See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/260121090

EMG, Heart Rate, and Accelerometer as Estimators of Energy Expenditure in

Locomotion

Article  in  Medicine and science in sports and exercise · February 2014

DOI: 10.1249/MSS.0000000000000298 · Source: PubMed

CITATIONS READS

23 752

6 authors, including:

Olli Tikkanen Salme Kärkkäinen

University of Jyväskylä University of Jyväskylä
15 PUBLICATIONS   147 CITATIONS    22 PUBLICATIONS   139 CITATIONS   

SEE PROFILE SEE PROFILE

Mauri Kallinen
University of Oulu
85 PUBLICATIONS   2,881 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

InPACT: Tailored Counseling to Increase non-exercise Physical Activity in adults with a sedentary job and in their young children: A year-long randomized controlled
trial View project

CHIPASE-Children’s physical activity spectrum: daily variations in physical activity and sedentary patterns related to school indoor physical environment View
project

All content following this page was uploaded by Salme Kärkkäinen on 29 December 2015.

The user has requested enhancement of the downloaded file.

EMG, Heart Rate, and Accelerometer as
Estimators of Energy Expenditure in Locomotion
OLLI TIKKANEN1, SALME KÄRKKÄINEN2, PIIA HAAKANA1, MAURI KALLINEN3,
TEEMU PULLINEN1, and TAIJA FINNI1
1
Neuromuscular Research Center, Department of Biology of Physical Activity, University of Jyväskylä, Jyväskylä, FINLAND;
2
Department of Mathematics and Statistics, University of Jyväskylä, Jyväskylä, FINLAND; and 3Department of Medical
Rehabilitation, Oulu University Hospital, Oulu, FINLAND

ABSTRACT
TIKKANEN, O., S. KÄRKKÄINEN, P. HAAKANA, M. KALLINEN, T. PULLINEN, and T. FINNI. EMG, Heart Rate, and Accel-
erometer as Estimators of Energy Expenditure in Locomotion. Med. Sci. Sports Exerc., Vol. 46, No. 9, pp. 1831–1839, 2014. Purpose:
Precise measures of energy expenditure (EE) during everyday activities are needed. This study assessed the validity of novel shorts measuring
EMG and compared this method with HR and accelerometry (ACC) when estimating EE. Methods: Fifty-four volunteers (39.4 T 13.9 yr)
performed a maximal treadmill test (3-min loads) including walking with different speeds uphill, downhill, and on level ground and one
running load. The data were categorized into all, low, and level loads. EE was measured by indirect calorimetry, whereas HR, ACC, and
EMG were measured continuously. EMG from quadriceps (Q) and hamstrings (H) was measured using shorts with textile electrodes. Validity
of the methods used to estimate EE was compared using Pearson correlations, regression coefficients, linear mixed models providing Akaike
information criteria, and root mean squared error (RMSE) from cross-validation at the individual and population levels. Results: At all loads,
correlations with EE were as follows: EMG(QH), 0.94 T 0.03; EMG(Q), 0.91 T 0.03; EMG(H), 0.94 T 0.03; HR, 0.96 T 0.04; and ACC,
0.77 T 0.10. The corresponding correlations at low loads were 0.89 T 0.08, 0.79 T 0.10, 0.93 T 0.07, 0.89 T 0.23, and 0.80 T 0.07, and at level
loads, they were 0.97 T 0.03, 0.97 T 0.05, 0.96 T 0.04, 0.95 T 0.08, and 0.99 T 0.02, respectively. Akaike information criteria ranked the
methods in accordance with the individual correlations. Conclusions: It is shown for the first time that EMG shorts can be used for EE
estimations across a wide range of physical activity intensities in a heterogeneous group. Across all loads, HR is a superior method of
predicting EE, whereas ACC is most accurate for level loads at the population level. At low levels of physical activity in changing terrains,
thigh muscle EMG provides more accurate EE estimations than those in ACC and HR if individual calibrations are performed. Key Words:
TEXTILE ELECTRODES, EMG CLOTHING, ENERGY EXPENDITURE, PHYSICAL ACTIVITY, INACTIVITY, ACCELERATION

T
he ability to accurately and objectively track energy
expenditure (EE) and physical activity is a topic of
great importance for two main reasons. First, physical
activity plays an important role in sustaining a healthy life-
style. Physical activity has beneficial effects on cardiore-
spiratory fitness, lipid profiles, blood pressure, weight control,
and chronic conditions such as cardiovascular diseases and
diabetes (17). Muscle function, specifically muscle strength, is
also important for maintenance of health and well-being, es-
pecially among middle-age and elderly people (20). Second,
there is compelling evidence that physical inactivity and sit-
ting are harmful for health even in people considered to be
physically active (5), and interestingly, exercise for fitness
does not decrease the total daily inactivity time (10). Low
Address for correspondence: Olli Tikkanen, M.Sc., Neuromuscular Research
Center, Department of Biology of Physical Activity, University of Jyväskylä,
PO Box 35, 40014 Jyväskylä, Finland; E-mail: olli.m.tikkanen@jyu.fi.
Submitted for publication May 2013.
Accepted for publication January 2014.
0195-9131/14/4609-1831/0
MEDICINE & SCIENCE IN SPORTS & EXERCISEÒ
Copyright Ó 2014 by the American College of Sports Medicine
DOI: 10.1249/MSS.0000000000000298
levels of everyday muscle activity, however, have positive
effects on glucose uptake and insulin sensitivity and thus may
prevent type 2 diabetes, for example (15).
It is also important that activities of low-to-moderate in-
tensity be tracked accurately because in the normal physical
activity level (PAL) range, the distribution of time spent on
low- and moderate-intensity activities determines the activity
level of a person (34). For example, it has been shown that
sedentary time and activities related to transportation and com-
muting, such as walking and cycling, contribute significantly
to average PAL (4) whereas high-intensity activity does not
have much effect on PAL in normal populations (34). It is
clear that people spend more time sleeping and performing
APPLIED SCIENCES
sedentary and light activities than that for vigorous activities,
and hence, accurate estimation of EE at low intensities is
especially important in long-term measurements (8).
Accelerometers are used extensively to estimate EE and
physical activity, although they are generally validated in lab-
oratory settings, limiting the generalizability of the results to
free-living situations (8). Several accelerometer equations work
well for classifying moderate activity but fail elsewhere, whereas
ActiGraph and Actical regressions tend to overestimate walking
and sedentary activities and underestimate most other ac-
tivities (8). Recently, accelerometers storing raw data for

