the unstructured context. The interaction terms from each model are listed in the fourth column of Table 3 (upper portion) and are plotted in Figure 2(A). The simple effects associated with each recording context are listed in columns 2 and 3 in Table 3. Note that because the unstructured context was coded as 0, the interaction effect plus the unstructured simple effect yields the structured simple effect. The simple effect coefficients indicated that for these six variables, structured sensor recordings were better able to index motor coordination compared to the unstructured sensor recordings. The peak magnitude from both sides, the variance from both sides, the average jerk from both sides, the ND magnitude, and the bilateral magnitude associated better with the MABC-2 score when taken from the structured recording context compared to the unstructured recording context. It is notable that, despite the structured task requiring dominant limb underhand throwing, the associations are very similar for the non-dominant sensor variables. To visualize these relationships, the simple effects in each recording context are shown in Figure 3 for the D and ND peak magnitude variables. One can see much stronger relationships (black lines) from the structured recordings on both the dominant arm (used to throw during the lab task) as well as the non-dominant arm. Running the models with the caregiver-reported DCDQ score instead of the MABC score yields similar negative interactions across peak magnitude, variance, and jerk from both limbs indicating again that, for these sensor variables, the structured context is better able to capture motor coordination compared to the unstructured context. A comparison of the sensor variable variances by recording context showed that 8 of the 11 variables had statistically greater variance in the structured context compared to the unstructured context. These comparisons are shown in Table 4.

Figure 2. Standardized interaction coefficients. (A) These interaction coefficients represent the difference in sensor variable by MABC-2 score relationship between the unstructured and structured recording contexts. Negative coefficients favor the structured recording context. (B) These interaction coefficients represent the difference in sensor variable by coordination score relationship between the indirect (DCDQ) and direct (MABC-2) coordination assessment. Negative coefficients favor the direct coordination assessment. Bars = 95% confidence interval. Triangles = adjusted p value ≤0.05. Circles = adjusted p value >0.05.

Figure 3. Example interactions between recording context and coordination score. (A) Scatter plot of MABC total standard score by D peak magnitude. (B) Scatter plot of MABC-2 total standard score by ND peak magnitude. Grey = unstructured recording context. Black = structured recording context. Lines are fitted values from the hierarchical linear model. Axes are in original units, with the y-axis for both panels in activity counts (1 activity count = 0.001664 g).

Table 3. Interaction and simple effects for each sensor variable model.

Hypothesis 1: Sensor Variable ∼ MABC + Recording Context + (MABC × RC)
Sensor variable	Simple effect of MABC in the unstructured context:	Simple effect of MABC in the structured context:	Interaction effect (MABC × RC)
D peak magnitude	0.04	−0.52	−0.56
ND peak magnitude	0.20	−0.41	−0.61
D magnitude	0.10	−0.19	−0.29
ND magnitude	0.15	−0.14	−0.29
Bilateral magnitude	0.13	−0.18	−0.31
D variance	0.18	−0.33	−0.51
ND variance	0.26	−0.28	−0.54
D average jerk	0.16	−0.31	−0.47
ND average jerk	0.16	−0.27	−0.43
D mean of frequency spectrum	−0.06	−0.26	−0.2
D variance of frequency spectrum	−0.08	−0.33	−0.25
Hypothesis 2: Coordination Score ∼ Sensor Variable + Assessment Type + (SV × AT)
Sensor variable	Simple effect of sensor variable on the DCDQ	Simple effect of sensor variable on MABC	Interaction effect (SV × AT)
D peak magnitude	−0.32	−0.52	−0.21
ND peak magnitude	−0.23	−0.41	−0.18
D magnitude	−0.03	−0.16	−0.12
ND magnitude	−0.02	−0.12	−0.10
Bilateral magnitude	−0.03	−0.15	−0.13
D variance	−0.32	−0.32	0.001
ND variance	−0.19	−0.28	−0.09
D average jerk	−0.19	−0.29	−0.11
ND average jerk	−0.15	−0.25	−0.10
D mean of frequency spectrum	−0.06	−0.25	−0.20
D variance of frequency spectrum	−0.09	−0.33	−0.24

All coefficient values are standardized. Top: Comparison of the structured vs. unstructured sensor variable associations with the MABC total standard score. Bottom: Comparison of the indirect and direct coordination score associations with the structured sensor variables. Bold = adjusted p value ≤0.05.

