ArticlePDF Available

Cardiorespiratory Fitness Estimation in Free Living Using Wearable Sensors

Authors:
  • HRV4Training

Abstract and Figures

Objective: In this paper we propose artificial intelligence methods to estimate cardiorespiratory fitness (CRF) in free-living using wearable sensor data. Methods: Our methods rely on a computational framework able to contextualize heart rate (HR) in free-living, and use context-specific HR as predictor of CRF without need for laboratory tests. In particular, we propose three estimation steps. Initially, we recognize activity primitives using accelerometer and location data. Using topic models, we group activity primitives and derive activities composites. We subsequently rank activity composites, and analyze the relation between ranked activity composites and CRF across individuals. Finally, HR data in specific activity primitives and composites is used as predictor in a hierarchical Bayesian regression model to estimate CRF level from the participant's habitual behavior in free-living. Results: We show that by combining activity primitives and activity composites the proposed framework can adapt to the user and context, and outperforms other CRF estimation models, reducing estimation error between 10.3% and 22.6% on a study population of 46 participants. Conclusions: Our investigation showed that HR can be contextualized in free-living using activity primitives and activity composites and robust CRF estimation in free-living is feasible.
Content may be subject to copyright.
Cardiorespiratory fitness estimation in free-living using
wearable sensors
Marco Altini
a
, Pierluigi Casale
b
, Julien Penders
b
, Oliver Amft
c
a
altini.marco@gmail.com +31 6 46375742, Signal Processing and Systems, Eindhoven
University of Technology, Den Dolech 2, Eindhoven, NL and Bloom Technologies,
Agoralaan Building Abis 2.13, 3590 Diepenbeek, Belgium
b
imec The Netherlands, High Tech Campus 31, 5656AE, Eindhoven, The Netherlands
c
Chair of Sensor Technology, University of Passau, Innstrasse 41, 94032, Passau, Germany
Abstract
Objective: In this paper we propose artificial intelligence methods to esti-
mate cardiorespiratory fitness (CRF) in free-living using wearable sensor data.
Methods: Our methods rely on a computational framework able to contex-
tualize heart rate (HR) in free-living, and use context-specific HR as predictor
of CRF without need for laboratory tests. In particular, we propose three es-
timation steps. Initially, we recognize activity primitives using accelerometer
and location data. Using topic models, we group activity primitives and derive
activities composites. We subsequently rank activity composites, and analyze
the relation between ranked activity composites and CRF across individuals.
Finally, HR data in specific activity primitives and composites is used as pre-
dictor in a hierarchical Bayesian regression model to estimate CRF level from
the participant’s habitual behavior in free-living.
Results: We show that by combining activity primitives and activity com-
posites the proposed framework can adapt to the user and context, and outper-
forms other CRF estimation models, reducing estimation error between 10.3%
and 22.6% on a study population of 46 participants.
Conclusions: Our investigation showed that HR can be contextualized in
free-living using activity primitives and activity composites and robust CRF
estimation in free-living is feasible.
Keywords: context recognition, topic models, Bayesian models,
Preprint submitted to Journal of L
A
T
E
X Templates February 16, 2016
cardiorespiratory fitness
1. Introduction
In the past few years, ubiquitous sensing technologies showed unprecedented
insights into the relation between physical activity and health [1]. Wearable sen-
sors are getting more and more widespread due to improvements in miniatur-
ization, battery capacity and user experience design, reaching ubiquitousness in5
the quantified-self community and being rapidly adopted by the general popula-
tion. Due to fast paced technological developments and increased availability of
multivariate data streams acquired from wearable sensors (e.g. accelerometer,
physiological data), new computational and artificial intelligence applications
and techniques have been developed. When deployed in unsupervised free-living10
settings computational and artificial intelligence techniques can help shedding
light on the complex relation between human behavior and health, ultimately
driving behavioral change and better health outcomes [2, 3, 4, 5].
Wearable sensors have great potential for accurate physical activity moni-
toring in daily life [3, 2]. However, artificial intelligence capabilities of current15
systems and devices are limited, with almost all solutions focusing on behavioral
aspects of physical activity such as steps, activity type and energy expenditure
[6, 7]. On the other hand, cardiorespiratory fitness (CRF) is a marker of cardio-
vascular and cardiorespiratory health, and therefore is a key health parameter
that could be estimated using state of the art technologies and computational20
methods [8, 9, 10].
CRF is defined as the ability of the circulatory and respiratory systems to
supply oxygen during sustained physical activity and is considered among the
most important determinants of health and wellbeing. CRF is not only an
objective measure of habitual physical activity, but also a useful diagnostic and25
prognostic health indicator for patients in clinical settings, as well as healthy
individuals [8]. Epidemiological research has shown that in both individuals
affected by disease [11] and healthy individuals [12, 13] higher level of CRF
2
resulted in better outcomes in term of slower disease progression, lower risk
of cardiovascular disease as well as lower risk of all cause mortality. Thus,30
knowledge of CRF can be key in managing a healthy lifestyle.
Current practice for CRF measurement is direct measurement of oxygen
volume (V O
2
in ml/min) during maximal exercise (i.e.V O
2
max), the gold stan-
dard. However, V O
2
max tests are affected by multiple limitations. Medical
supervision is required and the test can be risky for individuals in non-optimal35
health conditions. Less risky submaximal tests have also been developed [14].
Submaximal tests to estimate CRF typically require measuring heart rate (HR)
while running at a certain speed or biking at a certain intensity. The inverse
relation between HR at a certain exercise intensity, fixed by the strict exer-
cise protocol that has to be sustained, and fitness, is the rationale behind this40
approach.
In this work, we propose to use artificial intelligence methods to estimate
CRF using wearable sensor data acquired in free-living. We rely on the inverse
relation between HR and fitness, but without the need for specific exercise
protocols in laboratory settings. We aim at using computational techniques to45
automatically determine contexts in which HR can be interpreted, without any
supervision from the user, and in free-living. Our hypothesis is that physiological
data, for example HR, in free-living settings is not only affected by activity
primitives such as walking, but by a combination of activity primitives and
more abstract activity composites such as social interactions, working, etc. We50
define context as a combination of activity primitives and activity composites.
Thus, we propose a method to determine both activity primitives and activity
composites, to contextualize HR. Finally, after determining the user’s context,
we use contextualized HR to estimate person-specific CRF in a hierarchical
Bayesian model. By using a non-nested hierarchical Bayesian model, parameters55
can vary depending on the activity performed, therefore being more flexible
than models requiring specific activities. This paper provides the following
contributions:
3
1. We propose a context recognition framework to contextualize HR and es-
timate CRF based on contextualized HR in free-living. First, we use topic60
models (TMs) to derive activity composites. Secondly, we rank activ-
ity composites to determine which activity composites are best suited for
CRF estimation. Finally, we use HR data in specific contexts (i.e. activity
primitives, walking speeds and activities composites) as a predictor in a
hierarchical Bayesian model to estimate CRF.65
2. We show the effectiveness of the proposed approach to estimate CRF on
a dataset including 14 days of unsupervised free-living recordings from
46 participants and reference V O
2
max acquired in laboratory conditions.
CRF estimation error was reduced between 10.3% and 22.6% compared
to alternative methods.70
2. Related work
2.1. Wearable sensors and artificial intelligence to monitor physical activity
Energy expenditure is the most commonly used metric to quantify physical
activity. Accelerometers and HR monitors are the most commonly used sin-
gle sensor devices in epidemiologic studies and consumer products. Different75
methods have been developed in the past to monitor physical activity using
such accelerometer and HR monitors. Typically, accelerometer-based systems
rely on the relation between motion intensity close to the body’s center of mass
and energy expenditure. Using a single accelerometer prevents discriminating
upper and lower body movement, e.g. biking and arm exercises, leading to large80
estimation error for activities not involving whole body motion. For example,
Crouter et al. [15] had to remove biking activities from their evaluation, due to
the inability of their system to capture physical activity when there is limited
motion close to the body’s center of mass.
Recent work showed that introducing artificial intelligence methods, activity85
type can be reliably detected with wearable sensors, opening new opportunities
for physical activity monitoring [17, 18, 19, 20, 21]. Similarly, several activity
4
trackers and wearable sensors have been released on the market in the recent
past, typically providing users with estimates of calories burnt and steps taken
(e.g. Fitbit). While activity type, energy expenditure, steps, etc. are im-90
portant, they reflect only individual behavior, but do not provide insights on
the individual’s actual health status. CRF can potentially provide more infor-
mation on an individual’s health status, being a marker of cardiovascular and
cardiorespiratory health, and therefore a key health parameter [8, 9, 10]. Thus,
our work aims at moving beyond quantification of human behavior, and towards95
quantification of health status as derived by CRF.