1831

Copyright © 2014 by the American College of Sports Medicine. Unauthorized reproduction of this article is prohibited.
 long periods have become available, making it possible to
carry out long-term monitoring and use more sophisticated
and advanced data analysis for more accurate categoriza-
tion of physical activities (28). In addition, new commer-
cial accelerometers like ActivPAL, ActiGraph (GT3X), and
IDEAA can estimate lying, sitting, and standing postures,
although the discrimination of ActiGraph between standing
and sitting has been questioned (28). Another relatively easy
assessment of EE of physical activity is heart rate (HR). An
elevation in HR has been considered the prime response to
physical activity, and HR recordings are typically used to
evaluate cardiovascular responses. The relation between HR
and EE is linear only within a narrow range of approximately
90–150 bpm (25) and is subject to both intra- and inter-
individual variability (16). HR may be partially dissociated
from EE by factors such as environmental conditions, emo-
tion, and posture (16). During light activity or inactivity, there
is almost no slope in the relation (25).
By definition, physical activity is bodily movement that is
produced by the contractile activity of skeletal muscles that
substantially increases EE. Muscle activation is a prerequi-
site for physical activity, but its role in quantifying PAL has
been neglected. The development of garments with embed-
ded textile electrodes has made surface EMG measurements
convenient even for whole-day field measurements. These
garments enable muscle activity recordings in a valid and
reliable way without skin preparation and the problem of
wires hanging around the body (11,33). Therefore, the pur-
pose of this study was to assess the validity of estimating EE
with EMG shorts during treadmill walking and running and
to compare this method with the widely used HR and triaxial
accelerometry (ACC).
METHODS
Subjects. Volunteers were recruited through public ad-
vertisements and e-mail lists of local companies and in-
stitutes in Central Finland (staff from hospitals, construction
companies, paper manufacturers, University of Jyväskylä,
and the city of Jyväskylä). In total, 216 people volunteered.
All were screened for exclusion criteria (cardiovascular dis-
ease, current or previous injury in the lower extremity or lower
back, or any condition that could affect leg muscle activation)
and inclusion criteria of the group, being as heterogeneous
as possible in terms of sex, age, body composition, and PAL.
APPLIED SCIENCES
Finally, 63 people were measured in the laboratory after sign-
ing an informed consent. Nine subjects were lost because of
methodological reasons, resulting in 54 subjects from whom
the results are presented. The study was approved by the
ethics committee of the University of Jyväskylä.
Procedure. Subjects came twice to the laboratory, once
for tests done in a fasting state (12 h without eating and
drinking) and once for activity tests. In the fasting state, the
resting metabolic rate, body composition, and resting ECG
(Cardiofax V Ecaps12 ECG EKG machine; Nihon Kohden,
Tokyo, Japan) were measured and questionnaires regarding
1832 Official Journal of the American College of Sports Medicine
Copyright © 2014 by the American College of Sports Medicine. Unauthorized reproduction of this article is prohibited.
physical activity history and lifestyle were completed. Sub-
jects older than 45 yr were screened by a medical doctor. In
the second laboratory visit, subjects completed maximal
voluntary contraction (MVC) tests and a maximal treadmill
test. A subgroup of five participants did all the measure-
ments twice on different days, and these data were used to
establish test–retest reliability.
Subjects performed isometric MVC in bilateral knee ex-
tension and flexion (David 220; David Health Solutions,
Helsinki, Finland) at a knee angle of 140-. After a proper
warm-up, at least three maximal trials were done, with 2- to
5-min rest between trials. If torque increased more than 5%
in the third attempt, more trials were done. Each trial lasted
3–5 s, and loud verbal encouragement was used to encour-
age maximal performance.
In the treadmill test, each load lasted 3 min except the last
load, which was done until exhaustion. The treadmill incli-
nation was 0- unless otherwise reported. The first six loads
were the same for all subjects: 4 kmIhj1, 5 kmIhj1, 5 kmIhj1
with 4- descent, 5 kmIhj1 with 4- ascent, 6 kmIhj1, and
7 kmIhj1. Subsequently, participants under the age of 30
performed one running load (10 kmIhj1 for females and
12 kmIhj1 for males). The next step for all participants
was walking 5 kmIhj1 with 8- ascent. After walking for 3 min
with this load, how close participants were to their maximal
oxygen consumption was estimated. If two of three of the
following criteria were fulfilled, participants continued with the
same load: 1) V̇O2max G85% of the estimated maximum, 2) HR
G90% of the estimated maximum, and 3) Borg RPE G16. If
two of three criteria were not fulfilled, participants continued
the walking test at 7 kmIhj1 with 10- ascent until exhaustion.
This nonstandard test protocol was chosen to simulate
daily activities better than the traditional V̇O2max test pro-
tocols because uphill and downhill walking are rather com-
mon activities in normal daily life. The test ended with
walking (instead of the more commonly used running) be-
cause many of the subjects were unaccustomed to running
and safety could have been compromised especially in older
subjects. Twenty six subjects performed the running load.
Seven subjects did not do the last walking load (7 kmIhj1
with 10- ascent) because they were already exhausted from
the previous load (5 kmIhj1 with 8- ascent).
Measurements. Body mass, fat percentage, skeletal mus-
cle mass, fat mass, and visceral fat area were assessed using
bioimpedance (InBody 720; BioSpace Corp. Ltd., Seoul, South
Korea), and height of the subjects was measured with an in-
elastic tape measure attached to a wall. Questionnaires were
used to assess participants’ physical activity history and lifestyle.
During the treadmill test, a large treadmill (4
2 m,
model OJK-1; Telineyhtymä, Kotka, Finland) was used to
enable unobstructed walking and running. Ventilatory gases
were measured with Jaeger Oxycon Pro with the LabManager
3.0 software (Viasys Healthcare Gmbh, Hoechberg, Germany),
which was calibrated before every participant (the device has
accuracy of 2% for the volume (50 mL), 3% for flow
(70 mLIsj1), 0.2% for O2 analyzer, and 0.05% for CO2
http://www.acsm-msse.org
analyzer). Oxycon Pro has been shown to be both valid
(7,26) and reliable (7) to measure parameters of respiration.
Nasal respiration was blocked with a nose clip (Nose Clip
Disposable Series 9014; Hans Rudolph, Inc., Shawnee,
KS). HR was recorded with a Suunto T6 wrist computer
and HR belt (Suunto Oy, Vantaa, Finland). Ventilatory
gases, HR, triaxial ACC (HM120; Alive Technologies Pty.
Ltd., Queensland, Australia), and EMG were measured
continuously during the test. In the last 30 s of each load,
RPE was assessed using the Borg scale (6–20).
Thigh muscle activities were measured with EMG shorts
made of knitted fabric similar to elastic sports clothes or
functional underwear, with the exception of the capability to
measure EMG from the skin surface of the thigh areas
(Myontec Ltd., Kuopio, Finland) (11,32,33). To obtain the
average rectified value of EMG (aEMG), the shorts are
equipped with conductive electrodes and wires are inte-
grated into the fabric, which transfer the EMG signals from
the electrodes to the electronics module. The textile elec-
trodes were sewn onto the internal surface of the shorts and
located such that the bipolar electrode pair lies on the distal
part of the thigh, and reference electrodes, longitudinally
along the lateral sides. The shorts measure EMG from the
quadriceps femoris (vastus lateralis, medialis, intermedius,
and rectus femoris) and the hamstring muscles (biceps femoris,
semimembranosus, and semitendinosus).
Before the EMG shorts could be worn, electrolyte gel (Elec-
trolyte Creme REDUX; Parker Laboratories, Inc., Fairfield, NJ)
was applied onto the electrode surfaces. Skin–electrode im-
pedance was assessed from the connectors that attach to the
electronics module. Impedance was less than 5 k6 in all
subjects. The electronics module contains signal amplifiers, a
microprocessor with embedded software, data memory, and a
personal computer (PC) interface. In the module, the EMG
signal is measured in its raw form with a sampling frequency