Table 4. Comparison of sensor variable variances across recording contexts.

Sensor variable	Structured variance	Unstructured variance	Adjusted p value
D peak magnitude	205.80	182.90	0.12
ND peak magnitude	217.50	160.02	0.002
D magnitude	37.44	15.67	<0.0001
ND magnitude	40.77	15.22	<0.0001
Bilateral magnitude	69.31	30.26	<0.0001
D variance	27.81	21.17	0.014
ND variance	27.24	21.18	0.097
D average jerk	1.04	0.64	0.0002
ND average jerk	1.07	0.65	<0.0001
D mean of frequency spectrum	0.38	0.32	0.47
D variance of frequency spectrum	0.35	0.14	<0.0001

The variances appear in the variable’s original units. Comparison was made using Levene’s test. Bold = adjusted p value ≤0.05.

Associations Between Sensor Variables and Coordination Did Not Vary by Assessment Type

Our second hypothesis was not supported. After adjustment for multiple comparisons, none of the 11 sensor variables had a significant interaction with assessment type. While there were some significant simple effects, the lack of significant interactions indicates that the relationship between structured sensor variables and coordination did not statistically differ across the assessment types. The interaction terms from each model are listed in the fourth column in Table 3 (lower portion) and plotted in Figure 2(B). The simple effects associated with each type of coordination assessment are listed in columns 2 and 3 in Table 3. Note that because the indirect assessment was coded as 0, the interaction effect plus the indirect simple effect yields the direct simple effect. To visualize these relationships, the simple effects for each coordination assessment type are shown in Figure 4 for the D and ND peak magnitude. One can see the association with the direct assessment (black line) closely matches the association with the indirect assessment (grey line). These results utilized the structured sensor variables. Running the models with the unstructured variables yielded insignificant interaction terms for all sensor variables and all of the simple effects between unstructured variables and assessment type were small and insignificant.

Figure 4. Example interactions between coordination assessment type and structured sensor variables. (A) Scatter plot of coordination score by D peak magnitude. (B) Scatter plot of coordination score by ND peak magnitude. Grey = indirect coordination assessment (DCDQ). Black = direct coordination assessment (MABC-2). Lines are fitted values from the hierarchical linear model. The y-axis for both panels are in standardized units (mean = 0, SD = 1) because the coordination measures use different scales. x-axis is in activity counts (1 activity count = 0.001664 g).

DISCUSSION

This study examined the differential associations between sensor variables and coordination across two recording contexts and two coordination assessment types in a cohort of children aged 6–12. The coordination assessments quantified overall competence in goal-directed tasks, while the sensor variables captured aspects of upper-limb movement. Contrary to our first hypothesis, the majority (6/11) of the structured sensor variables had greater associations with coordination compared to the unstructured sensor variables. Contrary to our second hypothesis, the associations between the sensor variables and coordination did not interact with assessment type, with similar levels of association being observed between sensor variables and the direct and indirect assessments.

We hypothesized that recording unconstrained, free-living, bilateral upper-limb movements would yield sensor variables that diverged across participants according to motor coordination level. The promise of wearable sensors is often thought to be their ability to passively collect movement data in the participant’s natural environment. Here, we found greater associations between coordination and structured sensor variables which suggests that wearable sensors may more easily capture divergent coordination levels when school-age children perform a common structured motor task. This may be because the environmental constraints (same physical environment for all participants) and task constraints (same instructions and equipment for all participants) were held constant, allowing individual variation in coordination patterns to emerge in ways that could be quantified by the sensor variables. This interpretation is supported by the comparison of variances (Table 4). All of the sensor variables had greater variance in the structured context and 8/11 of those were significant when tested with Levene’s Test. The environmental and task constraints were lacking in the unstructured recording context. The participants performed heterogeneous activities in their daily life over a long duration recording which may have produced values among this set of sensor variables that did not diverge by coordination level to the same extent as the constrained structured sensor variables. It may be that other sensor variables could be extracted from the data are capable of capturing a meaningful coordination signal from the unstructured recordings.

