ArticlePDF Available

Cardiorespiratory fitness estimation using wearable sensors: Laboratory and free-living analysis of context-specific submaximal heart rates

Authors:
  • HRV4Training

Abstract and Figures

In this work, we propose to use pattern recognition methods to determine submaximal heart rate (HR) during specific contexts, such as walking at a certain speed, using wearable sensors in free-living, and use context-specific HR to estimate cardiorespiratory fitness (CRF). CRF of 51 participants was assessed by a maximal exertion test (VO2max). Participants wore a combined accelerometer and HR monitor during a laboratory based simulation of activities of daily living and for two weeks in free-living. Anthropometrics, HR while lying down and walking at predefined speeds in laboratory settings were used to estimate CRF. Explained variance (R(2)) was 0.64 for anthropometrics, and increased up to 0.74 for context-specific HR (0.73 to 0.78 when including fat-free mass). Then, we developed activity recognition and walking speed estimation algorithms to determine the same contexts (i.e. lying down and walking) in free-living. Context-specific HR in free-living was highly correlated with laboratory measurements (Pearson's r = 0.71-0.75). R(2) for CRF estimation was 0.65 when anthropometrics were used as predictors, and increased up to 0.77 when including free-living context-specific HR (i.e. HR while walking at 5.5 km/h). R(2) varied between 0.73 and 0.80 when including fat-free mass among the predictors. RMSE was reduced from 354.7 ml/min to 281.0 ml/min by the inclusion of context-specific HR parameters (21% error reduction). We conclude that pattern recognition techniques can be used to contextualize HR in free-living and estimated CRF with accuracy comparable to what can be obtained with laboratory measurements of HR response to walking.
Content may be subject to copyright.
Title:
Cardiorespiratory fitness estimation using wearable sensors: laboratory and free-living analysis of context-
specific submaximal heart rates.!
Authors:
Marco Altini1, Pierluigi Casale2, Julien Penders2, Gabrielle ten Velde3, Guy Plasqui3 and Oliver Amft4
Authors’ contribution:
Marco Altini: study design, data analysis and paper writing
Pierluigi Casale: support during data analysis
Julien Penders: support during study design and paper writing
Gabrielle ten Velde: support during data collection and paper writing
Guy Plasqui: support during data collection and paper writing
Oliver Amft: support during study design and paper writing
Affiliations:
1 Eindhoven University of Technology, The Netherlands and Bloom Technologies, Diepenbeek, Belgium
2 Holst Centre/imec, Eindhoven, The Netherlands
3 Human Biology department, Maastricht University, Maastricht, The Netherlands
4 University of Passau, Germany, and Eindhoven University of Technology, The Netherlands
Running head:
Estimating cardiorespiratory fitness in free-living
Address for correspondence:
Marco Altini, Eindhoven University of Technology, Den Dolech 2, 5612AZ, email:
altini.marco@gmail.com, phone: 0031 6 46375742, fax: 0031 40 246 3120
Abstract:
In this work, we propose to use pattern recognition methods to determine submaximal heart rate (HR)
during specific contexts, such as walking at a certain speed, using wearable sensors in free-living, and use
context-specific HR to estimate cardiorespiratory fitness (CRF). CRF of 51 participants was assessed by a
maximal exertion test (VO2max). Participants wore a combined accelerometer and HR monitor during a
laboratory based simulation of activities of daily living and for two weeks in free-living. Anthropometrics,
HR while lying down and walking at predefined speeds in laboratory settings were used to estimate CRF.
Explained variance (R2) was 0.64 for anthropometrics, and increased up to 0.74 for context-specific HR
(0.73 to 0.78 when including fat-free mass). Then, we developed activity recognition and walking speed
estimation algorithms to determine the same contexts (i.e. lying down and walking) in free-living. Context-
specific HR in free-living was highly correlated with laboratory measurements (Pearson’s r = 0.71-0.75).
R2 for CRF estimation was 0.65 when anthropometrics were used as predictors, and increased up to 0.77
when including free-living context-specific HR (i.e. HR while walking at 5.5 km/h). R2 varied between
0.73 and 0.80 when including fat-free mass among the predictors. RMSE was reduced from 354.7 ml/min
to 281.0 ml/min by the inclusion of context-specific HR parameters (21% error reduction). We conclude
that pattern recognition techniques can be used to contextualize HR in free-living and estimated CRF with
accuracy comparable to what can be obtained with laboratory measurements of HR response to walking.
New and noteworthy:
Many methods have been developed to estimate VO2max using data collected under supervised laboratory
conditions or following strict protocols. However, to the best of our knowledge, this is the first work, which
proposes pattern recognition methods to contextualize heart rate (HR) in free-living and use context-
specific HR to predict VO2max. The proposed method does not require laboratory tests or specific
protocols, showing error reductions up to 21% compared to VO2max estimates derived using
anthropometrics only.
Keywords:
Cardiorespiratory fitness, wearable sensors, heart rate, physical activity, context recognition
Introduction:
Cardiorespiratory fitness (CRF) is a diagnostic and prognostic health indicator for patients in clinical
settings, as well as healthy individuals and can be adopted as a proxy of cardiovascular and
cardiorespiratory health (10, 26). Thus, CRF is a marker of training status that can be considered one of the
most important determinants of health and wellbeing. While recent developments in wearable sensor
technologies improved the accuracy of physical activity monitoring devices in daily life, almost all
solutions focus on behavioral aspects such as steps, activity type and energy expenditure (EE) (4, 6). Steps
or EE are relevant markers of an individual’s health, however they mainly reflect the individual’s behavior,
instead of the individual’s health status. CRF estimation using wearable sensors could provide more
insights on an individual’s health status, non-invasively, and therefore help clinicians and individuals
coaching or leading a healthier lifestyle.
Currently, the gold standard for CRF measurement is performed by direct measurement of oxygen
consumption during maximal exercise (i.e.VO2 max) (30, 31). However, VO2max measurements require
medical supervision and can be risky for individuals where exercise until maximal exertion is contra-
indicated. Despite the indubitable importance of CRF in health, measurements of VO2max are therefore
rare (21) and less risky submaximal tests have been developed. Non-exercise CRF estimation models use
easily accessible measures such as age, gender and a self-reported physical activity level (14, 20).
However, for individuals with similar anthropometric characteristics, CRF levels cannot be discriminated
accurately. Alternatively, submaximal tests have been introduced to estimate VO2max during specific
protocols while monitoring HR at predefined workloads (5, 11). The strict workload imposed by the
protocol is used to exploit the inverse relation between HR in a specific context (e.g. while running or
biking at a specific intensity) and VO2max. However the need for laboratory equipment and the necessity to
re-perform the test to detect changes in CRF limit the practical applicability of such techniques. Ideally, we
would like to estimate CRF in free-living during activities of daily living, thus without the need for specific
laboratory tests or exercise protocols. Estimating CRF using wearable sensors data acquired during regular
activities of daily living could provide continuous assessment without the need for specific tests or
protocols.
Miniaturized wearable sensors combining accelerometer and HR data provide a way to investigate the
relation between physical activity, HR and VO2max in free-living. Additionally, advances in signal
processing and machine learning techniques, recently provided new methods to accurately recognize
contexts in which HR can be analyzed, such as activity type, walking speed and EE (1, 6, 28), in free-
living.
The relation between submaximal HR during activities of daily living simulated in laboratory settings and
VO2max has been evaluated by different research groups (2, 23, 28, 30). Tonis et al. (29) explored different
parameters to estimate CRF from HR and accelerometer data during activities of daily living simulated in
laboratory settings. However VO2max reference and free-living data were not collected. Others (2, 7, 23)
measured HR parameters representative of CRF in the context of improving estimates of energy
expenditure, showing how inter-individual differences in HR could be accounted for by surrogates of
fitness such as measured or estimated sub-maximal HR. In free-living conditions, the relation between
physical activity as expressed by a step counter, and CRF was investigated by Cao et al. (9). While steps
could provide useful insights, the relation between HR and VO2 at a certain exercise intensity cannot be
exploited using only motion based sensors. Plasqui et al. (22) showed that a combination of average HR
and physical activity over a period of 7 days correlates significantly with VO2max. However, the relation
between average HR and activity counts depends on the amount of activity performed, and therefore could
also be affected by behavioral correlates of CRF. Many studies showed strong links between sub-maximal
HR during simulated activities of daily living and CRF, thus motivating our research.
In this study, we aimed at investigating the relation between submaximal HR in specific contexts as
recorded by wearable sensors in free-living, and CRF, and to predict VO2max using free-living data. To
this aim, we hypothesized that isolating the same contexts in laboratory settings and in free living using
pattern recognition methods, could yield to similar relations between context-specific HR and VO2max.
Methods:
Participants:
Participants were 51 (24 male, 27 female) healthy adults. Anthropometric characteristics and CRF level are
reported in Table 1. Written informed consent was obtained by each participant. The study was approved
by the medical ethics committee of Maastricht University.
Table 1: Participants’ characteristics
Parameter
Mean ± SD
n
51 (24 male, 27 female)
Age (y)
25.1 ± 6.0
Body weight (kg)
68.4 ± 10.8
BMI (kg/m2)
22.7 ± 2.5
Fat free mass (kg)
52.6 ± 9.2
VO2max (ml/min)
3037.5 ± 671.6
ECG and accelerometer device:
The sensor platform used was an ECG Necklace. The ECG Necklace (1) is a low power wireless ECG
platform. The system relies on an ultra-low-power ASIC for ECG read-out, and it is integrated in a
necklace, providing ease-of-use and comfort while allowing flexibility in lead positioning and system
functionality. It achieves up to 6 days autonomy on a 175 mAh Li-ion battery. For the current study, the
ECG Necklace was configured to acquire one lead ECG data at 256 Hz, and accelerometer data from a tri-
axial accelerometer (ADXL330) at 64 Hz. The ADXL330 accelerometer provides a ±3g range and high
sensitivity (300 mV/g), and was digitalized to 12 bits input by the ECG Necklace. The x, y, and z axes of
the accelerometer were oriented along the vertical, mediolateral, and antero-posterior directions of the
body, respectively. The ECG Necklace was not attached to the body, to improve user comfort during free-
living. Two gel electrodes were placed on the participant’s chest, in the lead II configuration. Data were
recorded on the on-board SD card to ensure integrity.
The ECG Necklace was previously validated as a reliable physical activity monitor able to quantify
different physical activity parameters with high accuracy, such as activity type, walking speed and EE (1,
3). A continuous wavelet transform based beat detection algorithm was used to extract RR intervals from
ECG data (24). Segments of data identified as lying or sedentary (no or limited movement) as well as flat
ECG signal or inaccurate HR were treated as “monitor not worn”. Inaccurate HR was identified as periods
where consecutive RR intervals varied more than 20%, as typically performed in clinical practice for heart
rate variability analysis.
Indirect Calorimetry:
The gas-analysis was performed with an open-circuit indirect calorimeter in diluted flow mode, meaning
that the subject could freely breathe in an airstream. The flow past the subject mouth was set at 400 l/min.