2.2. CRF estimation in laboratory settings
V O
2
max is regarded as the most precise method for determining CRF [22].
Despite the indubitable importance of CRF in health, measurements of V O
2
max
are rare since they require specialized personnel and expensive equipment. The100
high motivation demand and exertion of the participants makes the test unfea-
sible in many patients groups [23]. As an alternative, many non-excercise and
2000
3000
4000
55 60 65 70
Body weight − kg
Measured VO2max − ml/min
Sex
Female
Male
a) VO2max vs Body Weight
1500
2000
2500
3000
3500
4000
80 90 100 110 120
HR Walking 5 km/h − bpm
Measured VO2max − ml/min
Sex
Female
Male
b) VO2max vs HR Walking 5 km/h
Figure 1: Relation between body weight, HR and CRF for participants with similar body size
(weight and height) characteristics. a) Positive relation between V O
2
max and body weight
disappears when participants with similar body size characteristics are considered. b) Negative
relation between V O
2
max and HR while walking holds on a subset of participants with similar
body size, and can potentially be used to discriminate CRF levels.
submaximal models have been developed. Non-exercise modellaboratory-baseds
5
of CRF use easily accessible characteristics such as age, gender and self-reported
physical activity [24, 25]. However, for individuals with similar characteristics,105
CRF levels cannot be discriminated, as shown in Fig. 1. Submaximal tests have
been developed to estimate V O
2
max during specific protocols while monitoring
HR at predefined workloads [14]. Contextualized HR, e.g. HR while performing
a specific activity in laboratory settings, is discriminative of CRF levels between
individuals with similar characteristics, due to the inverse relation between HR110
and CRF [26] (see Fig. 1). Commercial devices, for example some sport watches
paired to HR monitors [27, 28] (e.g. Garmin or Polar devices), provide CRF
estimation using a regression model including HR at a predefined running speed
as predictor. However, submaximal tests are still affected by limitations; the
test should be re-performed every time CRF needs to be assessed, often requires115
laboratory infrastructure and specific activities to be performed [29].
While some devices and methods were developed to provide CRF estimation
while performing intense exercise or under laboratory settings, very few systems
and algorithms developed up to now focus on providing CRF estimation in free-
living settings [26]. Using wearable sensor in free-living to estimate V O
2
max120
is a novel approach. The estimation could be applied to a larger population
compared to maximal or submaximal laboratory tests. Individuals not per-
forming sports could still benefit from knowing more about their health status,
via estimates of CRF, and potentially be motivated to take up a more active
lifestyle.125
2.3. CRF estimation in free-living
Preliminary work explored the relation between physical activity as ex-
pressed by a step counter, and CRF [30]. While number of steps can provide
useful insights, the relation between HR and oxygen uptake at a certain exercise
intensity cannot be exploited using motion based sensors. Plasqui et al. [26]130
showed that a combination of average HR and level of motion over a period of
seven days correlates significantly with V O
2
max. However, the relation between
average HR and activity counts depends on the amount of activity performed
6
[26]. Tonis et al. [31] explored different parameters to estimate CRF from HR
and accelerometer data during activities of daily living simulated in laboratory135
settings. However V O
2
max reference and free-living data were not collected.
0.00
0.02
0.04
0.06
60 80 100 120
HR − bpm
Density
ActivityComposite
cleaning
social
work
Figure 2: Density plot of HR data during the activity primitive sedentary, occurring in differ-
ent activity composites, i.e. cleaning, social, work. Although the activity primitive sedentary
occurs in all activity composites, HR differs consistently across activity composites. Thus,
detecting activity composites can improve interpretation of HR in free-living, and therefore
provide more accurate CRF estimation. Activities composites were manually annotated.
When moving towards free-living settings, HR is more difficult to interpret,
since activities vary depending on the different lifestyles people adopt. How-
ever, contextualizing HR in free-living settings using pattern recognition and
artificial intelligence methods opens an opportunity to bring sub-maximal tests140
to uncontrolled free-living conditions.
2.4. Artificial intelligence for context recognition
We hypothesized that HR in free-living settings is not only affected by ac-
tivity primitives but by a combination of activity primitives and more abstract
activity composites. Thus, we consider as context a combination of activity145
primitives and activity composites. For example, HR during the activity prim-
itive sedentary changes substantially depending on the context in which such
activity is performed. HR during social interactions is higher than during work
for sedentary activities, possibly due to the higher physiological stress involved
in talking and interacting with other people, as shown in Fig. 2. Thus, CRF esti-150
mation models might benefit from inclusion of activity composites representing
7
additional factors present in free-living (e.g. psychological stress, interactions
with other people, etc.).
Various pattern recognition and artificial intelligence methods have been
proposed to determine context and activities in literature. Typically, activities155
are thought of in a hierarchical manner, starting from activity primitives, to
more abstract activity composites [32].
An example of activity primitives can be a set of postures and locomotion
activities, such as: lying down, sedentary, dynamic, walking, biking and run-
ning, as determined using supervised methods in previous research [16]. On160
the contrary, higher level contextual information, such as activity composites,
require a different recognition approach. Such activities are personal and need
unsupervised methods able to discover different patterns in each individual,
depending on their behavior. A possible solution is the use of TMs [33]. In
activity recognition, TMs were applied to discover activity composites from ac-165
tivity primitives [34]. Recent work investigated the impact of multiple TMs
(in particular LDA, latent Dirichlet allocation) parameters for activity compos-
ites discovery, showing promising results [35] for recognition of abstract activity
composites.
In our previous work [36], we proposed a method to determine which activity170
composites are better suited to interpret HR for one individual. For example,
we determined in which activity composites HR was more representative of
HR normalization parameters used to personalize EE estimates. Our approach
consisted of ranking activity composites based on features in order to compare
them across participants. In this work, we extend our method to the relation175
between HR during activity composites and V O
2
max. We aim at finding for
each individual specific contexts where HR is representative of CRF in free-
living, using an unsupervised approach. Then, we use contextualized HR to
predict CRF without the need for laboratory tests or specific exercises.
8
3. Approach180
Following a top down approach, CRF y
CRF
was estimated from contextual-
ized HR HR
ctx
and anthropometric characteristics by a hierarchical Bayesian
regression model, as shown in Fig. 3. Contextualized HR HR
ctx
refers to HR
during specific activity primitives, speeds and relevant activity composites. We
used features from accelerometer X
acc
, HR X
hr
, location X
coo
and anthropo-185
metrics X
ant
as input to our context recognition and CRF estimation models.
Activity primitives c were used together with stay regions sr as input for LDA
topic discovery to obtain activity composites. Activity composites were ranked
to find the most relevant ones for CRF estimation, referred to as relevant activity
composites (see Sec. 3.3 for details). The procedure to determine activity prim-190
itives, speeds, activity composites, and therefore contextualized HR HR
ctx
is
shown in Fig. 4.
In the remaining of this section, we detail the approach and provide an
example. We consider walking at 3 and 5 km/h as exemplary activity primitives
and speeds. Thus, to determine contextualized HR, we consider HR data while195
walking at 3 and 5 km/h during relevant activity composites.
3.1. CRF estimation
The CRF estimation y
CRF
was derived by a hierarchal Bayesian regression
model. Parameters modeling the relations between HR
ctx
and y
CRF
vary
depending on the context ctx. We denote the estimation model as:
y
CRF
p
N(X
CRF
p
β
CRF
+ X
ctx[p]
β
ctx[p]
, σ
2
CRF
), (1)
ctx = 1, . . . , R p = 1, . . . , np
X
CRF
p
= [1, X
ant
p
] R
np×(D+1)
, p = 1, . . . , np
X
ctx
= [HR
ctx
] R
np×1
p = 1, . . . , np
where matrix X
CRF
p
is of dimension np × (D + 1). np is the number of partici-
pants, while D the number of anthropometric characteristics X
ant
p
for a person
9
p, which includes body weight, height, age and sex. The associated parameters200
β
CRF
do not vary by context ctx since they are relative to a person and remain
the same across different activities. Contexts ctx are a set R representing a
combination of activity primitives and speeds during relevant and activity com-
posites, as shown in Fig. 3. In our example, contexts are R = 2, i.e. walking
at 3 or 5 km/h during relevant activity composites, and control the parameters205
β
ctx
for the predictor HR
ctx
. By letting the parameters β
ctx
vary, users are not
constrained to one specific activity. Instead, the model will provide a CRF es-
timate y
CRF
depending on the available activity primitives and speeds. Details
on the model parameters estimation procedure are reported in Sec. 5.