Lower coordination scores were associated with greater peak magnitudes, more variance, and increased jerk (less smoothness). These associations were seen in both limbs (see next paragraph). The results lend support to the premise that sensor variables from structured recordings can quantify variation in coordination in school-age children. The pattern of associations found here can be viewed as consistent with the motor characteristics of children with developmental coordination impairments. The peak magnitude may reflect the participant’s degree of control over upper-limb movements. Children with DCD and ASD are known to have impaired control over neuromuscular parameters such as force production and timing (Fong et al., 2015; Hyde & Wilson, 2011; King et al., 2012). For example, while reaching to targets 6–10 year-old children with DCD show prolonged agonist activity and delayed antagonist activity relative to typically developing controls (Huh et al., 1998). Children with DCD also have increased reliance on visual feedback (Coats et al., 2015) and impaired predictive forward modeling (Wilson et al., 2013) which may lead to prolonged or excess agonist activity during the throwing motion and therefore higher magnitude activity counts. Variance quantifies the degree of variability (or consistency) in the upper-limb movements. Children with DCD are often found to have greater variability in parameters of movement (Sekaran et al., 2012; Smits-Engelsman et al., 2008). In a seated reaching task with 7–9-year-olds, the DCD group had higher variability in reaching parameters compared to controls (Biancotto et al., 2011). Jerk quantifies the smoothness movement. Jerk decreases in childhood as the motor system matures (Yan et al., 2000) and in a sample of children with ASD performing a reach-to-grasp task jerk was greater compared to the typically developing controls (Yang et al., 2014).

We were surprised to see that the structured sensor variables from the non-dominant upper limb had associations with coordination that were very similar to those from the dominant upper limb despite the fact that the structured throwing task was performed with the dominant hand only. The non-dominant limb was not used to throw and success on the task did not require non-dominant limb activity. Nonetheless, the non-dominant limb exhibited motor characteristics that were recorded by the sensors and related to the child’s overall coordination. The structured sensor data consisted of any movement that occurred during the bean bag throws, the time between throws, or the short breaks between blocks. The consistency seen here across the dominant and non-dominant limbs matches previous findings that limb movements of children with poor coordination are coupled to a greater degree, as demonstrated by higher correlations between joint angles in a catching task compared to controls (Utley et al., 2007). Again, the common environmental and task constraints of a discrete motor task may have facilitated the emergence of individual coordination differences that were captured by the sensor variables for both limbs. The similarity across the dominant and non-dominant limbs is promising because it implies that the ability of wearable sensors to index coordination is independent of the specific requirements of the structured motor task. The throwing task used here was chosen as part of another aim of this study and was not intended to be a proxy for coordination. It may be that any discrete, structured motor tasks could be used. The most important feature may be that children perform the same task under the same constraints. Standard developmental motor tests, such as the MABC-2, are valuable assessment tools but one of their limitations is that they require an expert clinician to administer the test in a clinical or research environment. Wearable sensors offer an alternative assessment that could be administered by support personnel. Custom hardware and software could be developed in combination with a simple structured motor task protocol performed in various locations (e.g. clinics, schools, homes). Future research could evaluate additional motor tasks to confirm that task-specificity is less important than having a common task with similar constraints imposed on all participants. Alternative structured tasks could include fine motor tasks (such as the peg-hole task or a tracing activity) or tasks that require visual-motor integration (such as ball catching). A lower-extremity task, such as ball kicking, could also be assessed to determine how integral a functional upper-limb activity is as a way to quantify coordination with sensor variables. It is also important to determine the minimum duration of the structured activity necessary to generate variables that reliably index coordination, as this has implications for clinical feasibility. A longitudinal study concurrent with motor skill interventions could determine if the sensor variables are sensitive to changes in coordination skills over time.

The finding that the association between sensor variables and coordination did not vary by assessment type refutes our second hypothesis. The interaction terms were found to be small and insignificant which indicated there was no difference in the association between coordination and sensor variables across the two coordination assessments. This may be because the two assessments capture similar variance in coordination level and therefore their relations with the sensor variables are similar. In this sample, the two assessments correlate at 0.56. This positive relationship is consistent with the previously mentioned correlation of 0.55 between the DCDQ and MABC (Wilson et al., 2009). The simple effects determine if the association between a particular assessment type and the sensor variable statistically differs from zero. We did find that for 6/11 sensor variables there was a significant simple effect of the direct coordination assessment while 2/11 sensor variables had a significant simple effect of the indirect coordination assessment. The more modest association between indirect coordination assessment and coordination is similar to previous relationships found between indirect assessments and human behaviors (Rice et al., 2013; Sanchez & Constantino, 2020; Wehrmann & Müller, 2015; Wilson et al., 2009; Zysset et al., 2018). Both the wearable sensors and the MABC-2 directly assess movement behaviors according to specified standards. For example, peak magnitude is measured in activity counts across all participants. Likewise, the tracing item on the MABC-2 is measured in seconds for all participants and then assigned an age-standardized score. These specific, objective criteria do not exist for indirect, caregiver-reported questionnaires and it may be wise to consider caregiver-reported scores as screening tools, while more direct measures of motor coordination are used to guide clinical care (Zysset et al., 2018).