This means subject breathing ventilation of up to 200 l/min can be measured without re-breathing except
for the volume of the applied face-mask. Total flow was measured and converted to STPD values with a
large dry-bellows flowmeter calibrated to 0.2% of used range by national standards bureau (5 point
calibration), and using calibrated temperature, humidity and pressure sensors. Gas-samples taken from the
flow are filtered, dried, pressurized and fed into high resolution O2 and CO2 analyzers made by ABB-
Hartmann&Braun (OA2020 and Easyline 19” rack units) and Servomex (Servomex 4100 and Servopro
5400 19” rack units) with a resolution 0.001% absolute. The analyzers are mounted separately to exclude
both vibration and climate-variation as confounding factors. Ranges for the analyzers are set to 0-21% O2
and 0-1% CO2 yet only limited to 25% and 2.5% respectively. No specific smoothing is applied as the
result is updated each 5 seconds while breathing is mechanically averaged in the ± 30 liter internal volume
and the dilution gasflow, resulting in a time constant of 4.5 second at the 400 l/min setting. The calorimeter
is validated by gas-infusion or burning fuel (methanol) over its full range (200 to 7000 ml.min-1) with a
1±2 % avg ± SD result. VO2 max was reached when a plateau in VO2 was observed and/or an RQ of 1.1 or
higher. VO2max was calculated as the highest average VO2 over 30 seconds (6 consecutive values).
Study design:
The ECG Necklace was worn during laboratory protocols and free-living.
Laboratory protocols: participants reported at the lab on three separate days and after refraining
from drinking, eating and smoking in the two hours before the experiment. Two laboratory
protocols were performed, while the third day was used for anthropometric measurements
including the participant’s body weight, height and body fat.
o The first protocol included simulated activities performed while connected to an indirect
calorimeter (Omnical, Maastricht University, The Netherlands), to determine context-
specific HR during activities of daily living simulated in laboratory settings. Activities
included: lying down, sitting, sit and write, standing, cleaning a table, sweeping the floor,
walking (treadmill flat at 2.5, 3, 3.5, 4, 4.5, 5, 5.5, 6 km/h) and running (treadmill flat at
7, 8, 9, 10 km/h). Activities were carried out for a period of at least 4 minutes.
o The second protocol was a VO2max test providing reference data for biking and CRF.
VO2max was determined during an incremental test on a cycle ergometer according to the
protocol of Kuipers et al. (17). After a 5-min warm-up at 100 W for men and 75 W for
women, workload was increased by 50 W every 2.5 min. When the HR reached 35 bpm
below the age-predicted maximal HR (208 0.7 x age) or the respiratory quotient
exceeded 1, workload was increased by 25 W every 2.5 min until exhaustion. Expired air
was continuously analyzed for O2 consumption and CO2 production using indirect
calorimetry.
Free-living protocol: participants wore the ECG necklace for 14 consecutive days in free-living
while carrying out their normal activities of daily living. Participants were instructed to wear the
ECG necklace during day and night, except during showering, water activities or charging of the
ECG necklace, since the ECG Necklace is not waterproof. Charging was performed daily for 1
hour. Participants were also instructed to change electrodes daily or after physical exercise.
Data processing:
Context-specific HR in laboratory settings was determined as the mean HR during scripted activities
performed by the participant and combined with anthropometrics in a regression model to predict VO2max.
The regression model was analyzed to validate the assumption that submaximal context-specific HR can be
used to estimate CRF level. Activity type recognition and walking speed models were built using data from
laboratory settings, and used in free-living. For each participant, models were built using only data from
other participants. Therefore, all models were non-individualized and no laboratory data from the
participant to be validated was used for model building. The procedure used for model building and
evaluation is shown in Figure 1. For the beat detection we relied on methods developed by the research
community in the past as these models are standard components that are already available in many sensor
devices today. More details on the validation procedures are reported in the Statistics and performance
measures Section. Context-specific HR in free-living was used in a multiple regression model to estimate
VO2max without the need for laboratory protocols and analyzed with respect to results obtained using
submaximal context-specific HR acquired during activities of daily living simulated in laboratory settings.
Figure 1: Block diagram of the proposed approach and validation procedure. Activity recognition and walking speed
estimation models are built and validated using supervised laboratory recordings. Then, models are deployed in free-
living. Activity recognition and walking speed estimation are used to determine HR in specific contexts in free-living.
Finally, HR in specific contexts (e.g. HR while lying down or walking at a certain speed) are used as predictors for
Building(blocks(later(
deployed(in(free(living(
to(determine(context8
specific(HR(
Free(living8derived(
parameters(used(as(
predictors(for(CRF(
es>ma>on(
Laboratory(validated.building.blocks.
model& reference&
Ac,vity&recogni,on&
Walking&speed&
Ac,vity&type&
manually&annotated&
Treadmill&walking&
speed&
deployed&in&free&living&
HR&in&specific&contexts&
Free.living.parameter.es8ma8on.
Laboratory@validated&CRF&es,ma,on&
CRF(es>ma>on(from(
parameters(
determined(from(free(
living(data(
CRF&es,ma,on& VO2max&
model& reference&
Laboratory(validated..CRF.es8ma8on.
VO2max estimation, effectively estimating CRF level from free-living data. All models are validated using leave one
subject out cross validation, i.e. no data used for model validation was used for model building, as described in the
Statistics section. An example of activity recognition and walking speed estimation models output is shown in Figure 5.
Activity type and walking speed: The raw acceleration signal was downloaded and processed for two
purposes. The first purpose was to develop an activity recognition algorithm using data acquired during
simulated activities of daily living in the laboratory protocols. The activity recognition algorithm was then
used to detect the activity types performed during the free-living protocol. Secondly, the raw acceleration
signal was processed to determine walking speed for activities recognized as walking. The acceleration
signal was segmented in non-overlapping intervals of 5 seconds. This segment length was selected based
on previous studies (28). Segmented data were separately filtered by two filters to create different feature
sets. One feature set included accelerometer data band-pass filtered between 0.1 and 10 Hz, to isolate the
dynamic component due to body motion, while the second feature set included accelerometer data low-pass
filtered at 1 Hz, to isolate the static component, due to gravity. The selected cut-off frequencies were based
on previous research (28) and are not complementary (i.e. they are not the same cut-off for both filters) due
to the fact that there is no clear cut-off frequency to choose, and the two frequencies chosen shown to be
ideal in discriminating static gravitational acceleration and body motion, as shown in Fig. 4. Fig. 4 shows
an example of raw data, low-passed data and band-passed data for one participant during one of the
laboratory protocols. Features used for activity recognition were: mean of the absolute signal, inter-quartile
range, median, variance, main frequency peak and low frequency band signal power. All accelerometer
features but the median, were derived from band-pass filtered data. These features were derived and
selected based on our previous work (1), using a different dataset. We report details on the mathematical
formulas defined to extract accelerometer features in Table 2.
Table 2: Accelerometer features used for activity classification and walking speed estimation. N indicates
the number of samples in a 5 seconds window, i.e. 160 (32 samples per second). LP and BP stand for low
pass and band pass filtered data. Qn represents the nth quartile.
Computation
Description
1
𝑁
!|𝑎!"#|
!
!!!
Represents motion intensity independently of
the axis or orientation, similarly to activity
counts
Q3-Q1 of 𝑎!"
Represents motion intensity, can be less
prone to outliers with respect to, e.g. range
Middle value of the ordered 𝑎!" array
Represents posture (gravitational vector)
1
𝑁
!(𝑎!!𝜇)!
!
!!!
where
𝜇=!
1
𝑁
!𝑎!
!
!!!
Represents variability in detected motion,
which might be discriminative of activity
type (28)
1. Apply Hamming window to
reduce spectral leakage.
2. Compute FFT
3. Determine main frequency
peak of the power spectrum
Provides information about the repetitiveness
of motion, e.g. during walking (28)
1. Apply Hamming window to
reduce spectral leakage.
2. Compute FFT
3. Sum signal power between 0
and 0.7 Hz (25)
Shown to be discriminative of sedentary and
walking activities in previous research (25)
1. Apply Hamming window to
reduce spectral leakage.
2. Compute FFT
3. Sum signal power between 0.7
and 10 Hz (25)
Shown to be discriminative of sedentary and
walking activities in previous research (25)
HR was extracted from RR intervals, and averaged over 15 seconds windows. Features for the multiple
linear regression model used to estimate walking speed were: mean of the absolute signal, inter-quartile
range, variance, main frequency peak, high frequency band signal power and height of the participant, and
were also based on our previous work (2). All accelerometer features used for the walking speed models
were derived from band-pass filtered data. Coefficients for the linear regression model used to estimate
walking speed are shown in Table 3.
Table 3: Coefficients of the linear regression model used to estimate walking speed. During validation all
models were evaluated using leave-one-participant-out cross-validation, the coefficients shown here
include all data.
Parameter
Estimate coefficient
P value
Intercept
-1.000e+00
<2e-16
Mean of the absolute signal
9.936e00
<2e-16
Variance
-!2.963e00
<2e-16
Quartile (X axis)
3.256e00
<2e-16
Quartile (Y axis)
-!2.475e00
<2e-16
High frequency band signal power (X axis)
-!5.920e-04
<2e-16
High frequency band signal power (Z axis)
-!6.320e-04
<2e-16
Main frequency peak (X axis)
-!1.323e-01
<2e-16
Participant height
1.439e-02
<2e-16
Laboratory activities were grouped into six clusters to be used for activity classification. The six clusters
were lying (lying down), sedentary (sitting, sit and write, standing), dynamic (cleaning the table,
sweeping the floor), walking, biking and running. Activities were derived using pattern recognition
methods, in particular a Support Vector Machine (SVM). SVMs are classifiers that showed good results in
classifying activities in our previous research (1, 2, 3, 4). The principle behind using pattern recognition
methods and accelerometer data for activity classification is that different activities clusters (e.g. lying
down, walking) result in different accelerometer patterns as collected by on-body sensors. By capturing
such accelerometer patterns using the features listed in Table 2, a classifier can be trained to distinguish
activities clusters with high accuracy (1-4, 28, 29). As an example, two features used for the classification
of the six activity clusters are analyzed in Fig. 2. We limited the features space to two dimensions to
provide a visualization that is easily human readable. Fig. 2.a shows the mean of the absolute acceleration
signal, a measure representative of motion intensity. The mean of the absolute acceleration is particularly
helpful in discriminating high intensity activities (e.g. running), average intensity activities (e.g. walking or
biking) and low intensity activities (e.g. lying, sedentary), as show in Fig. 2.a. Fig. 2.b shows the median of
the low-pass filtered accelerometer X-axis signal, a feature representative of body posture given our sensor
on-body positioning. During training, the SVM classifier takes as input multiple features (see Table 2) and
determines the optimal discrimination boundary between the activity clusters, i.e. the widest separation
between samples of different activity clusters (i.e. accelerometer features belonging to different activities).
The distinct colored regions in Fig. 2c illustrate that the two shown features provide relevant information to
discriminate the activities clusters. Hence already two features are sufficient to separate most - but not all -
activity clusters in this study.
Figure 2. Example of extracted features and multi-dimensional features space used for activity classification of the six
activity clusters included in this study. a-b) Histograms of two accelerometer features (mean of the absolute signal and
median of the X axis). c) Two dimensional features space showing clear separations between most activity clusters.
