Physical activity is defined as any bodily movement produced by skeletal muscles that results in energy expenditure (EE) [
Valid and reliable instruments are necessary, when examining the dose-response relationship between physical activity and health-related fitness [
Estimation of physical activity and energy expenditure in children is difficult since children show physical activities of varying intensity and short of duration [
Direct observation is considered a gold standard for the assessment of physical activity. Gold standard methods to assess activity related energy expenditure (AEE) are doubly labelled water and indirect calorimetry [
Accelerometry is objective as well as cost-effective and less invasive. Accelerometers have evolved from simple mechanical instruments to electronically three-dimensional instruments to assess physical activity and energy expenditure. An accelerometer estimates accelerations produced by movement of a body segment or limb parts [
In literature different prediction models are described to assess AEE in children and adolescents. The aim of this study is to review the validity and generalizability of accelerometry based prediction models to estimate AEE in children and adolescents.
Electronic bibliographic databases CINAHL, EMBASE, PsycINFO, and PubMed/MEDLINE were searched till April 2009. The following MeSH terms and text words were used:
Studies (written as full reports) were included in this review if their main purpose was to develop and/or validate an accelerometry based prediction model for the estimation of AEE in healthy children and/or adolescents (6–18 years). The AEE predicted by the model, had to be compared with a criterion measure of AEE as doubly labelled water or indirect calorimetry. Studies written in Dutch, English, and German were included. Studies concerning pedometers were excluded.
One researcher (SdG) performed the search strategy. The first selection regarding relevance, based on title and abstract, was performed by two independent researchers (SdG and MS). Furthermore, the included articles were judged on full-text by these two independent researchers. References of the included articles were screened for additional eligible studies.
To evaluate and compare the studies, data were extracted. Two reviewers independently extracted the data (SdG and MS). Disagreements between the two reviewers regarding a study’s eligibility were resolved by discussion until consensus was reached or, when necessary, a third person (JdG) acted as adjudicator.
The data extraction was based on items that have an impact on the range and generalizability of a prediction model according to Puyau et al. [
An existing checklist [
The literature search identified 438 studies, after judgement based on title and abstract 39 studies remained (see Figure
Selection process for studies included in the review.
In total eight studies were selected as eligible [
All included studies had a cross-sectional research design. In total twenty-eight different prediction models were described. Two studies assessed the generalizability of previously published prediction models [
The included studies described six different accelerometers; two omnidirectional (Actical, Actiwatch), two uniaxial (Actigraph/CSA, Caltrac) and one triaxial (RT3). For the Actiheart this property could not be retrieved from the included studies.
The score on the checklist regarding methodological issues ranged from 5.0 to 8.0 (median
Prediction models ordered by accelerometer.
Accelerometer | Activities | Criterion | Prediction models & Statistics |
---|---|---|---|
Flat walking, graded walking, and running on a treadmill. | Indirect calorimetry | Corder et al. [ | |
- Three sitting activities: handwriting, card sorting, and Video game playing. | Indirect calorimetry | Heil et al. [ | |
Playing Nintendo, using a computer, cleaning, aerobic exercise, ball toss, treadmill walking, and running. | Room respiration calorimetry 4 h, Indirect calorimetry 1 h. | Puyau et al. [ | |
Flat walking, graded walking, and running on a treadmill. | Indirect calorimetry | Corder et al. [ | |
Six activities, each activity lasted 5 minutes: | Indirect calorimetry | Corder et al. [ | |
Free-living; Two school weeks, 14 consecutive days, the children wore the monitor during daytime following their normal living. Exceptions were during water activities such as swimming and bathing. | Doubly labelled water | Ekelund et al. [ | |
Sedentary: Nintendo, arts and crafts, playtime 1 | Room respiration calorimetry | Puyau et al. [ | |
Field conditions; flat oval indoor track. Normal walking, brisk walking, easy running, fast running. | Indirect calorimetry | Trost et al. [ | |
Flat walking, graded walking, and running on a treadmill (protocol). | Indirect calorimetry | Corder et al. [ | |
Six activities, each activity lasted 5 minutes: | Indirect calorimetry | Corder et al. [ | |
