Calibration of Accelerometer Output
PAMTY FREEDSON', DAVID POBER', and KATHLEEN F. JANZ2
'Department of Exercise Science, University of Massachusetts, Amherst, MA; and 2Department of Health and Sports
Studies, University of Iowa, Iowa City, IA
FREEDSON, P., D. POBER, and K. F. JANZ. Calibration of Accelerometer Output for Children. Med. Sci. Sports Exerc., Vol. 37, No.
1 I(Suppl), pp. S523-S530, 2005. Understanding the determinants of physical activity behavior in children and youths is essential to
the design and implementation of intervention studies to increase physical activity. Objective methods to assess physical activity
behavior using various types of motion detectors have been recommended as an alternative to self-report for this population because
they are not subject to many of the sources of error associated with children's recall required for self-report measures. This paper
reviews the calibration of four different accelerometers used most frequently to assess physical activity and sedentary behavior in
children. These accelerometers are the ActiGmph, Actical, Actiwatch, and the RT3 Triaxial Research Tracker. Studies are reviewed
that describe the regression modeling approaches used to calibrate these devices using directly measured energy expenditure as the
criterion. Point estimates of energy expenditure or count ranges corresponding to different activity intensities from several studies are
presented. For a given accelerometer, the count cut points defining the boundaries for 3 and 6 METs vary substantially among the
studies reviewed even though most studies include walking, running and free-living activities in the testing protocol. Alternative data
processing using the raw acceleration signal is recommended as a possible alternative approach where the actual acceleration pattern
is used to characterize activity behavior. Important considerations for defining best practices for accelerometer calibration in children
and youths are presented. Key Words: PHYSICAL ACTIVITY MEASUREMENT, MOTION SENSORS, YOUTHS
he accurate measurement of free-living physical ac-
tivity (counts) is essential for research studies in
which physical activity (PA) is an outcome or expo-
sure of interest. PA surveillance and observational studies of
the association between PA and health outcomes require a
robust activity measure to establish accurate estimates of the
dose of activity needed for specific outcomes. PA interven-
tion projects also require an accurate activity measure to
determine the effectiveness of the intervention. Finally, ac-
curate assessments of PA are necessary if the physiologic
mechanisms linking PA and health are to be completely
elucidated. Although self-report methods are often the mea-
surement of choice, particularly in large-scale epidemiolog-
ical investigations, they are frequently prone to error. As a
result, researchers have explored a variety of objective
methods to assess PA in order to validate self-reports and
reduce the errors associated with these types of measure-
ment schemes, particularly when children are the subjects of
The most popular objective measurement device has been
the accelerometer, which evaluates both PA quantity and
quality. Accelerometer devices are designed with large
memory storage so that several days or weeks of activity can
Address for correspondence: Patty Freedson, Department of Exercise Sci-
ence, University of Massachusetts/Amherst, 30 Eastman Lane, Amherst,
MA 01003; E-mail: firstname.lastname@example.org.
MEDICINE & SCIENCE IN SPORTS & EXERCISE,
Copyright © 2005 by the American College of Sports Medicine
be assessed in small sampling intervals (e.g., seconds and
minutes). These monitors are relatively small and light-
weight making them unobtrusive and practical for extended
measurement periods. The small size makes them particu-
larly appealing for use in children. Typically they are worn
on the wrist, the lower leg, or the hip where they are
attached to a belt or band.
In recent years, accelerometers have been widely used to
characterize PA behavior in children. However, minimal
attention has been directed toward standardizing methods of
data collection, processing, and interpretation. One of the
fundamental questions critical to understanding the meaning
of PA assessed by accelerometry is how to translate and
interpret the accelerometer signal into meaningful data
linked to physiological outcomes or, in some cases, behav-
Acceleration is defined as the change in velocity over
time, and, as such, it quantifies the volume and intensity of
movement. Volume and intensity are dimensions of PA
needed by researchers to understand dose response. How-
ever, the raw acceleration signal is typically not used to
directly quantify PA. In most cases the raw acceleration
signal is translated, or calibrated, into a metric that is an-
chored to some biological variable (e.g., energy expendi-
ture, HR) or to specific PA patterns (e.g., stationary or
ambulatory). This approach gives the raw signal biological
or behavioral meaning. Generally, the biological meaning
for the raw acceleration signal is energy expenditure or
oxygen consumption. Typically, calibration studies derive
point estimates of energy expenditure from activity counts
using regression modeling. Alternatively, regression analy-
sis is used to establish ranges of accelerometer counts (cut
points) corresponding to predefined intensity levels. These
methods of data interpretation provide the type of informa-
tion needed for meaningful application in PA research.