The SVM trained in this paper determines decision boundaries (or separating hyperplanes) that can be used
later on to classify new accelerometer feature samples into activity clusters. The decision boundaries are
0
3000
6000
9000
0.0 0.5 1.0 1.5 2.0
Mean of the absolute signal
Count
Activities
biking
household
lying
running
sedentary
walking
a) Mean of the absolute signal
0
2500
5000
7500
10000
0.5 0.0 0.5 1.0
Median of the X axis
Count
b) Median of the X axis
0.0
0.5
1.0
0.0 0.5 1.0 1.5 2.0
Mean of the absolute signal
Median of the X axis
c) Two dimensional features space (median of the X axis and mean of the absolute signal)
optimal in the sense that the algorithm determines the maximal margin between training samples and the
decision boundary. Without maximizing the margin, various decision boundaries could be found. An
example of a linear separation of two classes using a SVM is shown in Fig. 3. Fig 3.a. shows multiple
example decision boundaries that separate the example data points, while Fig 3.b. shows the separating
hyperplane that maximizes the margin to the example data points, as determined by the SVM.
Figure 3. Example of linear decision boundaries to classify two classes. Example data points of the classes are
illustrated in different gray tones. On the left; different example decision boundaries. On the right; optimal hyperplane
obtained by maximizing the margin between the decision boundary and the closest data points of each class. Samples
on the maximal margin lines are called support vectors.
CRF estimation: CRF was estimated using multiple linear regression models. First, we investigated the
relation between HR in specific contexts as acquired during activities of daily living simulated in laboratory
settings, and VO2max. We predicted VO2max by combining anthropometric characteristics and HR while
lying down and while walking at 3.5 and 5.5 km/h. We chose lying down and walking at 3.5 and 5.5 km/h
as specific contexts since lying down and walking are activities of daily living commonly performed by
healthy individuals in most environments. Additionally, the average walking speeds in healthy individuals
was reported in previous studies between 5 and 6 km/h (5.3 km/h in (8) and 5 ± 0.8 km/h in (19)). Given
the estimation error of our walking speed estimation model and the variability of free-living walking, we
selected data segments with detected speed higher than 3 km/h and lower than 4 km/h as segments to be
considered of an average walking speed of 3.5 km/h. Similarly, we selected data segments with detected
speed higher than 5 km/h and lower than 6 km/h as segments to be considered of an average walking speed
of 5.5 km/h.
Then, we analyzed the relation between context-specific HR during activities of daily living simulated in
laboratory settings, and context-specific HR during the same activities as detected by our activity
recognition and walking speed models, in free-living. The analysis of the relation between context-specific
HR in laboratory settings and free-living consisted of computing the correlation coefficient and relative
differences between HR in laboratory settings and free living. This analysis is merely to provide some
perspective on context-specific HR with respect to laboratory measurements. However, free-living
regression models are built and evaluated using free-living data only.
Finally, we predicted VO2max by combining anthropometric characteristics and HR while lying down and
while walking at 3.5 and 5.5 km/h as determined from free-living data, to evaluate the ability of the
context-specific HR detected using pattern recognition methods to estimate CRF.
Figure 4: Raw accelerometer data (top), low-pass filtered data (center) and band-pass filtered data (bottom). The
gravity component is isolated when using low-pass filtered data, as shown in the center plot. This information is
particularly useful to distinguish postures. Band-pass filtered data isolates the accelerometer component due to body
motion, showing increased values for higher intensity motions. Band-pass filtered data is particularly useful to
distinguish ambulatory activities and walking speeds. Data were downsampled for visualization purposes.
Statistics:
Activity recognition and walking speed estimation models were derived using laboratory data and
evaluated using leave-one-participant-out cross validation. The same training set, consisting of data from
1
0
1
2
3
0 1000 2000
Minutes
g
a) Raw accelerometer data
0.0
0.5
1.0
0 1000 2000
Minutes
g
b) Lowpass filtered accelerometer data
1
0
1
2
0 1000 2000
Minutes
g
c) Bandpass filtered accelerometer data
all participants but one, was used to build feature selection, activity recognition and walking speed
estimation and CRF estimation models. The remaining data was used for validation. The procedure was
repeated for each participant and results were averaged. Performance of the activity recognition models was
evaluated using the class-normalized accuracy, using laboratory recordings. Results for walking speed
estimation were reported in terms of Root-mean-square error (RMSE), where the outcome variable was
speed in km/h. The relation between HR and CRF were reported using Pearson’s correlation coefficient (r)
for both activities simulated in laboratory settings and free-living data. The relation between context-
specific HR during activities of daily living simulated in laboratory settings and in free-living as detected
by pattern recognition methods was reported using Pearson’s correlation coefficient (r) and the mean and
standard deviation of the difference between context-specific HR in laboratory settings and in free-living.
Results for CRF estimation models were reported in terms of explained variance (R2). The Bland-Altman
plot was used to determine the agreement between measured and predicted CRF. Finally, subject-
independent evaluation for CRF estimation models was also performed, using leave one participant out
cross-validation. Regression models including different HR parameters (e.g. HR while lying down or HR
while walking at different speeds) were compared using the likelihood ratio. More specifically, we
compared two models, the first one including anthropometrics and HR while lying down, and a second one
including anthropometrics, HR while lying down as well as HR while walking. We compared likelihood
ratios for both laboratory recordings and free-living data. We reported results for subject independent CRF
estimation in terms of RMSE, where the outcome variable was VO2max in ml/min as measured in
laboratory conditions. Paired t-tests were used to compare results. Significance was set at α < 0.05.
Results:
Descriptive statistics:
The dataset considered for this work contained 491 days of data collected from 51 participants in free-
living, thus about 10 days per participant, including accelerometer and ECG data. Eighty-three hours of
laboratory recordings including reference VO2, VCO2, acceleration, ECG and VO2max were collected for
model building and evaluation. Laboratory measurements were discarded for two participants where we
observed measurement errors such as unusable ECG data due to excessive noise or bad lead attachment.
Anthropometric characteristics and CRF level for the participants are reported in Table 1. Fig. 5 shows an
exemplary output of the walking speed and activity recognition models for one participant during 24 hours
of free-living recordings. Context-specific HR as identified using activity recognition and walking speed
models in free-living is also shown in Fig. 5.
Figure 5: Exemplary output of the models used to contextualize HR in free-living in this work, for one participant. a)
Recognized activity types. Commuting by bike, training (running), sleep and a mostly sedentary job during waking
hours can be easily identified from this plot. b) Estimated walking speeds when the activity type algorithm identifies the
walking activity. c) HR and contextualized HR. Contextualized HR, i.e. in this example the HR while walking at 5.5
km/h, is highlighted in black.
CRF estimation from context-specific submaximal HR during simulated activities of daily living:
HR during activities of daily living simulated in laboratory settings was 66.2 ± 12.3 bpm for lying, 91.0 ±
15.3 bpm for walking at 3.5 km/h and 107.8 ± 17.7 bpm for walking at 5.5 km/h. Pearson’s correlation
0
2
4
6
0 5 10 15 20 25
Time (hours)
Activity ID
Activity
Biking (Activity ID = 5)
Dynamic (Activity ID = 3)
Lying (Activity ID = 1)
Running (Activity ID = 6)
Sedentary (Activity ID = 2)
Walking (Activity ID = 4)
a) Activity Type
0.0
2.5
5.0
0 5 10 15 20 25
Time (hours)
km/h
b) Walking Speed
0
50
100
150
0 5 10 15 20 25
Time (hours)
bpm
Context
Other
Walking at 5.5 km/h
c) Heart Rate
between context-specific submaximal HR as measured during activities of daily living simulated in
laboratory settings and CRF was -0.43 for lying down, -0.47 for walking at 3.5 km/h and -0.51 for walking
at 5.5 km/h. Thus, confirming the hypothesis that submaximal HR is inversely related to CRF. Explained
variance (adjusted R2) for multiple regression models including sex, body weight and age as predictors of
CRF, was 0.64. Adjusted R2 increased when including context-specific HR, and was 0.69 for lying, 0.72 for
walking at 3.5 km/h and 0.74 for walking at 5.5 km/h. Thus, confirming that activities of higher
submaximal intensities explain more of the variance in the model. Results are reported in Table 4 while
Fig. 6 shows scatterplots of reference against fitted values as well as Bland-Altman plots. When including
more advanced anthropometrics, such as fat free mass instead of body weight, R2 was 0.73 when no HR
was used among the predictors, 0.74 for lying, 0.76 for walking at 3.5 km/h and 0.78 for walking at 5.5
km/h. We computed the likelihood ratio between regression models including anthropometrics data and HR
while lying down with respect to regression models including anthropometrics data and HR while walking
at 3.5 and 5.5 km/h. The likelihood ratio showed that for both walking speeds, including HR while walking
significantly improved the model fit (p=0.044 when including the HR while walking at 3.5 km/h and
p=0.0048 when including the HR while walking at 5.5 km/h).
Table 4: Multiple linear regression models for VO2max estimation from activities of daily living simulated
in laboratory settings. N = 49. For each predictor, detailed information (model coefficient, p-value) are
indicated.
Model description
Predictors
R2
Anthropometric characteristics
only
Intercept (1431.686, p = 0.00322) Body weight
(18.510, p = 0.00645), age (-1.888, p = 0.84595),
sex (798.561, p = 7.48e-07)
0.64
Context-specific HR
Intercept (2849.294, p = 3.86e-05) HR while lying
down in laboratory settings (-14.268, p = 0.00344),
body weight (13.862, p = 0.02862), age (-8.170, p
= 0.37335), sex (803.669, p = 1.20e-07)
0.69
Intercept (3183.644, p = 5.67e-06), HR while
walking at 3.5 km/h in laboratory settings (-13.636,
p = 0.000496), body weight (14.921, p =
0.013229), age (-12.046, p = 0.183661), sex
(777.430, p = 1.03e-07)
0.72
Intercept (3367.865, p = 9.65e-07), HR while
walking at 5.5 km/h in laboratory settings (-13.044,
p=8.21e-05), body weight (15.234, p = 0.00856),
age (-13.222, p = 0.13055), sex (754.772, p =
8.95e-08)
0.74
Context recognition; activity type and walking speed:
Laboratory recordings with reference activity type were used to determine accuracy of the models used in
free-living. Accuracy of the SVM activity recognition classifier was 94.1%. More specifically, the accuracy
was 96.4% for lying, 95.6% for sedentary activities, 83.3% for dynamic, 98.2% for walking, 91.4% for
biking and 99.7% for running. The confusion matrix for the subject independent results of the activity
recognition model is shown in Table 5. The explained variance for the walking speed model was 0.85 (R2).
Walking speed estimation RMSE for subject independent analysis was 0.37 km/h across all speeds.
Table 5. Confusion matrix showing the normalized performance of the activity recognition model, in
percentage.