Sedentary: Nintendo, arts and crafts, playtime 1 | Room respiration calorimetry | Puyau et al. [ | |
Playing Nintendo, using a computer, cleaning, aerobic exercise, ball toss, treadmill walking and running. | Room respiration calorimetry 4 h, Indirect calorimetry 1 h. | Puyau [ | |
Free-living; Three days, including one weekend day. The subjects began wearing the Caltrac upon waking in the morning and continued until just before going to sleep at night. The Caltrac was taken off for activities involving water, such as swimming or bathing. | Doubly labelled water | Johnson et al. [ | |
Indoor: laying down, sitting relaxed, writing, standing relaxed, sitting and standing (alternating every 5 s), cycling, stepping up and down, walking. The speed of treadmill was predetermined so that most children could complete jogging on the treadmill. | Indirect calorimetry | Sun et al. [ |
Abbreviations: AC: Accelerometer Counts, adj.: adjusted, AEE: Activity related Energy Expenditure, B&A: Bland & Altman, bpm: beats per minute, CI: Confidence Interval, FFM: Fat Free Mass, FM: Fat Mass, g: gram, h: hour, Hz: hertz, J: Joule, kcal: Kilocalorie, kg: kilogram, min: minute, r= correlation coefficient, RMSE: Root Mean Squared Error, SEE: Standard Error of the Estimate, SSE: Sum of Squared Errors.
Eleven prediction models regarding the Actical were derived in laboratory settings based on activities as handwriting, cleaning, playing a video game/Nintendo, and walking at different speeds and grades (treadmill and indoor track) [
For the Actigraph/CSA accelerometer, seven models were derived. One model was based on free-living data compared to doubly labelled water [
The model by Ekelund et al. [
The studies of Corder et al. [
The studies of Puyau et al. [
The study of Johnson et al. [
Two prediction models were derived by the study of Sun et al. [
Corder et al. [
Trost et al. [
Mean bias on ratio scale was 1.33 (a difference between measured and predicted AEE of +33%). The corresponding limits of agreement were 0.44–2.22. Thus for any individual in the population, AEE values predicted by the Puyau et al. [
This review shows that accelerometer-based prediction models can explain up to 91% [
The difference found between laboratory-based models and free-living models might be explained by the derivation activities and the limitations of accelerometers. AEE predicted by a linear model, is likely to be more accurate when this model is derived and applied on a limited set of structured activities such as running and walking [
Our findings suggest that the accuracy of the prediction model seems improved when a triaxial accelerometer is used. A triaxial accelerometer captures more movements than uniaxial and omnidirectional accelerometers. In the review of Westerterp [
Models that included heart rate explained 86–90% of the variance in measured AEE [
The generalizability of the models is however limited and seems mainly dependent on the derivation activities. Nilsson et al. [
Free-living studies are most likely to represent actual daily activities performed by children. The laboratory-placed studies, included in this review, attempted to represent these activities by including locomotion activities [
Free-living studies estimate AEE by subtracting REE from total energy expenditure (TEE) provided by the doubly labelled water method [
Measurement is more accurate since REE can vary with age, maturation, body mass, and level of physical activity [
Implication for clinicians is that previously published prediction models have limited applicability. Laboratory-based models can be used, on a group level, to predict AEE during specific activities, similar to the derivation activities. The use of a model combining accelerometer counts and heart rate, or a model combining triaxial accelerometer counts with body weight enhances validity. Generalizability of the models during free-living, however, is very limited. This is a significant limitation because measurement during free living is important to examine the dose-response relationship between physical activity and health-related fitness. The model derived by Ekelund et al. can be used, on a group level, for the prediction of AEE during free-living in 9-year-old children [
Future development of prediction models applicable to free-living scenarios is needed. Future free-living studies should concern prediction models combining accelerometer counts and heart rate, or the counts of a triaxial accelerometer. As stated by Corder et al. especially the combination of accelerometer counts and heart rate might provide a more accurate and widely applicable model [
Regarding the reporting of findings, future recommendation is the description of the correlation between counts and measured AEE, since the counts are part of the prediction model. The limitations of the accelerometer itself may cause less accuracy, and therefore a less accurate prediction of AEE by the model.
