ABSTRACT
Objectives
We aim to quantify the performance of accelerometry in objectively measuring physical activity (PA) intensity among infants and toddlers.
Methods
Thirty‐eight 6‐ to 24‐month‐olds participated in a 30‐min, semistructured lab visit. Twenty‐three (61%) children could walk independently. Children engaged in a variety of activities while wearing accelerometers on each ankle and at the waist. Visits were video recorded, and study team members independently coded the first three 5‐s epochs of each minute for PA intensity using a 5‐level scale ranging from 1 = completely sedentary to 5 = moderate‐to‐high intensity. Interrater agreement for PA classifications was excellent (median kappa per child = 0.85). A series of logistic regression models were fit to find the vector magnitude threshold per 5‐s epoch that differentiated activity intensity above each PA level with ≥ 80% sensitivity.
Results
Analyses included 3191 epochs; a median of 88 epochs per child. The classification performance applying all thresholds concurrently for the five PA intensity levels was poor for each wear location (agreement < 50%, kappa < 0.25). Classification improved when concatenating intensity levels, with the best performance comparing sedentary (levels 1–2) to nonsedentary (levels 3–5) and using data from the left ankle device: agreement ≥ 77.6%, kappa ≥ 0.44. Applying those novel thresholds to predict the total time spent in level 3–5 activities over all coded epochs using linear regression performed as well as using the sum of vector magnitude across epochs when using data from the left ankle device. Overall, the performance of the left ankle wear location was similar to the right ankle wear location and superior to the waist location.
Conclusions
Ankle‐worn accelerometry had adequate validity to classify in‐the‐moment nonsedentary behaviours and total time spent nonsedentary over a time interval among this sample of infants and toddlers. While caution is warranted when generalizing these lab‐based findings to naturalistic settings, findings provide insight into objective measures of PA for this age range.
Keywords: accelerometry, ankle, infants, physical activity, toddlers
Abbreviations
- CARS
Children Activity Rating Scale
- MVPA
moderate to vigorous physical activity
- PA
physical activity
- VM
vector magnitude
1. Introduction
Engaging in physical activity and play has insurmountable health benefits during childhood, including gross motor skill development, coordination and muscle strength (Physical Activity Guidelines Advisory Committee 2018; Timmons et al. 2012). Importantly, greater movement (Benjamin‐Neelon et al. 2020; Li et al. 1995) and less time spent restrained (Wells and Ritz 2001) during infancy may be protective against excess adiposity gain. These early years may be an important timeframe to encourage physical activity because physical activity behaviours in infancy and toddlerhood track into the preschool years (Del Pozo‐Cruz et al. 2019).
While accelerometry is the gold standard for objectively measuring physical activity in naturalistic settings, there is a considerable lack of data on the performance of accelerometry among infants and young toddlers. Briefly, accelerometers quantify change in acceleration per a desired cycling period (e.g., Hertz), filter the measurements to exclude noise and create a metric reflecting ‘counts’ per specified measurement epoch (Cliff, Reilly, and Okely 2009). Counts are unitless, and methods have been developed to calibrate the count data against a gold standard to reflect a meaningful metric of physical activity intensity among toddlers able to walk independently. Specifically, both Trost et al (Trost et al. 2012). and Costa et al (Costa et al. 2014). have assessed the validity of hip‐based accelerometry using ActiGraph accelerometers among 14‐ to 36‐month‐old children in controlled lab settings against visually observed physical activity intensity. Performance of the Trost cut‐points was derived via area under the ROC curve, and values were 0.74 (95% CI: 0.71–0.76) for classifying sedentary behaviour [‘acceptable’ per common guidelines (Hosmer, Lemeshow, and Sturdivant 2013)] and 0.90 (95% CI: 0.88–0.92) for classifying moderate‐to‐vigorous PA (‘outstanding’ per guidelines). Performance of the Costa cut points was also acceptable. For example, there was 88.5% agreement and kappa = 0.76 when classifying sedentary behaviours (vs. nonsedentary) using the accelerometer vector magnitude, and 82.2% agreement and kappa = 0.43 when classifying MVPA (vs. non‐MVPA) (Costa et al. 2014).
However, while the hip wear location is most often used for accelerometry among older children and adults, the ankle‐wear location is more desirable among infants and younger toddlers for a few reasons (Benjamin‐Neelon et al. 2020; Prioreschi and Micklesfield 2016; Pate et al. 2020). Devices worn at the hip can be uncomfortable for infants who may spend considerable time on their back or tummy. Limb movement can be a considerable source of activity for infants not able to walk yet, and a device worn at the hip would miss such movements that do not also engage the truck. Furthermore, validating methods to measure physical activity intensity at the ankle location can facilitate 24‐h measurements of movement and sleep with one device that can be worn continuously, as ankle‐worn accelerometry has been validated against polysomnography to measure sleep metrics among infants (Sadeh et al. 1995), toddlers and preschool‐age children (Bélanger et al. 2013; Van Kooten et al. 2021).
