Energy expenditure estimation in beta-blocker medicated cardiac patients by
combining heart rate and body movement data
Jos J Kraala, Francesco Sartorb, Gabriele Papinib, Wim Stutb, Niels Peekc, Hareld M.C.
Kempsd, Alberto G Bonomib
a Department of Medical Informatics, Academic Medical Center Amsterdam, University
of Amsterdam, Amsterdam, The Netherlands
b Personal Health department, Philips Research, Eindhoven, The Netherlands
c Health eResearch Centre, Farr Institute for Health Informatics Research, University of
Manchester, Manchester, United Kingdom
d Department of Cardiology, Máxima Medical Center Veldhoven, Veldhoven, The
Netherlands
Word count: 5196
Corresponding author:
Jos J Kraal, Department of Medical Informatics, Academic Medical Center,
Meibergdreef 9, 1105 AZ Amsterdam, The Netherlands
Email: [email protected]
Abstract
Background: Accurate assessment of energy expenditure provides the opportunity to
monitor physical activity during cardiac rehabilitation. However, available assessment
methods, based on the combination of heart rate (HR) and body movement data, are not
applicable for patients using beta-blocker medication. Therefore, we developed an
energy expenditure prediction model for beta-blocker medicated cardiac rehabilitation
patients.
Methods: Sixteen male cardiac rehabilitation patients (age: 55.8±7.3 years, weight:
93.1±11.8 kg) underwent a physical activity protocol with 11 low to moderate intensity
common daily life activities. Energy expenditure was assessed using a portable indirect
calorimeter. HR and body movement data were recorded during the protocol using
unobtrusive wearable devices. In addition, patients underwent a symptom limited
exercise test and resting metabolic rate assessment. Energy expenditure estimation
models were developed using multivariate regression analyses based on HR, body
movement data and/or patient characteristics. In addition, a HR-flex model was
developed.
Results: The model combining HR, body movement and patient characteristics showed
the highest correlation and lowest error (R2=0.84, RMSE=0.834 kcal/min) with total
energy expenditure. The method based on individual calibration data (HR-flex) showed
lower accuracy (R2=0.83, RMSE=0.992 kcal/min).
Conclusions: Our results show that combining HR and body movement data improves
the accuracy of energy expenditure prediction models in cardiac patients, similar to
2
methods developed for healthy subjects. The proposed methodology does not require
individual calibration and is based on data available in clinical practice.
Word count: 228
Keywords: Accelerometry, heart rate monitoring, physical activity, energy expenditure,
cardiac rehabilitation, beta-blockers.
3
Introduction
Physical inactivity is an expanding problem and is related to several chronic diseases
including cardiovascular diseases, diabetes and obesity 1. Lack of physical activity may
lead to a chronic imbalance between energy intake and energy expenditure, which is
associated with all-cause and cardiovascular mortality 2,3. After a cardiac incident,
patients are recommended to participate in cardiac rehabilitation to improve their
physical fitness level and engage in a healthy lifestyle 4,5. Cardiac rehabilitation is a
multidisciplinary intervention to improve physical, psychological and social well being
of patients after a cardiac incident 6. However, the majority of the cardiac rehabilitation
programs focus on the improvement and enhancement exercise capacity and quality of
life, but do not target sedentary time and overall physical activity behavior in daily life.
This may be related to the fact that to date, no standardized methods exist to assess
physical activity accurately and unobtrusively in cardiac patients in daily life 7.
Physical activity is a multi-dimensional human behavior characterized by factors
such as type, pattern, duration and intensity of physical tasks and can be quantified by
determining energy expenditure 8. Gold standard methods for the assessment of energy
expenditure are indirect calorimetry and doubly labelled water. Although these methods
were shown to be accurate, they are costly and obtrusive, making them unsuitable for
population-based studies or integration in clinical practice 9.
Over the past decade, the use of wearable movement sensors for measuring
physical activity in free-living conditions has become widespread. Heart rate (HR)
monitors provide insight into the intensity of the aerobic work during physical activity.
During activities with moderate to vigorous intensity, HR and oxygen consumption
4
(VO2) are linearly related 10,11. Therefore energy expenditure, which is dependent on
VO212, can be predicted using HR recordings. However, in free-living conditions energy
expenditure assessment using HR has some important limitations13. First, the HR vs.
energy expenditure relation is not linear during rest and low-intensity activities and is
substantially influenced by internal and external factors (i.e. temperature, emotional
stress, caffeine) 14. In addition, inter-individual variation in HR is large, suggesting the
need for individual calibration processes to build a valid HR-based energy expenditure
model 15.