Accelerometer technology is based on biomechanical
principles while energy expenditure and oxygen consump-
tion are biological measures; therefore, it is not surprising
that calibrating acceleration signals to these latter variables
is challenging. This task is particularly difficult in children
and adolescents because growth and maturation confound
associations. For example, the metabolic cost of movement,
expressed relative to body mass (mL.kg-1.min-'), de-
creases as children mature (4). Activity energy expenditure
(kcal.d-1) is influenced by body mass and the metabolic
economy of movement resulting in heavier children expend-
ing more energy at the same speed of movement than peers
(29). In addition, the sensitivity of the acceleration signal is
influenced by the interplay of stride length and step fre-
quency so that at the same speed of movement the signal is
lower when the step frequency is higher (2). Even the
distance of the accelerometer from the center of gravity
when the same positioning strategy is employed (e.g., at the
hip) influences the acceleration signal. That is, the acceler-
ation signal is greater when the accelerometer is further
from the center of gravity, as is the case with heavier
children when compared with peers (32). One solution to
address these multiple issues of confounding is population-
specific equations. However, a more parsimonious solution
is modeling that incorporates easily measured variables that
proxy for more difficult to measure attributes. In many
situations, age can serve as a viable proxy for maturity while
height and weight can proxy for stride length and distance
from center of mass.
Using energy expenditure as the criterion for calibrating
raw accelerometry signals is appropriate when the research
examines obesity, heart disease risk factors, type 2 diabetes,
and other metabolic risks. However, when bone outcomes
are of interest, researchers should consider calibrating the
raw signal to ground reaction forces (newtons) measured
using force plates. Because acceleration is equal to force
divided by mass, accelerometer devices might be particu-
larly well suited for capturing bone strain and aid in our
understanding of how mechanical overload contributes to
healthy bone development in children.
This paper reviews and critiques the literature on accel-
erometer calibration in children and discusses the implica-
tions of these findings for future research on objective
assessment of PA in children.
Calibration studies have been conducted using several
accelerometers but the majority of the literature has focused
on three accelerometer devices: ActiGraph (formerly known
as Computer Science and Applications (CSA) and Manu-
facturing Technology, Inc. (MTI) (ActiGraph, LLC, Fort
Walton Beach, FL)), Actical and Actiwatch (Mini Mitter
Co., Inc., Bend, OR), and the RT3 Triaxial Research
Tracker (formerly known as the Tritrac-R3D (StayHealthy,
Official Journal of the American College of Sports Medicine
Inc., Monrovia, CA)). The ActiGraph is a uniaxial acceler-
ometer, the Actical and Actiwatch devices are omnidirec-
tional accelerometers and the RT3 is a triaxial accelerome-
ter. This paper focuses on these three devices because they
have been the primary monitors used in calibration studies
in children and adolescents.
ActiGraph model 5032. Janz (12) published one of
the first quasi-calibration studies of an accelerometer in 7- to
15-yr-old children. This investigation used the first gener-
ation of the ActiGraph accelerometer (Model 5032). The
approach Janz employed to establish the accelerometer
counts representing vigorous PA was to examine the distri-
bution of counts per minute data for the sample. The 80th
percentile (256 counts per minute) was defined as the lower
bound for vigorous PA. The average number of minutes in
this predefined vigorous activity category ranged from 117
to 144 min, and the number of minutes where HR was above
150 bpm ranged from 20 to 29 min during 3 d of monitoring.
Thus, it appears that using the 80th percentile for counts per
minute to represent vigorous activity may not have been
appropriate. Nevertheless, correlations between minutes
where HR was above 150 bpm and minutes of counts per
minute above 256 ranged from r = 0.50 to 0.70. This study
was the first to employ the cut point method to establish a
count threshold for intensity of activity and set the stage for
future calibration studies in children.
Caltrac personal activity computer (Muscle Dy-
namics, Torrence, CA). As suggested earlier, in most
instances, accelerometers should be calibrated using oxygen
consumption or energy expenditure as the criterion variable.
In fact, the original accelerometer used in PA research was
the Caltrac, which used measured energy expenditure as the
criterion (17). Algorithms were developed for adults in the
laboratory to translate acceleration into estimates of activity
energy expenditure. Sallis et al. (25) used oxygen consump-
tion measured during treadmill walking and running to
calibrate the Caltrac for children (N = 15). They reported
that one Caltrac count was equivalent to a net energy ex-
penditure of 0.101 kcal.kg-1 and the correlation between
activity counts and net energy expenditure was r = 0.89.
ActiGraph model 7164. The first children's calibra-
tion study on this monitor was conducted by Trost et al.
(31). They developed an equation to predict energy expen-
diture from activity counts using laboratory treadmill exer-
cise. Thirty children between the ages of 10 and 14 wore the
ActiGraph accelerometer and completed treadmill exercise
at 3, 4, and 6 mph with oxygen consumption serving as the
criterion measure. The equation was developed on 20 sub-
jects and was cross-validated on the remaining 10 subjects
(r = 0.93, SEE = 0.93 kcal.min-').