Unfortunately, data validating the performance of ankle‐based accelerometry in measuring physical activity intensity is sparse for infants and toddlers. Hager et al (Hager et al. 2016). reported acceptable validity of ankle‐based accelerometry among 14‐ to 35‐month‐olds. In that study, accelerometry was calibrated against direct observation using a protocol similar to that used by Trost and Costa, and the sensitivity and specificity for classifying sedentary (81.8% and 77.5%, respectively), light PA (61.7% and 84.7%, respectively) and MVPA (85.7% and 88.4%, respectively) were acceptable, although low for classifying light PA. Preliminary evidence from a study of 10 infants and young toddlers (aged 31 to 433 days old) (Ghazi et al. 2024) supports the utility of accelerometry thresholds to distinguish between physical activity intensity levels based on ankle‐worn accelerometry, although data were limited to when children were in the supine position. Additional data are needed to understand if accelerometry can be used in more naturalistic settings to distinguish between thresholds of infant physical activity intensity specifically, or if accelerometry is most valid for quantifying sedentary versus nonsedentary behaviours more generally.
In summary, while current methods to classify physical activity intensity among young children have acceptable validity, methods have not been fully examined for infants and young toddlers who are not able to walk independently. This is a considerable gap to fill given the importance of physical activity at such a young age for physical health and gross motor skill development.1,2 3,4 Accelerometry is currently being used to measure physical activity among nonambulatory infants and toddlers (Benjamin‐Neelon et al. 2020; Pate et al. 2020), and thus, validation metrics are needed. Therefore, the goal of the current study was to assess the performance of accelerometry in measuring physical activity intensity for 6–24‐month‐olds when engaging in naturalistic activities. Children wore an accelerometer near the hip with an elastic belt at the waist and one on each ankle to compare the performance across wear locations.
2. Methods
2.1. Participants
The study enrolled children aged 6‐ to 24‐month‐olds and one parent each to participate in a semistructured laboratory visit lasting 30 min. Participants were recruited via our laboratory's database of families interested in research studies supplemented with recruitment via paper flyers and email ‘listserv’ announcements targeted to towns within a 30‐min drive of our laboratory. The towns included reflected varying income levels, with the town‐level median household income ranging from $53 697 to $111 958; additionally, the percent of each town's population identifying as white, non‐Hispanic ranged from 71.7% to 91.9% based on 2018–2022 U.S. Census estimates. Recruitment and enrolment occurred between June 2022 and September 2022. Children with significant developmental delays or medical conditions that limited movement or gross motor skills were excluded. Parents needed to read and converse in English. Recruitment was stratified by child age (< 15 vs. ≥ 15 months) in an attempt to enrol an equal number of children who could walk independently (herein defined as ambulatory) and who could not. The final ambulatory status of children was defined at the lab visit by the study team, where ambulatory children were those who could walk three full strides independently without falling. Parents provided signed informed consent. Parents also completed a questionnaire to report on child, parent and household characteristics. Parents were provided with a $40 gift card to one of two online retailers for participating and an additional $10 for gas costs if they travelled more than 15 miles to the study visit. Visits were scheduled between 8 am and 5 pm on weekdays or weekends. All study procedures were reviewed and approved of by the Dartmouth College Committee for the Protections of Human Subjects, IRB approval 32 173.
2.2. Study Design
Study visits took place in a large (1047 ft2) indoor laboratory that was one large open room. The study team members (2 or 3), the child, the parent(s) and sometimes a child's sibling were in the laboratory during the visits. The laboratory included books, children's musical instruments, several toys for gross and fine motor use (e.g., balls and blocks), larger props for play including a small slide and a flexible cloth tunnel to crawl into and a speaker for playing music. Two high‐performance video cameras (GoPro, Los Angeles, CA) were placed in different locations in the room to record child movements. Children were first allowed to explore the space to get comfortable. Study team members then placed accelerometers on each ankle of the child with Velcro straps and one near the hip with an elastic belt purchased from the device manufacturer (herein referred to a waist‐worn device for specificity given how the device was worn by children in practice). The devices on the right ankle and hip were both ActiGraph devices, model wGT3X and collected data at 100 Hz. The device on the left ankle was an ActiGraph model GT9X and collected data at 100 Hz. The GT9X device had the IMU features activated (e.g., accelerometer plus gyroscope measurements) for the purpose of addressing a separate research question; that additional data was not used in this current study. Importantly, evidence supports that differences in sampling rates (from 30 to 100 Hz) have no discernable impact on count data per 5‐s epoch for ActiGraph accelerometers (Clevenger et al. 2022).
Study team members encouraged the child to participate in a variety of activities following a semistructured protocol with the study team allowing the children to lead the play. Specifically, the study team encouraged play with the child and between the child and the parent at different levels: sedentary or low‐level intensity activities for the first 10 min (e.g., reading a book, holding toys), then light intensity activities for the second 10 min (e.g., walking, slow dancing movements), then moderate‐to‐vigorous intensity activities for the last 10 min (e.g., running, rapid dance movements). The activities study team members encouraged did not specifically map onto each of the 5 levels in the CARS coding used for the gold standard (as described below). During a pilot testing phase of the study protocol, we found that our semistructured approach focusing on sedentary, light and moderate‐to‐vigorous intensity activities, versus a more structured approach that prescribed specific activities in each CARS level, was more naturalistic and effective at encouraging children to be active as it provided flexibility for each child. Children's length and weight were measured at the end of the study visit by the study team. Height was measured for several toddlers who refused to lay down for length measurements.