Figure 3: Hierarchical Bayesian model in plate notation. Parameters β
ctx
vary by context ctx
and model the relation between contextualized HR HR
ctx
and CRF y
CRF
.
3.2. Context recognition210
In this section we introduce our context recognition architecture to deter-
mine contextualized HR
HR
ctx
, as shown in Fig. 4. Activity composites were
discovered using LDA. LDA is a generative probabilistic model which discovers
K activity composites, from S time windows of N words y
n
. For activity recog-
nition, words y
n
are typically basic building blocks for activity composites, such
as activity primitives. In our implementation we used stay regions and activity
primitives (see Sec. 5) as words y
n
. Accelerometer features X
acc
were used to
derive activity primitives c
i
combining a Support Vector Machines (SVM) clas-
sifier and subsequent Hidden Markov Models (HMM) used to smooth transitions
10
Figure 4: Proposed approach to determine contextualized HR HR
ctx
. LDA uses histograms
of activity primitives c and stay regions sr to discover a set of activity composites, which are
ranked to determine relevant activity composites. Contextualized HR HR
ctx
is shown in the
top block, and is determined by combining activity primitives, activity composites and speed.
HR
ctx
is used as input for the CRF estimation model detailed in Fig. 3.
between activities. The hidden states corresponded to the real activity compos-
ites, c
i
, while the observable states are the ones recognized by the SVM. Stay
regions were derived from GPS coordinates X
coo
using time and distance thresh-
olds (see Sec. 5). According to the generative process, for each word y
n
, we
first draw the activity composite z
n
. z
n
is a scalar z 1 : K indicating the
activity composite for time window n. Each assigned activity composite z
n
is
derived from a multinomial distribution defined by the parameter θ
s
. θ
s
is the
distribution over activity composites for time window s:
θ
s
Dir(α) 1 s S (2)
z
n
Mult(θ
s
) 1 s S, 1 n N (3)
LDA defines θ
s
as a Dirichlet distribution with hyperparameter α. Then,
another multinomial is used to choose a word y
n
, conditioned on the activity
composite z
n
, p(y
n
|z
n
):
y
n
Mult(β
z
n
) 1 n N (4)
11
Where β is defined as the probability of each word n 1 : N for topic z.
The joint distribution can be specified as:
p(y, z, θ, φ|α, β) =
S
Y
s=1
Z
p(θ
s
, α)
N
Y
n=1
K
X
z=1
p(z
sn|θ
s
)p(y
sn
|z
sn
, β)
s
(5)
We were interested in estimating the distributions of the parameter θ
s
. Mul-
tiple activity composites were derived by LDA in each time window s, each
activity composite being assigned a probability. For each time window we con-
sidered only the activity composite maximizing θ
s
, indicated hereafter as z
s
, the
window’s main activity composite.215
3.3. Relevant activity composites
During the training phase, the HR for activity primitives and speeds was
computed for each main activity composite z
s
and participant par. Accelerome-
ter features X
acc
were used to estimate walking speed as y
speed
= X
speed
β
speed
+
, X
speed
= {X
acc
, X
ant
}. The resulting matrix HR
ctx
is of dimension K ×npar,
where K is the number of activity composites and npar is the number of partici-
pants. LDA-derived activity composites do not include semantics and cannot be
compared across participants. To overcome the problem of comparing activity
composites, we characterized them with a set of features T which we used to
rank activity composites, as in [36]. In order to provide a generalized method
that is applicable to new participants, we chose features T that are indepen-
dent of a person’s lifestyle, for example, T
1
T could be the relative time spent
sedentary in each activity composite for the different participants. Regardless
of what a person’s lifestyle is, it will always be possible to order LDA-derived
activity composites by feature T
1
, e.g. the relative time spent sedentary in
each activity composite. Then, HR
ctx
was ranked by feature T
1
, providing a
way to investigate the relation between the HR in different activity composites
and CRF, across participants. The ranking orders HR
ctx
by values of T
1
from
12
maximum to minimum. Since we are interested in highlighting commonalities
across activities composites, ranked HR
ctx
are smoothed by a moving average,
resulting in HR
ctx
. As a result, we obtain an array of k ranked HR values per
participant. We conclude the training phase by determining which feature in T
maximizes Pearson’s correlation between HR
ctx
and CRF. We define the vector
of correlations r
T
for a set of T N features in a context ctx. Thus, for each
context ctx, we have:
r
T
= {r
rank
T 1
, . . . , r
rank
T N
}, (6)
r
rank
i
= r(HR
ctx
par={1,...,npar},i
, CRF
par={1,...,npar}
) (7)
Where r
rank
i
is the correlation between the vector of contextualized HR HR
ctx
and CRF, among all participants par for a feature T
i
in a context ctx. The
activity composite providing the highest correlation was selected, i.e. the first
element of the HR
ctx
vector across individuals and CRF, to determine which220
feature T
i
results in activity composites most representative of CRF. Thus, the
feature T
i
= maxr
T ctx
showing the highest correlation between HR
ctx
and CRF
is chosen to determine relevant activities composites.
As an example, we consider as contexts ctx walking at 5 km/h during ac-
tivity composites with the maximum relative time spent sedentary, i.e. relevant225
activity composites, as shown in Fig. 5. We first determine the vector of k
elements HR
ctx
, representing the mean HR while walking at 5 km/h in each
LDA-discovered activity composite. Then, HR
ctx
are ranked based on the fea-
ture T
i
maximizing the correlation on our training set (i.e. the relative time
spent sedentary in each activity composite), to determine HR
ctx
. The first230
element of the ranked and smoothed HR
ctx
vector, is the contextualized HR
HR
ctx
, used as input for CRF estimation.
13
00:00 06:00 12:00 18:00
00:00 06:00 12:00 18:00
00:00 06:00 12:00 18:00
00:00 06:00 12:00 18:00
Walking
Other activities
5 km/h
Other speeds
Contextualized HR
Relevant activity composites
Other activity composites
Figure 5: Exemplary diagram of the procedure to determine contextualized HR HR
ctx
.
Plots show 24 hours of free-living data for one participant. For this illustration, we selected
as activity primitive and speed walking at 5 km/h during relevant activity composites, and
highlighted them in red. a) Recognized activity primitives, as detected by the combined SVM
and HMM classifier. b) Walking speed y
s
, determined when walking is detected, using a linear
regression model. c) Activity composites determined by LDA and defined by the distribution
of activity primitives and stay regions over 15 minutes windows. Relevant activity composites
are determined using the procedure detailed in Sec. 3.3, maximizing the correlation between
HR and CRF. d) Contextualized HR HR
ctx
is determined as the mean HR while walking at
5 km/h during relevant activity composites in this example, and highlighted in red. HR
ctx
is used to estimate CRF, as shown in Fig. 3. Between 17 and 18 hours no data are present
since the sensor was being charged.
14
4. Evaluation study
4.1. Participants and data acquisition
Participants were 46 (21 male, 25 female), age 24.7± 4.9 years, weight 68.6±235
10.9 kg, height 172.8±8.9 cm, BMI 22.9±2.5 kg/m
2
and V O
2
max 3020.8±668.9
ml/min. Written informed consent was obtained, and the study was approved
by the ethics committee of Maastricht University. The sensor platform used was
an ECG Necklace, a platform configured to acquire one lead ECG data at 256
Hz, and three-axial accelerometer data at 32 Hz. The ECG Necklace was worn240
on the chest, close to the body’s center of mass. The ECG Necklace was worn
during laboratory protocols and free-living. Additionally, during free-living each
participant carried a Samsung Galaxy S3 used to record GPS coordinates at 5
minutes intervals. Reference CRF was determined as V O
2
max, by means of
an incremental test on a cycle ergometer [37] using a indirect calorimeter that245
analyzed O
2
consumption and CO
2
production. The dataset considered for this
work contains 507 days of data collected from 46 participants in free-living, thus
for each participant we acquired about 11 days of accelerometer, ECG and GPS
data. Compared to the two-week protocol, available data per participants varied
between 7 and 14 days, due to participant availability, sensor failures and power250
outages, as participants forgot to recharge, causing data losses. Nevertheless,
we consider that the recordings were sufficient to capture the behavior of each
participant. 75 hours of laboratory recordings including reference V O
2
, V CO
2
,
acceleration, ECG and V O
2
max were also obtained for model validation.