Freedson et al. (9) developed a regression equation to
estimate METs from counts and age where 6- to 18-yr-old
children and adolescents completed two treadmill walking
speeds and one running speed. Respiratory gas exchange
was measured using indirect calorimetry and the ActiGraph
was worn on the hip and programmed to collect minute-by-
minute counts. Resting energy expenditure was estimated
0L A £A A
tritrac total magnaitude
1o0o0 12000 50700
FIGURE l-Relationship between activity counts and oxygen con-
sumption. Reprinted with permission from Eston et at. (7).
from age specific prediction equations to derive the meta-
bolic equivalent of MET intensity levels. The equation was
METs =2.757 + (0.00 15 * counts per minute) - (0.08957 * age (yr))
- (0.000038 * counts per minute * age (yr))
R2= 0.74 SEE= 1.1 METs
Accelerometer calibration studies should be performed in a
setting where a variety of activities are included to closely
represent the broad repertoire of activities that children typi-
cally perform. For example, Eston et al. (7) had children (ages
8.2-10.8) walk and run on a treadmill, play catch, play hop-
scotch, and color with crayons. Oxygen consumption was
measured, and PA was assessed with the ActiGraph acceler-
ometer. Although this study was designed as a validity evalu-
ation, data were presented in a style where calibration issues
could be examined. In Figure 1, the relationship between
accelerometer counts and oxygen consumption is illustrated.
The oxygen consumption data are scaled to body weight°'75,
making it difficult to determine MET values corresponding to
selected count values. However, using the average body weight
of 29.8 kg, the upper boundaries for 3, 6, and 9 METs corre-
spond to approximately 500, 4000, and 7600 counts per minute
for the ActiGraph.
In another field-based study, a broad range of intensities of
11 physical activities (sedentary to vigorous) were used to
calibrate the ActiGraph in 74 girls ages 13-14 (30). Energy
ACTIVITY MEASUREMENT IN CHILDREN AND YOUTHS
Sed Light Mod
FIGURE 2-Activity count cut points (counts per minute) for seden-
tary, moderate, and vigorous activity for the CSA (NITI) ActiGraph.
Cut points are the upper boundaries for the intensity categories.
expenditure was assessed at rest and during the activity proto-
col with a portable indirect calorimetry data acquisition unit
while an ActiGraph was worn on each hip. A combination of
regression and examination of error rates (e.g., low counts for
moderate-intensity activity or high counts for low-intensity
activity) were used to establish final cut points designating
activity intensity categories. These cut points are illustrated in
Figure 2. Also shown in Figure 2 are the cut points for intensity
categories from a study by Puyau et al. (20) for 26 children
between the ages of 6 and 16. The boundaries for defining
intensity categories are higher than those reported by Treuth et
al. (30), particularly for the upper limit for sedentary and
moderate activity. The calorimetry methods, the age range of
the subjects, the number of subjects per age range, and the
menu of activities between these two studies are quite different
and may explain the dissimilarity in cut point values. In
addition to exploring movement count to energy expendi-
ture relationships in free-living situations, both studies are
notable in that they directly measured resting energy expen-
diture to establish MET levels. This is important because
children and adolescents have a higher resting metabolic
rate (RMR) than adults; therefore, using the adult standard
of 3.5 mL.kg- 1.min- 1 to represent RMR would be expected
to introduce systematic error.
In addition to providing accelerometer count ranges to
define intensity levels, Treuth et al. (30) and Puyau et al.
(20) derived prediction equations to estimate METs or ac-
tivity energy expenditure from counts. The equations and
pertinent statistics are:
METs = 2.01 + 0.000856 (counts per minute)
R'=0.84 SEE =1.36 METs 
Activity energy expenditure or AEE (kcal'kg-'min-)
0.0183 + 0.000010 (counts per minute)
R2 = 0.75 SEE = 0.0172 kcal'kg-l"min-'
Figure 3 presents the relationship between oxygen con-
sumption and ActiGraph counts, including all of the activ-
ities from the Treuth et al. (30) and Puyau et al. (20) studies.
Medicine & Science in Sports & Exercisee
FIGURE 3-Relationship between Acti-
Graph counts per minute and oxygen con-
sumption for Treuth et al. (30) and Puyau et
*Trueth, at al
mPuyau, et al!
Because the data of Puyau et al. (20) were presented in
energy expenditure units, oxygen consumption was esti-
mated using a 5 kcal.L-1 V02 constant and their mean body
mass of 40 kg. Using the combined data, we developed a
composite regression equation between counts and V02.We
did not include the data point for cycling from Treuth et al.
(30) because the count value was extremely low for a
relatively high V02. This was expected given the poor
association between cycling and accelerometer output when
devices are worn on the hip. We predicted counts at MET
levels of 3, 6, and 9 using I MET = 3.8 mL.kg-'.min-'.
The resting V,02 of 3.8 mL.kg- 1-min-' directly measured in
Treuth et al. (30) was used as the baseline V/02. The equa-
tion to predict •/02 (mL.kg-1.min-1) is as follows:
O2(mL.kg-'.min-') = 7.7104 + 0.002631974 (counts per minute) 
Using these data, the counts per minute values (rounded off)
corresponding to 3, 6, and 9 METs are 1400, 5700, and
Schmitz et al. (27) reported that the slope of the line
describing the relationship between ActiGraph counts and
energy expenditure decreases as activity intensity increases.
Thus, when accelerometer count data are used to derive
point estimates of energy expenditure using simple linear
models, a systematic under estimation of energy expenditure
during low-intensity activities and an overestimation of en-
ergy expenditure during high-intensity exercise will occur.