Classification results
True
Activities
Lying
Sedentary
Dynamic
Walking
Biking
Running
Lying
96%
4%
0%
0%
0%
0%
Sedentary
0%
96%
4%
0%
0%
0%
Dynamic
0%
10%
83%
0%
7%
0%
Walking
0%
0%
0%
98%
2%
0%
Biking
0%
1%
4%
5%
90%
0%
Running
0%
0%
0%
0%
0%
100%
Activities in free-living over the complete dataset were recognized as follows: 44.4% lying, 36.4%
sedentary, 9.5% dynamic, 5.4% walking, 3.8% biking and 0.4% running. Average walking speed was 3.6 ±
1.5 km/h. Participants spent on average 77,7 minutes per day walking, 11.9 minutes of which were at 3.5
km/h and 11.6 minutes of which were at 5.5 km/h.
Relation between context-specific submaximal HR during activities of daily living simulated in laboratory
settings and in free-living:
Pearson’s correlation between context-specific submaximal HR measured during activities of daily living
simulated in laboratory settings and in free-living as detected by pattern recognition methods was 0.71 for
lying down, 0.71 for walking at 3.5 km/h and 0.75 for walking at 5.5 km/h. Mean difference between
context-specific HR in laboratory settings and free-living was 2.9 ± 8.7 for lying (mean HR while lying
down was 63.2 bpm in free-living and 66.2 bpm in laboratory settings), 8.7 ± 11.2 for walking at 3.5 km/h
(mean HR while walking at 3.5 km/h was 99.9 bpm in free-living and 91.0 bpm in laboratory settings) and
-2.7 ± 11.5 for walking at 5.5 km/h (mean HR while walking at 5.5 km/h was 106.3 bpm in free-living and
107.8 bpm in laboratory settings). Thus, all differences were below 10%. Histograms of the differences and
scatterplots of context-specific HR in laboratory settings and free-living are shown in Fig. 7.
Figure 6. Accuracy of the prediction models for CRF estimation. Regression plots and Bland-Altman plots are shown
for models using as predictors anthropometrics and context-specific HR during activities of daily living simulated in
laboratory conditions. R2 is also reported.
R2=0.64
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Fitted CRF
Reference CRF
Anthropometrics only
R2=0.69
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Fitted CRF
Reference CRF
Lying HR
R2=0.72
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Fitted CRF
Reference CRF
Walking 3.5 km/h HR
R2=0.74
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Fitted CRF
Reference CRF
Walking 5.5 km/h HR
500
0
500
2500 3000 3500 4000
Mean CRF, (Reference + Fitted)/2
Residuals
Anthropometrics only
500
0
500
2000 2500 3000 3500 4000
Mean CRF, (Reference + Fitted)/2
Residuals
Lying HR
500
0
500
2000 2500 3000 3500 4000
Mean CRF, (Reference + Fitted)/2
Residuals
Walking 3.5 km/h HR
500
0
500
2000 2500 3000 3500 4000
Mean CRF, (Reference + Fitted)/2
Residuals
Walking 5.5 km/h HR
Figure 7. Top row: histograms of differences between context-specific HR in laboratory settings and free-living.
Bottom row: scatterplots showing the relation between context-specific HR in laboratory settings and free-living.
CRF estimation from context-specific submaximal HR in free-living:
HR during specific contexts in free-living was 63.2 ± 9.3 bpm for lying, 99.9 ± 11.6 bpm for walking at 3.5
km/h and 106.3 ± 11.8 bpm for walking at 5.5 km/h. Pearson’s correlation between context-specific
submaximal HR as measured in free-living and CRF was -0.54 for lying down, -0.52 for walking at 3.5
km/h and -0.60 for walking at 5.5 km/h. Thus, confirming the hypothesis that submaximal HR is inversely
related to CRF. Adjusted R2 increased from the case where no HR was included (R2 = 0.65), when
including context-specific HR. More specifically R2 was 0.73 for lying, 0.74 for walking at 3.5 km/h and
0.77 for walking at 5.5 km/h. Thus, confirming that activities of higher submaximal intensities explain
more of the variance in the model, even when carried out in free-living. Results for all models are reported
in Table 6 and Bland-Altman plots for all models are shown in Fig. 8. When including more advanced
anthropometrics, such as fat free mass instead of body weight, R2 was 0.73 when no HR was used among
the predictors, 0.77 for lying and 0.80 for walking at 3.5 km/h and 5.5 km/h. We computed the likelihood
0
3
6
9
40 20 0 20 40
Beats per minute (bpm)
Count
Lying HR
60
80
100
50 60 70 80
Free living HR (bpm)
Laboratory HR (bpm)
Gender
Female
Male
Lying HR
0.0
2.5
5.0
7.5
40 20 0 20 40
Beats per minute (bpm)
Count
Walking at 3.5 km/h
60
80
100
120
70 80 90 100 110 120
Free living HR (bpm)
Laboratory HR (bpm)
Walking at 3.5 km/h
0
2
4
6
40 20 0 20 40
Beats per minute (bpm)
Count
Walking at 5.5 km/h
90
110
130
150
80 90 100 110 120 130
Free living HR (bpm)
Laboratory HR (bpm)
Walking at 5.5 km/h
ratio between regression models including anthropometrics data and HR while lying down with respect to
regression models including anthropometrics data and HR while walking at 3.5 and 5.5 km/h. The
likelihood ratio showed that for both walking speeds, including HR while walking, significantly improved
the model fit (p=0.0047 when including the HR while walking at 3.5 km/h and p=0.00027 when including
the HR while walking at 5.5 km/h).
Table 6: Multiple linear regression models for VO2max estimation from free-living data. N = 51. For each
predictor, detailed information (model coefficient, p-value) are indicated.
Model description
Predictors
R2
Anthropometric characteristics
only
Intercept (1403.603, p = 0.00326), Body weight
(19.531, p = 0.00355), age (-3.184, p = 0.73931),
sex (803.869, p = 4.59e-07)
0.65
Context-specific HR
Intercept (2914.307, p = 6.31e-06), HR while lying
down in free-living (-22.118, p = 0.000554), body
weight (21.150, p = 0.000511), age (-9.027, p =
0.298184), sex (634.875, p = 1.32e-05)
0.73
Intercept (4175.338, p = 2.19e-06), HR while
walking at 3.5 km/h in free-living (-20.798, p =
0.000136), body weight (16.106, p = 0.005611),
age (-20.240, p = 0.032176), sex (738.579, p =
1.55e-07)
0.74
Intercept (4647.138, p = 1.03e-07), HR while
walking at 5.5 km/h in free-living (-23.884, p =
7.03e-06), body weight (16.801, p = 0.0022), age (-
21.322, p = 0.0156), sex (668.687, p=5.02e-07)
0.77
Cross-validation of VO2max estimates:
VO2max estimation models derived from free-living data were cross-validated using the leave-one-out
technique. Results are reported in Table 7 and 8. Cross-validation of VO2max estimates using as predictors
context-specific HR as measured during activities of daily living simulated in laboratory settings: RMSE
for the model including anthropometric characteristics only as predictors was 358.3 ml/min (R2 was 0.66).
RMSE was reduced when including HR in specific contexts among the predictors, with RMSE = 314.3
ml/min (R2 = 0.73) for lying down, RMSE = 310.0 ml/min (R2 = 0.75) for walking at 3.5 km/h, and RMSE
= 284.7 ml/min (R2 = 0.78) for walking at 5.5 km/h as specific contexts. Thus, RMSE was reduced up to
21% when including context-specific HR among the predictors. Cross-validation of VO2max estimates
using as predictors context-specific HR as derived by pattern recognition methods in free-living: RMSE for
the model including anthropometric characteristics only as predictors was 354.7 ml/min (R2 was 0.67).
RMSE was reduced when including HR in specific contexts among the predictors, with RMSE = 309.4
ml/min (R2 = 0.75) for lying down, RMSE = 305.91 ml/min (R2 = 0.76) for walking at 3.5 km/h, and
RMSE = 281.0 ml/min (R2 = 0.79) for walking at 5.5 km/h as specific free-living contexts. Thus, RMSE
was also reduced up to 21% when including context-specific HR as determined from pattern recognition
methods, among the predictors.
Figure 8. Accuracy of the prediction models for CRF estimation. Regression plots and Bland-Altman plots are shown
for models using as predictors anthropometrics and context-specific HR in free-living. R2 is also reported.
R2=0.65
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Fitted CRF
Reference CRF
Anthropometrics only
R2=0.73
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Fitted CRF
Reference CRF
Lying HR
R2=0.74
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Fitted CRF
Reference CRF
Walking 3.5 km/h HR
R2=0.77
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Fitted CRF
Reference CRF
Walking 5.5 km/h HR
500
0
500
2500 3000 3500 4000
Mean CRF, (Reference + Fitted)/2
Residuals
Anthropometrics only
500
0
500
2000 2500 3000 3500 4000
Mean CRF, (Reference + Fitted)/2
Residuals
Lying HR
500
0
500
2000 2500 3000 3500 4000
Mean CRF, (Reference + Fitted)/2
Residuals
Walking 3.5 km/h HR
500
0
500
2000 2500 3000 3500 4000
Mean CRF, (Reference + Fitted)/2
Residuals
Walking 5.5 km/h HR
Table 7: Cross validation of multiple linear regression models for VO2max estimation using as predictors
context-specific HR as measured during activities of daily living simulated in laboratory settings.
Model description
Predictors
RMSE
ml/min
R2
Anthropometric
characteristics only
Body weight, age, sex
358.3
0.66
Context-specific HR
HR while lying down in laboratory settings,
body weight, age, sex
314.3
0.73
HR while walking at 3.5 km/h in laboratory
settings, body weight, age, sex
310.0
0.75
HR while walking at 5.5 km/h in laboratory
settings, body weight, age, sex
284.7
0.78
Table 8: Cross validation of multiple linear regression models for VO2max estimation using as predictors
context-specific HR as detected by pattern recognition methods in free-living.
Model description
Predictors
RMSE
ml/min
R2
Anthropometric
characteristics only
Body weight, age, sex
354.7
0.67
Context-specific HR
HR while lying down in free-living, body
weight, age, sex
309.4
0.75
HR while walking at 3.5 km/h in free-living,
body weight, age, sex
305.9
0.76
HR while walking at 5.5 km/h in free-living,
body weight, age, sex
281.0
0.79
Discussion:
In this work, we proposed a method to estimate VO2max in free-living, without the need for laboratory tests
or specific protocols. While many methods have been developed to estimate VO2max using data collected
under supervised laboratory conditions or following strict protocols, limited work tried to estimate CRF
using wearable sensors and data collected under unsupervised settings in free-living (9, 22). We adopted
pattern recognition techniques to determine specific contexts, e.g. low intensity activities of daily living
such as lying down and walking at predefined speeds, to contextualize submaximal HR without the need
for a strict exercise protocol. We first validated the effectiveness of submaximal context-specific HR as a
predictor of VO2max during activities of daily living simulated in laboratory settings. Then we analyzed the
correlation and relative differences between context-specific HR during activities simulated in the lab and
context-specific HR as detected by pattern recognition methods deployed in free-living. Finally, we used
context-specific HR in free-living to estimate CRF. Our results showed that VO2max estimation using as
predictors context-specific HR in free living provides accuracy comparable with laboratory derived models.
In particular, RMSE for VO2max estimation could be reduced up to 21% compared to anthropometric
characteristics only, by using as predictors HR in specific contexts as determined by pattern recognition
methods in free-living.