4.2. Experiment design and validation procedure255
We collected data in free-living and laboratory settings and evaluated four
approaches to CRF estimation. All approaches were evaluated with respect to
reference CRF measured by means of a V O
2
max test carried out on a cycle
ergometer. In the remaining of this paper, we will use the following terminology
to characterize the four estimation conditions that were used for comparison;260
a) anthropometrics: no HR data was used, b) no-context: HR in free-living was
15
Figure 6: ECG Necklace and Samsung S3, the wearable sensor and phone used to collect ac-
celerometer ECG and GPS data in this study. The ECG Necklace was worn during laboratory
protocols and free-living recordings close to the body’s center of mass. The Samsung S3 was
carried during free-living only.
used directly to estimate CRF, c) primitives: HR in free-living was contextual-
ized using activity primitives and speed, d) composites: HR in free-living was
contextualized using activity primitives, speed and relevant activity composites.
Two laboratory protocols were designed and implemented for each partic-265
ipant on two separate days to avoid the maximal fitness test to affect physi-
ological parameters during less intense activities and vice versa. Additionally,
each participant wore the ECG Necklace in free-living for 14 days. All results on
CRF estimation were obtained from the free-living data, whereas the laboratory
data was used to derive the models, as detailed in the next Sections.270
Data from laboratory protocols were used to develop supervised methods for
activity type recognition and walking speed estimation. Activity type recogni-
tion and walking speed estimation models were deployed in free-living and used
as building blocks to contextualize HR. Additionally reference V O
2
max was
collected under laboratory protocols to validate the proposed CRF estimation275
models. Data collected in free-living were used to determine contextualized HR
and use contextualize HR as predictor for CRF estimation. CRF estimation
models including contextualized HR as predictor relied on; laboratory-validated
activity type recognition and walking speed estimation models, stay regions de-
termined without supervision in free-living (see Sec. 5) and activity composites280
determined using LDA, in free-living.
16
4.2.1. 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. The first protocol included simulated285
activities performed while wearing a portable indirect calorimeter. 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. The second protocol was a V O
2
max test providing290
reference data for biking and CRF. The third day was used for anthropometric
measurements including the participant’s body weight, height and body fat
assessed using doubly labelled water [38].
4.2.2. Free-living protocol
Participants worn the ECG necklace for 14 consecutive days in free-living and295
manually annotated their activity composites in a paper diary. Participants were
instructed to annotate activity composites as they occurred during the day and
to annotate only activity composites such as going to work, sleeping, commuting,
etc. Annotated activity composites were not used for model development since
activity composites were derived using LDA, and therefore without supervision300
from activity primitives, as detailed in Sec. 3 and Sec. 5. The annotations
were only used to interpret the LDA and CRF estimation results as detailed
in the discussion, Sec. 7. Activity composites can only be determined from
free-living data, since they cannot be simulated under laboratory conditions.
Participants carried a Samsung S3 phone and were instructed to charge both305
the ECG Necklace and phone and to change electrodes daily.
4.2.3. Statistics and performance measures
All models were derived using leave-one-participant-out cross validation.
The same training set, consisting of data from all participants but one, was
used to build feature selection, activity recognition, walking speed estimation310
17
and CRF estimation models. The remaining data was used for validation. The
procedure was repeated for each participant and results were averaged. LDA
models were built on data from the participant to be validated, since no refer-
ence or training set are necessary. Performance of the activity recognition mod-
els was evaluated using the class-normalized accuracy, in laboratory recordings.315
Results for walking speed estimation and CRF estimation are reported in terms
of Root-mean-square error (RMSE) and Pearson’s correlation (r), where the
outcome variables were speed in km/h and CRF in ml/min respectively. Paired
t-tests were used to compare RMSE between models.
5. Implementation320
5.1. Context recognition
5.1.1. Features
Accelerometer data from the three axes were segmented in 5 s windows,
band-pass filtered between 0.1 and 10 Hz, to isolate the dynamic component
due to body motion, and low-pass filtered at 1 Hz, to isolate the static compo-325
nent, due to gravity. Feature selection for activity type recognition was based
on results from our previous work [19], using a different dataset. Selected fea-
tures were: mean of the absolute signal, inter-quartile range, median, variance,
main frequency peak, low frequency band signal power. Accelerometer features
for walking speed estimation were: mean of the absolute signal, inter-quartile330
range, variance, main frequency peak, high frequency band signal power. HR
was determined from RR intervals extracted from raw ECG data and averaged
over 15 seconds windows.
5.1.2. Activity primitives
Laboratory activities were grouped into six clusters to be used for classifica-335
tion of activity primitives. The six clusters were lying (lying down), sedentary
(sitting, sit and write, standing), dynamic (cleaning the table, sweeping the
floor), walking, biking and running. Activity primitives were derived combining
18
a SVM and HMM. For the SVMs, we used a Gaussian radial basis kernel (cost
function parameter C = 1). Parameters were set based on previous work from340
our group [21]. The HMM is defined by parameters λ = (π, A, B); where π are
the initial state probabilities, A is the transition probability matrix, defining
the probability of transitioning between one activity to the other at time inter-
val t. The HMM states corresponded to activity primitives. B is the emission
matrix, which defines the probability of getting an emission at time t, given345
the state. We implemented the emission matrix B as b
ij
= 0.5 i = j,
b
ij
= 0.1 i 6= j, while transitions probabilities A between actual states
were derived from training data. Training data was the SVM classification result
obtained with reference activity primitives manually annotated in laboratory
settings.350
5.1.3. Walking speed
Walking speed was estimated using a multiple regression model using as
predictors the features listed in Sec. 5.1.1, together with the participant’s height.
Laboratory recordings on a treadmill while walking at different speeds were used
to build participant-independent walking speed models.355
5.1.4. Stay regions
Stay regions were computed from GPS coordinates according to time and
distance thresholds, which were set to 60 minutes and 1 km according to pre-
vious literature [39]. The time threshold ensures that each stay region is a
location where the participants spent a significant amount of time, while the360
distance threshold ensures that noisy recordings do not result into a multitude
of stay regions being detected. GPS data was collected at 5 minutes intervals to
conserve battery power. The relatively wide distance and time thresholds were
chosen due to the low frequency of the GPS recordings.
5.1.5. Relevant activity composites365
Input primitives for LDA were occurrences histograms of stay regions and
activity primitives in each time window s. LDA hyperparameter α was set to
19
0.01, while segment size and number of topics k were set to 15 minutes and 20
topics respectively, based on results obtained in previous research [35]. Param-
eters were optimized using an implementation of the variational expectation-370
maximization algorithm proposed in [33]. HR during activities composites
HR
ctx
was ranked according to different features T : amount of time spent in
each activity composite, relative amount of time spent in each activity primitive
for an activity composite, with respect to the total time spent in the same activity
primitive across all activities composites and relative time spent in each activity375
primitive per activity composite. These features were chosen since they can be
computed across participants and activities composites regardless of the partic-
ipant lifestyle or activity composite semantics. Ranked HR
ctx
were correlated
with CRF to determine which activities composites features were more repre-
sentative of CRF. Ranking of HR
ctx
values was smoothed by a moving average380
of 2 elements, i.e. over the first two ranked activity composites. The relevant
activity composites discovery procedure was also evaluated independently of the
participant. Contextualized HR HR
ctx
was ranked and correlated with CRF for
np1 participants. The feature resulting as the most representative of CRF, i.e.
the one for which correlation was maximized, was used to determine relevant385
activity composites for the left out participant. The procedure was repeated np
times, where np was the number of participants.
5.2. CRF estimation
Hierarchical Bayesian models for CRF estimation introduced in Sec. 3 were
implemented using R and JAGS. Posterior parameters estimations were per-390
formed by Gibbs sampling with 3 chains and 10000 iterations. The first 500
iterations were discarded (burn-in period). We consider reference V O
2
max as
CRF. We chose walking at different speeds as activity primitives normally car-
ried out by most of the population. We evaluated our V O
2
max estimation
models using as predictor HR contextualized over a broad range of walking395
speeds, from 2.5 to 6 km/h. The hierarchical Bayesian model to estimate CRF
also included the participant’s weight, age, sex and height as predictors. We im-
20
0.4
0.5
0.6
0.7
All 2.5 km/h 3 km/h 3.5 km/h 4 km/h 4.5 km/h 5 km/h 5.5 km/h 6 km/h
Walking speeds
Absolute correlation r
Context
No−context
Primitives
Composites
Figure 7: Correlation between HR and V O
2
max. Correlation is lowest for No-context and
were highest when activity composites (Composites) were used, compared to the condition
were only activity primitives (Primitives) were considered. HR data during activity primitives
and composites was acquired in free-living settings.
plemented the models listed in Sec. 4 for comparison, thus estimating V O
2
max
using anthropometric characteristics only (case anthropometrics), HR in free-
living (case no-context ), HR while walking at a certain speed (case primitives),400
and HR while walking at a certain speed relevant activity composites (case com-
posites).