Yet, combining HR and accelerometer data may reduce these limitations 16. Tri-
axial accelerometers have been used to quantify body movement and predict energy
expenditure using linear models in healthy subjects and patients with a chronic disease
17–19. Accelerometers are small and minimally obtrusive, with battery life and storage
capacity enabling long-term monitoring from several days to weeks. Brage et al.
described a branched equation model in which a combination of HR and accelerometer
recordings was used to estimate physical-activity related energy expenditure in healthy
subjects 20. Other studies showed that models combining HR and accelerometer data
significantly improved energy expenditure estimation accuracy as compared to single-
input only models in children and healthy adults 21,22.
Although combining HR and accelerometer data for energy expenditure
assessment seems to lead to higher estimation accuracy in healthy adults, these
prediction models cannot be translated directly to cardiac patients. Beta-blocker therapy,
which aims at reducing myocardial oxygen demand by lowering heart rate and blood
pressure 23, plays an important role in the treatment of cardiovascular diseases and the
5
prevention of future cardiac events. The heart rate-lowering effect of beta-blockers,
which is highly variable between subjects, may impact the ability of HR to predict
energy expenditure. Therefore, the main objective of this study was to develop an
energy expenditure prediction model for beta-blocker medicated cardiac rehabilitation
patients based on both HR and accelerometer data. For this aim, multivariate linear
regression models were developed, and different combinations of features were
evaluated. In addition, a method based on a calibration protocol for model
personalization, called the flex-HR method 24, was evaluated as a benchmark solution.
6
Methods
Study population
Patients were eligible for participation when they were admitted to cardiac rehabilitation
two to four weeks after hospitalisation for myocardial infarction, unstable angina, or a
revascularization procedure (percutaneous coronary intervention or coronary artery
bypass grafting) at the Máxima Medical Center, Veldhoven (The Netherlands). High
risk of further events was an exclusion criterion (i.e. patients with symptomatic heart
failure, complex congenital heart disease, severe depression, arrhythmias or co-
morbidity limiting exercise performance). Recruitment was performed between
November 2013 and August 2014. Sixteen male patients that were participating in
exercise-based cardiac rehabilitation were recruited. Fourteen patients were taking beta-
blockers (metoprolol) at the time of the test. Each participant signed an informed
consent form. The study was approved by the local Medical Ethical Committee of
Máxima Medical Center, Veldhoven, The Netherlands.
Symptom limited exercise test
A symptom limited exercise test was carried out on a cycle ergometer in an upright-
seated position on an electromagnetically braked cycle ergometer (Lode Corrival, Lode
BV, Groningen), using an individualized ramp protocol aiming at a test duration of 8-12
minutes. Ventilatory parameters were measured breath-by-breath (Masterscreen CPX,
CareFusion, Hoechberg, Germany). The test was ended when the patient was not able to
maintain the required pedaling frequency. PeakVO2 was recorded as the final 30-second
averaged value of the test.
7
Physical activity protocol
The physical activity protocol, as described in Table 1, consisted of a randomly ordered
set of 11 daily living activities of low to moderate intensity. Resting HR (RHR) was
measured in the beginning of the test, and after each activity the patient had a recovery
break, which lasted until the HR reached the resting value.
Energy expenditure measurements
Oxygen uptake (VO2) and carbon dioxide production (VCO2) were measured for the
entire duration of the activity protocol by the Cosmed K4 b2 (Cosmed, Rome, Italy)
portable metabolic system, which is a breath-by-breath pulmonary gas exchange
measurement system consisting of a face mask, an analyzer unit, and a battery. Total
energy expenditure (TEE) was then obtained using the Weir equation from the breath-
by-breath measurements of O2 and CO2, averaged on a minute-by-minute basis and
divided by resting metabolic rate (RMR) to determine the physical activity level (PAL).
PAL is a popular parameter used to represent energy expenditure adjusted for RMR,
which allows to compare measurements of TEE for subjects with different body size
and composition. PAL can be used to determine activity intensity. Light, moderate, and
vigorous intensity activities are characterized by a PAL < 3, < 6, and ≥ 6, respectively
25. The average daily PAL for the healthy adult population is 1.7, a PAL of 1.2
represents the activity level of a bed-bound subjects 26, while an athlete participating in
the Tour de France can reach a PAL of 5 27.