To assess feasibility, authors should also report the amount missing and lost data due to malfunctioning of the motion sensor. Regarding free-living studies is additionally the refusal rate, or compliance rate with wearing the motion sensor interesting for clinicians.
Accelerometry based prediction models may provide an accurate estimate of AEE in children on a group level. The estimation of AEE is more accurate when the model is derived (and used) in a laboratory setting. The best results are retrieved when the model combines accelerometer counts with heart rate or when a triaxial accelerometer is used. Generalizability of the models during free-living however is limited. Future development of equations applicable during free-living is needed.
There are no professional relationships with companies or manufacturers who will benefit from the results of the present study.
Evaluation checklist for studies on prediction models for accelerometers
Source; Clinimetric review of motion sensors in children and adolescents. Journal of Clinical Epidemiology 59 (2006) 670–680. De Vries SI, Bakker I, Hopman-Rock M, Hirasing RA, van Mechelen W.
Modified by S. de Graauw, 2009.
See Tables
Items concerning study design.
1 | ||
0.5 | 4-5 sample characteristics are described | |
0 | ||
1 | Information on setting, activities, duration (days or hours), and period of wearing the motion sensor | |
0.5 | Information on period of wearing the motion sensor is missing | |
0 | Not clear at all | |
1 | Complete information on motion sensor (type, output, epoch, placement) and reference method(s) (type, output) | |
0.5 | Some information on motions sensor (type, output, epoch, placement) and reference method(s) (type, output) is missing | |
0 | Very limited information on motion sensor (type, output, epoch, placement) and reference method(s) (type, output) | |
1 | Complete information on statistical analysis (tests, subgroup analysis), statistical software package and | |
0.5 | Some information on statistical analyses (tests, subgroup analysis), statistical software package and | |
0 | Very limited information on statistical analysis (tests, subgroup analysis), statistical software package and |
Items concerning validity.
1 | Yes | |
0 | No | |
1 | Sensitivity | |
1 | Specificity | |
1 | Pearson’s product-moment correlation coefficient | |
1 | Spearman’s rank order correlation coefficient | |
0.5 | 95% limits of agreement (Bland and Altman) | |
0 | Other measure | |
+ | ||
1 | Yes | |
0 | No | |
1 | Intraclass correlation coefficients | |
1 | 95% limits of agreement (Bland and Altman) | |
1 | Cohen’s Kappa | |
1 | Standard error of measurement | |
1 | Coefficient of variation | |
0 | Pearson’s product-moment correlation coefficient | |
0 | Spearman’s rank order correlation coefficient | |
0 | Kendall’s tau | |
0 | Other measure | |
+ | ||
Items concerning feasibility.
1 | Yes | |
0 | No | |
+ | ||
1 | Yes | |
0 | No | |
+ | ||
15–30% | ||
See Table
Data of included studies.