Context-specific HR during activities of daily living simulated in laboratory settings: the main assumption
behind this study was that submaximal HR is inversely related to VO2max, and that the correlation is higher
during submaximal activities of higher intensity. Our laboratory recordings confirm this assumption.
Pearson’s correlation between context-specific HR and VO2max went from -0.43 to -0.51 for lying and
walking activities. Multiple regression models showed higher explained variance (R2 between 0.64 and
0.74) when including context-specific HR. Increasing activity intensity, i.e. from lying to slow walking (3.5
km/h) to faster walking (5.5 km/h) further improved R2. Finally, the likelihood ratio showed that model fit
improved significantly when including in the regression models not only HR while lying down, but also
HR while walking at different speeds. These results are in agreement with a significant body of literature
relying on submaximal HR for VO2max estimation during more intense activities, such as biking or
running, compared to the low intensity activities used in this study (30).
Context recognition in free-living: We deployed activity recognition and walking speed estimation
algorithms in free-living, in order to contextualized submaximal HR without the need for strict exercise
protocols or laboratory tests. Our activity recognition model showed high accuracy in detecting lying and
walking activities (96.4-98.2%), given the characteristic accelerometer fingerprints of such activities,
characterized either by different accelerometer orientation with respect to other activities or very specific
repetitive movements. The activities chosen as free-living contexts were lying down and walking, for the
following reasons. First, those are common activities performed by healthy individuals in most
environments. Secondly, the inverse relation between HR at rest or sleeping HR and CRF was already
shown in previous research, highlighting how this parameter can be valuable for VO2max estimation (13,
30). Finally, walking activities can be discriminated in intensity, by detecting walking speed, using simply
an accelerometer. This is an important factor when trying to detect specific context in free-living, since
detecting only activity type, if the activity can be carried out at different intensities, would not be sufficient
to determine the same context for each individual. However, walking is an activity that can be accurately
quantified in terms of both type (i.e. walking) and intensity (i.e. speed). The proposed activities are low
intensity and were performed daily by the participants involved in our study, as shown by the analysis of
free-living data. Our study population spent on average 44.4% of the free-living time lying down and 5.4%
of the free-living time walking. Of the time spent walking, 11.9 minutes daily were spent at 3.5 km/h, while
11.6 minutes daily were spent at 5.5 km/h, the two speeds used by our models to contextualize HR.
Considering that many fitness tests require protocols shorter than 11 minutes (e.g. the common 6-minutes
walking test), we believe a total of 10 minutes daily is a sufficient amount of data for prediction of
VO2max, at least in the population of healthy adults considered in this study. We could evaluate activity
recognition and walking speed models only under laboratory conditions, where reference was present.
Among the recognized activities, the dynamic activity cluster was recognized with accuracy below average
(see confusion matrix). We interpret that activities with high variability in movement and execution
between participants and using a single chest-worn sensor resulted in higher classifier confusions.
However, the high accuracy of walking speed estimation models and activity recognition for walking
provide confidence for the free-living detection of activities used to contextualize HR. Additionally, from
the cross-validation analysis results we can see how subject independent models built using activities of
daily living simulated in laboratory settings (RMSE were 314.3 ml/min, 310.0 ml/min and 284.7 ml/min for
lying, walking at 3.5 km/h and walking at 5.5 km/h were respectively) are similar to RMSE results obtained
contextualizing HR using pattern recognition methods in free-living (309.4 ml/min, 305.9 ml/min and
281.0 ml/min for lying, walking at 3.5 km/h and walking at 5.5 km/h respectively). These results can serve
as indirect validation of the accuracy of activity recognition and walking speed estimation in properly
detecting the relevant contexts in free-living.
Context-specific HR in free-living: Context-specific HR in free-living showed relations with VO2max
similar to what we reported in laboratory settings. The inverse relation between HR at a certain workload
and VO2max is the key principle behind laboratory based submaximal CRF tests and this relation showed
to be valid not only in laboratory settings but also in free-living as well. The correlation between HR while
lying down in free-living and VO2max was -0.54 and it was increased up to -0.60 when the HR while
walking at 5.5 km/h in free-living was used, highlighting how activities of higher intensity result a stronger
link between submaximal HR and VO2max. Explained variance also increased, between 0.65 when
anthropometrics characteristics only were used to estimate VO2max, and 0.77 when using context-specific
HR. Finally, the likelihood ratio showed that model fit improved significantly when including in the
regression models not only HR while lying down, but also HR while walking at different speeds. We also
analyzed the relation between HR during the same activities carried out in laboratory settings and free-
living. We expected differences in HR due to the different settings, e.g. walking in free-living might
include carrying weights, walking on inclined surfaces, or other factors that might raise HR. On the other
hand, lying down in laboratory settings might be more stressful than sleeping, therefore lowering HR with
respect to laboratory conditions. Additionally, a single laboratory measurement might be affected by factors
such as the previous days physical activity, while free-living recordings averaged over multiple days might
provide more stable representations of a participant’s physiology. On the other hand, free-living data might
include more bouts of fragmented walking and therefore HR might not always reach steady state. Thus, the
relation between HR during activities simulated in laboratory conditions and between HR and free-living
activities is most likely different and models deployed in free-living should be developed using free-living
data, as proposed by our methodology. However, analyzing the relation between laboratory and free-living
HR in the same contexts can be useful to determine to what extent laboratory recordings can be reproduced
in free-living as well as the ability of pattern recognition methods to detect differences between contexts
such as lying down or walking at different speeds, in unsupervised free-living conditions. The relatively
high correlation between laboratory and free-living HR (0.71-0.75), as well as similar mean values and
consistent differences between conditions (i.e. higher HR for walking at higher speed, or higher intensity,
in our case HR for laboratory activities and free-living was 66.2 bpm and 63.2 bpm for lying, 91.0 and 99.9
for walking at 3.5 km/h and 107.8 and 106.3 for walking at 5.5 km/h) are all promising results that free-
living data can be used as a reliable substitute of laboratory recordings for context-specific submaximal
HR.
Fat free mass: Analysis of VO2max estimation including fat free mass instead of body weight among the
predictors resulted in higher accuracy, as expected and previously shown in literature (22). In particular, R2
was increased between 0.74 and 0.78 for laboratory based measurements and between 0.77 and 0.80 for
context-specific HR determined in free-living. However, since the aim of our work is to provide VO2max
estimation outside of the laboratory environment, we focus on simple anthropometrics only (i.e. body
weight, age and sex) in the remaining of our discussion.
Cross-validation of VO2max estimates: We also performed cross validation using subject independent
models for VO2max estimation as our aim was to validate the proposed methods using state of the art
techniques able to validate the model on unseen data. Results for cross validation were consistent with what
was shown before. Our results confirm that when estimating CRF, the individual’s anthropometric
characteristics are not sufficient to provide an accurate estimate. Differences in CRF among participants
with similar body size (e.g. similar body weight and height) are not distinguishable if no physiological data
is used in the models. Thus, the lower RMSE showed by VO2max estimation models including HR as
predictor shows the ability of submaximal context-specific HR to discriminate between such participants
with similar anthropometric characteristics and further reduce VO2max estimation error. As expected,
contextualizing HR using more intense activities, such as walking at 5.5 km/h instead of lying, provides
better results. It is interesting to note that subject independent analysis RMSE was reduced consistently
between models using anthropometrics only and context-specific HR (for any activity), both in laboratory
settings and free-living. However, increasing the intensity of the specific context analyzed, e.g. from lying
down to walking at 3.5 km/h to walking at 5.5 km/h did not consistently reduce RMSE. RMSE for models
including HR while lying down and slow walking (i.e. walking at 3.5 km/h) were similar, highlighting that
the physiological responses to exercise we are interested in monitoring, might require a certain level of
intensity for the model to benefit beyond what can be already achieved using lying HR as predictor. These
findings are valid both in laboratory settings using HR during simulated activities of daily living and in
free-living using HR as detected by pattern recognition methods.
Comparison with prior work: Little work was reported in literature on protocol-free VO2max estimation.
Previous studies aiming at estimating VO2max in free-living conditions were either limited to using
physical activity-related parameters, such as steps, as proposed by Cao et al. (9), HR normalized by activity
intensity, as proposed by Plasqui et al. (22), or requiring intense exercise such as running (32). Results for
VO2max estimation reported in terms of R2 or RMSE cannot be easily compared between studies, due to
the dependency of these parameters on the study’s participants characteristics, for example body weight
and VO2max levels. However, we report in this section R2 results as typically reported by other studies to
put ours in perspective with current state of the art in VO2max estimation. For some studies, e.g. 21,
participants had similar characteristics to our study, and therefore comparisons can be meaningful. We
reported R2 of 0.79 for our subject independent analysis. Results reported by Plasqui et al. on a cross-
validation sample for his method showed that using as predictor HR divided by activity counts, a measure
of motion intensity, VO2max could be predicted with R2 = 0.72. The populations in the two studies are
comparable, and therefore further contextualizing HR in free-living (i.e. using as predictor HR while
walking at a certain speed) seems beneficial. Other protocols involving more intense activities, such as
running, did not provide better results. For example, by combining the ratio of inverse foot-ground contact
time and HR during steady state running, Weyand et al. (32) reported R2 = 0.74 in the experimental group
and R2 = 0.67 in the cross-validation group.
By using context-specific HR in free-living as predictor, we obtained results comparable to or better than
previous free-living studies and are also comparable to what was reported using similar metrics in
laboratory settings or while performing strict protocols (25). For example, ninety-two different VO2max
protocols were reviewed in a recent analysis by Sartor et al. (27). Additionally to the free-living studies
here discussed, the authors suggested that many other sub-maximal tests could be performed in free-living,
without laboratory infrastructure. However, most of these tests require intense activities and strict
protocols, for example the most commonly used 2-mile run (Mello et al. (18), R2 = 0.81), Canadian aerobic
fitness test (Jette et al. (15), R2 = 0.82), or YMCA (Santo et al. (26), R2 = 0.56). The accuracy of the best
performing tests is comparable to our free-living estimation. However, the approach proposed in this work
does not require intense activities, and is therefore suitable on a wider population. Additionally, the
proposed approach does not require a specific test, and therefore VO2max could be continuously assessed
longitudinally over time, and not only re-assessed when the test is performed. The effectiveness of context-
specific HR as derived in free-living with respect to laboratory based protocols was also validated in our
own analysis, showing comparable RMSE and R2 when including laboratory derived HR or free-living HR.
Other studies investigate the relation between easily accessible measures such as HR or HR variability at
rest and VO2max (12). However, these studies typically reported low levels of accuracy (Esco et al. (12), R2
= 0.29), showing that single measurements or “spot” measurements of physiological parameters and limited
levels of context are insufficient for a reliable VO2max estimate. A possible explanation for the better
performance of the proposed approach compared to both single spot checks (12) and more intense protocols
that can be carried out in free-living, is that by contextualizing HR over multiple days, our proposed
approach is less prone to the day-to-day variability typical of physiological measurements.