To overcome this limitation, Schmitz et al. (27) developed
a model that includes both linear and quadratic terms to
account for the change in the trajectory of the slope.
Energy expenditure (kJ.min-') = 7.6628 + 0.1462 [(counts per minute
- 3000)1100] + 0.2371 (body weight (kg))
- 0.00216 [(counts per minute - 3000)/100]2
+ 0.004077 [((counts per minute - 3000)/100).(body weight (kg))]
Model concordance correlation coefficient = 0.85
SEE = 5.61 kJ.min-1
Official Journal of the American College of Sports Medicine
6000 8000 10000
Counts per HUMo
The above equations may work reasonably well for pre-
dicting group minutes in different activity intensity catego-
ries. However, predicting an individual's energy expendi-
ture from activity counts works only moderately well at
best. One factor that contributes to the imprecise point
estimates of individual energy expenditure is that the rela-
tionship between activity counts and energy expenditure
exhibits large individual differences. This point is illustrated
in the data of Ekelund et al. (5), who showed large vari-
ability in ActiGraph activity counts for children walking at
standardized speeds. For example, ActiGraph counts ranged
from approximately 400 to 2600 counts per minute at 4
km.h-', and from 1000 to 5000 counts per minute at 6
km.h- 1. In this study, the coefficients of variation for counts
per minute were 21-40% at given velocities of walking.
This variability contributes to large errors in predicting
individual-level energy expenditure. This observation was
confirmed by McMurray et al. (16). In addition, accelerom-
eters are obviously not capturing all movement all the time
and thus estimates of total daily energy expenditure from
movement counts underestimate actual energy expenditure.
Caution should be exercised in using these energy expen-
diture prediction equations to estimate individual-level en-
To account for individual differences in the relationship
between ActiGraph counts and energy expenditure, Ekelund
et al. (5) proposed using the activity-related time equivalent
based on accelerometry (ArteACC) method to calibrate PA
behavior. This method uses ActiGraph counts to establish
time at a given energy expenditure based on the energy
expenditure of selected reference activities. ArteACC
(min.d-1) is equal to total daily activity counts divided by
reference activity counts (counts per minute). The ArteACC
index was shown to be valid surrogate measure of total PA
and if combined with a direct measure of RMR and gender
it can predict total energy expenditure (5).
Another approach for calibration used by Ekelund et al.
(6) established ActiGraph cut points in 500-counts-per-
minute increments and defined sedentary activity as less
than 500, light activity from 500 to 1999, moderate activity
from 2000 to 2999, and vigorous activity as over 3000
counts per minute. Although the purpose of this study was
not to calibrate an activity monitor, these cut point standards
were used to examine the association between body fatness
and time spent in moderate-to-vigorous PA (MVPA) in
more than 1200 children between the ages of 9 and 10.
Children spending more than 2 h.d-1 in MVPA had a
significantly lower sum of skinfolds than children who spent
less than 60 min in MVPA. This approach has some pre-
dictive validity, but lacks the more precise calibration fea-
tures used by others.
Actiwatch. Puyau et al. (21) calibrated the Actiwatch
on 26 children between the ages of 6 and 16 using direct
calorimetry as the criterion measure. Three sedentary, two
light, three moderate, and six vigorous activities were per-
formed in controlled or free-living settings. The Actiwatch
was worn on the hip and the leg. Counts per minute (hip)
ranged from 6 (resting) to 2647 (jogging) with derived VOQ 2
values ranging from 3.2 mL-kg-1.min-1 (rest) to 19.2
mL.kg -. min-] (jogging). The counts per minute cutoffs
for sedentary, light, and moderate activity were, respec-
tively, 99, 899, and 2199 for the hip measurement and 199,
1799, and 4299 for the leg measurement. Point estimates of
activity energy expenditure (kcal.kg-l.min-1) are deter-
mined from the following equations:
AEE (kcal-kg--rmin-1) = 0.0144 + 0.000038 (counts per minute (hip))
R2=0.81 SEE = 0.0147 kca]'kg`-min`
AEE (kcal'kg-'-min-1) = 0.0143 + 0.000020 (counts per minute (leg))
R2 = 0.71
Equations were also generated for total energy expenditure
where total energy expenditure where basal metabolic rate
(BMR) was included (20). In these equations, age is a
significant predictor of total energy expenditure because of
the age-related changes in RMR. However, a study by
Lopez-Alarcon et al. (15) on 4- to 6-yr-old children revealed
no relationships between activity measured with the Acti-
watch and total daily energy expenditure measured with
doubly labeled water.
Actical. Puyau et al. (21) reported an equation to predict
activity energy expenditure using the Actical. To develop
this equation, 32 children between the ages of 7 and 18
performed a variety of activities where directly measured
kcal.kg- .min-1. The equation is
AEE (kcal'kg-'-min-1) = 0.00423 + 0.0003 Pactivity counts
R2= 0.81 SEE= 0.0111 kcal'kg-1'min-1
Heil (10) conducted a calibration study on 14 boys (ages
8-16) in which energy expenditure was assessed using a
portable indirect calorimetry system and Actical units were
secured to the ankle, wrist, and hip. Ten different activities
were performed including sedentary activities, houseclean-
ing tasks, and locomotion activities (walking and jogging).