Power for the study was computed as the 95% confidence interval for classifying moderate‐to‐vigorous physical activity with 80% sensitivity using data from 40 children, including 3 epochs per minute across 30 min per child (3600 total epochs) and assuming one‐third of epochs are spent in moderate‐to‐vigorous physical activity based on the semistructured study protocol, which resulted in a 95% confidence interval of 77.6% to 82.2%.
2.3. Measures
2.3.1. Video Coding of Physical Activity Intensity
Accelerometry data was synced to the videos. Specifically, a study team member handled the accelerometer devices in a standardized way (e.g., devices still for > 5 s, devices shaken vigorously for 5 s, then devices still for > 5 s) at the start of each lab visit to enable syncing between the time‐stamped accelerometer data and the time‐stamped video footage. The syncing of all video footage was completed by a study team member not involved in the video coding process and reviewed by the study lead to ensure the accuracy of the syncing across the entire lab visit per child. The first three 5‐s epochs of each minute were visually coded, using the recorded video, by two study team members who each had professional experience working with infants and toddlers in early childcare settings. Selecting a subset of epochs for validation is consistent with the methods of previous validation studies among preschool‐age children (Pate et al. 2006). We decided on using 5‐s epochs, as done in a previous accelerometry validation study among 2‐ to 3‐year‐olds (Costa et al. 2014), to capture the sometimes‐rapid changes in momentary movements exhibited by infants and young toddlers while also minimizing the need of study team members to average child behaviours over a longer epoch. That selection process would produce 7.5 min of coded footage total per child, which is similar to the total time included per child in previous accelerometer validation studies of preschool‐age children (Pate et al. 2006) and toddlers (Hager et al. 2016). In this current study, epochs were excluded if the child was playing with or moving the accelerometer devices or if the child was being picked up by an adult.
An a priori defined codebook was used that was based on the validated Children's Activity Rating Scale (Hager et al. 2016; Puhl, Greaves, and Hoyt 1990). Briefly, the CARS classifies child physical activities by intensity level, with a score that ranges from 1 (stationary with no movement) to 5 (translocation, very fast/strenuous) (Puhl, Greaves, and Hoyt 1990). We further incorporated modifications to capture movements specific to infants as previously defined by Li et al (Li et al. 1995). Our initial scale included six levels of physical activity intensity that incorporated differences for infants and toddlers (Table S1). An epoch was coded based on the activity that lasted for the majority of the 5‐s epoch (3 s or more). The study team members did not see the accelerometer device data during the video coding process. After data coding, only one epoch across all children was coded as a level 6 activity (i.e., vigorous), and that epoch was thus combined with level 5 for all analyses. Median interrater agreement across all children was excellent (weighted kappa, k = 0.85; IQR: 0.79, 0.90) and was higher among children who were ambulatory (k = 0.89; IQR: 0.84, 0.92) than nonambulatory (k = 0.79; 0.69, 0.85). Epochs with disagreements were reviewed and adjudicated by the two coders and the study PI as a group.
2.3.2. Accelerometry Data
For each device, the vector magnitude (VM) of the count data over the X, Y and Z axes was computed per 5‐s epoch. This current analysis only included accelerometer data corresponding to the visually coded epochs.
2.4. Statistical Analysis
Participant characteristics were summarized overall and by ambulatory status of the child. The primary analysis included a set of logistic regression models to classify physical activity intensity above, versus at, each level (e.g., > level 1 activities vs. level 1 activities) with the modified CARS coding as the gold standard. Models were run in a stepped manner, such that level 1 activities were excluded when defining > level 2 versus at level 2, and levels 1 and 2 activities were excluded when defining > level 3 versus at level 3, etc. That stepwise process was to specifically assess how well higher intensity levels could be differentiated. The threshold for each level was defined from ROC curves as the threshold that produced 80% sensitivity with maximum sensitivity. The area under the ROC curve (AUC), accuracy (total % agreement), and kappa values were computed when comparing accelerometer data using the set of thresholds to the gold standard. Additional models using the same modelling approach were completed, using concatenated physical activity levels instead of the original CARS levels.
Finally, because studies have used aggregate accelerometry metrics (e.g., sum of VM counts per day) to measure infant and young toddler activity (Benjamin‐Neelon et al. 2020; Pate et al. 2020), we fit a set of linear regression models using aggregate metrics (i.e., time spent in higher‐level activities using the study's thresholds, sum of VM, mean VM per epoch, median VM per epoch) to model the time spent in higher‐intensity activities (dependent variable) based on the gold standard; those models only used data from the epochs that were visually coded for physical activity intensity. The coefficient of determination (R2) was used to assess model fit, the square root of the R2 was computed to demonstrate the standardized correlation, and the root mean squared error (RMSE) was reported to compare model fit across models. All analyses were completed with the R Language and Environment for Statistical Computing (version 4.2.0). Statistical significance was set to p < 0.05.