6. Results
6.1. Activity primitives and walking speed
Activity primitives and walking speed were validated in laboratory settings.405
Class-normalized accuracy of the SVM-HMM activity recognition classifier was
95.8%. More specifically, accuracy was 98.2% for lying, 98.9% for sedentary,
83.5% for dynamic, 99.4% for walking, 96.5% for biking and 98.4% for running.
Walking speed estimation RMSE was 0.37 km/h.
21
6.2. Relevant activity composites410
Fig. 7 shows the absolute value of the correlation between HR and V O
2
max
for different contexts. HR in free-living was moderately correlated with V O
2
max
(comparison case no-context, r = 0.43). Correlation between HR and V O
2
max
in free-living was stronger for walking activity primitives, compared to no-
context, ranging from 0.55 to 0.63. Correlation had a tendency to increase415
as speed increased, reaching the highest value for walking at 6 km/h. Fig. 8
shows results for V O
2
max estimation models. RMSE between estimated and
predicted V O
2
max when no HR data was used (case anthropometrics) was 322.5
ml/min. The relation between contextualized HR HR
ctx
(i.e. including relevant
200
250
300
350
None All 2.5 km/h 3 km/h 3.5 km/h 4 km/h 4.5 km/h 5 km/h 5.5 km/h 6 km/h
Walking speeds
RMSE
model
Anth.
No−context
Primitives
Composites
Figure 8: RMSE of CRF estimation in free-living against V O
2
max reference. Error bars
represent standard error. RMSE is highest for Anth, followed by No-Context, showing that
not using HR data or using HR data without context produces larger errors in V O
2
max esti-
mation. A combination of activity primitives and activity composites (condition Composites)
shows optimal results, i.e. the lowest RMSE across different walking speeds, compared to the
condition were only activity primitives (Primitives) were considered. HR data used as predic-
tors was acquired during activity primitives and composites performed without supervision in
free-living settings.
22
activity composites) and V O
2
max was maximized ranking activities composites420
by feature T
i
= relative time spent sedentary within an activity composite. Cor-
relation ranged between 0.57 and 0.71, reaching the highest value for walking
at 6 km/h. Thus, correlation was consistently improved when a combination of
activity primitives and relevant activity composites was used to contextualized
HR, compared to no-context and activity primitives only, as shown in Fig. 7.425
R2=0.67
RMSE=321.7
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Estimated VO2max
Measured VO2max
a) Anthropometrics
R2=0.7
RMSE=285.6
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Estimated VO2max
Measured VO2max
b) No−context
R2=0.73
RMSE=267.4
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Estimated VO2max
Measured VO2max
c) Primitives
R2=0.76
RMSE=249.4
2000
2500
3000
3500
4000
2000 2500 3000 3500 4000
Estimated VO2max
Measured VO2max
d) Composites
Figure 9: Estimated and measured V O
2
max for the four conditions compared in this work.
R
2
is increased and RMSE is reduced when adding more levels of contexts. The best results
when V O
2
max is estimated using HR contextualized by activity primitives and composites,
as shown in d). HR data used as predictors was acquired during activity primitives and
composites performed without supervision in free-living settings.
6.3. CRF estimation
RMSE was reduced to 286.3 ml/min (11.3% error reduction) when includ-
ing free-living HR as predictor but no contextual information (case no-context).
Estimation error was further reduced for case primitives, i.e. using the HR
while walking at a certain speed as predictors. More specifically, RMSE var-430
ied between 287.3 and 267.6 ml/min, depending on walking speed. RMSE was
reduced by 17.0% and 6.5% compared to case anthropometrics and no-context
respectively, when the best model was used (i.e. walking at 6 km/h). Contex-
tualizing HR by a combination of activity primitives and activity composites
provided better accuracy than any other model. RMSE varied between 268.9435
ml/min and 249.5 ml/min, depending on walking speed. A combination of activ-
ity primitives and activity composites always outperformed activity primitives
alone, as shown in Fig. 8.
23
Activity primitives in free-living were recognized as follows: 44.5% lying,
36.4% sedentary, 9.5% dynamic, 5.4% walking, 3.8% biking and 0.4% running.440
The average walking speed in free-living over the entire dataset was 3.5 ± 1.5
km/h. Participants spent 71 ± 27 minutes per day in walking activities, 7 ± 5.4
minutes walking at 6 km/h.
Overall, combining activity primitives and activities composites provided
error reductions up to 22.6%, 12.8% and 10.3% compared to anthropometrics,445
no-context and primitives respectively. Fig. 9 shows estimated and measured
V O
2
max for the four models compared in this study. Explained variance (R
2
)
and RMSE are reported, showing increased R
2
and reduced error as more con-
text is included. For the latter figure, only the best performing models is shown
for cases primitives and composites.450
7. Discussion
Many methods have been developed to estimate V O
2
max using data col-
lected under supervised laboratory conditions or following strict protocols. How-
ever, to the best of our knowledge, this is the first work, which proposes to
use pattern recognition and artificial intelligence methods to determine activity455
primitives and activity composites as contextual information, and then inter-
pret HR data in free-living. We showed RMSE reductions of 22.6% compared
to estimates derived using anthropometric characteristics only, and RMSE re-
ductions up to 10.3% compared to estimates derived using activity primitives,
highlighting the benefit of our context recognition method.460
We hypothesized that the presence of a multitude of factors such as psycho-
logical stress, interactions with other people, etc. in free-living required a novel
approach over the prior estimation attempts used in laboratory settings. In par-
ticular, HR in free-living is not only affected by activity primitives - as shown
in the lab - but by both activity primitives and activity composites. Thus, de-465
veloping computational methods able to incorporate knowledge of contextual
information beyond activity primitives could potentially improve interpretation
24
of HR in free-living. Our results confirm the importance of activity composites
in free-living. RMSE was consistently reduced over a broad range of walking
speeds, as shown in Fig. 8. We translated the need for contextual informa-470
tion into a hierarchical model. In our previous work we introduced relevant
activity composites for energy expenditure estimation [36]. We established rel-
evant activity composites to relate discovered activity composites for which no
supervised information exists, to behaviour-related HR.
In this work, we discovered individual activity composites of each user with-475
out identifying them. Measuring the composite discovery performance was not
needed since we assessed the CRF estimation based on the LDA output. We
determined which activity composites were better suited for CRF estimation by
ranked them according to the correlation of selected features and HR. The ap-
proach simplified our study methodology as no activity composite annotations480
were required. However, discovered activity composite do not provide semantics
and comparison between participants is challenging. Typically, activity com-
posite of interest are isolated and further classified using supervised methods
[34, 35], thus requiring prior knowledge of the activity composites to discover,
effectively limiting the unsupervised nature of the method. Ranking allowed485
for comparison of activity composite specific features across participants, thus
making the approach unsupervised and generalizable to new participants.
We found a strong relation between the relative time spent sedentary in each
activity composite and CRF. A possible explanation for the relation between HR
contextualized by activity composites ranked by relative time spent sedentary490
in each activity composite and CRF is that activities composites in which peo-
ple spend most of their time sedentary are typically representative of a stable
physiological condition, which might be more representative of their CRF level.
On the contrary, short or infrequent activities might involve more stressful situ-
ations as well as more intermittent HR, causing cardiovascular responses which495
are not as reliable for HR interpretation [40]. An example of an activity com-
posite that maximizes the relative time spent sedentary is working at the office.
While most of the time while working at the office an individual is probably
25
sedentary, there can still be many periods of walking, that are therefore used
to contextualize HR. In such periods, HR might be less affected by for example500
carrying loads, effects of previously performed intense exercise, walking hills,
etc.) and therefore be more representative of CRF.