8
RMR measurement
Prior to the RMR assessment, patients were instructed to fast for 12 hours and to avoid
sports activities for 24 hours. During the assessment, patients were asked to stay awake.
RMR was measured using the Cosmed metabolic equipment with the Canopy option
using the mean values of VO2 and VCO2 data collected over a five minutes stability
interval according to Matarese’s recommendations 28.
Physical activity and heart rate measurements
Physical activity was measured using an ActiGraph wGT3X+ activity monitor
positioned at the waist (ActiGraph, Pensacola, Florida), which is a tri-axial
accelerometer with dynamic range of ±6G and sampling frequency set to 40Hz.
Acceleration data were used to determine activity counts per minute (ACmin) according
to the following equation:
where, for each i-sample in a minute period (T) the vector magnitude signal (v) is
subtracted to the mean v over that minute to determine ACmin. The v signal is obtained
from the Euclidean norm of the x, y, z signal representing each sensing axis of the
accelerometer. HR was recorded on a second-by-second frequency using a
commercially available ECG chest belt (Garmin Ltd.) that wirelessly transmitted HR
data to the ActiGraph accelerometer unit for synchronous data storage. Minute-by-
9
minute average for HR and ACmin were computed to predict energy expenditure. HR
above rest (HRnet) was obtained by subtracting RHR from each value of the HR. Data
were visually inspected and annotated to determine start and stop for each activity over
the recorded acceleration signals. The recorded signals were processed and organized
using MatLab (Matlab, Mathworks, Cambridge, MA).
Energy expenditure modeling
TEE and PAL estimation models were developed using multivariate linear regression
analysis. This method allowed for automatically and iteratively selecting which
independent variables to include in the prediction models. Three different groups of
independent variables were considered: i) patient characteristics which included age,
body weight, height, BMI, RMR, Beta-blocker dose (i.e. metoprolol 0, 50, 75 or
100mg/day), and peakVO2; ii) Body movement as described by ACmin, and the
logarithm transform of ACmin; and iii) HR and HRnet information. Five different
energy expenditure prediction models were derived using the following combination of
independent variables: a) Body movement features; b) Body movement + patient
characteristics; c) HR features; d) HR features + patient characteristics; e) Body
movement + HR features + patient characteristics. Patient characteristics were included
in the modeling process to account for between-individual variability in the relationship
between HR or ACmin and energy expenditure (TEE or PAL) 29. HR and body
movement features were included to describe the within-individual variability in energy
expenditure.
10
Additionally, we used the HR-flex method as alternative solution to account for
between-patient variability in the energy expenditure prediction models 24. This method
deploys HR and energy expenditure measurements during a set of cycling activities at
increasing intensity (0W to 70W) to determine parameters of a linear prediction model
applicable to any other activity for HR values above the HR-flex. HR-flex is defined as
the lowest HR recorded during cycling at very low effort (0W). Below the HR-flex
energy expenditure is set to resting values. The HR-flex method leads to development of
an energy expenditure prediction equation by personalizing model parameters using
data from a cycling calibration protocol.
Statistics
Data are presented as mean standard deviation. Pearson correlation coefficient was
calculated to describe the association between ACmin, HR features and TEE or PAL for
each individual patient data or for the entire dataset. Student t-test was used to assess
significant differences in the estimation error of the energy expenditure prediction
models between the development and validation samples. Development of the energy
expenditure prediction model was carried out with data from 11 patients, while data
from the remaining 5 patients were used for validation. For the HR-flex model, patient
data from both groups (from the development and validation group) were used for
developing the patient-specific prediction model. Significance level was set to p < 0.05.
All statistical analyses were carried out using RStudio (version 0.98.507, R
Development Core Team, Free Software Foundation Boston, MA, USA).
11
12
Results
Sixteen patients agreed to participate. The dataset included a set of static variables
indicating the individual characteristics as well as a set of time-varying features like
TEE, PAL, HR and ACmin for an average of 64 minutes/patient (total 1027 data points
for each time-varying variables). Patients were divided in two groups for model
development and validation. Baseline characteristics are presented in Table 2. On the
entire study population, a significant association was found between peakVO2 and
resting VO2 (R = 0.65, p < 0.01) and between peakVO2 and RMR (R = 0.52, p < 0.05).