3. Results
Our enrollment target was 40 children. While we obtained consent from 40 parents, two did not attend the study visit, and thus, our final analysis includes 38 children. Characteristics of enrolled children (n = 38) are presented in Table 1. The sample included more ambulatory (n = 23, 60.5%) children than nonambulatory (n = 15, 39.5%). More than half (63.2%) of children were female, most (86.8%) were non‐Hispanic White reflecting the local population, and mean weight‐to‐length percentile was 65.9%. The median number of epochs coded per child was 88, corresponding to 7.3 min. On average, children spent 33% of the coded epochs in level 2 activities, 18% in level 3 and 20% level 4. Median VM per epoch increased over physical activity levels for both the ankle (p < 0.001) and waist (p < 0.001) wear locations (Figure 1, Table S2). The correlation between the VM per 5‐s epoch between the left and right ankle worn devices was moderate (Spearman r = 0.82; p < 0.001) and similar for ambulatory (Spearman r = 0.84; p < 0.001) yet lower for nonambulatory (Spearman r = 0.74; p < 0.001) children. The correlation between the VM per 5‐s epoch between the left ankle and waist worn devices was lower (Spearman r = 0.62; p < 0.001) and similar for ambulatory (Spearman r = 0.61; p < 0.001) and nonambulatory (Spearman r = 0.58; p < 0.001) children.
TABLE 1.
Sample characteristics and summary of CARS coding results.
| All children (n = 38) | Ambulatory children (n = 23) | Nonambulatory children (n = 15) | |
|---|---|---|---|
| Age (months), mean (SD) | 15.7 (5.3) | 19.0 (3.5) | 10.6 (3.0) |
| Female sex, n (%) | 24 (63.2%) | 13 (56.5%) | 11 (73.3%) |
| White, non‐Hispanic | 33 (86.8%) | 20 (86.9%) | 13 (86.7%) |
| Weight‐for‐length percentile, a , b mean (SD) | 65.9 (32.2) | 69.2 (28.5) | 60.5 (38.1) b |
| Physical activity coding | |||
| Number of 5‐s epochs coded per child, median (IQR) | |||
| Overall | 88 (82, 89) | 89 (87, 90) | 81 (73, 87) |
| By level | |||
| Level 1 | 14 (10, 24) | 12 (10, 21) | 20 (14, 27) |
| Level 2 | 29 (21, 37) | 24 (19, 32) | 36 (30, 48) |
| Level 3 | 16 (10, 22) | 18 (13, 23) | 11 (6, 21) |
| Level 4 | 18 (9, 27) | 23 (19, 32) | 5 (2, 9) |
| Level 5 | 3 (1, 6) | 4 (2, 8) | 1 (1, 2) |
| Interrater agreement (κ) per child, median (IQR) | 0.85 (0.79, 0.90) | 0.89 (0.84, 0.92) | 0.79 (0.69, 0.85) |
Note: Kappa was computed using quadratic weights.
Abbreviation: IQR, interquartile range.
Age‐ and sex‐adjusted weight‐for‐length was computed using the 2000 CDC growth charts.
The length measurement was missing for one nonambulatory child.
FIGURE 1.
Vector magnitude per 5‐s epoch by physical activity level. Presented for all children and stratified by ambulatory status.
The algorithm to classify the five separate physical activity levels using the thresholds defined by the series of stepped logistic regression models performed poorly, with final percent agreements < 50% and kappa values ≤ 0.25 for each wear location for all children combined (Table 2) and when stratified by ambulatory or nonambulatory status (Tables S3a–c). The classification performance improved for all wear locations when concatenating physical activity levels into binary classifications of higher‐ (levels 3–5) versus lower‐ (levels 1–2) level activity (Table 3). Models using from the left ankle device performed the best for all children and once stratified by ambulatory status (agreement ≥ 77.6%, kappa ≥ 0.44), although results were similar when using data from the right ankle device (agreement ≥ 76.6%, kappa ≥ 0.45). Models using data from the waist did not perform as well (agreement ≥ 70.3%, kappa ≥ 0.28). A set of sensitivity analyses were completed to repeat those later models with PA intensity concatenated as levels 4–5 intensity versus levels 1–3 intensity. Results were largely similar (Table S4) with slightly lower kappa values across the models.
TABLE 2.
Performance when classifying physical activity level using accelerometer data from each wear location among all children, ambulatory and nonambulatory combined.
| AUC | VM threshold | Sensitivity | Specificity | |
|---|---|---|---|---|
| Wear location: left ankle | ||||
| PA classification | ||||
| > 1 vs. 1 | 0.84 | 53 | 0.80 | 0.77 |
| > 2 vs. 2 | 0.81 | 284 | 0.80 | 0.62 |
| > 3 vs. 3 | 0.68 | 514 | 0.80 | 0.42 |
| > 4 vs. 4 | 0.79 | 653 | 0.80 | 0.65 |
| Final model performance | Accuracy = 41.3%, kappa = 0.22 | |||
| Wear location: right ankle | ||||
| PA classification | ||||
| > 1 vs. 1 | 0.84 | 80 | 0.80 | 0.73 |
| > 2 vs. 2 | 0.80 | 270 | 0.80 | 0.61 |
| > 3 vs. 3 | 0.67 | 506 | 0.80 | 0.36 |
| > 4 vs. 4 | 0.80 | 584 | 0.79 | 0.69 |
| Final model performance: | Accuracy = 38.2%, kappa = 0.18 | |||
| Wear location: waist | ||||
| PA classification | ||||
| > 1 vs. 1 | 0.83 | 83 | 0.80 | 0.71 |
| > 2 vs. 2 | 0.75 | 218 | 0.80 | 0.51 |
| > 3 vs. 3 | 0.59 | 375 | 0.80 | 0.25 |
| > 4 vs. 4 a | 0.65 | 403 | 0.79 | 0.31 |
| Final model performance: | Accuracy = 34.9%, kappa = 0.13 | |||
Note: Kappa was computed using quadratic weights. The accelerometer data and PA levels are each measured per 5‐s epoch. Analyses included 3191 coded epochs in total.