Activity energy expenditure categories defining sedentary
ACTIVITY MEASUREMENT IN CHILDREN AND YOUTHS
and light, moderate, and vigorous were <0.05, >-0.06-
0.09, and ->0.10 kcal,kg-1.min-t. A unique feature of the
regression modeling approach was the use of two linear
models, one for the sitting and cleaning activities and one
for the locomotion activities, because the pattern of the
relationship between energy expenditure and activity counts
was quite different between these two sets of activity.
Tritrac and RT3. The original commercially available
triaxial accelerometer was the Tritrac. McMurray et al. (16)
developed regression models to predict energy expenditure
from the Tritrac vector magnitude counts on 308 children
between the ages of 8 and 18. Nine activities across a broad
range of intensities comprised the activity regime and en-
ergy expenditure was assessed using a portable metabolic
measurement system. The equation to estimate energy ex-
penditure from the Tritrac vector magnitude was:
902 mL'min-i = 0.32 (vector magnitude counts per minute)
+ 6.97 (height (cm)) + 6.19 (body weight (kg)) - 857.86
r = 0.81
SD =306 mLmin-'
Stayllealthy, Inc. (Monrovia, CA) purchased the rights to
the Tritrac design and repackaged it into a smaller device
called the RT3. Rowlands et al. (24) conducted a calibration
study using the triaxial accelerometer in which they com-
pared the Tritrac to the RT3 and developed count cut points
corresponding to different intensities of activity. Nineteen
boys (mean age 9.5) participated in treadmill walking (two
speeds) and running (two speeds), hopscotch, ball kicking,
and sedentary activities. Oxygen consumption was mea-
sured using Douglas bags with the Tritrac and RT3 secured
to the right and left hip. In general, the RT3 counts were
higher than the Tritrac counts. The counts per minute values
corresponding to 3 and 6 METs when all activities were
included in the analysis were 970 and 2333 counts per
minute. When only the treadmill activities were used the
corresponding cut points were 1806 and 3022 counts per
minute. Thus, it appears the pattern of the activity monitor
response to all activities was different from the pattern with
treadmill exercise alone and resulted in different cut points
to represent the same exercise intensities. This observation
confirms the results of Heil (10) where a two-component
regression was reportedly required to improve the accuracy
of prediction. Table 1 presents a summary of the calibration
studies that have used these devices to establish accelerom-
eter count cut points for this population.
Other approaches to accelerometer calibration.
Exercise physiologists and activity epidemiologists have
endorsed and emphasized the need to calibrate accelerom-
eters to some physiological criterion such as oxygen con-
sumption or energy expenditure. However, in some cases, it
may be best to use the raw activity counts or even the raw
acceleration signal as the outcome measure because use of
the raw signal eliminates the errors associated with the
regression model methods described earlier. This may be
particularly important if accelerometer-based activity mea-
sures are used to track PA over time in longitudinal designs
such as surveillance studies or intervention studies. Jackson
Medicine & Science in Sports & Exercise,
TABLE 1. Accelerometer count cut points defining exercise intensities.
Cut Point (counts
Cut Point (counts
Age or Age
W, R, FL
W, R, FL
W, R, FL
W, R, FL
W, R, FL
W, R, FL
W, R, FL
W, R, FL
McMurray et aL (16)
Eston et al. (7)
Rowlands et al. (24)
Rowlands et al. (24)
Treuth et al. (30)
Puyau et al. (20)
Eston et at. (7)
Puyau et al. (21)
Cut point values represent the lower boundaries for the intensity category.
W, walking; R, running; FL, free-living.
et al. (11) used this approach to describe PA levels in a 1-yr
longitudinal study of preschool children. However, a weak-
ness of this approach is that intensity cannot be evaluated.
A limitation of using the raw counts as the outcome
measure, however, was that Actiwatch activity counts (worn
on the ankle, summed over 7 d) was not associated with total
daily energy expenditure assessed using doubly labeled wa-
ter (14). Twenty-nine children between the ages of 4 and 6
comprised the sample in this investigation. The correlation
between total activity counts and total daily energy expen-
diture was r = 0.27 (P > 0.05) during free-living activities.
It was suggested that the reason for the low correlation was
the prolonged time interval of measurement and the diverse
types of activities included. Further study on this particular
measurement problem is warranted as the goal is to establish
standards of practice in using the accelerometer to charac-
terize habitual PA behavior.