We relied on the inverse relation between HR at a certain workload and
V O
2
max, as often reported for laboratory protocols. However, by using a non-
nested hierarchical approach, where parameters varied based on the activities,505
we did not constrain the participant in performing specific activities or walk-
ing at predefined speeds. Instead, based on the participant’s preferred walking
speed in free-living, the optimal parameters were used. The activity primitives
chosen as free-living contexts were lying down and walking, for the following
reasons. First, we aimed at activities commonly performed by healthy individ-510
uals. Secondly, the inverse relation between HR at rest or sleeping HR and
CRF was already shown in previous research [41, 42], highlighting how this
parameter can be valuable for V O
2
max estimation. Finally, walking activities
can be discriminated in intensity, by detecting walking speed, using simply an
accelerometer. Determining the specific intensity of an activity is an important515
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). On our free-living dataset, partici-520
pants spent more than an hour per day walking (71 ± 27 minutes), and about
10% of walking activities involved walking at 6 km/h (7 ± 5.4 minutes). Thus,
walking confirmed to be a common activity of daily life, and a good candidate
to contextualize HR for CRF estimation. Noteworthy, RMSE for V O
2
max es-
timation was not consistently reduced by including in the models HR collected525
while walking at higher speeds. Thus, highlighting the additional complexity of
analyzing HR data in free-living.
Being able to accurately determine the user context in terms of activity type
and intensity allows us to bring the principle used in laboratory-based submax-
26
imal tests (i.e. the inverse relation between HR measured while performing an530
exercise at a certain intensity, such as biking at a fixed power on a cycle er-
gometer, and V O
2
max) to free-living settings. Contextualizing HR by means
of activity primitives and speed improved correlation between free-living HR
and CRF. Our approach builds on LDA and a hierarchical Bayesian model to
discover activity composites and relating V O
2
max to behavior in a probabilistic535
framework. As a result, RMSE for CRF estimation against VO2max reference
was reduced up to 22.6%.
The proposed CRF estimation model could be used to provide accurate
information about an individual’s health without the need for laboratory infras-
tructure or specific tests. Estimating CRF continuously in free-living creates a540
feedback loop from measurements to lifestyle. CRF estimates may provide the
basis for many adaptive applications supporting behavior change. Moreover,
CRF is not only important for fitness applications, but to health and patient
care too, as CRF has known associations to disease risk. Our investigation
showed that robust CRF estimation in free-living is feasible, thus confirming545
that the applications described above are realistic.
Acknowledgment
The authors would like to thank G. Plasqui, G. Schiavone, G. ten Velde and
S. Camps for support during data collection.
References550
[1] S. J. Marshall, E. Ramirez, Reducing sedentary behavior a new paradigm in
physical activity promotion, American Journal of Lifestyle Medicine 5 (6)
(2011) 518–530.
[2] M. Chan, D. Est`eve, J.-Y. Fourniols, C. Escriba, E. Campo, Smart wearable
systems: Current status and future challenges, Artificial intelligence in555
medicine 56 (3) (2012) 137–156.
27
[3] F. Buttussi, L. Chittaro, Mopet: A context-aware and user-adaptive wear-
able system for fitness training, Artificial Intelligence in Medicine 42 (2)
(2008) 153–163.
[4] W. H. Wu, A. A. Bui, M. A. Batalin, L. K. Au, J. D. Binney, W. J.560
Kaiser, Medic: Medical embedded device for individualized care, Artificial
Intelligence in Medicine 42 (2) (2008) 137–152.
[5] C. C. Bennett, K. Hauser, Artificial intelligence framework for simulating
clinical decision-making: A markov decision process approach, Artificial
intelligence in medicine 57 (1) (2013) 9–19.565
[6] C. V. Bouten, K. T. Koekkoek, M. Verduin, R. Kodde, J. D. Janssen, A
triaxial accelerometer and portable data processing unit for the assessment
of daily physical activity, IEEE Transactions on Biomedical Engineering
44 (3) (1997) 136–147.
[7] H. Vathsangam, A. Emken, E. T. Schroeder, D. Spruijt-Metz, G. S.570
Sukhatme, Determining energy expenditure from treadmill walking using
hip-worn inertial sensors: An experimental study, IEEE Transactions on
Biomedical Engineering 58 (10) (2011) 2804–2815.
[8] D.-c. Lee, E. G. Artero, X. Sui, S. N. Blair, Review: Mortality trends in the
general population: the importance of cardiorespiratory fitness, Journal of575
Psychopharmacology 24 (4 suppl) (2010) 27–35.
[9] G. Plasqui, K. R. Westerterp, Accelerometry and heart rate as a measure
of physical fitness: proof of concept., Medicine and science in sports and
exercise 37 (5) (2005) 872–876.
[10] F. Sartor, G. Vernillo, H. M. de Morree, A. G. Bonomi, A. La Torre, H.-P.580
Kubis, A. Veicsteinas, Estimation of maximal oxygen uptake via submaxi-
mal exercise testing in sports, clinical, and home settings, Sports medicine
43 (9) (2013) 865–873.
28
[11] L. Thorsen, E. Skovlund, S. B. Strømme, K. Hornslien, A. A. Dahl, S. D.
Foss˚a, Effectiveness of physical activity on cardiorespiratory fitness and585
health-related quality of life in young and middle-aged cancer patients
shortly after chemotherapy, Journal of Clinical Oncology 23 (10) (2005)
2378–2388.
[12] M. Wei, J. B. Kampert, C. E. Barlow, M. Z. Nichaman, L. W. Gibbons,
R. S. Paffenbarger Jr, S. N. Blair, Relationship between low cardiorespi-590
ratory fitness and mortality in normal-weight, overweight, and obese men,
Jama 282 (16) (1999) 1547–1553.
[13] S. N. Blair, J. B. Kampert, H. W. Kohl, C. E. Barlow, C. A. Macera, R. S.
Paffenbarger, L. W. Gibbons, Influences of cardiorespiratory fitness and
other precursors on cardiovascular disease and all-cause mortality in men595
and women, Jama 276 (3) (1996) 205–210.
[14] P. O.
˚
Astrand, I. Ryhming, A nomogram for calculation of aerobic capac-
ity (physical fitness) from pulse rate during submaximal work, Journal of
Applied Physiology 7 (2) (1954) 218–221.
[15] S. E. Crouter, D. R. Bassett, A refined 2-regression model for the actigraph600
accelerometer, Medicine & Science in Sports & Exercise 42 (5) (2010) 1029–
1037.
[16] L. Bao, S. Intille, Activity recognition from user-annotated acceleration
data, in: Pervasive ’04, Vol. 3001, 2004, pp. 1–17.
[17] A. G. Bonomi, Improving assessment of daily energy expenditure by iden-605
tifying types of physical activity with a single accelerometer., Journal of
Applied Physiology 107 (3) (2009) 655–661.
[18] E. Tapia, Using machine learning for real-time activity recognition and
estimation of energy expenditure, in: PhD thesis, MIT, 2008.
[19] M. Altini, J. Penders, O. Amft, Energy expenditure estimation using wear-610
able sensors: A new methodology for activity-specific models, in: Proceed-
29
ings of the Conference on Wireless Health, WH ’12, ACM, New York, NY,
USA, 2012, pp. 1:1–1:8.
[20] M. Altini, J. Penders, R. Vullers, O. Amft, Estimating energy expenditure
using body-worn accelerometers: a comparison of methods, sensors number615
and positioning, Biomedical and Health Informatics, IEEE Journal of 19 (1)
(2015) 219–226.
[21] M. Altini, P. Casale, J. Penders, O. Amft, Personalized cardiorespiratory
fitness and energy expenditure estimation using hierarchical bayesian mod-
els, Journal of biomedical informatics 56 (2015) 195–204.620
[22] L. Vanhees, J. Lefevre, R. Philippaerts, M. Martens, W. Huygens,
T. Troosters, G. Beunen, How to assess physical activity? how to as-
sess physical fitness?, European Journal of Cardiovascular Prevention &
Rehabilitation 12 (2) (2005) 102–114.
[23] V. Noonan, E. Dean, Submaximal exercise testing: clinical application and625
interpretation, Physical Therapy 80 (8) (2000) 782–807.
[24] A. S. Jackson, S. N. Blair, M. T. Mahar, L. T. Wier, R. M. Ross, J. E.
Stuteville, Prediction of functional aerobic capacity without exercise test-
ing., Medicine and Science in Sports and exercise 22 (6) (1990) 863–870.
[25] B. M. Nes, I. Janszky, L. J. Vatten, T. Nilsen, S. T. Aspenes, U. Wisløff,630
Estimating vo2peak from a nonexercise prediction model: the hunt study,
norway, Medicine and Science in Sports and Exercise 43 (11) (2011) 2024–
30.