RHR was not significantly correlated with any of the patient characteristics. Max HR at
the symptom limited exercise test was negatively associated with age (R = - 0.57, p <
0.05) and positively associated with peakVO2 (R = 0.63, p < 0.01).
A significant correlation was found between HR or body movement data (HR,
HRnet, ACmin) and measures of energy expenditure (Table 3). On a group level HRnet
showed the strongest association with energy expenditure as compared to ACmin or
HR. On an individual level, the average correlation between HR and energy expenditure
sharply increased, indicating a strong between-subject difference in the relationship
between HR and TEE or PAL.
Results from the stepwise multivariate regression analysis are shown in Table 4
and Figure 1. Model e) including a combination of body movement, HR features, and
patient characteristics showed the lowest estimation error and the largest correlation
with both TEE and PAL (> 76% and > 83%). When HR features were omitted as input
to the prediction model (Models a and b) energy expenditure estimates showed larger
error (Table 4).
13
The HR-flex model showed the energy expenditure estimation accuracy was
comparable with some of the models based on a combination of HR and body
movement features (Table 4). However, accuracy was lower than the most accurate
prediction model based on HR, body movement and patient characteristics (Figure 2).
The HR-flex model accuracy was not possible to test using a hold-out group of patients
(like the validation group) since the method required individual calibration to set the
value of constituting parameters.
14
Discussion
In this study we developed an energy expenditure prediction model for cardiac
rehabilitation patients using beta-blockers, based on body movement and HR data. A
multivariate regression model shows the lowest estimation error in the estimation of
PAL and TEE when HR, body movement and patient characteristics are included.
Personalization of the prediction model using the HR-flex model with a cycling protocol
for calibration resulted in somewhat lower accuracy than the most accurate multivariate
model.
To our knowledge, this study is the first to validate energy expenditure
estimation models based on acceleration and HR in cardiac patients taking beta-
blockers. The observed correlations between body movement features and TEE as well
as HR features and TEE were comparable to studies in healthy subject 30,31. In addition,
a prediction model using individual HR features or body movement features showed
comparable coefficients with studies in healthy subjects 24,32. As discussed in previous
literature, accuracy of energy expenditure estimation can improve when HR and body
movement data are combined 20,33. A validation study in healthy subjects demonstrated a
lower RMSE and higher correlation in a HR-body movement model estimating free-
living PAEE (r2 = 0.78) as compared to a model using estimates from HR or body
movement alone (r2 = 0.59 and r2 = 0.61 respectively) 20. We observed similar results for
the individual HR-body movement regression models in cardiac patients (TEE: r2 = 0.56
and r2 = 0.64 respectively), and a similar improvement in r2 and RMSE in the regression
model when body movement, HR and patient characteristics were combined (r2 = 0.76
15
for TEE, r2 = 0.83 for PAL). Studies validating HR-body movement models in free-
living conditions showed comparable results 34,35.
Our results revealed that the accuracy of energy expenditure estimation improves
when specific patient characteristics are included in the model. Although HR and
oxygen consumption are linearly related, our results showed that variation in individual
patient characteristics (e.g. fitness, age and gender) cause a high between-subject
variation. Adding these patient characteristics and using individual calibration protocols
(e.g. step-test, cycling protocol) can capture these variations and improve the accuracy
of TEE/PAL estimation. In particular, the HR-Flex method, using HR data and a cycling
protocol for calibration proved to be more accurate in TEE and PAL estimation than the
models combining body movement and HR data only. These results are in line with the
literature, indicating the beneficial effects of individual calibration to capture between-
individual variances 34,36. However, individual calibration also requires resource-
demanding procedures, and is therefore unsuitable for large-scale studies and clinical
practice. In addition, branched equation models, used in several studies to distinguish
between low- and high intensity activities based on the transition between walking-
running 20, are unsuitable for cardiac patients with limited physical fitness levels.
Therefore, we aimed at developing an energy expenditure model that could be
applicable to patient data that are available at the start of a rehabilitation program to
successfully monitor physical activity in these patients in daily life. Currently, general
patient characteristics (i.e. age, body weight, BMI, medication use) are available in the
electronic patient record and a symptom limited exercise test is performed at the start of
each cardiac rehabilitation program to determine physical fitness (peakVO2). With this
16
data available, only RMR data are required before the most accurate multivariate
regression model described in our results (Table 4, model e) can be implemented.