Calibration of accelerometers to assess seden-
tary behavior. Some studies exploring the determinants
of inactivity are using accelerometers to assess sedentary
behavior. Additionally, studies examining factors influenc-
ing obesity may benefit by quantifying inactivity as well as
activity. Reilly et al. (23) suggests that accelerometers can
be used to assess sedentary behavior and developed and
validated a count cut point of <1100 counts per minute for
the ActiGraph to define the upper bound for inactivity in 3-
to 4-yr-old children. What is unique about this work is the
use of observation to calibrate activity counts. When re-
searchers are interested in patterns and determinants of PA,
calibrating activity counts to observational systems is a
more direct alternative. Using this approach, accelerometers
are used to classify sets of activities (e.g., stationary or slow
trunk movement). The observation criterion avoids interpre-
tation errors associated with METs and errors associated
with extrapolation from treadmill exercise to free-living
behaviors (23). The value of having an accelerometer cut
point to define inactive behavior has yet to be realized but
the potential application to studies related to adiposity re-
bound and physical inactivity is significant. Behaviorally
based approaches for calibrating accelerometry output may
be particularly useful when studying young children where
measurement and interpretation of energy expenditure data
are difficult tasks.
Uniaxial versus triaxial accelerometers.
itively it seems that a triaxial or omnidirectional accelerom-
Official Journal of the American College of Sports Medicine
eter should be better at capturing children's activity in
particular because children tend to move in all directions all
the time. However, no direct evidence exists to suggest that
a triaxial accelerometer is better for detecting PA than a
uniaxial monitor. Analysis of the activity monitor calibra-
tion studies presented in this review does not indicate su-
periority of one monitor over the other and all seem to work
reasonably well. Other papers in this supplement address
this issue in more detail.
An issue that requires careful consideration is whether
linear regression is the correct approach to calibrate activity
monitor calibration, given that it does not appear to always
do what we want it to do. For example, examination of the
data in Treuth et al. (30) reveals a clear violation of the
typical regression assumptions (i.e., systematic errors, par-
ticularly at low count values, and an increase in variance of
METs as activity intensity increases). Although the bet-
eroscedasticity of the errors may have been handled in
mixed, there remains the systematic error issue remains.
Also, it appears that the relationship between counts and
energy expenditure may be nonlinear in certain ranges (e.g.,
0-500 counts per minute). It is possible, that a generalized
(rather than general) linear model, such as the polytomous
regression model, may handle some of these problems, and
thus should be investigated.
Bassett et al. (1) indicate that "no single regression equa-
tion appears to accurately predict energy expenditure based
on acceleration score for all activities." Other investigators
have identified this problem, and it has been suggested that
separate regression relationships between accelerometer
output and PA energy expenditure should be determined for
all activities of interest. Because the typical data-processing
methods do not identify specific activities from accelerom-
eter data, it is currently not feasible to apply separate re-
gression equations on an activity-by-activity basis. This
issue can be addressed by improving the methods used to
process accelerometer data.
A potential means of reducing error in accelerometer-
based estimates is to adopt a new methodology for data
processing. In particular, there is an opportunity to use the
pattern of accelerometer counts rather than total accelera-
tion. Several classes of stochastic models are available to
identify patterns in data and use those patterns to provide
information about the underlying process that generated the
data. As early as 1970, it was suggested that these types of
models be applied in processing data from studies of human
movement (28). Due to the computational complexity, how-
ever, these types of methods have been implemented infre-
Kiani et al. (13,14) have reported the use of a neural network
approach to processing motion data in a clinical setting. They
used accelerometer and goniometer signals to determine
whether a patient was supine, seated, standing, or locomoting.
Zhang et al. (34,35) used pattern recognition algorithms to
extract information about the type, duration, and intensity of
activity from a new monitor based on an array of accelerom-
eters. These results suggest that sophisticated modeling tech-
niques may be useful in measuring PA. The challenge is to
develop a method applicable to data that can be obtained
without unduly burdening the subjects or investigators. These
investigations relied on data from instruments that have limited
utility for application in field settings due to expense, memory
capabilities, and/or complexity.
A particular class of stochastic model for which the
theory and application are well developed in a pattern-
recognition setting is the hidden Markov model (HMM).
These models have been used with considerable success in
the speech recognition literature, where they are used in
natural language recognition algorithms that identify words
in human speech (22). HMM also appear in analyses of
neuron firing patterns, DNA sequences, analysis of patterns
in viral mutations, and many other natural phenomena (e.g.,
Our group recently reported success in applying BMM to
the analysis of accelerometer data (19). Specifically, we
developed an HMM that could accurately categorize two
activities (uphill walking and vacuuming) as vigorous and
moderate-intensity activities, respectively, using data from
the ActiGraph. These activities are misclassified as moder-
ate and sedentary, respectively, by the regression cut point
method of categorizing accelerometer data. Thus, it appears
that pattern recognition-based approaches in general, and
HMM in particular, have good potential to improve the
accuracy of objective assessment of PA. Research into al-
ternative activity monitor calibration approaches such as the
HMM are greatly needed to improve our ability to assess PA
Several different models of accelerometers have been
used to evaluate PA in children and youths. The calibration
of these devices typically involves using walking and run-
ning alone or in combination with free-living activities to
estimate the ranges of accelerometer counts corresponding
to predefined intensity levels or to estimate energy expen-
diture. Recent advances have refined prediction models to
improve estimates by incorporating nonlinear components
into the equations. The following recommendations are in-
tended to help move accelerometer calibration studies for-
ward and to provide researchers with guidelines to establish
standards of practice for calibration studies in children.