[26] G. Plasqui, K. R. Westerterp, Accelerometry and heart rate as a measure
of physical fitness: cross-validation., Medicine and science in sports and635
exercise 38 (8) (2006) 1510–1514.
[27] M. R. Esco, E. M. Mugu, H. N. Williford, A. N. McHugh, B. E. Bloomquist,
Cross-validation of the polar fitness test via the polar f11 heart rate monitor
in predicting vo2max, Journal of Exercise Physiology 14 (2011) 31–37.
30
[28] S. Crumpton, H. N. Williford, S. O’Mailia, M. S. Olson, L. E. Woolen,640
Validity of the polar m52 heart rate monitor in predicting vo2max, Medicine
& Science in Sports & Exercise 35 (5) (2003) S193.
[29] J. R. Ruiz, J. Ramirez-Lechuga, F. B. Ortega, J. Castro-Pinero, J. M. Ben-
itez, A. Arauzo-Azofra, C. Sanchez, M. Sj¨ostr¨om, M. J. Castillo, A. Gutier-
rez, et al., Artificial neural network-based equation for estimating vo 2max645
from the 20m shuttle run test in adolescents, Artificial intelligence in
medicine 44 (3) (2008) 233–245.
[30] Z.-B. Cao, N. Miyatake, M. Higuchi, K. Ishikawa-Takata, M. Miyachi,
I. Tabata, Prediction of vo2max with daily step counts for japanese adult
women, European journal of applied physiology 105 (2) (2009) 289–296.650
[31] T. onis, K. Gorter, M. Vollenbroek-Hutten, H. Hermens, Comparing
vo2max determined by using the relation between heart rate and accelerom-
etry with submaximal estimated vo2max., The Journal of sports medicine
and physical fitness 52 (4) (2012) 337–343.
[32] O. Amft, C. Lombriser, T. Stiefmeier, G. Toster, Recognition of user ac-655
tivity sequences using distributed event detection, in: Smart Sensing and
Context, Springer, 2007, pp. 126–141.
[33] D. M. Blei, A. Y. Ng, M. I. Jordan, Latent dirichlet allocation, the Journal
of machine Learning research 3 (2003) 993–1022.
[34] T. Huynh, M. Fritz, B. Schiele, Discovery of activity patterns using topic660
models, in: Proceedings of the 10th international conference on Ubiquitous
computing, ACM, 2008, pp. 10–19.
[35] J. Seiter, O. Amft, M. Rossi, G. Toster, Discovery of activity composites
using topic models: An analysis of unsupervised methods, Pervasive and
Mobile Computing 15 (2014) 215–227.665
31
[36] M. Altini, P. Casale, J. Penders, O. Amft, Personalization of energy ex-
penditure estimation in free living using topic models, IEEE Journal of
Biomedical and Health Informatics 19 (5) (2015) 1577–1586.
[37] H. Kuipers, F. Verstappen, H. Keizer, P. Geurten, G. Van Kranenburg,
Variability of aerobic performance in the laboratory and its physiologic670
correlates, International journal of sports medicine 6 (04) (1985) 197–201.
[38] K. Westerterp, L. Wouters, L. W. van Marken, The maastricht protocol for
the measurement of body composition and energy expenditure with labeled
water., Obesity research 3 (1995) 49–57.
[39] Y. Zheng, L. Zhang, X. Xie, W.-Y. Ma, Mining interesting locations and675
travel sequences from gps trajectories, in: Proceedings of the 18th interna-
tional conference on World wide web, ACM, 2009, pp. 791–800.
[40] E. Redding, M. Wyon, J. Shearman, L. Doggart, Validity of using heart rate
as a predictor of oxygen consumption in dance, Journal of Dance Medicine
& Science 8 (3) (2004) 69–72.680
[41] A. Loimaala, H. Huikuri, P. Oja, M. Pasanen, I. Vuori, Controlled 5-mo aer-
obic training improves heart rate but not heart rate variability or baroreflex
sensitivity, Journal of Applied Physiology 89 (5) (2000) 1825–1829.
[42] N. Uth, H. Sørensen, K. Overgaard, P. K. Pedersen, Estimation of vo2max
from the ratio between hrmax and hrrest–the heart rate ratio method,685
European journal of applied physiology 91 (1) (2004) 111–115.
32
... Scaling fitness prediction. Despite some promising studies which attempt to infer VO 2 max from data collected during free-living conditions, these mostly stem from small-scale cohorts with less than 50 participants and use contextual data from treadmill activity, which again limits their application in real-world contexts (Altini et al., 2016). In this thesis, we employ data from the Fenland Study, the largest study of its kind, following more than 10,000 participants for a week and almost a decade later to assess the change of fitness. ...
... Due to the inability of algorithms back then to work directly on the raw pixels of an image (raw sensors in our case), researchers published inventive methods that were called feature descriptors. Seminal papers of that time like the Scale Invariant Feature Transform (SIFT) (Lowe, 2004), or the Histogram of Oriented Gradients (HOG) (Dalal and Triggs, 2005), are handcrafted algorithms that extract interest points from an image based on geometry 8 . The turning point was in 2012 when the Imagenet study (Krizhevsky et al., 2012) showed that better results are possible with deep learning that does not need all these extra hand-crafted features. ...
... Although certain commercial devices have shown stronger results than others, many tend to rely on detailed activity intensity measurements paired with speed monitoring through GPS and require users to reach heart rates that are close to their maximum capabilities, limiting the application to self-selecting, fitter individuals (Cooper and Shafer, 2019; Lucio et al., 2018). Despite some promising studies which attempt to infer VO 2 max from data collected during free-living conditions, these mostly stem from small-scale cohorts with less than 50 participants and use contextual data from treadmill activity, which again limits their application in real-world contexts (Altini et al., 2016). In this work, we use data from the largest study of it's kind, by over two orders of magnitude, and use purely free-living data to predict VO 2 max, with no requirement for context-awareness. ...
Thesis
The widespread adoption of smartphones and wearables has led to the accumulation of rich datasets, which could aid the understanding of behavior and health in unprecedented detail. At the same time, machine learning and specifically deep learning have reached impressive performance in a variety of prediction tasks, but their use on time-series data appears challenging. Existing models struggle to learn from this unique type of data due to noise, sparsity, long-tailed distributions of behaviors, lack of labels, and multimodality. This dissertation addresses these challenges by developing new models that leverage multi-task learning for accurate forecasting, multimodal fusion for improved population subtyping, and self-supervision for learning generalized representations. We apply our proposed methods to challenging real-world tasks of predicting mental health and cardio-respiratory fitness through sensor data. First, we study the relationship of passive data as collected from smartphones (movement and background audio) to momentary mood levels. Our new training pipeline, which combines different sensor data into a low-dimensional embedding and clusters longitudinal user trajectories as outcome, outperforms traditional approaches based solely on psychology questionnaires. Second, motivated by mood instability as a predictor of poor mental health, we propose encoder-decoder models for time-series forecasting which exploit the bi-modality of mood with multi-task learning. Next, motivated by the success of general-purpose models in vision and language tasks, we propose a self-supervised neural network ready-to-use as a feature extractor for wearable data. To this end, we set the heart rate responses as the supervisory signal for activity data, leveraging their underlying physiological relationship and show that the resulting task-agnostic embeddings can generalize in predicting structurally different downstream outcomes through transfer learning (e.g. BMI, age, energy expenditure), outperforming unsupervised autoencoders and biomarkers. Finally, acknowledging fitness as a strong predictor of overall health, which, however, can only be measured with expensive instruments (e.g., a VO2max test), we develop models that enable accurate prediction of fine-grained fitness levels with wearables in the present, and more importantly, its direction and magnitude almost a decade later. All proposed methods are evaluated on large longitudinal datasets with tens of thousands of participants in the wild. The models developed and the insights drawn in this dissertation provide evidence for a better understanding of high-dimensional behavioral and physiological data with implications for large-scale health and lifestyle monitoring.
... For a realistic data set, required for level 5 and progression to level 7 or 8, the observation period and recording duration were specifically important, as we found in 12 studies. Three studies used an observation period of 24 hours [23,32,64]; one for a week [17], one for 2 weeks [27], and one for 90 days [16]. Overall, 2 studies implied an observation period of months but did not explicitly report it [19,20]. ...
... One considered recordings of at least eight hours [19] and one reported an average recording duration of 11.3 hours [20]. Finally, only one [27] fully used the potential of wearables and reported a (near-) continuous recording duration. ...