of 1000 Hz and a frequency band 50–200 Hz (j3 dB) and is
first rectified and then averaged over 100-ms nonoverlapping
intervals. The averaged data were stored in ASCII format
in the memory of the module from which the data were
downloaded to a PC using the specifically designed HeiMo
PC software (Myontec Ltd., Kuopio, Finland).
Processing of data. Resting metabolic rate was deter-
mined as the average oxygen uptake over 10 min that was
preceded by 20 min of quiet bed rest. HR was recorded con-
tinuously, and resting HR was determined as the lowest value
within a 5-s window. From the treadmill test, HR, acceler-
ometer data, EMG data, and respiratory gases from the last
60 s of each load were averaged and further analyzed. EE was
calculated from oxygen consumption (V̇O2) and respiratory
exchange ratio (RER) with the following equation:
R suggest that the individual data can be summarized into a
EE ðkcalRminj1 Þ ¼ ð1:2 RER þ 3:85ÞðVO2 =1000 Þ
Absolute HR (beats per minute) values were used in the
analysis. The EMG signal from each of the four muscles
ENERGY EXPENDITURE ESTIMATION IN LOCOMOTION
Copyright © 2014 by the American College of Sports Medicine. Unauthorized reproduction of this article is prohibited.
(right quadriceps, right hamstring, left quadriceps, and left
hamstring) was normalized to the maximal EMG values
(EMGMVC), taken as an average from a 1-s period during
MVC for the given muscle. Right and left quadriceps values
were averaged to form the quadriceps variable (aEMG(Q)),
right and left hamstrings, to form the hamstring variable
(aEMG(H)), and all channels, to form the variable for quad-
riceps and hamstrings (aEMG(QH)). ACC signals were fil-
tered with a 0.5- to 11-Hz band-pass filter with a finite
impulse response to decrease the effects of gravity and high-
frequency artefacts. After filtering, the signal was rectified
and a moving average with a 10-s window was applied. Ac-
tivity counts were calculated as area under the curve (in-
tegration) from 60-s periods. Activity counts were first
calculated separately for x, y, and z values, and later, a sum
vector of x, y, and z [(x2 + y2 + z2)2] was calculated to
represent the physical activity from a given 60-s period.
Categorization of loads. The treadmill test was di-
vided into the following functional categories: 1) ‘‘all loads’’
include all measured conditions comprising resting meta-
bolic rate and all walking and running loads, 2) ‘‘low loads’’
include resting metabolic rate and walking loads up to 6 kmIhj1
including loads with inclinations, and 3) ‘‘level loads’’ in-
clude resting metabolic rate and loads without inclination
(also including running load).
Data exclusion. Nine participants were lost because
of methodological reasons (four because of problems with
ventilatory gas measurements and five because of problems
with EMG measurements), leaving 54 participants for fur-
ther analysis. In these 54 subjects, some EMG channels or
data points were excluded because of nonphysiological sig-
nals. From 28% of subjects (13/54), one hamstring channel
was excluded from the analysis, and from 4% (2/54), both
hamstring channels were excluded. One quadriceps channel
was excluded from 11% of subjects (6/54). From 11% of
subjects (6/54), some data points were excluded from the
analysis. The criteria for exclusion were set beforehand and
are described in detail in the study of Tikkanen et al (32).
The validity of the calculated aEMG due to missing chan-
nels was simulated and tested; excluded quadriceps channels
resulted in greater aEMG, and excluded hamstring channels
resulted in lower aEMG compared with aEMG from all
four channels.
Statistical analysis. The main focus of the study was
the comparison of different methods aEMG(H), aEMG(Q),
APPLIED SCIENCES
aEMG(QH), HR, and ACC when estimating EE. First, Pear-
son correlation coefficient (R value) and regression co-
efficients (intercept and slope) were calculated between
EE and each method for each subject and the means and
SD of the correlation coefficients were reported for each
method. This approach follows Diggle et al. (9) who
derived variable such as a correlation coefficient or re-
gression coefficients (intercept and slope). Then, the correla-
tion of measurements within each subject can be avoided and
the standard ANOVA or Student’s t-test can be used to test
Medicine & Science in Sports & Exercised 1833
 the differences between methods with the significance level
P G 0.05.
Besides the derived variable approach for each method,
we compared the methods using linear mixed models. Mixed
models were used to model the variation between subjects
when having repeated measurements. In the modeling besides
the fixed factors, random coefficients, subject-specific inter-
cept (u0), and regression coefficient (u1) were tested with
maximum likelihood ratio tests. The final model was calcu-
lated using a restricted maximum likelihood method. When
comparing significant but nonnested linear mixed models, a
maximum likelihood ratio test cannot be used. Then, Akaike
information criteria (AIC), which is a measure for goodness
of fit, is one convenient tool for a comparison (1). It is cal-
culated as follows:
AIC ¼ 2 log½Lðestimated parametersÞ þ 2p
where L is the likelihood of the fitted linear mixed model at
the estimated parameter values (maximum value) and p is the
number of parameters in the model. The value of AIC de-
creases if the value of the likelihood increases or the number
of parameters decreases. Thus, we preferred the method giv-
ing the model with the smallest AIC within a certain load
categorization. In addition to the comparison, the fitted linear
mixed model could also be used for prediction purposes both
at the population level and at the individual level (AIC mea-
sures the prediction at the individual level). Having the pre-
diction formulas at the population level (fixed coefficients
only), we are able to give the predicted EE for a new indi-
vidual. To evaluate the prediction accuracy of the equations,
we first performed a leave-one-out cross-validation where all
measurements of an individual were, one at a time, moved
from the data to a test set. For the measurements of the test
set, the EE were predicted at the population level by using the
fixed part of the model fitted to the subset of the data. From
the errors between the predicted EE and the observed EE,
we estimated the root mean squared error (RMSE), defined
as follows:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 ffi
RMSE ¼ mean EEp j EEo
coefficients and typical errors. Characteristics of the subjects
are reported as means and SD.
RESULTS
The characteristics of the subjects are shown in Table 1.
Variables other than age, fat mass, hip circumference, and
HR were significantly different between males and females
(P G 0.001–0.05). Means of individual correlations between
EE and the different methods for the entire study sample are
shown in Table 2. When all loads were included in the
analysis, HR was the best predictor of EE, according to both
mean correlations and AIC. At low loads, hamstring aEMG
was the best predictor, and at level loads, ACC provided the
best prediction for EE when the individual calibration is
added to the model.
Table 3 shows the means and SD of intercepts (a) and
slopes (b) of the fitted regression lines for all methods. With
such a variation of regression coefficients between the in-
dividuals, the linear mixed model enables us to model this
variation using random regression coefficients (individual
intercept and slope) besides the population intercept and
slope. For all methods, the fitted linear mixed models in-
cluded a random regression coefficient, except HR at all,
low, and level loads, hamstring EMG at all loads, and ACC
at level loads, which also had a random constant. Using the
obtained AIC values shown in Table 2, we observed similar
ranking as with mean R values of individuals.
For prediction purposes, we added covariates into the
model. Prediction equations to estimate momentary EE
consisted of the same random coefficients as the models of
Table 2. Level loads were excluded for simplicity because
natural environments contain uphills, downhills, stairs, etc.,
and for practical reasons, only variables that need no laboratory
measurements were included (age and sex). Prediction equa-
tions are shown in Table 4 and include only variables that were
significant or interacted significantly. Furthermore, Table 4
shows the calculated accuracies, as follows: the RMSE when
using 1) cross-validation for the prediction at the population
level (fixed coefficients used), 2) cross-validation for the