However, RMR can be calculated using the Harris-Benedict equation 37, although RMR
assessment as performed in this study is probably more accurate.
Clinical interpretation and future directions
With the introduction of accurate energy expenditure models based on HR and body
movement for cardiac patients using beta-blockers, physical activity behavior can
become a more eminent topic in cardiac rehabilitation. First, objective feedback on
physical activity monitored during the intervention provides awareness concerning a
patient’s current physical activity behavior. With feedback and motivational coaching
based on the objective physical activity data, patients are able to develop self-
management skills and improve and/or maintain their physical activity behavior 38.
Second, accurate assessment of physical activity is essential for evaluating the
effectiveness of an intervention program and for studying the dose-response relation
between physical activity levels and health outcomes. Therefore, it is essential that
accurate energy expenditure models are developed, using data that are routinely
available at the start of cardiac rehabilitation without the need of additional procedures
and resources. Although a slight improvement in the TEE and PAL estimation error was
observed with individual calibration processes, generalizability is limited. With the
developments in wearable sensors, HR and body movement data collection will become
more accurate, comfortable and feasible in the home environment. Whereas this study
showed that TEE and PAL can be accurately predicted using HR and body movement
17
data in combination with patient characteristics and data derived from a symptom
limited exercise test, the inclusion of activity recognition and improvement of
technology may lead to an even further reduce the energy expenditure estimation error
39. Future studies can use our results as a roadmap for accurate physical activity
monitoring with wearable sensors. If the methodology is confirmed in a larger cardiac
rehabilitation population with mixed conditions, it will be more applicable in the cardiac
rehabilitation setting.
Strengths and limitations
In this study we developed energy expenditure estimation models using HR and body
movement in cardiac patients using beta-blocker medication. Nevertheless, generalizing
the results to the entire cardiac population (e.g. chronic heart failure) is not
straightforward considering the composition of the study population. Furthermore, our
results are based on laboratory investigation using an activity protocol including
treadmill walking and ergometer cycling as modes of activity. It is not known how well
the results apply to free-living conditions. The results of our study are based on a
limited number of cardiac rehabilitation patients. Therefore, the results cannot be
translated directly to a cardiac rehabilitation population with mixed cardiac conditions.
Conclusion
In conclusion, we developed an energy expenditure estimation model for beta-blocker
medicated cardiac rehabilitation patients that showed the highest accuracy when HR,
body movement data and patient characteristics were combined. Personalization of the
18
model using patient characteristics available at the start of the cardiac rehabilitation
program, results in an accurate and feasible model to estimate energy expenditure and
physical activity levels of beta-blocker medicated patients. Additional personalization
using a cycling protocol did not result in a substantial improvement in the estimation of
energy expenditure, indicating that this additional personalization protocol has little
benefit in clinical practice.
Acknowledgements
This study received funding from ZonMW, The Netherlands Organisation for Health
Research and Development (ZonMw project number 837001003) and was supported in
kind by Philips Research.
Conflict of interest
Authors FS, GP, WS and AB are employed by Philips Research.
Author contributions
All authors contributed to the conception and design of the work. JK, FS, GP and AB
contributed to the acquisition, analysis, or interpretation of data for the work. JK and
AB drafted the manuscript. All authors critically revised the manuscript, gave final
approval and agree to be accountable for all aspects of work ensuring integrity and
accuracy.
19
References
1. Lee IM, Shiroma EJ, Lobelo F, et al. Effect of physical inactivity on major non-
communicable diseases worldwide: An analysis of burden of disease and life
expectancy. Lancet [Internet]. 2012;380(9838):219–229.
2. Myers J, Kaykha A, George S, et al. Fitness versus physical activity patterns in
predicting mortality in men. Am J Med. 2004;117(12):912–918.
3. Paffenbarger RS, Brand RJ, Sholtz RI, et al. Energy expenditure, cigarette
smoking, and blood pressure level as related to death from specific diseases. Am
J Epidemiol. 1978;108(1):12–18.