The clear advantage of the current approach is the ability to provide estimates during normal activities of
daily living, as carried out by individuals. We validated our models independently on the participant, using
cross-validation and the leave-one-out technique. Additionally, for all our models, we also computed
results using as predictor body weight instead of fat-free mass. Thus, providing estimates from easily
accessible measures that can be acquired without complex and expensive laboratory infrastructure. Our
results are extendable to new participants without the need of re-training the models or other laboratory
protocols. The current implementation could be directly deployed to new studies in free-living conditions.
Limitations and future work: A limitation of this study is the validation on healthy adults only, with similar
lifestyles in a Dutch setting. Future work should investigate if the proposed CRF estimation model is
suitable for other groups such as the obese and persons affected by chronic disease, and if the proposed
activity recognition system or other activity recognition systems trained to recognize only the relevant
activities to contextualize HR (e.g. lying and walking) can be suitable for these populations. In non-healthy
populations changes in CRF could provide an additional marker of disease progression. Additionally, future
work should address the ability of the proposed method not only to estimate CRF for an individual, but to
track changes in CRF over time, e.g. by means of a physical activity intervention. In this study, we assumed
VO2max to remain constant over a period of two weeks, since participants were not implementing changes
to their lifestyle, and typical interventions to modify VO2max are of much longer duration (e.g. 3 months to
1 year (16)). Finally, in this study we used a wearable sensor prototype (the ECG Necklace) to collect data.
The ECG Necklace provided raw accelerometer and ECG data streams that were processed to determine
activity type, and HR. While the heart beat detection and activity recognition algorithms are not detailed in
this paper, these basic processing components are replaceable and well known in literature (1, 24, 28) and
the novelty of our contribution is in the methodology of using the components to contextualize HR in free
living so that we could validate our hypothesis of estimating VO2max using only free-living data. Thus, this
study can be completely replicated by using off-the-shelf sensors for accelerometer and HR recordings
instead of the ECG Necklace prototype, as many wearable sensors able to detect activities and HR are
available on the market today. Especially heart activity sensors today mostly provide HR data and not
ECG, simplifying the analysis procedure.
CRF is a strong and independent predictor of all-cause and cardiovascular mortality. When evaluating the
suitability and practical applicability of a new test, many parameters should be accounted for. The cost,
convenience and infrastructure required are current barriers to widespread VO2max measurements, despite
the well-known relevance in healthcare. The proposed CRF estimation model is applicable to a wide
population, since it does not require intense physical exercise, and requires accelerometer and HR data
only. Such measures, are becoming more and more widespread due to mainstream availability of wearable
technology, including combined accelerometer and HR monitors. Similarly, the processing capabilities of
modern mobile phones are sufficient for practical deployment of machine learning methods (4).
Conclusions: In conclusion, this work showed that contextualized HR in free-living can be used to provide
VO2max estimation with accuracy comparable to other methods relying on submaximal HR measured in
laboratory settings. This is the first study utilizing pattern recognition methods to automatically
contextualized HR in free living and predict CRF. We showed that considering context-specific HR
provides better CRF estimates, and including context-specific HR at higher intensities (e.g. while walking)
further reduces estimation error. Additionally, we show increased accuracy depending on activity intensity.
When including HR while walking in the estimation model, we did not consider relevant including lying
HR too, since the information that we are trying to capture is already present in the model as represented by
walking HR (and even better represented, given the higher intensity of walking with respect to lying down).
Moreover, if we were to include both HR parameters in the regression model, the sleeping HR parameter
would be non-significant, given the weaker link between sleeping HR and CRF with respect to walking HR
and CRF, as shown by the lower correlation. The proposed approach could be used to provide more
information about an individual’s health without the need for laboratory infrastructure or specific tests.
Building up on the proposed approach, new opportunities for applications targeted at inducing behavioral
change could be developed. For example, by creating a feedback loop between objectively measured
physical activity, and changes in CRF and associated reduced risk of disease.
Acknowledgments:
The authors would like to thank Giuseppina Schiavone and Stefan Camps for their support during data
collection.
Disclosure:
This work was funded by Holst Centre/imec.
References:
[1] Altini M, Penders J, Amft O. Energy expenditure estimation using wearable sensors: a new methodology for
activity-specific models. In Proceedings of the conference on Wireless Health, WH '12, pages 1:8, New York, NY,
USA, 2012. ACM.
[2] Altini M, Penders J, and Amft O. Personalizing energy expenditure estimation using a cardiorespiratory fitness
predicate. In Pervasive Computing Technologies for Healthcare (PervasiveHealth), 2013 7th International Conference
on, pp. 65-72. IEEE, 2013.
[3] Altini M, Penders J, Vullers R, Amft O. Estimating Energy Expenditure Using Body-Worn Accelerometers: a
Comparison of Methods, Sensors Number and Positioning. IEEE Journal of Biomedical and Health Informatics, no. 99,
p. 1, 2014.
[4] Altini M, Penders J, Vullers R, Amft O. Personalized physical activity monitoring on the move. In Proceedings of
the 4th Conference on Wireless Health, ser. WH ’13. New York, NY, USA: ACM, 2013, pp. 8:18:2.
[5] Astrand PO, Ryhming I. A nomogram for calculation of aerobic capacity (physical fitness) from pulse rate during
submaximal work. J Appl Physiol, vol. 7, no. 2, pp. 218221, 1954.
[6] Bonomi AG, Plasqui G, Goris AH, Westerterp KR. Improving assess- ment of daily energy expenditure by
identifying types of physical activity with a single accelerometer. J Appl Physiol 107: 655661, 2009.
[7] Brage, S., Ekelund, U., Brage, N., Hennings, M.A., Froberg, K., Franks, P.W. and Wareham, N.J., 2007. Hierarchy
of individual calibration levels for heart rate and accelerometry to measure physical activity. Journal of Applied
Physiology, 103(2), pp.682-692.
[8] Browning RC, Kram R. Energetic cost and preferred speed of walking in obese vs. normal weight women. Obesity
Research, vol. 13,
no. 5, pp. 891899, 2005.
[9] Cao ZB, Miyatake N, Higuchi J, Ishikawa-Takata K, Miyachi M, Tabata I. Prediction of vo2max with daily step
counts for Japanese adult women. European journal of applied physiology, vol. 105, no. 2, pp. 289296, 2009.
[10] Lee DC, Artero EG, Sui E, Blair SN. Review: Mortality trends in the general population: the importance of
cardiorespiratory fitness. Journal of Psychopharmacology, vol. 24, no. 4 suppl, pp. 2735, 2010.
[11] Ebbeling CB, Ward A, Puleo EM, Widrick J, Rippe JM. Development of a single-stage submaximal treadmill
walking test. Med Sci Sports Exerc, vol. 23, no. 8, pp. 966973, 1991.
[12] Esco MR, et al. Cross-validation of the polar fitness test TM via the polar f11 heart rate monitor in predicting vo2
max. Age (yrs) 24 (2011): 5-1.
[13] Loimaala, A., Huikuri, H., Oja, P., Pasanen, M., & Vuori, I. (2000). Controlled 5-mo aerobic training improves
heart rate but not heart rate variability or baroreflex sensitivity. Journal of Applied Physiology, 89(5), 1825-1829.
[14] Jackson AS, Blair SN, Mahar MT, Wier LT, Ross RM, Stuteville JE. Prediction of functional aerobic capacity
without exercise testing. Med Sci Sports Exerc, vol. 22, no. 6, pp. 863870, 1990.
[15] Jetté M, et al. The Canadian Home Fitness Test as a predictor of aerobic capacity. Canadian Medical Association
Journal 114.8 (1976): 680.
[16] Katzel, L. I., Bleecker, E. R., Colman, E. G., Rogus, E. M., Sorkin, J. D., & Goldberg, A. P. (1995). Effects of
weight loss vs aerobic exercise training on risk factors for coronary disease in healthy, obese, middle-aged and older
men: a randomized controlled trial. Jama, 274(24), 1915-1921.
[17] Kuipers H, Verstappen F, Keizer H, Geurten P, Van Kranenburg G. Variability of aerobic performance in the
laboratory and its physiologic correlates. International journal of sports medicine, vol. 6, no. 04, pp. 197201, 1985.
[18] Mello RP, Murphy MM, Vogel JA. Relationship Between a Two Mile Run For Time and Maximal Oxygen
Uptake. The Journal of Strength & Conditioning Research 2, no. 1 (1988): 9-12.
[19] Minetti AE, Boldrini L, Brusamolin L, Zamparo P, McKee T. A feedback-controlled treadmill (treadmill-on-
demand) and the spontaneous speed of walking and running in humans. Journal of Applied Physiology, vol. 95, no. 2,
pp. 838843, 2003.
[20] Nes BM, Janszky I, Vatten LJ, Nilsen T, Aspenes ST, Wisløff U, Estimating vo2peak from a nonexercise
prediction model: the hunt study, Norway. Med Sci Sports Exerc, vol. 43, no. 11, pp. 202430, 2011.
[21] Noonan V, Dean E. Submaximal exercise testing: clinical application and interpretation. Physical Therapy, vol.
80, no. 8, pp. 782807, 2000.
[22] Plasqui G, Westerterp KR. Accelerometry and heart rate as a measure of physical fitness: cross-validation. Med
Sci Sports Exerc, vol. 38, no. 8, pp. 15101514, 2006.
[23] Rennie, K.L., Hennings, S.J., Mitchell, J. and Wareham, N.J., 2001. Estimating energy expenditure by heart-rate
monitoring without individual calibration. Medicine and science in sports and exercise, 33(6), pp.939-945.
[24] Romero, I., Grundlehner, B., & Penders, J. (2009, September). Robust beat detector for ambulatory cardiac
monitoring. In Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual International Conference of
the IEEE (pp. 950-953). IEEE.
[25] Rothney, M. P., Neumann, M., Béziat, A., & Chen, K. Y. (2007). An artificial neural network model of energy
expenditure using nonintegrated acceleration signals. Journal of applied physiology, 103(4), 1419-1427.
[26] Santo AS, Golding LA. Predicting maximum oxygen uptake from a modified 3-minute step test. Research
quarterly for exercise and sport 74.1 (2003): 110-115.
[27] Sartor F, Vernillo G, de Morree HM, Bonomi AG, La Torre A, Kubis HP, Veicsteinas A. Estimation of maximal
oxygen uptake via submaximal exercise testing in sports, clinical, and home settings. Sports medicine, vol. 43, no. 9,
pp. 865873, 2013.
[28] Tapia E. Using machine learning for real-time activity recognition and estimation of energy expenditure. In PhD
thesis, MIT, 2008.
[29] Tonis T, Gorter K, Vollenbroek-Hutten M, Hermens H. Comparing vo2max determined by using the relation
between heart rate and accelerometry with submaximal estimated vo2max. The Journal of sports medicine and physical
fitness, vol. 52, no. 4, pp. 337343, 2012.
[30] Uth, N., Sørensen, H., Overgaard, K., & Pedersen, P. K. (2004). Estimation of VO2max from the ratio between
HRmax and HRrestthe heart rate ratio method. European journal of applied physiology, 91(1), 111-115.
[31] Vanhees L, Lefevre J, Philippaerts R, Martens M, Huygens W, Troosters T, Beunen G. How to assess physical
activity? how to assess physical fitness?. European Journal of Cardiovascular Prevention & Rehabilitation, vol. 12,
no. 2, pp. 102114, 2005.