Abbreviations: AUC, area under the receiver operator curve; PA, physical activity; VM, vector magnitude per 5‐s epoch.
There were only six epochs in level 5 for nonambulatory children, and thus, those observations were combined with level 4 for all analyses.
TABLE 3.
Performance when classifying concatenated physical activity levels 3–5 versus 1–2 from each wear location.
| PA classification | VM thresholds per 5‐s epoch | AUC | Sensitivity | Specificity | Accuracy a | Kappa a | |
|---|---|---|---|---|---|---|---|
| Wear location: left ankle | |||||||
| Levels 3–5 vs. 1–2 | All children | 0–209, > 210 | 0.86 | 0.80 | 0.74 | 77.6% | 0.54 |
| Ambulatory | 0–201, > 201 | 0.87 | 0.80 | 0.76 | 77.9% | 0.56 | |
| Nonambulatory | 0–218, > 218 | 0.82 | 0.80 | 0.71 | 77.6% | 0.44 | |
| Wear location: right ankle | |||||||
| Levels 3–5 vs. 1–2 | All children | 0–210, > 210 | 0.85 | 0.80 | 0.72 | 76.6% | 0.52 |
| Ambulatory | 0–200, > 200 | 0.86 | 0.80 | 0.74 | 77.3% | 0.55 | |
| Nonambulatory | 0–226, > 226 | 0.82 | 0.80 | 0.70 | 77.7% | 0.45 | |
| Wear location: Waist | |||||||
| Levels 3–5 vs. 1–2 | All children | 0–177, > 177 | 0.81 | 0.80 | 0.62 | 72.5% | 0.43 |
| Ambulatory | 0–193, > 193 | 0.81 | 0.80 | 0.62 | 70.3% | 0.41 | |
| Nonambulatory | 0–155, > 155 | 0.74 | 0.80 | 0.51 | 73.0% | 0.28 | |
Abbreviations: AUC, area under the receiver operator curve; PA, physical activity; VM, vector magnitude per 5‐s epoch.
Accuracy and kappa are comparing the model classification to the gold standard of observed PA level.
Additional analyses were completed to compare aggregate metrics of accelerometer data over the entire study visit (including coded epochs only) to the time spent in higher‐level activities (levels 3–5) per the gold standard (Table 4). Models used the time spent in levels 3–5 activities as defined using this study's novel thresholds in Table 3 and well as the sum, mean and median of all VM over the coded epochs. Among all children combined, models using the time spent in levels 3–5 activities and the sum of VM performed equally well, with a better performance for the ankle devices. Model performance varied once stratified by ambulatory status of the child, yet performance was consistently better for the ankle versus waist worn device. A sensitivity analysis was completed to repeat those models, using time spent in levels 4–5 activities (Table S5). Overall, model performances were similar or worse.
TABLE 4.
Performance of aggregate accelerometer metrics in modelling time spent in physical activity levels 3–5 for each wear location.
| Wear location: left ankle | Wear location: right ankle | Wear location: waist | ||||
|---|---|---|---|---|---|---|
| R 2 | RMSE | R 2 | RMSE | R 2 | RMSE | |
| All children (n = 38) | ||||||
| Time spent in levels 3–5 activities using the new thresholds a | 0.68* | 10.9 | 0.69* | 10.7 | 0.66* | 11.2 |
| Sum of VM | 0.69* | 10.6 | 0.68* | 10.8 | 0.66* | 11.2 |
| Mean VM | 0.66* | 11.2 | 0.64* | 11.5 | 0.64* | 11.5 |
| Median VM | 0.62* | 11.8 | 0.62* | 11.8 | 0.68* | 10.9 |
| Ambulatory children (n = 23) | ||||||
| Time spent in levels 3–5 activities using the new thresholds a | 0.63* | 8.4 | 0.70* | 7.6 | 0.51* | 9.7 |
| Sum of VM | 0.64* | 8.3 | 0.63* | 8.4 | 0.46* | 10.2 |
| Mean VM | 0.63* | 8.4 | 0.62* | 8.5 | 0.45* | 10.2 |
| Median VM | 0.66* | 8.1 | 0.66* | 8.1 | 0.54* | 9.4 |
| Nonambulatory children (n = 15) | ||||||
| Time spent in levels 3–5 activities using the new thresholds a | 0.66* | 7.0 | 0.58* | 7.7 | 0.19 | 10.8 |
| Sum of VM | 0.67* | 6.9 | 0.65* | 7.1 | 0.28** | 10.2 |
| Mean VM | 0.75* | 6.0 | 0.72* | 6.4 | 0.27** | 10.2 |
| Median VM | 0.68* | 6.8 | 0.65* | 7.1 | 0.30** | 10.0 |
Note: Aggregate metrics (the independent variables) were aggregated over all of the coded 5‐s epochs in the lab session per child. The gold standard of time spent in physical activity levels 3–5 (the dependent variable) was based on visual coding.
Abbreviations: R 2, coefficient of determination from each linear regression model; RMSE, root mean squared error; VM, vector magnitude.