where EEp is the predicted value and EEo is the observed

value of the same individual. Having the prediction equation TABLE 1. Anthropometric and physiological characteristics of participants for males and
females (mean (SD)).
at the individual level (fixed and random coefficients), we can
Males (n = 28) Females (n = 26)
predict the EE for a new observation of the individual in-
APPLIED SCIENCES

cluded in the data. We evaluated the prediction accuracy using Age (yr) 39.2 (13.7) 39.6 (14.2)
Height (cm) 180.1 (7.2)* 166.7 (5.1)
a leave-one-out cross-validation where one observation of Weight (kg) 79.4 (13.0)* 62.7 (7.7)
an individual was moved to the test set and all other mea- Body mass index (kgImj2) 24.4 (3.1)** 22.5 (2.4)
Fat percentage (%) 16.9 (5.3)*** 21.6 (7.6)
surements formed the training data. The RMSE was calcula- Skeletal muscle mass (kg) 36.9 (6.0)* 27.1 (2.9)
ted similarly as shown previously. Furthermore, for selected Fat mass (kg) 14.6 (7.8) 13.9 (6.9)
Visceral fat (cm2) 91.1 (31.0)*** 66.9 (30.4)
models, the SE of the estimates of fixed coefficients and the Hip circumference (cm) 98.6 (6.4) 96.9 (5.3)
SD of random coefficients and residual were reported. The Waist circumference (cm) 87.5 (9.8)* 76.8 (8.5)
estimation of individual correlations, individual regression Resting HR (bpm) 53.2 (10.5) 57.8 (9.4)
HRmax (bpm) 182.6 (14.7) 182.6 (12.4)
coefficients, and mixed models was performed using the R V̇O2 max (mLIkgj1Iminj1) 49.7 (8.7)** 43.9 (9.6)
software (R Core Team 2013, Vienna, Austria). Furthermore, EEmax (kJIminj1) 19.9 (3.5)* 13.9 (2.8)
reliability of test–retest was described with correlation *P G 0.001; **P G 0.05; ***P G 0.01.

1834 Official Journal of the American College of Sports Medicine http://www.acsm-msse.org

Copyright © 2014 by the American College of Sports Medicine. Unauthorized reproduction of this article is prohibited.
TABLE 2. Means of individual R values (mean (SD)) and AIC values between EE, aEMG(QH) from quadriceps and hamstrings (n = 52), aEMG(Q) from quadriceps (n = 54), aEMG(H) from
hamstrings (n = 52), HR (n = 54), and ACC (n = 51).
Correlation between EE and aEMG(QH) aEMG(Q) aEMG(H) HR ACC
All loads

R values 0.94 (0.03) 0.91 (0.07) 0.94 (0.03) 0.96 (0.04) 0.77 (0.10)
AIC 2437 2658 2382 1817 2842
Low loads

R values 0.89 (0.08) 0.79 (0.10) 0.93 (0.07) 0.89 (0.23) 0.80 (0.07)
AIC 1120 1251 989 1133 1151
Level loads

R values 0.97 (0.03) 0.97 (0.05) 0.96 (0.04) 0.95 (0.08) 0.99 (0.02)
AIC 983 991 1045 1066 683
Both correlation and AIC represent individual-based prediction accuracies. The lower the AIC value, the better the method, whereas for correlation, R values closer to 1 represent a better
method in estimating EE.
All correlations between different methods were significantly different (at level of P G 0.001–0.05) unless marked with gray lines to the table (NS, not statistically significant). All correlations
between different load categories (all, low, and level loads) were significantly different (at level of P G 0.001–0.05), except aEMG(H) between all loads and low loads and HR between all loads
and level loads.