4. Leon AS, Franklin B a, Costa F, et al. Cardiac rehabilitation and secondary
prevention of coronary heart disease: an American Heart Association scientific
statement from the Council on Clinical Cardiology (Subcommittee on Exercise,
Cardiac Rehabilitation, and Prevention) and the Council on Nut. Circulation
[Internet]. 2005 Jan 25;111(3):369–76.
5. Piepoli MF, Corrà U, Benzer W, et al. Secondary prevention through cardiac
rehabilitation: from knowledge to implementation. A position paper from the
Cardiac Rehabilitation Section of the European Association of Cardiovascular
Prevention and Rehabilitation. Eur J Cardiovasc Prev Rehabil. 2010;17:1–17.
6. Müller-Riemenschneider F, Meinhard C, Damm K, et al. Effectiveness of
nonpharmacological secondary prevention of coronary heart disease. Eur J
Cardiovasc Prev Rehabil [Internet]. 2010 Dec;17(6):688–700.
7. Kaminsky LA, Brubaker PH, Guazzi M, et al. Assessing Physical Activity as a
20
Core Component in Cardiac Rehabilitation. J Cardiopulm Rehabil Prev
[Internet]. 2016;36(4):217–229.
8. Caspersen CJ, Powell KE, Christenson GM. Physical activity, exercise, and
physical fitness: definitions and distinctions for health-related research. Public
Health Rep. 1985;100(2):126–131.
9. Speakman JR. The history and theory of the doubly labeled water technique. Am
J Clin Nutr [Internet]. 1998;68(4):932S–938S.
10. Strath SJ, Swartz a M, Bassett DR, et al. Evaluation of heart rate as a method for
assessing moderate intensity physical activity. Med Sci Sports Exerc. 2000;32(9
Suppl):S465–S470.
11. Astrand P-O, Ryhming I. A Nomogram for Calculation of Aerobic Capacity
(Physical Fitness) From Pulse Rate During Submaximal Work. J Appl Physiol.
1954 Sep;7(2):218–221.
12. Weir JB. New methods for calculating metabolic rate with special reference to
protein metabolism. 1948;109(5):1–9.
13. Hay DC, Wakayama A, Sakamura K, et al. Improved estimation of energy
expenditure by artificial neural network modeling. 2008;1222:1213–1222.
14. Livingstone MBE. Heart-rate monitoring: the answer for assessing energy
expenditure and physical activity in population studies? Br J Nutr [Internet].
1997;78(6):869–871.
15. Vanhees L, Lefevre J, Philippaerts R, et al. How to assess physical activity? How
to assess physical fitness? Eur J Cardiovasc Prev Rehabil [Internet]. 2005 Apr
1;12(2):102–114.
21
16. Haskell W, Yee MC, Evans A, et al. Simultaneous measurement of heart rate and
body motion to quantitate physical activity. Med Sci Sport Exerc.
1993;25(1):109–115.
17. Freedson PS, Melanson E, Sirard J. Calibration of the Computer Science and
Applications, Inc. accelerometer. Med Sci Sports Exerc. 1998;30(5):777–781.
18. Plasqui G, Westerterp KR. Physical activity assessment with accelerometers: an
evaluation against doubly labeled water. Obesity (Silver Spring).
2007;15(10):2371–2379.
19. Alharbi M, Bauman A, Neubeck L, et al. Validation of Fitbit-Flex as a measure
of free-living physical activity in a community-based phase III cardiac
rehabilitation population. Eur J Prev Cardiol [Internet]. 2016;1–10.
20. Brage S, Brage N, Franks PW, et al. Branched equation modeling of
simultaneous accelerometry and heart rate monitoring improves estimate of
directly measured physical activity energy expenditure. J Appl Physiol [Internet].
2004 Jan;96(1):343–51.
21. Corder K, Brage S, Wareham NJ, et al. Comparison of PAEE from combined and
separate heart rate and movement models in children. Med Sci Sports Exerc.
2005;37(10):1761–1767.
22. Johansson HP, Rossander-Hulthén L, Slinde F, et al. Accelerometry combined
with heart rate telemetry in the assessment of total energy expenditure. Br J Nutr
[Internet]. 2007 Mar 8;95(03):631.
23. López-Sendón J, Swedberg K, McMurray J, et al. Expert consensus document on
beta-adrenergic receptor blockers. Eur Heart J. 2004;25(15):1341–1362.