Time spent in levels 3–5 activities is the sum of all 5‐s epochs defined as levels 3–5 activities using the thresholds presented in Table 3. Thresholds are specific to the wear location and ambulatory status.
Model slope (beta coefficient) is statistically different than 0 at the p < 0.001 level.
Model slope (beta coefficient) is statistically different than 0 at the p < 0.05 level.
4. Discussion
This study assessed the performance of ankle and waist‐based accelerometry among 6‐ to 24‐month‐old children in a semistructured laboratory study. Children engaged in a variety of age‐appropriate activities while wearing accelerometers at the ankle and waist, and the gold standard of physical activity intensity was defined using a validated rating scale with age‐appropriate modifications. Using a series of stepped logistic regression models, we first defined thresholds to classify physical activity intensity for each 5‐s epoch. The performance of the algorithm applying all thresholds concurrently was poor across all analyses, suggesting that accelerometry cannot distinguish between higher intensity activities when assessed as 5‐s epochs among 6‐ to 24‐month‐olds.
The performance of our classification algorithms improved after concatenating physical activity levels into higher‐ versus lower‐intensity levels. Specifically, the models performed best when considering levels 3–5 versus levels 1–2 activities and using data from the ankle wear location. The performance of the models was similar when using data from the left or right ankle. Levels 3–5 activities include very light, light and moderate‐to‐high intensity activities. Thus, we propose this classification corresponds to nonsedentary behaviours. In a previous lab‐based study among 14–36‐month‐olds (Hager et al. 2016) who could walk independently (n = 24, 54.2% > 24 months), Hager et al. reported a sensitivity of 81.8% and specificity of 77.5% when classifying sedentary versus nonsedentary behaviours using an ankle worn accelerometer (Actical model) and with 15 s epochs. Thus, our current findings are similar to that previous study that included a purely ambulatory sample. However, our findings greatly add to the field by providing data on the validity of these methods to approximate physical activity intensity among nonambulatory children.
Because there are no widely accepted thresholds to classify physical activity intensity for infants and young toddlers, researchers have used aggregate accelerometer metrics including total counts per day or average counts per hour (Benjamin‐Neelon et al. 2020; Pate et al. 2020) to estimate total child activity. Thus, our study also considered aggregating data across all coded epochs to predict time spent in higher‐intensity activities. Specifically, models predicted time spent in levels 3–5 activities (i.e., very light to moderate/high intensity vs. sedentary). Including the accelerometer data as the sum of VM counts over all epochs resulted in the best fit, with models including time spent in higher‐level activities using the thresholds defined in this current study having similar and acceptable fits. Results support that the sum of VM counts over all epochs, as well as time spent in levels 3–5 activities, are acceptable metrics for estimating total time spent in nonsedentary behaviours over a specified time period for ambulatory and nonambulatory children in this age range.
While the performance of ankle accelerometry in our study was acceptable (e.g., similar to the performance of wrist‐worn accelerometry in measuring time spent sedentary among children [Van Loo et al. 2017)], it was not high. This is likely due to the considerable interchild variability in accelerometry data within intensity levels for young children, as we and other studies among toddlers (Trost et al. 2012; Hager et al. 2016) have observed. Such high interchild variability further explains while trends in median VM are observed across increasing CARS levels, classification at each level is poor. Another limitation is that the waist worn device was not secured directly above the right hip as done in past studies of preschool‐age children, which limits generalizability of our findings to a device specifically placed at the hip (Trost et al. 2012; Costa et al. 2014). However, our findings are generalizable to studies that use this device with an elastic waist band provided by the manufacturer.
Strengths of this study include enrolling infants and toddlers, the first accelerometry validation study to our knowledge to include among infants (≥ 6 months) and young toddlers who are not yet walking independently as they participated in naturalistic play activities. Our semistructured protocol encouraged a range of activities while having a high ecological validity regarding children's natural movements. Our stepped modelling approach reflects accelerometer performance relative to how physical activity is often classified in practice (i.e., sedentary, light, or moderate‐to‐vigorous). A limitation of our study is the small sample size of nonambulatory children enrolled (n = 15) and the limited number of epochs (77 epochs in total) in levels 4–5 activities for nonambulatory children. That may explain the reduced model performance among nonambulatory children when classifying levels 4–5 behaviours.
Aspects of our study design limit generalizability to naturalistic settings. Our study excluded epochs where children were playing with or moving the accelerometer devices and epochs where the child was being picked up by an adult. In practice, available accelerometry methods cannot differentiate those movements from the child's own movements. Future studies incorporating heart rate monitoring may offer new insights into infant and toddler physical exertion (Claiborne et al. 2023). It is further unknown how the performance of wrist‐worn accelerometry may compare to these study findings. Additionally, while model results were similar for the devices worn at the left and right ankle, it may be preferrable to use the left ankle for data collection at this age because this side is largely the nondominant side by 24 months of age (Nelson, Campbell, and Michel 2013).