"o Calibration should employ an appropriate biological or
behavioral standard. For example, accelerometers used
to examine the impact of activity on obesity should
relate output to energy expenditure while bone health
research should relate output to mechanical forces.
"* Study designs should include a wide variety of activ-
ities representing a broad spectrum of energy expen-
diture (low to vigorous intensity) and short duration of
movements. The latter suggests sampling epochs of no
longer than 1 min. More research is needed to identify
specific epoch length to ensure that children's short
bouts of activity are captured. For example, consider-
ing the sporadic and short bursts of vigorous activity
performed by children (18) epochs as short as 5 s may
"o Studies should include at least 10 subjects per age
group (e.g., 8-10, 10-12 yr). Because body mass is
known to be a significant factor in the relationship
between the accelerometer signal and energy expendi-
ture, calibration prediction equations should be popu-
"* Reports should present activity energy expenditure and
MET prediction equations. Appropriate calculation of
METs will require measuring resting energy expendi-
ture or the use of age-specific estimates of resting
"* Behavioral approaches (e.g., direct observation) for
calibrating accelerometry output may be particularly
useful when studying young children where measure-
ment and interpretation of energy expenditure data are
"o Depending on the purpose of the study, the most ap-
propriate accelerometer measure may be the raw
counts or raw acceleration (e.g., tracking of PA).
"o New studies should be conducted to establish calibra-
tion protocols to assess sedentary behavior and explore
classification systems to identify specific movements
based on the raw acceleration signal.
Use of these guidelines in conducting calibration studies
will improve our ability to compare study results and inter-
pret PA measurement schemes. The use of objective mea-
sures to assess PA will likely continue to expand and if
certain standards of practice guidelines are used, our under-
standing of the mechanisms by which PA affects health in
children will be improved.
The results of the present study do not constitute endorsement
by the authors or ACSM of the products described in this paper.
ACTIVITY MEASUREMENT IN CHILDREN AND YOUTHS
Medicine & Science in Sports & Exerclsea
1. BASSETT, D. R., JR., B. E. AINswoRTfi, A. M. SWARTZ, S. J. STRATI],
W. L. O'BRiEN, and G. A. KING. Validity of four motionsensors in
measuring moderate intensity physical activity. Med. Sci. Sports
Exerc. 32:S471-480, 2000.
2. BRAGE, S., N. WEDDERKoPp, L. B. ANDERSEN, and K. K. FROBERO.
Influence of step frequency on movement intensity predictions
with the CSA accelerometer: a field validation study in children.
Pediatr. Exerc. Sci. 15:277-287, 2003.
3. COOPER, B., and M. LIpsrrcH. The analysis of hospital infection
data using hidden Markov models. Biostatistics 5:223-237, 2004.
4. DAvIs, C. T. Metabolic cost of exercise and physical performance
in children with some observations on external loading. Eur.
J. AppL PhysioL 45:95-102, 1980.
5. EKELUNO, U., J. ANiAN, and K. WEMTERTERP. Is the ARTEACC
index a valid indicator of free-living physical activity in adoles-
cents? Obes. Res. 11:793-801, 2003.
6. EKELUND, U., L. B. SARDINHA, S. A. ANDERSSON, et al. Associations
between objectively assessed physical activity and indicators of
body fatness in 9- to 10-y-old European children: a population-
based study from 4 distinct regions in Europe (the European Youth
Heart Study). Am. J. Clin. Nuir. 80:584-590, 2004.
7. ESTON, R. G., A. V. ROWLANDS, and D. K. INGLEDEW. Validity of
HR, pedometry, and accelerometry for predicting the energy cost
of children's activities. J. AppL PhysioL 84:362-371, 1998.
8. FOULKES, A. S., and D. V. DE GRUTTOLA. Characterizing the
progression of viral mutations over time. J. Am. Stat. Assoc.
9. FREEDSON, P. S, J. SiRARD, E. DEBOLD, et al. Calibration of the
Computer Science and Applications, Inc. (CSA) accelerometer.
Med. Sci. Sports Exerc. 29(Suppl):S45,1997.
10. HEIL, D. Predicting activity energy expenditure in using the Ac-
tical activity monitor. Res. Q. Exerc. Sport. In press, 2005.
11. JACKSON, D. M., J. J. REILLY, L. A. KELLY, C. MONTGOMERY, S.
GRANT, and J. Y. PATON. Objectively measured physical activity in
a representative sample of 3- to 4-year-old children. Obes. Res.
12. JANZ, K. F. Validation of the CSA accelerometer for assessing
children's physical activity. Med. Sci. Sports Exere. 26:369-375,
13. KIANI, K., C. J. SNUDERS, and E. S. GELSEMA. Computerized anal-
ysis of daily life motor activity for ambulatory monitoring. Tech-
noL Health Care 5:307-318, 1997.
14. KIANI, K., C. J. SNUDERS, and E. S. GELSEIA. Recognition of daily
life motor activity classes using an artificial neural network. Arch.
Phys. Med. Rehabil. 79:147-154, 1998.