prediction at the individual level (fixed and random co-
efficients used), and 3) AIC of the original fitted model.
Table 5 provides the correlation coefficients and typical er-
rors for test–retest comparisons. In Table 6, the SE of fixed
coefficients (with P values) and the SD of random co-
efficients and residual are shown for HR, ACC, and
aEMG(H) at all and low loads, respectively.
DISCUSSION
In general, EMG, HR, and ACC all predicted EE rela-
tively well according to individually calibrated models. It is
shown for the first time that EMG shorts can be used in EE
estimations across a wide range of physical activity inten-
sities in a heterogeneous subject group. Hamstrings EMG
provide more accurate EE estimations than those in quadri-
ceps EMG in all and low loads according to correlations,
AIC, and RMSE (Tables 2 and 4). ACC was the best pre-
dictor of EE at level loads, but according to AIC and cor-
relations, its accuracy was considerably lower in all loads
that also included uphill and downhill (Table 2). Correla-
tions of HR were found to be considerably lower at low
loads than those at level and all loads (Table 2).
EE of muscles during locomotion is a mix of energy used
to generate force isometrically and to perform work, as
muscles perform a variety of functions by acting as tensile
struts, brakes, and motors during different periods of the
stride (14). Despite this complexity of muscle function,
the metabolic cost of walking should be proportional to the
magnitude and rate of generating force if muscles work with
consistent relative shortening velocities and efficiencies at
different walking speeds (14). Bourdin et al. (6) have shown
a close relation between EMG characteristics and the energy
cost of running per unit distance. Previously, EMG shorts
have been used to detect second ventilatory threshold during
incremental treadmill running (33) and to measure activity
and inactivity times in daily life (10,32).
Because people spend more time in light activities than
that in moderate and vigorous activities, estimation of EE at
low intensities is of great importance in long-term measure-
ments (34). The methods aEMG(H), aEMG(QH), and HR
had significantly higher means of correlation coefficients
(Table 2) and lower AIC and RMSE than those in ACC at
low loads (Table 2 and 4), indicating that measurement of
low-intensity activities seems to be more accurate with EMG
and HR than that with ACC when individual calibrations are
performed.
Hamstring aEMG had high mean correlations across
conditions (0.94, 0.93, and 0.96 for all, low, and level loads,
respectively) (Table 2). Hamstring aEMG predicted EE at
least as accurately as or better than aEMG from quadriceps
or the averaged quadriceps and hamstrings value. This might
be due to the fact that hamstrings are responsible for pro-
pulsion in walking (21), whereas the rectus femoris of
quadriceps is mainly responsible for leg swing, which ac-
counts for only 10% of the metabolic cost of walking on
level ground (13). Although hamstrings seem to provide a APPLIED SCIENCES
better estimation of EE in locomotion, it might be advisable
to use both muscle groups combined in measurement of all
daily activities, which is also supported by RMSE results in
Table 4 (first column, all loads). Because daily activities in-
clude activities other than locomotion, having more muscle

TABLE 3. Averages and SD of intercepts and slopes for fitted regression lines for all loads from combined thigh muscle activity (aEMG(QH)), quadriceps muscle activity (aEMG(Q)),
hamstring muscle activity (aEMG(H)), HR, and ACC.
Intercept (a) Slope (b)
aEMG(QH) aEMG(Q) aEMG(H) HR ACC aEMG(QH) aEMG(Q) aEMG(H) HR ACC
Average 1.11 0.71 1.80 j5.46 0.95 0.63 0.88 0.49 0.11 1.04
SD 1.12 1.11 1.21 2.15 1.29 0.31 0.50 0.23 0.03 0.38

ENERGY EXPENDITURE ESTIMATION IN LOCOMOTION Medicine & Science in Sports & Exercised 1835

Copyright © 2014 by the American College of Sports Medicine. Unauthorized reproduction of this article is prohibited.
 TABLE 4. Prediction equations for EE (kJIminj1) from combined thigh muscle activity (aEMG(QH)), quadriceps muscle activity (aEMG(Q)), hamstring muscle activity (aEMG(H)), HR, and
ACC and the RMSE using both the fitted models from leave-one-out cross-validation for the prediction at the population level and at the individual level and further AIC of the original
fitted model.
Regression RMSE Pop RMSE Ind AIC Ind
All loads
EE = 0.627 + 0.254
sex + 0.015
age + (0.968 + 0.179
sex j 0.011
age)
aEMG(QH) 4.21 (3) 1.65 (2) 2439 (3)
EE = 0.514 + 0.203
sex + 0.010
age + (1.375 + 0.265
sex j 0.017
age)
aEMG(Q) 4.78 (5) 2.14 (4) 2659 (4)
EE = 1.241 + 0.634
sex + 0.012
age + (0.682 + 0.121
sex j 0.007
age)
aEMG(H) 4.54 (4) 1.65 (2) 2391 (2)
EE = j4.832 j 0.659
sex + (0.094 + 0.030
sex)
HR 2.31 (1) 1.06 (1) 1808 (1)
EE = 1.249 + 0.503
sex + (0.774 + 0.353
sex)
ACC 3.48 (2) 3.16 (5) 2810 (5)
Low loads
EE = 0.739 j 0.001
age + 0.503
sex + (0.936 j 0.009
age)
aEMG(QH) 2.33 (4) 0.98 (3) 1129 (3)
EE = 1.398 j 0.007
age + 0.880
sex + (0.886 j 0.008
age)
aEMG(Q) 2.30 (3) 1.33 (5) 1253 (5)
EE = 1.141 j 0.005
age + (0.788 j 0.006
age)
aEMG(H) 2.69 (5) 0.81 (1) 1006 (1)
EE = j3.928 j 0.854
sex + (0.083 + 0.031
sex)
HR 1.93 (2) 0.90 (2) 1126 (2)
EE = 1.041 + 0.326
sex + (0.622 + 0.250
sex)
ACC 1.57 (1) 1.26 (4) 1134 (4)
Ranks for prediction accuracy are given in parentheses.
ACC, counts per minute; age, year; sex, 1 for men and 0 for women; aEMG, % EMGMVC; HR, beats per minute.
Ind, individual level; pop, population level.