In summary, study results are reassuring that ankle‐worn devices have acceptable validity in classifying epochs as sedentary or nonsedentary behaviour for young children from age 6 to 24 months, even among those not fully able to walk independently. These methods greatly add to the field by providing a validated method for research studies that aim to measure in‐the‐moment time spent sedentary among infants and young toddlers. In‐the‐moment activity can be critical to measure when, for example, considering precedents of child physical activity (Lopez et al. 2019). Results also support that accelerometry can approximate the total amount of time spent in higher‐intensity physical activities among infants and toddlers by aggregating accelerometer data over a defined time period, with the sum of VM over all epochs demonstrating the best performance in this current dataset. That time period was over the lab visit in our study but could be generalized to a day in naturalistic studies. Caution is warranted when extrapolating these findings to naturalistic settings when infants and toddlers may experience external forces (i.e., being carried or in a stroller). However, results offer more insight into the potential of accelerometry to measure physical activity behaviours at a young age.
Author Contributions
Hannah J. Dorris: writing – original draft, methodology. Delaina D. Carlson: project administration, formal analysis. Grace A. Ballarino: project administration, formal analysis. Nanette V. Lopez: methodology. Jennifer A. Emond: conceptualization, funding acquisition, writing – original draft, methodology, formal analysis, project administration, resources.
Disclosure
The funder had no role in the conceptualization, design, data collection, analysis, decision to publish, or preparation of the manuscript.
Conflicts of Interest
The authors declare no conflicts of interest.
Supporting information
Table S1. Physical activity level definitions as adapted from the CARS. Levels coded per each 5‐s epoch.
Table S2. Median vector magnitude per 5‐s epoch by physical activity level.
Table S3. a. Performance when classifying physical activity level using accelerometer data from the left ankle wear location, overall and stratified by ambulatory status. b. Performance when classifying physical activity level using accelerometer data from the right ankle wear location, overall and stratified by ambulatory status. c. Performance when classifying physical activity level using accelerometer data from the waist wear location, overall and stratified by ambulatory status.
Table S4. Performance when classifying concatenated physical activity levels 4–5 versus 1–3 from each wear location.
Table S5. Performance of aggregate accelerometer metrics in modelling time spent in physical activity levels 4–5: left versus right ankle wear location.
Acknowledgements
The authors thank the families who participated in this research and the journal reviewers for their insightful feedback.
Funding: This study was supported by a grant from the National Institute of Diabetes and Digestive and Kidney Diseases (R03DK131154).
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.
References
- Bélanger, M. È. , Bernier A., Paquet J., Simard V., and Carrier J.. 2013. “Validating Actigraphy as a Measure of Sleep for Preschool Children.” Journal of Clinical Sleep Medicine 9, no. 7: 701–706. 10.5664/jcsm.2844. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Benjamin‐Neelon, S. E. , Bai J., Østbye T., Neelon B., Pate R. R., and Crainiceanu C.. 2020. “Physical Activity and Adiposity in a Racially Diverse Cohort of US Infants.” Obesity 28, no. 3: 631–637. 10.1002/oby.22738. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Claiborne, A. , Williams A., Jolly C., et al. 2023. “Methods for Analyzing Infant Heart Rate Variability: A Preliminary Study.” Birth Defects Research 115, no. 10: 998–1006. 10.1002/BDR2.2177. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Clevenger, K. A. , Brønd J. C., Mackintosh K. A., Pfeiffer K. A., Montoye A. H. K., and McNarry M. A.. 2022. “Impact of ActiGraph Sampling Rate on Free‐Living Physical Activity Measurement in Youth.” Physiological Measurement 43, no. 10: 105004. 10.1088/1361-6579/AC944F. [DOI] [PubMed] [Google Scholar]
- Cliff, D. P. , Reilly J. J., and Okely A. D.. 2009. “Methodological Considerations in Using Accelerometers to Assess Habitual Physical Activity in Children Aged 0‐5 Years.” Journal of Science and Medicine in Sport 12: 557–567. 10.1016/j.jsams.2008.10.008. [DOI] [PubMed] [Google Scholar]
- Costa, S. , Barber S. E., Cameron N., and Clemes S. A.. 2014. “Calibration and Validation of the ActiGraph GT3X+ in 2‐3 Year Olds.” Journal of Science and Medicine in Sport 17, no. 6: 617–622. 10.1016/j.jsams.2013.11.005. [DOI] [PubMed] [Google Scholar]
- Del Pozo‐Cruz, B. , Perales F., Parker P., et al. 2019. “Joint Physical‐Activity/Screen‐Time Trajectories During Early Childhood: Socio‐Demographic Predictors and Consequences on Health‐Related Quality‐of‐Life and Socio‐Emotional Outcomes.” International Journal of Behavioral Nutrition and Physical Activity 16, no. 1: 55. 10.1186/s12966-019-0816-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Ghazi, M. A. , Zhou J., Havens K. L., and Smith B. A.. 2024. “Accelerometer Thresholds for Estimating Physical Activity Intensity Levels in Infants: A Preliminary Study.” Sensors 24, no. 14: 4436. 10.3390/S24144436. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Hager, E. R. , Gormley C. E., Latta L. W., Treuth M. S., Caulfield L. E., and Black M. M.. 2016. “Toddler Physical Activity Study: Laboratory and Community Studies to Evaluate Accelerometer Validity and Correlates.” BMC Public Health 16, no. 1: 936. 10.1186/s12889-016-3569-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Hosmer, D. W. , Lemeshow S., and Sturdivant R. X.. 2013. Applied Logistic Regression. Third ed. Hoboken, NJ: John Wiley & Sons. 10.1002/9781118548387. [DOI] [Google Scholar]