[32] Weyand, P. G., Kelly, M., Blackadar, T., Darley, J. C., Oliver, S. R., Ohlenbusch, N. E., ... & Hoyt, R. W. (2001).
Ambulatory estimates of maximal aerobic power from foot-ground contact times and heart rates in running
humans. Journal of Applied Physiology, 91(1), 451-458.
... As the demand for wearable devices for healthcare continues to rise, it has generated a booming market, and the companies are now seeing the opportunities of supplying wearable healthcare technologies to their consumers as beneficial. So far, wristwatches, gloves, patches, headbands, eyeglasses, and necklaces have been reported as types of wearable devices [8][9][10][11][12][13][14][15][16]. The wristwatch is the most affordable and widely invented wearable device type since it can provide good wearing comfort and obtain information from the skin. ...
Article
Full-text available
The objective of this work was to design a versatile readout circuit for patch-type wearable devices consisting of a Transimpedance Amplifier (TIA). The TIA performs Current to Voltage (I–V) conversion, the most widely used technique for amperometry and impedance measurement for various types of electrochemical sensors. The proposed readout circuit employs a digitally controllable feedback resistor (Rf) technique in the TIA to improve accuracy, which can be utilized in a variety of electrochemical sensors within a current range of 0.1 µA–100 µA. It is designed to accommodate multiple sensors simultaneously to track multiple target analytes for high accuracy and versatile usage. The readout circuit consists of low power operational amplifier (op–amp) and digital circuit blocks, is designed and fabricated with Magna 0.18 µm Complementary Metal Oxide Semiconductor (CMOS) technology, which provides low power consumption and a high degree of integration. The design has a small size of 0.282 mm2 and low power consumption of 0.38 mW with a 3.3 V power supply, which are desirable factors in wearable device applications.
... They include accelerometers, finger sensors, microfluidic sensors, as well as seismo-and ballistocardiography [26,27]. Attempts have also been made to use data derived from RM systems to estimate cardiorespiratory fitness and stress exposure-important parameters in quantifying the risk of heart failure and ischemic events, respectively [28,29]. ...
Article
Full-text available
The registration of physical signals has long been an important part of cardiological diagnostics. Current technology makes it possible to send large amounts of data to remote locations. Solutions that enable diagnosis and treatment without direct contact with patients are of enormous value, especially during the COVID-19 outbreak, as the elderly require special protection. The most important examples of telemonitoring in cardiology include the use of implanted devices such as pacemakers and defibrillators, as well as wearable sensors and data processing units. The arrythmia detection and monitoring patients with heart failure are the best studied in the clinical setting, although in many instances we still lack clear evidence of benefits of remote approaches vs. standard care. Monitoring for ischemia is less well studied. It is clear however that the economic and organizational gains of telemonitoring for healthcare systems are substantial. Both patients and healthcare professionals have expressed an enormous demand for the further development of such technologies. In addition to these subjects, in this paper we also describe the safety concerns associated with transmitting and storing potentially sensitive personal data.
... Polar devices), provide CRF estimation using a regression model including HR at a predefined running speed as predictor. A few methods have been recently proposed to estimate CRF using wearable sensor data acquired in free-living [7], [8], [14], [15], without the need for laboratory tests. Using wearable sensors in freeliving to estimate V O 2 max is a novel approach that could be applied to a larger population compared to maximal or sub-maximal laboratory tests. ...
... The emerging field of digital biomarkers may help in the future to provide solutions like this [44]. In the past, home-based VO2max estimates have been technically challenging to develop [45,46] though technology continues to improve [47] with projects ongoing (http://sagebionetworks.org/researchprojects/cardiorespiratory-fitness-module/). We also found that though we implemented a standardized home-based interval exercise testing program in order to try to control dose intensity of the intervention, the rigidity of the protocol was challenging from an adherence standpoint and may not have been optimally suited for participants of varying levels of baseline motivation and fitness. ...
Article
Full-text available
Allogeneic hematopoietic cell transplantation (alloHCT) is a life-saving technology that can cure otherwise incurable diseases, but imposes significant physiologic stress upon recipients. This stress leads to short-term toxicity and mid- to long-term physical function impairment in some recipients. Exercise interventions have demonstrated preliminary efficacy in preserving physical function in HCT recipients, but the role of these interventions prior to HCT (prehabilitative) is less known. We tested a 5- to 12-week, prehabilitative higher intensity home-based aerobic exercise intervention in a randomized study of alloHCT candidates. Of 113 patients screened, 34 were randomized to control or intervention groups, 16 underwent pre- and post-intervention peak oxygen consumption (VO2peak) testing, and 12 underwent pre- and post-intervention 6-min walk distance (6MWD) testing. No significant differences in VO2peak or 6MWD were seen pre- to post-intervention between intervention and control groups, but final numbers of evaluable participants in each group were too small to draw inferences regarding the efficacy of the intervention. We conclude that the design of our prehabilitative intervention was not feasible in this pilot randomized study, and make recommendations regarding the design of future exercise intervention studies in alloHCT.
Article
The expanding array and adoption of consumer health wearables is creating a new dynamic to the patient-health-care provider relationship. Providers are increasingly tasked with integrating the biometric data collected from their patients into clinical care. Further, a growing body of evidence is supporting the provider-driven utility of wearables in the screening, diagnosis, and monitoring of cardiovascular disease. Here we highlight existing and emerging wearable health technologies and the potential applications for use within sports cardiology. We additionally highlight how wearables can advance the remote cardiovascular care of patients within the context of the COVID-19 pandemic. Finally, despite these promising advances, we acknowledge some of the significant challenges that remain before wearables can be routinely incorporated into clinical care.
Article
Full-text available
Background: Finger pulse oximeters are widely used to monitor physiological responses to high-altitude exposure, the progress of acclimatization, and/or the potential development of high-altitude related diseases. Although there is increasing evidence for its invaluable support at high altitude, some controversy remains, largely due to differences in individual preconditions, evaluation purposes, measurement methods, the use of different devices, and the lacking ability to interpret data correctly. Therefore, this review is aimed at providing information on the functioning of pulse oximeters, appropriate measurement methods and published time courses of pulse oximetry data (peripheral oxygen saturation, (SpO2) and heart rate (HR), recorded at rest and submaximal exercise during exposure to various altitudes. Results: The presented findings from the literature review confirm rather large variations of pulse oximetry measures (SpO2 and HR) during acute exposure and acclimatization to high altitude, related to the varying conditions between studies mentioned above. It turned out that particularly SpO2 levels decrease with acute altitude/hypoxia exposure and partly recover during acclimatization, with an opposite trend of HR. Moreover, the development of acute mountain sickness (AMS) was consistently associated with lower SpO2 values compared to individuals free from AMS. Conclusions: The use of finger pulse oximetry at high altitude is considered as a valuable tool in the evaluation of individual acclimatization to high altitude but also to monitor AMS progression and treatment efficacy.
Article
Full-text available
Objective: Controlling the spread of the COVID-19 pandemic largely depends on scaling up the testing infrastructure for identifying infected individuals. Consumer-grade wearables may present a solution to detect the presence of infections in the population, but the current paradigm requires collecting physiological data continuously and for long periods of time on each individual, which poses limitations in the context of rapid screening. Technology: Here, we propose a novel paradigm based on recording the physiological responses elicited by a short (~2 minutes) sequence of activities (i.e. "snapshot"), to detect symptoms associated with COVID-19. We employed a novel body-conforming soft wearable sensor placed on the suprasternal notch to capture data on physical activity, cardio-respiratory function, and cough sounds. Results: We performed a pilot study in a cohort of individuals (n=14) who tested positive for COVID-19 and detected altered heart rate, respiration rate and heart rate variability, relative to a group of healthy individuals (n=14) with no known exposure. Logistic regression classifiers were trained on individual and combined sets of physiological features (heartbeat and respiration dynamics, walking cadence, and cough frequency spectrum) at discriminating COVID-positive participants from the healthy group. Combining features yielded an AUC of 0.94 (95% CI=[0.92, 0.96]) using a leave-one-subject-out cross validation scheme. Conclusions and Clinical Impact: These results, although preliminary, suggest that a sensor-based snapshot paradigm may be a promising approach for non-invasive and repeatable testing to alert individuals that need further screening.
Chapter
Ambulatory monitoring devices are enabling a new paradigm of healthcare by continuously collecting, processing, and interpreting long-term data aiming to provide reliable clinical diagnoses. These devices are becoming increasingly popular, both among healthcare practitioners and patients, for long-term continuous monitoring of cardiac diseases. Advancements in the fields of hardware technologies and software algorithms have enabled solutions that are both affordable and reliable, allowing monitoring of vulnerable populations from the comfort of their homes. These devices provide early detection of important physiological events, providing patients with timely alerts to seek medical attention. In this chapter, we aim to summarize recent developments and challenges in the area of ambulatory and remote monitoring solutions for cardiac and respiratory diagnostics. We present solutions based on wearable devices, smartphones/tablets, as well as implantable sensors. Finally, we present an overview of the limitations of current technologies, their effectiveness, and adoption by the general population, and we discuss some of the recently proposed methods that may help overcome those challenges.
Article
Full-text available
Background: Low cardiorespiratory fitness (CRF) increases risk of all-cause mortality and cardiovascular events. Periodic CRF assessment can have an important preventive function. Objective: To develop a protocol-free method to estimate CRF in daily life based on heart rate (HR) and body acceleration measurements. Methods: Acceleration and HR data were collected from 37 subjects (M=49%) while performing a standardized laboratory activity protocol (sitting, walking, running, cycling) and during a 5-days free-living monitoring period. CRF was determined by oxygen uptake (VO2max) during maximal exercise testing. A doubly-labeled water validated equation was used to predict total energy expenditure (TEE) from acceleration data. A fitness index was defined as the ratio between TEE and HR (TEE-pulse). Activity recognition techniques were used to process acceleration features and classify sedentary, ambulatory and other activity types. Regression equations based on TEE-pulse data from each activity type were developed to predict VO2max. Results: TEE-pulse measured within each activity type of the laboratory protocol was highly correlated to VO2max (r from 0.74 to 0.91). Averaging the outcome of each activity-type specific equation based on TEE-pulse from the laboratory data led to accurate estimates of VO2max (RMSE: 300.0 mlO2/min or 10%). The difference between laboratory and free-living determined TEE-pulse was 3.7 ± 11% (r =0.85). The prediction method preserved the prediction accuracy when applied to free-living data (RMSE: 367 mlO2/min or 12%). Conclusions: Measurements of body acceleration and HR can be used to predict VO2max in daily life. Activity-specific prediction equations are needed to achieve highly accurate estimates of CRF.