Study | Population | Score checklist (out of 10) | Setting | Accelerometer (placement) | Prediction model(s) | Conclusion authors |
---|---|---|---|---|---|---|
Corder et al. [ | 39 children aged | 7.5 | Laboratory setting. | Six prediction models were derived, one not consisting of accelerometer counts, this one was excluded. | Corder et al. concluded that the combined HR and activity monitor Actiheart is valid for estimating AEE in children during treadmill walking and running. The combination of HR and activity counts provides the most accurate estimate of AEE as compared with accelerometry measures alone. | |
Corder et al. [ | 145 children aged | 7.5 | Laboratory setting. | Five previously published prediction models (Coder et al. [ | Corder et al. concluded that the ACC and HR + ACC can both be used to predict overall AEE during these six activities in children; however, systematic error was present in all predictions. Although both ACC and HR + ACC provides accurate predictions of overall AEE, according to the activities in their study, Corder et al. concluded that AEE-prediction models using HR + ACC may be more accurate and widely applicable than those based on accelerometry alone. | |
Ekelund et al. [ | 26 children aged | 8.0 | Free-living. | One prediction model derived. | Ekelund et al. concluded that activity counts contributed significantly to the explained variation in TEE and was the best predictor of AEE. Their cross-validation study showed no significant differences between predicted and measured AEE. | |
Heil et al. [ | 24 children: 14 | 5.5 | Laboratory setting. | Nine prediction models derived. | Heil et al. concluded that the proposed algorithms for the Actical appeared to predict AEE accurately whether worn at the ankle, hip or wrist. Additionally they state that their results however, are clearly limited by the laboratory nature of the data collection and need to be validated under free-living conditions. | |
Johnson et al. [ | 31 children aged | 5.0 | Free-living. | Sallis et al. 1989 equation; originally validated against HR, thus excluded in this study. One prediction model derived. | According to Johnson et al. their study failed to find a significant correlation between either activity counts and AEE or Caltrac average calories with AEE. Their major finding was that the Caltrac accelerometer was not a useful predictor of AEE in the sample. | |
Puyau et al. [ | 26 children 14 | 5.5 | Laboratory setting. | Four prediction models were derived. | Puyau et al. concluded that the high correlations between the activity counts and AEE demonstrates that the CSA and Actiwatch monitors strongly reflected energy expended in activity. Given the large SEE of the regression of AEE on activity counts, they found the prediction of AEE from CSA of Actiwatch activity counts inappropriate for individuals. | |
Puyau et al. [ | 32 children aged 7–18 yr, 14♂, 18 | 5.5 | Laboratory setting. | Two models derived. | Puyau et al. concluded that activity counts accounted for the majority of the variability in AEE with small contributions of age, sex, weight, and height. Overall, Actiwatch equations accounted for 79% and Actical equations for 81% of the variability in AEE. Relatively wide 95% prediction intervals for AEE showed considerable variability around the mean for the individual observations. Puyau et al. suggest that accelerometers are best applied to groups rather than individuals. | |
Sun et al. [ | 27 children aged 12–14 yr, 21 | 8.0 | Laboratory setting. | Two models derived and manufacturer’s model was used. Since the manufacturer’s model is not revealed it was excluded. | Sun et al. concluded that the results of their study show that the RT3 accelerometer provides a valid method to examine physical activity patterns qualitatively and quantitatively for children. The moderate to high correlation coefficients between the physical activities in various lifestyle conditions from this device and the metabolic costs in simulated free-living conditions strongly supports, according to Sun et al., that the RT3 accelerometer serves as a valid, objective measure of physical activity of children, even in a tropical environment such as Singapore. | |
Trost et al. [ | 45 children aged | 5.5 | Laboratory setting. | Validation of three models. Two models not concerning AEE were excluded. The model by Puyau et al. [ | Trost et al. concluded that previously published ActiGraph equations developed specifically for children and adolescents do not accurately predict AEE on a minute-by-minute basis during overground walking and running. |
Abbreviations; ACC: Accelerometer, AEE: Activity related Energy Expenditure, HR: Heart Rate, SEE: Standard Error of the Estimate, TEE: Total Energy Expenditure, yr: year.
The authors would like to thank the following organizations for their financial support: Royal Dutch Society for Physical Therapy (KNGF), Bio Foundation (Stichting Bio Kinderrevalidatie), and Wilhelmina Children’s Hospital-Fund (WKZ-Fonds).