groups in the analysis seems well-grounded. For example,
it has been shown that for EMG threshold detection in run-
ning, it is better to use both quadriceps and hamstrings than
either alone (33).
As aEMG(Q) and EE correlations were significantly
higher at level loads than at low and all loads, it seems that,
especially at low loads, accuracy of aEMG(Q) for predicting
EE is considerably decreased (Table 2). One reason for this
finding could be that some subjects were unaccustomed
to treadmill walking and, especially, some older subjects
seemed uncertain when walking on a treadmill. Older sub-
jects maintaining balance at the lowest walking speeds could
be a reason for excessive coactivation of the quadriceps (27).
The lowest walking speeds were also performed first, leav-
ing less time for adaptation to treadmill walking, compared
with higher speeds.
Naturally, other muscles not assessed in this study are also
activated in locomotion and daily tasks. Activity of other
muscles not measured in this study may create bias in EE
estimations but, in theory, only in those instances in which
activity of muscles not assessed increases disproportionately
to activity of measured thigh muscles. Sousa and Tavares
(29) have shown that gait speed influences not only activity
levels but also relative activity patterns of leg muscles. Ac-
tivity of the rectus femoris, gluteus maximus, gastrocnemius
medialis, and biceps femoris increase (in decreasing order)
during stance below and above the preferred walking speed
of individuals. Nonetheless, it could be assumed that most of
APPLIED SCIENCES
(22). Back muscles keep the trunk erect during walking and
provide appropriate equilibrium between flexibility and
stiffness (30). Mean amplitudes and phasic activation in trunk
muscles generally increase with increasing walking speed,
although activation of back muscles remains almost un-
changed between walking at speeds of 4 and 6 kmIhj1 (2).
It has been shown that although arm swing is not entirely
passive, arm swing actually reduces the metabolic cost of
locomotion (23). With increases in uphill grade, hip, knee,
and ankle extensor muscle activities increase progressively
in the stance phase (12). In downhill walking, only knee
extensor activities increase with steeper downhill grade and
ankle extensor muscle activities decrease with increasing
downhill slope (12). The importance of thigh muscles either
in deceleration or ACC in uphill and downhill locomotion is
supported by the present finding that EMG of thigh muscles
expressed high correlations at all loads, justifying the use of
EMG for estimating EE in those activities.
When considering EE estimations by ACC, several arti-
cles have previously shown validity of ACC in predicting EE
in locomotion on level ground, which is also shown in the
present study. However, accuracy of ACC was lower when
locomotion also included uphills and downhills (i.e., at all
loads) (Tables 2 and 4). This finding is in line with studies
that found that standard analysis of body ACC cannot accu-
rately predict EE in uphill or downhill walking (31) or stair
ascent or descent (18). This is probably explained by the fact
that uphill walking requires extra positive work but does not

the time in nonexercise activities of normal daily life, people
walk close to their preferred speed.
Plantar flexor muscle activity increases at faster walking
speeds (3), and these muscles contribute greatly to me-
chanical energetic demands of walking and are the primary
contributors to propulsive ground reaction forces in walking
produce ACC proportional to the extra muscle work needed
(the opposite being true in downhill walking). Because hills
and stairs are part of normal daily environments, validity of
ACC is lower in real-life situations than that shown in
treadmill validation studies performed only with level loads.
Validity studies of ACC in free-living conditions also need to

TABLE 5. Correlation coefficients and typical errors (SEM) in two tests performed on different days (n = 5).
aEMG(QH) aEMG(Q) aEMG(H) HR ACC EE
R values 0.96 0.93 0.97 0.67 0.98 0.99
Typical error 1.98% EMGMVC 2.17% EMGMVC 2.23% EMGMVC 20.65 bpm 1.03 countsIminj1 0.35 kJIminj1
aEMG(H), hamstring muscle activity; aEMG(Q), quadriceps muscle activity; aEMG(QH), combined thigh muscle activity.

1836 Official Journal of the American College of Sports Medicine http://www.acsm-msse.org

Copyright © 2014 by the American College of Sports Medicine. Unauthorized reproduction of this article is prohibited.
TABLE 6. SE of the estimates of fixed coefficients with P values and the SD of random coefficients and residual for hamstring muscle activity (aEMG(H)), HR, and ACC in all and low loads.
Fixed Coefficients Random Coefficients and Residual
All loads
aEMG(H) Intercept aEMG(H) Sex Age aEMG(H)
sex aEMG(H)
age SD (u0) SD (u1) SD (e)
SE 0.488 0.088 0.308 0.012 0.055 0.002 0.595 0.189 1.805
P value 0.011 G0.001 0.045 0.303 0.030 G0.001
HR Intercept HR Sex HR
sex SD (u0) SD (u1) SD (e)
SE 0.395 0.005 0.557 0.007 1.567 0.022 1.146
P value G0.001 G0.001 0.242 G0.001
ACC Intercept ACC Sex ACC
sex SD (u1) SD (e)
SE 0.409 0.059 0.582 0.085 0.134 3.310
P value 0.002 G0.001 0.392 G0.001
Low loads
aEMG(H) Intercept aEMG(H) Age aEMG(H)
age SD (u1) SD (e)
SE 0.301 0.123 0.007 0.003 0.274 0.848
P value G0.001 G0.001 0.544 0.045
HR Intercept HR Sex HR
sex SD (u0) SD (u1) SD (e)
SE 0.612 0.008 0.876 0.011 2.471 0.035 0.978
P value G0.001 G0.001 0.334 G0.001
ACC Intercept ACC Sex ACC
sex SD (u1) SD (e)
SE 0.233 0.056 0.337 0.082 0.168 1.337
P value G0.001 G0.001 0.337 0.003