22
24. Spurr GB, Reina JC, Prentice a M, et al. Energy expenditure from minute-by-
minute recording : comparison with indirect calorimetry. Am J Clin Nutr
[Internet]. 1988;48:552–559.
25. Whaley MH, Brubaker PH, Otto RM, et al. ACSM’s guidelines for exercise
testing and prescription. 7th editio. 2006.
26. Black AE, Coward WA, Cole TJ PA. Human energy expenditure in affluent
societies: an analysis of 574 doubly-labelled water measurements. Eur J Clin
Nutr. 1996;50:72–92.
27. KR W, WH S, M van E, et al. Use of the doubly labeled water technique in
humans during heavy sustained exercise. J Appl Physiol. 1986;61:2162–2167.
28. Matarese LE. Indirect calorimetry: technical aspects. J Am Diet Assoc.
1997;97(10 suppl 2):154–160.
29. Bonomi AG, Goldenberg S, Papini G, et al. Predicting energy expenditure from
photo-plethysmographic measurements of heart rate under beta blocker therapy:
Data driven personalization strategies based on mixed models. Proc Annu Int
Conf IEEE Eng Med Biol Soc EMBS. 2015;2015-Novem:7642–7646.
30. Bouten C V, Westerterp KR, Verduin M, et al. Assessment of energy expenditure
for physical activity using a triaxial accelerometer. Med Sci Sports Exerc.
1994;26(12):1516–1523.
31. Ceesay SM, Prentice AM, Day KC, et al. The use of heart rate monitoring in the
estimation of energy expenditure: a validation study using indirect whole-body
calorimetry. Br J Nutr. 1989;61:175–186.
32. Westerterp KR, Bouten C V. Physical activity assessment: comparison between
23
movement registration and doubly labeled water method. Z Ernahrungswiss
[Internet]. 1997;36(4):263–7.
33. Rennie K, Rowsell T, Jebb S a, et al. A combined heart rate and movement
sensor: proof of concept and preliminary testing study. Eur J Clin Nutr
[Internet]. 2000 May;54(5):409–14.
34. Brage S, Westgate K, Franks PW, et al. Estimation of Free-Living Energy
Expenditure by Heart Rate and Movement Sensing: A Doubly-Labelled Water
Study. PLoS One [Internet]. 2015;10(9):e0137206.
35. Silva A, Santos D, Matias C, et al. Accuracy of a combined heart rate and motion
sensor for assessing energy expenditure in free-living adults during a double-
blind crossover caffeine trial using doubly labeled water as the reference method.
Eur J Clin Nutr. 2015;69(April 2014):20–27.
36. Brage S, Ekelund U, Brage N, et al. Hierarchy of individual calibration levels for
heart rate and accelerometry to measure physical activity. J Appl Physiol
[Internet]. 2007 Aug;103(2):682–92.
37. Harris J a, Benedict FG. A Biometric Study of Human Basal Metabolism. Proc
Natl Acad Sci [Internet]. 1918;4(12):370–373.
38. Izawa KP, Watanabe S, Omiya K, et al. Effect of the Self-Monitoring Approach
on Exercise Maintenance During Cardiac Rehabilitation. Am J Phys Med Rehabil
[Internet]. 2005 May;84(5):313–321.
39. Bonomi a G, Plasqui G, Goris a HC, et al. Improving assessment of daily energy
expenditure by identifying types of physical activity with a single accelerometer.
J Appl Physiol. 2009;107(3):655–661.
24
25
Table 1. Routine of activity tasks randomly performed by the patients and their duration
Activity task Duration
(minutes)
Sitting at rest 5
Standing at rest 2
Walking 4 km/h 3
Walking 4 km/h, slope 5% 3
Walking 5.5 km/h 3
Cycling 0W 70 rpm 3
Cycling 40W 70 rpm 3
Cycling 70W 70 rpm 3
Cleaning a table 3
Vacuuming 3
Putting dishes in a dish washer 3
W, Watt; rpm, rounds per minute
26
Table 2. Baseline characteristics.