15. LOpEz-ALARCON, M., J. MERRIFIELD, D. A. FIELDS, et al. Ability of
the Actiwatch accelerometer to predict free-living energy expen-
diture in young children. Obes. Res. 12:1859-1865, 2004.
16. MCMURRAY, R. G., C. D. BAGGE•I, J. S. HARRELL, M. L. PENNELL,
and S. I. BANGDIWALA. Feasibility of the Tritrac R3D accelerom-
eter to estimate energy expenditure in youth. Pediatr. Exerc. Sci.
17. MoNToYF, H. J., R. WASHBURN, S. SERVAIS, A. ERTL, J. G. WEBSTER,
and F. J. NAGLE. Estimation of energy expenditure by a portable
accelerometer. Med. Sci. Sports and Exerc. 15:403-407, 1983.
18. NILSSON, A., U. EKELUND, A. YNGvE, and M. SIos'ioOM. Assessing
physical activity among children with accelerometers using dif-
ferent time sampling intervals and placements. Pediatr. Exerc. Sci.
19. POBER, D., M. C. RAPHEAL, and P. S. FREEDSON. Novel technique
for assessing physical activity using accelerometer data. Med. Sci.
Sports Exerc. 36:S198, 2004.
20. PUYAU, M. R., A. L. ADOLPM, F. A. VouRA, and N. F. BtrrrE.
Validation and calibration of physical activity monitors in chil-
dren. Obes. Res. 10:150-157, 2002.
21. PUYAU, M. R., A. L. ADOLPH, F. A. VoURA, 1. ZAKERI, and N. F.
BUTrrTE. Prediction of activity energy expenditure using accelerom-
eters in children. Med. Sci. Sports Exerc. 36:1625-1631, 2004.
22. RABINER, L. R., and B. H. JUANY. Fundamentals of Speech Rec-
ognition. Englewood Cliffs, NJ: PTR Prentice Hall, 1993.
23. REILLY, J. J., J. COYLE, L. KELLY, G. BURKE, S. GRANT, and J. Y.
PATON. An objective method for measurement of sedentary behav-
ior in 3-4-year olds. Obes. Res. 11:1155-1 158, 2003.
24. ROWLANDS, A. V., P. W. M. THOMAS, R. G. EsToN, and R. TOPPING.
Validation of the RT3 triaxial accelerometer for the assessment of
physical activity. Med. Sci. Sports Exerc. 36:518-524, 2004.
25. SALLIS, J. F., M. J. BUONO, J. ROBY, D. CARLSON, and J. A. NELSON.
The Caltrac accelerometer as a physical activity monitor for
school-age children. Med. Sci. Sports Exerc. 22:698-703, 1990.
26. SCIuEP, A., C. STEINHOFF, and A. SciONimrrii. Robust inference of
groups in gene expression time-courses using mixtures of HMMs.
Bioinfornatics 20(suppl 1):1283-1289, 2004.
27. ScImaZ, K. H., M. TREurT, P. HANNAN, et al. Predicting energy
expenditure from accelerometry counts in adolescent girls. Med.
Sci. Sports Exerc. 37:155-161, 2005.
28. Scturz, R. W. Stochastic processes: their nature and use in the
study of sport and physical activity. Res. Q. 41:205-213, 1970.
29. ScuuTz, Y., R. L. WEINSIER, and G. R. HuNTEIR.
free-living physical activity in humans: an overview of currently
available and proposed new measures. Obes. Res. 9:368-378,
30. TREUTi, M. S., K. SCHMIiZ, D. J. CATELLIER, and R. G. McMuRRAY,
et al. Defining accelerometer thresholds for activity intensities in
adolescent girls. Med. Sci. Sports Ererc. 36:1259-1266, 2004.
31. TROST, S. G., D. S. WARD, S. M. MOOREHEAD, P. D. WATSON, W.
RINER, and J. R. BURKE. Validity of the Computer Science and
Applications (CSA) activity monitor in children. Med. Sci. Sports
Exerc. 30:629-633, 1998.
32. WESTERTERP, K. R. Physical activity assessment with accelerom-
eters. Int. J. Obes. 23(suppl):45-49, 1999.
33. Wu, W., M. J. BLACK, D. MUSiFORD, Y. GAU, E. BiENENsTOCK, and
J. P. DONOGHUE. Modeling and decoding motor cortical activity
using a switching Kalman filter. IEEE Trans. Biomed. Eng. 51:
34. ZHANG, K., F. X. PI-SUNYER, and C. N. BOOZER. Improving energy
expenditure estimation of physical activity. Med. Sci. Sports -r-
erc. 36:883-889, 2004.
35. ZHANG, K., P. WERNER, M. SUN, F. X. PI-SUNYER, and C. N.
BOOZER. Measurement of human daily physical activity. Obes.
Res. 11:33-40, 2003.
Official Journal of the American College of Sports Medicinehttp://www.acsm-msse.org
The magazine publisher is the copyright holder of this article and it
is reproduced with permission. Further reproduction of this article in
violation of the copyright is prohibited.
Copyright 1982-2005 The H.W. Wilson Company. All rights reserved.