Article
Full-text available
The purpose of this study was to cross-validate the Polar Fitness Test™ via the Polar F11 heart rate (HR) monitor in predicting VO 2 max in apparently healthy male volunteers. Fifty male subjects (age = 24.0 ± 5.1 yrs) volunteered to participate in the study. Each subject was instructed to assume a supine position for 10 min while the Polar F11 HR monitor predicted their VO 2 max (pVO 2 max) by way of the Polar Fitness Test. Criterion VO 2 max (aVO 2 max) was determined via a maximal graded exercise test on a treadmill. The mean values for pVO 2 max and aVO 2 max were 45.4 ± 11.3 mL.kg -1.min -1 and 47.4 ± 9.1 mL.kg -1.min-1, respectively, which were not significantly different (p > 0.05). The validity statistics for pVO 2 max versus aVO 2 max were r = 0.54 (p < 0.05), CE = -1.93 mL.kg -1.min-1, SEE = 7.69 mL.kg -1.min-1, TE = 10.04 mL.kg -1.min-1 and limits of agreement ranged from -18.0 mL.kg -1.min-1 to 21.8 mL.kg -1.min- Due to the moderate validation coefficient and large individual differences between the predicted and criterion VO 2 max, the Polar Fitness Test™ is limited. Therefore, one should take caution when using this method to predict VO 2 max.
Conference Paper
Full-text available
Accurate estimation of Energy Expenditure (EE) in ambulatory settings is a key element in determining the causal relation between aspects of human behavior related to physical activity and health. We present a new methodology for activity-specific EE algorithms. The proposed methodology models activity clusters using specific parameters that capture differences in EE within a cluster, and combines these models with Metabolic Equivalents (METs) derived from the compendium of physical activities. We designed a protocol consisting of a wide set of sedentary, household, lifestyle and gym activities, and developed a new activity-specific EE algorithm applying the proposed methodology. The algorithm uses accelerometer (ACC) and heart rate (HR) data acquired by a single monitoring device, together with anthropometric variables, to predict EE. Our model recognizes six clusters of activities independent of the subject in 52.6 hours of recordings from 19 participants. Increases in EE estimation accuracy ranged from 18 to 31% compared to state of the art single and multi-sensor activity-specific methods.
Conference Paper
Full-text available
Accurate Energy Expenditure (EE) estimation is key in understanding how behavior and daily Physical Activity (PA) patterns affect health. Mobile phones and wearable sensors (e.g. accelerometers (ACC) and heart rate (HR) monitors) have been widely used to monitor PA. In this paper we present a real-time implementation of activity-specific EE estimation algorithms, using an Health Patch and an iPhone. Our approach to continuous monitoring of PA targets personalized behavior and health status assessment, by automatically accounting for a person's cardiorespiratory fitness level (CRF), which is the main cause of inter-individual variation in HR during moderate to vigorous activities. The proposed system opens new opportunities for personalized health assessment in daily life, using ubiquitous devices.
Article
Full-text available
Several methods to estimate Energy Expenditure (EE) using body-worn sensors exist, however quantifications of the differences in estimation error are missing. In this paper, we compare three prevalent EE estimation methods and five body locations to provide a basis for selecting among methods, sensors number and positioning. We considered (a) counts-based estimation methods, (b) activity-specific estimation methods using METs lookup, and (c) activity-specific estimation methods using accelerometer features. The latter two estimation methods utilize subsequent activity classification and EE estimation steps. Furthermore, we analyzed accelerometer sensors number and on-body positioning to derive optimal EE estimation results during various daily activities. To evaluate our approach, we implemented a study with 15 participants that wore five accelerometer sensors while performing a wide range of sedentary, household, lifestyle, and gym activities at different intensities. Indirect calorimetry was used in parallel to obtain EE reference data. Results show that activity-specific estimation methods using accelerometer features can outperform counts-based methods by 88% and activity-specific methods using METs lookup for active clusters by 23%. No differences were found between activityspecific methods using METs lookup and using accelerometer features for sedentary clusters. For activity-specific estimation methods using accelerometer features, differences in EE estimation error between the best combinations of each number of sensors (1 to 5), analyzed with repeated measures ANOVA, were not significant. Thus, we conclude that choosing the best performing single sensor does not reduce EE estimation accuracy compared to a five sensors system and can reliably be used. However, EE estimation errors can increase up to 80% if a nonoptimal sensor location is chosen.
Conference Paper
Full-text available
Accurate Energy Expenditure (EE) estimation is key in understanding how behavior and daily physical activity (PA) patterns affect health, especially in today's sedentary society. Wearable accelerometers (ACC) and heart rate (HR) sensors have been widely used to monitor physical activity and estimate EE. However, current EE estimation algorithms have not taken into account a person's cardiorespiratory fitness (CRF), even though CRF is the main cause of inter-individual variation in HR during exercise. In this paper we propose a new algorithm, which is able to significantly reduce EE estimate error and inter-individual variability, by automatically modeling CRF, without requiring users to perform specific fitness tests. Results show a decrease in Root Mean Square Error (RMSE) between 28 and 33% for walking, running and biking activities, compared to state of the art activity-specific EE algorithms combining ACC and HR.
Article
Full-text available
Assessment of the functional capacity of the cardiovascular system is essential in sports medicine. For athletes, the maximal oxygen uptake [Formula: see text] provides valuable information about their aerobic power. In the clinical setting, the [Formula: see text] provides important diagnostic and prognostic information in several clinical populations, such as patients with coronary artery disease or heart failure. Likewise, [Formula: see text] assessment can be very important to evaluate fitness in asymptomatic adults. Although direct determination of [Formula: see text] is the most accurate method, it requires a maximal level of exertion, which brings a higher risk of adverse events in individuals with an intermediate to high risk of cardiovascular problems. Estimation of [Formula: see text] during submaximal exercise testing can offer a precious alternative. Over the past decades, many protocols have been developed for this purpose. The present review gives an overview of these submaximal protocols and aims to facilitate appropriate test selection in sports, clinical, and home settings. Several factors must be considered when selecting a protocol: (i) The population being tested and its specific needs in terms of safety, supervision, and accuracy and repeatability of the [Formula: see text] estimation. (ii) The parameters upon which the prediction is based (e.g. heart rate, power output, rating of perceived exertion [RPE]), as well as the need for additional clinically relevant parameters (e.g. blood pressure, ECG). (iii) The appropriate test modality that should meet the above-mentioned requirements should also be in line with the functional mobility of the target population, and depends on the available equipment. In the sports setting, high repeatability is crucial to track training-induced seasonal changes. In the clinical setting, special attention must be paid to the test modality, because multiple physiological parameters often need to be measured during test execution. When estimating [Formula: see text], one has to be aware of the effects of medication on heart rate-based submaximal protocols. In the home setting, the submaximal protocols need to be accessible to users with a broad range of characteristics in terms of age, equipment, time available, and an absence of supervision. In this setting, the smart use of sensors such as accelerometers and heart rate monitors will result in protocol-free [Formula: see text] assessments. In conclusion, the need for a low-risk, low-cost, low-supervision, and objective evaluation of [Formula: see text] has brought about the development and the validation of a large number of submaximal exercise tests. It is of paramount importance to use these tests in the right context (sports, clinical, home), to consider the population in which they were developed, and to be aware of their limitations.
Article
Obesity is now considered a global epidemic and is predicted to become the number one preventive health threat in the industrialized world. Presently, over 60% of the U.S. adult population is overweight and 30% is obese. This is of concern because obesity is linked to leading causes of death, such as heart and pulmonary diseases, stroke, and type 2 diabetes. The dramatic rise in obesity rates is attributed to an environment that provides easy access to high caloric food and drink and promotes low levels of physical activity. Unfortunately, many people have a poor understanding of their own daily energy (im)balance: the number of calories they consume from food compared with what they expend through physical activity. Accelerometers offer promise as an objective measure of physical activity. In prior work they have been used to estimate energy expenditure and activity type. This work further demonstrates how wireless accelerometers can be used for real-time automatic recognition of physical activity type, intensity, and duration and estimation of energy expenditure. The parameters of the algorithms such as type of classifier/regressor, feature set, window length, signal preprocessing, sensor set utilized and their placement on the human body are selected by performing a set of incremental experiments designed to identify sets of parameters that may balance system usability with robust, real-time performance in low processing power devices such as mobile phones. The algorithms implemented are evaluated using a dataset of examples of 52 activities collected from 20 participants at a gymnasium and a residential home. The algorithms presented here may ultimately allow for the development of mobile phone-based just-in-time interventions to increase self-awareness of physical activity patterns and increases in physical activity levels in real-time during free-living that scale to large populations.
Article
Objective. —To compare the effects of weight loss vs aerobic exercise training on coronary artery disease risk factors in healthy sedentary, obese, middle-aged and older men.Design. —Randomized controlled trial.Subjects. —A total of 170 obese (body mass index, 30±1 kg/m2 [mean±SEM]), middle-aged and older (61 ±1 years) men.Interventions. —A 9-month diet-induced weight loss intervention, 9-month aerobic exercise training program, and a weight-maintenance control group.Main Outcome Measures. —Change in body composition, maximal aerobic capacity (Vo2max), blood pressure, lipoprotein concentrations, and glucose tolerance.Results. —Forty-four of 73 men randomized to weight loss completed the intervention and had a 10% mean reduction in weight (-9.5±0.7 kg; P<.001), with no change in Vo2max. Forty-nine of 71 men randomized to aerobic exercise completed the intervention, increased their Vo2max by a mean of 17% (P<.001), and did not change their weight, whereas the 18 men who completed in the control group had no significant changes in body composition or Vo2max. Weight loss decreased fasting glucose concentrations by 2%, insulin by 18%, and glucose and insulin areas during the oral glucose tolerance test (OGTT) by 8% and 26%, respectively (P<.01). By contrast, aerobic exercise did not improve fasting glucose or insulin concentrations or glucose responses during the OGTT but decreased insulin areas by 17% (P<.001). In analysis of variance, the decrement in fasting glucose and insulin levels and glucose areas with intervention differed between weight loss and aerobic exercise when compared with the control group (P<.05). Similarly, weight loss but not aerobic exercise increased high-density lipoprotein cholesterol levels (+13%) and decreased blood pressure compared with the control group. In multiple regression analyses, the improvement in lipoproteins and glucose metabolism was related primarily to the reduction in obesity.Conclusions. —These results suggest that weight loss is the preferred treatment to improve coronary artery disease risk factors in overweight, middle-aged and older men.(JAMA. 1995;274:1915-1921)
Article
The purpose of this study was to examine the relationship between a maximal effort two-mile run for time and maximal oxygen uptake ([latin capital V with dot above] O2 max) as measured by treadmill running. Subjects were 44 males (aged 20-51) and 17 females (aged 20-37) of various fitness and activity levels. All subjects performed a timed two mile run and a treadmill runing test for maximal oxygen uptake. Mean [latin capital V with dot above] O2 max values for men and women respectively were 50.4 +/- 7.7 and 42.0 +/- 6.0 ml/kg.min. Mean two mile run times (min:sec) for men and women were 14:44 +/- 2:06 and 17:26 +/- 3:01 respectively. Correlation between [latin capital V with dot above] O2 max and two mile run time for males and females were rm = -0.91 and rf= -0.89, respectively. The addition of such variables as age, height, weight, and percent body fat produced no significant improvement (p<.05) in the predictability of either equation. Inclusion of body weight in the male equation, however, did increase its predictive accuracy (SEE = 3.31 to 2.69). This study confirms the usefulness and validity of a timed two mile run test to indicate the level of aerobic fitness capacity when the test is properly supervised and the subjects are well-motivated. (C) 1988 National Strength and Conditioning Association