Wearable sensors have been used to assess many behaviors of interest (i.e. sleep, stereotypies, physical activity level). The optimal assessment duration and context to capture the desired behavior may vary and must be fit for purpose. For some behaviors, a short, structured context may be superior (i.e. gait parameters) while other behaviors may be best captured in long, unstructured context (i.e. Moderate to Vigorous Physical Activity). The choice of assessment context has the potential to influence the coordinated movement patterns that underlie these behaviors as those patterns emerge from the interaction of the task requirements, the environment, and the individual. Aspects of these movement patterns are then captured in the sensor variables. Our findings suggest that the best method of quantifying aspects of developmental coordination with wearable sensors may be a short assessment duration in a structured context.

Limitations

There are a few limitations to consider when interpreting the results of this study. First, our recruitment strategy was primarily based on word-of-mouth and social media advertising and may have led to parents self-selecting for concerns related to motor function. This led to a sample that was enriched for families with concerns about a child’s coordination skills, although most did not have a developmental diagnosis. This is reflected in the fact that 38% of the sample had a MABC-2 score <15th percentile which indicates they have or are at risk for DCD. While enrichment for motor deficits facilitated the evaluation of our hypotheses, the sample may lack generalizability to the wider population. Second, participant medical history was provided by caregiver-report only and was not verified. It is possible that there were other comorbid conditions that were not reported, again reducing generalizability of the findings. Third, most of our data collection took place in the summer months when many participants reported periods in which they did not wear the sensors due to swimming, which reduced the total data included in the analysis, although the average unstructured recording duration still exceeded the goal of 12 hours per day. It is possible that the association between coordination and the variables from the unstructured recordings would have been stronger if the movement that occurred during swimming had been included since swimming is typically a highly active, bilateral activity. This may be possible in the future as waterproofing of sensors improves.

CONCLUSIONS

Bilateral sensors worn at the wrists can quantify motor behavior and index motor coordination. Sensor variables calculated from a structured activity were superior in showing relationships with measures of motor coordination to variables calculated from free-living, unstructured activity. A pattern emerged across the sensor variables where lower motor coordination was associated with higher peak magnitude, higher variance, and less smoothness. This pattern was nearly identical across both limbs despite the fact that only the dominant limb was used for the throwing task performed during the structured recordings. The greater associations between structured sensor variables and coordination, regardless of limb dominance, have the potential to benefit the management of developmental populations that are characterized by motor coordination impairment. The finding implies that wearable sensors could be deployed in combination with discrete motor task protocols administer by support personnel to yield sensor variables that quantify motor coordination. For providers who assess developmental coordination impairments, wearable sensors could be an adjunct to standard clinical tests or subjective caregiver reports.

ETHICAL APPROVAL

This study was approved by the Human Research Protection Office of Washington University and conformed to the Declaration of Helsinki. Caregivers provided informed consent and participants provided verbal assent.

AUTHOR CONTRIBUTIONS

JDK: Conceptualization, Methodology, Investigation, Project administration, Formal Analysis, Writing – Original Draft. NM: Conceptualization, Writing – Review & Editing. KRL: Conceptualization, Formal Analysis, Writing – Review & Editing. KMT: Investigation, Writing – Review & Editing. CEL: Conceptualization, Methodology, Investigation, Formal Analysis, Writing – Original Draft.

Acknowledgements

The authors would like to thank Jennifer Brown for assistance with recruitment and the children and families who participated in the study.

DISCLOSURE STATEMENT

The authors report there are no competing interests to declare.

Additional information

Funding

This work was supported by a Promotion of Doctoral Studies I Scholarship from the Foundation for Physical Therapy Research and the NIH [R01MH123723, T32HD007434].