be interpreted with caution because many studies do not
mention partial correlations of activity counts when multiple
regressions are used with subject characteristics and activity
counts (24).
HR tends to be linearly related within an individual
throughout a large portion of the aerobic work range (19),
which was also seen in this study, as correlation values for
HR were very high for all and level loads. However, several
factors besides V̇O2 can affect HR including emotions, food
intake, body position, muscle groups used, type of muscle
activity, and ambient temperature (16). The effect of some of
these factors is probably demonstrated in our results, as
correlations were significantly lower for low loads (and SD
were considerably higher) than those for all and level loads.
Furthermore, test–retest reliability for HR was considerably
lower than that for EMG and ACC (Table 5).
Regarding the accuracy of EE prediction, individual cor-
relations and AIC ranked the methods in the same order
within a certain load categorization. It has to be noted that
AIC tells how well a model can predict EE in a relative
sense. Although individual correlations between EE and
different methods were rather high, correlations do not take
into account the slope of the curves among different in-
dividuals. Therefore, to account for the effect of the method
on EE, it is necessary to establish a linear regression equa-
tion for EE and the method used (e.g., HR, ACC, or EMG)
for each subject at several activity intensities at the first stage
and the linear mixed model at the second stage. This kind of
individual calibration will improve EE estimations accounting
for individual factors affecting EMG (e.g., walking technique,
muscle mass, thickness of subcutaneous fat tissue), HR (e.g.,
cardiovascular fitness, resting HR, and HRmax), and ACC
(e.g., walking technique and body mass). In practice, to
perform individual calibration, one needs to establish an
equation by measuring EMG and V̇O2 simultaneously, at
minimum, during two different workloads. It has to be noted
that this demands quite a bit of additional work and might be
inconvenient in some research and applied settings.
HR had low SD in slopes (b) at all loads, perhaps indi-
cating that for every increase in HR, EE increases by ap-
proximately the same amount independent of the individual.
However, even small deviations in HR can make the indi-
vidual regression coefficient significant in the linear mixed
model (see the SD of the random coefficients of HR in
Table 6). It has to be noted that the range of the HR is larger,
which means that the values of slope cannot be very large with
respect to the range of EE. High SD in intercept together with
significant individual intercept in the mixed model indicates
that absolute EE at a given HR differs between individuals
(Tables 3 and 6). SD of intercepts (a) and slopes (b) in all
EMG variables and ACC were of the same magnitude, indi-
cating that individual differences are quite similar for regres-
sion equations; linear mixed models of those variables only
had individual slope besides aEMG(H), which had also an
individual intercept at all loads but not at low loads (Table 6).
The equations to predict instantaneous EE provide a
convenient way to estimate EE in long-term EMG mea-
surements. The accuracy of the prediction equations were
evaluated in three ways: the RMSE of cross-validation at the
population and individual level and AIC of the original fitted
model. Thus, RMSE at the individual level rank the accu-
racies of five methods almost similarly as AIC, as expected
(Table 4). At the population level, HR seems to work best at
all loads and ACC, at low loads. At all and low loads, we
can see that the SD of residual in ACC model is larger when
APPLIED SCIENCES
compared with that in others (Table 6), meaning that pre-
diction accuracy was not improved much by applying indi-
vidual calibration (Table 4), whereas the prediction with HR
and especially with aEMG is improved considerably when
individual calibration is performed. This is due to the ad-
vantage of the use of variations in the individual slopes and
regression coefficients when included in the model (Table 6).
At low loads, for example, the SD of the random coefficient
(u1) for aEMG(H) is larger and the SD of residual (e) is
smaller when compared with that of ACC. HR has further
random constant with quite larger deviation. It is important to

ENERGY EXPENDITURE ESTIMATION IN LOCOMOTION Medicine & Science in Sports & Exercised 1837