- Li, R. , O'Connor L., Buckley D., and Specker B.. 1995. “Relation of Activity Levels to Body Fat in Infants 6 to 12 Months of Age.” Journal of Pediatrics 126, no. 3: 353–357. 10.1016/S0022-3476(95)70447-7. [DOI] [PubMed] [Google Scholar]
- Lopez, N. V. , Yang C. H., Belcher B. R., Margolin G., and Dunton G. F.. 2019. “Within‐Subject Associations of Maternal Physical Activity Parenting Practices on Children's Objectively Measured Moderate‐to‐Vigorous Physical Activity.” Journal of Pediatric Psychology 44, no. 3: 300–310. 10.1093/JPEPSY/JSY100. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Nelson, E. L. , Campbell J. M., and Michel G. F.. 2013. “Unimanual to Bimanual: Tracking the Development of Handedness From 6 to 24 Months.” Infant Behavior & Development 36, no. 2: 181–188. 10.1016/J.INFBEH.2013.01.009. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Pate, R. R. , Almeida M. J., McIver K. L., Pfeiffer K. A., and Dowda M.. 2006. “Validation and Calibration of an Accelerometer in Preschool Children.” Obesity 14, no. 11: 2000–2006. 10.1038/oby.2006.234. [DOI] [PubMed] [Google Scholar]
- Pate, R. R. , Frongillo E. A., Cordan K., et al. 2020. “Linking Activity, Nutrition, and Child Health (LAUNCH): Protocol for a Longitudinal Cohort Study of Children as They Develop From Infancy to Preschool Age.” BMC Public Health 20, no. 1: 931. 10.1186/s12889-020-09023-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Physical Activity Guidelines Advisory Committee . 2018. ‘2018 Physical Activity Guidelines Advisory Committee Scientific Report‘, accessed April 22, 2024. https://health.gov/our‐work/nutrition‐physical‐activity/physical‐activity‐guidelines/current‐guidelines/scientific‐report.
- Prioreschi, A. , and Micklesfield L. K.. 2016. “A Scoping Review Examining Physical Activity Measurement and Levels in the First 2 Years of Life.” Child: Care, Health and Development 42, no. 6: 775–783. 10.1111/cch.12382. [DOI] [PubMed] [Google Scholar]
- Puhl, J. , Greaves K., and Hoyt M.. 1990. “Children's Activity Rating Scale (Cars): Description and Calibration.” Research Quarterly for Exercise and Sport 61, no. 1: 26–36. 10.1080/02701367.1990.10607475. [DOI] [PubMed] [Google Scholar]
- Sadeh, A. , Acebo C., Seifer R., Aytur S., and Carskadon M. A.. 1995. “Activity‐Based Assessment of Sleep‐Wake Patterns During the 1st Year of Life.” Infant Behavior & Development 18, no. 3: 329–337. 10.1016/0163-6383(95)90021-7. [DOI] [Google Scholar]
- Timmons, B. W. , Leblanc A. G., Carson V., et al. 2012. “Systematic Review of Physical Activity and Health in the Early Years (Aged 0‐4 Years).” Applied Physiology, Nutrition, and Metabolism 37, no. 4: 773–792. 10.1139/H2012-070. [DOI] [PubMed] [Google Scholar]
- Trost, S. G. , Fees B. S., Haar S. J., Murray A. D., and Crowe L. K.. 2012. “Identification and Validity of Accelerometer Cut‐Points for Toddlers.” Obesity 20, no. 11: 2317–2319. 10.1038/oby.2011.364. [DOI] [PubMed] [Google Scholar]
- Van Kooten, J. A. M. C. , Jacobse S. T. W., Heymans M. W., De Vries R., Kaspers G. J. L., and Van Litsenburg R. R. L.. 2021. “A Meta‐Analysis of Accelerometer Sleep Outcomes in Healthy Children Based on the Sadeh Algorithm: The Influence of Child and Device Characteristics.” Sleep 44, no. 4: zsaa231. 10.1093/SLEEP/ZSAA231. [DOI] [PubMed] [Google Scholar]
- Van Loo, C. M. T. , Okely A. D., Batterham M. J., et al. 2017. “Wrist Accelerometer Cut Points for Classifying Sedentary Behavior in Children.” Medicine and Science in Sports and Exercise 49, no. 4: 813–822. 10.1249/MSS.0000000000001158. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Wells, J. C. K. , and Ritz P.. 2001. “Physical Activity at 9‐12 Months and Fatness at 2 Years of Age.” American Journal of Human Biology 13, no. 3: 384–389. 10.1002/ajhb.1062. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Table S1. Physical activity level definitions as adapted from the CARS. Levels coded per each 5‐s epoch.
Table S2. Median vector magnitude per 5‐s epoch by physical activity level.
Table S3. a. Performance when classifying physical activity level using accelerometer data from the left ankle wear location, overall and stratified by ambulatory status. b. Performance when classifying physical activity level using accelerometer data from the right ankle wear location, overall and stratified by ambulatory status. c. Performance when classifying physical activity level using accelerometer data from the waist wear location, overall and stratified by ambulatory status.
Table S4. Performance when classifying concatenated physical activity levels 4–5 versus 1–3 from each wear location.
Table S5. Performance of aggregate accelerometer metrics in modelling time spent in physical activity levels 4–5: left versus right ankle wear location.
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.