Development Group (n = 11)
Mean ± Sd Min - Max
Age, year 57.1 ± 6.6 45 - 64
Height, cm 182.8 ± 10.6 173 - 191
Weight, kg 95.2 ± 9.5 82 - 108
BMI, kg/m2 28.6 ± 3.5 23.8 - 33.7
RHR, bpm 54.4 ± 6.7 42 - 66
RMR, kcal/min 1.2 ± 0.1 1.0 - 1.4
Resting VO2, ml/min 324.7 ± 43.2 266.7 - 418.8
PeakVO2, ml/min 2326.8 ± 730.8 1429.3 - 3908.6
Max HR, bpm 138.5 ± 23.5 96.3 - 177.2
Beta-blocker dose, mg/day 48.3 ± 32.1 0 - 100
Validation Group (n = 5)
Mean ± Sd Min - Max
Age, y 53.124 ± 8.782 42 - 61
Height, m 175.5 ± 6.6 164 - 182
Weight, kg 94.8 ± 12.2 80 - 110
BMI, kg/m2 30.7 ± 3.1 25 - 33.2
RHR, bpm 59.1 ± 5.3 52 - 68
RMR, kcal/min 1.2 ± 0.2 1.0 - 1.4
Resting VO2, ml/min 351.8 ± 48.9 315.8 - 443.8
PeakVO2, ml/min 2369.3 ± 450.0 1839.4 - 3175.1
Max HR, bpm 151.7 ± 3.9 148.1 - 156.7
Beta-blocker dose, mg/day 80.3 ± 39.9 0 - 100
27
BMI; body mass index; RHR, resting heart rate; bpm, beats per minute; RMR, resting metabolic rate;
VO2, oxygen consumption; PeakVO2, maximum oxygen consumption; max HR, maximum heart rate;
Table 3. Correlation between energy expenditure measures and body movement or HR
features for data from the entire study population (Group) or from each individual
patient (Individual).
Correlation TEE vs ACmin TEE vs HR TEE vs HRnet
p < 0.05 Mean Sd Mean Sd Mean Sd
Group 0.47 0.55 0.68
Individual 0.53 0.15 0.84 0.12 0.84 0.12
PAL vs ACmin PAL vs HR PAL vs HRnet
Mean Sd Mean Sd Mean Sd
Group 0.49 0.49 0.56
Individual 0.53 0.15 0.84 0.12 0.84 0.12
TEE, total energy expenditure; ACmin, activity counts per minute; HR, heart rate; HRnet, (heart rate –
resting heart rate); PAL, physical activity level
28
Table 4. Accuracy of the energy expenditure prediction models.
TEE Development Validation
R2
RMSE
(kcal/
min)
Bias
(kcal/
min) R2
RMSE
(kcal/
min)
Bias
(kcal/
min)
Model a) Movement features 0.64 0.984 0 0.65 1.031 -0.135
Model b) Movement + Patient
features 0.63 0.99 0 0.57 1.156 -0.249
Model c) HR features 0.56 1.084 0 0.76 0.922 0.107
Model d) HR + Patient features 0.60 1.032 0 0.79 0.856 -0.079
Model e) Movement + HR +
Patient features 0.76 0.799 0 0.79 0.799 -0.126
HR-flex model 0.79 1.11 -0.22 - - -
PAL Development Validation
R2
RMSE
(kcal/
min)
Bias
(kcal/
min) R2
RMSE
(kcal/
min)
Bias
(kcal/
min)
Model a) Movement features 0.61 1.262 0 0.62 1.278 -0.076
Model b) Movement + Patient
features 0.67 1.161 0 0.68 1.178 0.008
Model c) HR features 0.74 1.026 0 0.72 1.123 0.22
Model d) HR + Patient features 0.75 1.006 0 0.79 0.986 0.145
Model e) Movement + HR +
Patient features 0.83 0.835 0 0.84 0.834 0.118
HR-flex model 0.80 0.846 -0.13 - - -
TEE, total energy expenditure; RMSE: root mean squared error; Bias: mean error; HR, heart rate; PAL,
physical activity level. The HR-flex model accuracy was not possible to test using a hold-out group of
29
patients since the method required individual calibration to set the value of constituting parameters.
Results from the development group were showed to demonstrate the large generalization properties of
the models not affected by overfitting.
30
Figure 1. Correlation between measured and predicted TEE for all the developed
models based on a combination of HR, ACmin and patient characteristics.
31
Figure 2. Measured and predicted PAL from a representative patient according to the
HR-flex method and the multivariate model based on HR, body movement features and
patient characteristics.
Black line represents the measured PAL, while red and blue lines the predicted PAL
according to the prediction models.
32