Copyright © 2014 by the American College of Sports Medicine. Unauthorized reproduction of this article is prohibited.
 note that prediction equations for HR contain a negative
constant that is nonphysiological. Therefore, the prediction
equation using HR is limited for HR values larger than 44.28
for men and 51.40 for women at all loads and for HR values
larger than 41.94 for men and 47.32 for women at low loads
(in the accuracy studies, negative values were included). Note
that predictions are also affected by the accuracy of the fixed
coefficients (SE in Table 6). Accuracies of the methods at the
population level can be further improved by calibration of the
methods for a new individual. In this study, the subject group
was intentionally selected to be as heterogeneous as possible
for wide generalizability of the results. However, validity of
the equations should also be verified in another study with
different subjects.
This study used an unconventional treadmill protocol,
which had its shortcomings. Although the test protocol might
underestimate the V̇O2max values (because local muscle fa-
tigue might terminate the test prematurely), those were of
secondary importance in this study. In addition, because of
the treadmill protocol used, at the end of the test, some sub-
jects were doing a substantial amount of anaerobic work in-
dicated by respiratory quotient values of greater than 1.0
(0.93 T 0.10 and 1.04 T 0.13 for men and women, respec-
tively) but this was only at the last load and probably had a
minimal effect on results. On the other hand, this unconven-
tional test protocol relates more closely to normal daily lo-
comotion because extra stress of locomotion usually comes
from ascent rather than higher speed. Excluded data also
present a limitation to this study. One EMG channel in 33%
(18/54) of subjects and two channels in 3.7% (2/54) of sub-
jects were excluded from the analysis. In addition, 1.3% of
data points were excluded from the remaining channels.
REFERENCES
1. Akaike H. A new look at the statistical model identification. IEEE
Trans Automatic Control. 1974;19(6):716–23.
2. Anders C, Wagner H, Puta C, Grassme R, Petrovitch A, Scholle
HC. Trunk muscle activation patterns during walking at different
speeds. J Electromyogr Kinesiol. 2007;17(2):245–52.
3. Bartlett JL, Kram R. Changing the demand on specific muscle
groups affects the walk–run transition speed. J Exp Biol. 2008;
211(Pt 8):1281–8.
4. Bonomi AG, Plasqui G, Goris AH, Westerterp KR. Aspects of
activity behavior as a determinant of the physical activity level.
Scand J Med Sci Sports. 2012;22(1):139–45.
5. Booth FW, Roberts CK, Laye MJ. Lack of exercise is a major
APPLIED SCIENCES
cause of chronic diseases. Compr Physiol. 2012;2:1143–211.
6. Bourdin M, Belli A, Arsac LM, Bosco C, Lacour JR. Effect of
vertical loading on energy cost and kinematics of running in
trained male subjects. J Appl Physiol (1985). 1995;79(6):2078–85.
7. Carter J, Jeukendrup A. Validity and reliability of three commer-
cially available breath-by-breath respiratory systems. Eur J Appl
Physiol. 2002;86(5):435–41.
8. Crouter SE, Churilla JR, Bassett DR Jr. Estimating energy expendi-
ture using accelerometers. Eur J Appl Physiol. 2006;98(6):601–12.
9. Diggle P, Liang K, Zeger S. Analysis of longitudinal data. Oxford:
Clarendon Press; 1994;122–3.
10. Finni T, Haakana P, Pesola AJ, Pullinen T. Exercise for fitness
does not decrease the muscular inactivity time during normal daily
life. Scand J Med Sci Sports. 2014;24(1):211–9.
1838 Official Journal of the American College of Sports Medicine
Copyright © 2014 by the American College of Sports Medicine. Unauthorized reproduction of this article is prohibited.
In general, EMG measurements can be more prone to
measurement errors than many other methods to measure
physical activity. EMG measurements become more difficult
with increasing obesity because EMG signal is attenuated by
fat tissues between the muscle and the measuring electrodes.
There can also be more subcutaneous tissue movement (e.g.,
due to impact forces to ground contact) in obese people, and
this might affect the quality of the recorded EMG signal. In
addition, one of the shortcomings of EMG shorts is that people
of different sizes naturally need different sizes of shorts.
In conclusion, this study shows that EMG shorts can be
used in EE estimations across a wide range of physical ac-
tivity intensities within a heterogeneous subject group.
Overall, EMG, HR, and ACC all predicted EE well in level
loads when individual calibrations were included. At all
loads, HR was the best predictor in all methods studied. At
low loads of physical activity in changing terrains, ACC was
the best predictor at the population level, but when indi-
vidual calibrations are performed, hamstring muscle EMG
provides more accurate EE estimations than those in ACC.
This study was supported by the Academy of Finland (#128643,
T. F.), Ministry of Education and Culture (42/627/2010, T. F),
Urheiluopistosäätiö (Finnish Sport Institute Foundation, O. T.), Ellen ja
Artturi Nyyssösen Säätiö (O. T.), and the University Alliance Finland
project STATCORE (S. K.).
Myontec Ltd. and Suunto Oy are acknowledged for providing
equipment and technical support. The authors gratefully acknowl-
edge Gauthier Perez, M.Sc., for data collection and analysis, Dr.
Timo Rantalainen for his valuable help in data analysis, and Dr. Neil
Cronin for revision of English language. We would also like to thank
Prof. Tapio Nummi, Prof. Antti Penttinen, and Dr. Lauri Mehtätalo for
fruitful discussions and suggestions.
The authors declare no conflict of interest.
The results of this study do not constitute endorsement by the
American College of Sports Medicine.
11. Finni T, Hu M, Kettunen P, Vilavuo T, Cheng S. Measurement of
EMG activity with textile electrodes embedded into clothing.
Physiol Meas. 2007;28(11):1405–19.
12. Franz JR, Kram R. The effects of grade and speed on leg muscle
activations during walking. Gait Posture. 2012;35(1):143–7.
13. Gottschall JS, Kram R. Energy cost and muscular activity required
for leg swing during walking. J Appl Physiol. 2005;99(1):23–30.
14. Griffin TM, Roberts TJ, Kram R. Metabolic cost of generating
muscular force in human walking: insights from load-carrying and
speed experiments. J Appl Physiol (1985). 2003;95(1):172–83.
15. Hamilton MT, Hamilton DG, Zderic TW. Role of low energy ex-
penditure and sitting in obesity, metabolic syndrome, type 2 dia-
betes, and cardiovascular disease. Diabetes. 2007;56(11):2655–67.
16. Hebestreit H, Bar-Or O. Influence of climate on heart rate in
children: comparison between intermittent and continuous exer-
cise. Eur J Appl Physiol Occup Physiol. 1998;78(1):7–12.
17. Kesaniemi YK, Danforth E Jr, Jensen MD, Kopelman PG,
Lefebvre P, Reeder BA. Dose-response issues concerning physical
activity and health: an evidence-based symposium. Med Sci Sports
Exerc. 2001;33(6 Suppl):S351–8.
18. Kozey SL, Lyden K, Howe CA, Staudenmayer JW, Freedson PS.
Accelerometer output and MET values of common physical ac-
tivities. Med Sci Sports Exerc. 2010;42(9):1776–84.
19. Livingstone MB. Heart-rate monitoring: the answer for assessing
energy expenditure and physical activity in population studies? Br
J Nutr. 1997;78(6):869–71.
http://www.acsm-msse.org
20. Mitchell WK, Williams J, Atherton P, Larvin M, Lund J, Narici M.
Sarcopenia, dynapenia, and the impact of advancing age on human
skeletal muscle size and strength; a quantitative review. Front
Physiol. 2012;3:260.
21. Neptune RR, Clark DJ, Kautz SA. Modular control of human
walking: a simulation study. J Biomech. 2009;42(9):1282–7.
22. Neptune RR, Sasaki K, Kautz SA. The effect of walking speed on
muscle function and mechanical energetics. Gait Posture. 2008;
28(1):135–43.
23. Ortega JD, Fehlman LA, Farley CT. Effects of aging and arm swing
on the metabolic cost of stability in human walking. J Biomech.
2008;41(16):3303–8.
24. Plasqui G, Westerterp KR. Physical activity assessment with ac-
celerometers: an evaluation against doubly labeled water. Obesity
(Silver Spring). 2007;15(10):2371–9.
25. Rennie KL, Hennings SJ, Mitchell J, Wareham NJ. Estimating
energy expenditure by heart-rate monitoring without individual
calibration. Med Sci Sports Exerc. 2001;33(6):939–45.
26. Rietjens GJ, Kuipers H, Kester AD, Keizer HA. Validation of a
omputerized metabolic measurement system (Oxycon-Pro) during
low and high intensity exercise. Int J Sports Med. 2001;22(4):291–4.
27. Schmitz A, Silder A, Heiderscheit B, Mahoney J, Thelen DG.
Differences in lower-extremity muscular activation during walking
ENERGY EXPENDITURE ESTIMATION IN LOCOMOTION
Copyright © 2014 by the American College of Sports Medicine. Unauthorized reproduction of this article is prohibited.
View publication stats
between healthy older and young adults. J Electromyogr Kinesiol.
2009;19(6):1085–91.
28. Skotte J, Korshoj M, Kristiansen J, Hanisch C, Holtermann A.
Detection of physical activity types using triaxial accelerometers.
J Phys Act Health. 2014;11(1):76–84.
29. Sousa ASP, Tavares JM. Effect of gait speed on muscle activity
patterns and magnitude during stance. Motor Control. 2012;16(4):
480–92.
30. Swinnen E, Baeyens JP, Meeusen R, Kerckhofs E. Methodology
of electromyographic analysis of the trunk muscles during walking
in healthy subjects: a literature review. J Electromyogr Kinesiol.
2012;22(1):1–12.
31. Terrier P, Aminian K, Schutz Y. Can accelerometry accurately
predict the energy cost of uphill/downhill walking? Ergonomics.
2001;44(1):48–62.
32. Tikkanen O, Haakana P, Pesola A, et al. Muscle activity and
inactivity periods during normal daily life. PLoS ONE. 2013;
8(1):e52228.
33. Tikkanen O, Hu M, Vilavuo T, Tolvanen P, Cheng S, Finni T.
Ventilatory threshold during incremental running can be estimated
using EMG shorts. Physiol Meas. 2012;33(4):603–14.
34. Westerterp KR, Plasqui G. Physical activity and human energy
expenditure. Curr Opin Clin Nutr Metab Care. 2004;7(6):607–13.
APPLIED SCIENCES
Medicine & Science in Sports & Exercised 1839