... Another study predicted vascular age and 10-year cardiovascular disease risk [34]. The third assigned a cardiorespiratory fitness score [27]. Notably, only the first 2 studies constructed a prognostic model. ...
Preprint
Full-text available
BACKGROUND Wearable technology has the potential to improve cardiovascular health monitoring using machine learning. It enables remote health monitoring and allows for diagnosis and prevention. In addition to detection of cardiovascular disease, it can exclude a diagnosis in symptomatic patients, preventing unnecessary hospital visits. Furthermore, early warning systems can aid the cardiologist in timely treatment and prevention. OBJECTIVE We systematically assessed literature on detecting and predicting outcomes of cardiovascular disease with data obtained from wearables to gain insights in the current challenges and limitations. METHODS We searched PubMed, Scopus and IEEE Xplore on September 26, 2020 with no restrictions on the publication date using keywords: wearables, machine learning and cardiovascular disease. Methodologies were categorized and analyzed according to machine learning based technology readiness levels (TRLs) that score studies on their potential to be deployed in an operational setting from 1 to 9 (most ready). RESULTS After removal of duplicates, applying exclusion criteria, and full-text screening, 55 eligible studies remained for analysis, covering a variety of cardiovascular diseases. None of the studies were integrated into a health care system (TRL < 6), prospectively phase 2 and 3 trials were absent (TRL < 7 and 8) and group cross-validation was rarely used, limiting to demonstrate their effectiveness. Furthermore, there seems to be no agreement on the sample size needed to train these models, the size of the observation window used to make predictions, how long subjects should be observed and the type of machine learning model that is suitable for predicting cardiovascular outcomes. CONCLUSIONS Although current studies show the potential of wearables to monitor cardiovascular events, their deployment as a diagnostic and/or prognostic cardiovascular clinical tool is hampered by the lack of using a realistic dataset and a proper systematic and prospective evaluation.
... By narrowing the activity, e.g. walking at a given speed [3], or analysing selected exercises, a focused assessment condition could be realised and variability in movement parameters may be reduced. Since we extracting walking segments in free-living behaviour data, we did not require patients to perform specific test exercises. ...
... RMSE for increasing number of selected features(1,3,5). Plots show results for simulated rotations 0 • , 5 • , 10 • and 15 • estimated using the GLM. ...
Thesis
Patients after stroke often face long-term disability due to hemiparesis and thus require rehabilitation. With ageing societies, the stroke incidence is expected to increase, even among people who are in the workforce. Hence, costs for healthcare systems will rise. The current situation in stroke rehabilitation could intensify, more patients require treatment while at the same time a shortage of clinical personnel1 becomes apparent. Wearable motion sensors, including inertial measurement units (IMUs), have the potential to mitigate challenges in stroke rehabilitation and offer great potential for reshaping healthcare. With digital biomarkers derived from wearable sensors, e.g. to describe gait parameters or motion intensity, clinicians and patients could be supported during the rehabilitation. For example, objective movement quantification might help clinicians adapting therapies to the individual needs of a patient after stroke. State-of-the-art performance monitoring and evaluation is restricted to guided short-term measurements that follow defined assessment tasks in clinical environments, which are subjectively assessed by clinicians. Remote monitoring and evaluation of patients after stroke in free-living and the potential of wearable sensors is insufficiently addressed in research. Therefore, solutions for continuous and objective performance monitoring and evaluation using wearable motion sensors and algorithms are sought. The aim of this thesis is to devise and evaluate new solutions for longitudinal performance monitoring and evaluation in patients after stroke using algorithms and digital biomarkers, which could be used in freeliving. We test the following hypotheses: 1. IMUs, (machine learning) algorithms, and digital biomarkers are viable for longitudinal performance monitoring in patients after stroke. 2. Motion performance differences in the affected and less-affected upper and lower body-sides can be evaluated during therapies and free-living using IMU data and digital biomarkers. To test the hypotheses, a six month, longitudinal clinical observation study with eleven hemiparetic patients after stroke was implemented. In a novel study design, outpatients were followed by the examiner and more than 620 hours of motion data were recorded and annotated using a smartphone application. In full-day recordings, patients followed their therapy and performed various activities of daily living while wearing six body-worn IMUs. In addition, we used digital twins for personalised movement analyses in two case studies, including athletes and patients after stroke. This thesis includes eight peer-reviewed scientific publications, addressing four specific goals: (1) to review wearable motion sensors and machine learning algorithms for clinical assessment score estimation, (2) to implement activity primitive extraction algorithms for clinical score estimation and trend analysis, (3) to develop and evaluate digital biomarkers for performance analysis, and (4) to investigate digital twins for movement analysis and the evaluation of wearable sensor systems and algorithms. Wearable motion sensors and machine learning algorithms for clinical score estimation were reviewed. The review showed that mainly accelerometers for measurements were deployed and that score estimation algorithms included classification or regression-based machine learning techniques. Rule-based algorithms for activity primitive extraction from continuous sensor data were implemented. We showed that the Extended Barthel Index (EBI) can be estimated with approx. 12% relative error on average using support vector regression and leave-one-participant-out cross-validation. Further, the analysis of activity primitives revealed patient-specific recovery trends. The convergence point (CP), a newly developed digital biomarker for longitudinal, bilateral trend analysis revealed patient-specific recovery trends. In addition, the physical activity (PA) and functional range of motion (fROM) was analysed. The CP, PA, and fROM confirmed that differences in affected and lessaffected upper and lower body can be quantified during rehabilitation, including therapies and free-living. Finally, we present a novel methodology based on biomechanical simulations and motion data synthesis for the systematic evaluation of wearable sensor systems, algorithms, and digital biomarkers using personalised digital twins.
... Although certain wearable devices show promise, they tend to rely on detailed physical activity intensity measurements, GPS-based speed monitoring, and require users to reach near-maximal HR values, which limits their use to fitter individuals 9 . Some studies attempt to estimate VO 2 max from data collected during free-living conditions, but these are typically from small-scale cohorts and use contextual data from treadmill activity, which restricts their application in population settings 10 . ...
Preprint
Full-text available
Cardiorespiratory fitness is an established predictor of metabolic disease and mortality. Fitness is directly measured as maximal oxygen consumption (VO2max), or indirectly assessed using heart rate response to a standard exercise test. However, such testing is costly and burdensome, limiting its utility and scalability. Fitness can also be approximated using resting heart rate and self-reported exercise habits but with lower accuracy. Modern wearables capture dynamic heart rate data which, in combination with machine learning models, could improve fitness prediction. In this work, we analyze movement and heart rate signals from wearable sensors in free-living conditions from 11,059 participants who also underwent a standard exercise test, along with a longitudinal repeat cohort of 2,675 participants. We design algorithms and models that convert raw sensor data into cardio-respiratory fitness estimates, and validate these estimates' ability to capture fitness profiles in a longitudinal cohort over time while subjects engaged in real-world (non-exercise) behaviour. Additionally, we validate our methods with a third external cohort of 181 participants who underwent maximal VO2max testing, which is considered the gold standard measurement because it requires reaching one's maximum heart rate and exhaustion level. Our results show that the developed models yield a high correlation (r = 0.82, 95CI 0.80-0.83), when compared to the ground truth in a holdout sample. These models outperform conventional non-exercise fitness models and traditional bio-markers using measurements of normal daily living without the need for a specific exercise test. Additionally, we show the adaptability and applicability of this approach for detecting fitness change over time in the longitudinal subsample that repeated measurements after 7 years.
... For heart health, a chest wearable to monitor physical and cardiac activity was developed where it involves ECG, accelerometers, sensors for temperature, a microcontroller, and a Bluetooth [34]. Other research estimated cardio-respiratory fitness via designing ML algorithms to explore data from heart rate, accelerometers, and Global Positioning System (GPS) [35]. For cognitive disorders, an intelligent ring with EDA, heart rate, accelerometers for movement, and sensors for dermal temperature was designed for monitoring the sympathetic nervous systems [36]. ...
Chapter
Full-text available
Conventional health systems are aiming to diagnose first and treat diseases afterward. The growing precision health seeks early diagnosing and customized therapy for a health condition, by ongoing tracking of individual health and well‐being. Portable and implantable devices can dynamically track and evaluate health, supplying a broad view of human living and health. Growing techniques are promising in powering the effect of precision health. For instance, the emerging monitoring techniques would continuously gather data and engage the customer. Novel techniques are steadily developing, hence the need for improved insights into the scenery of the current and future techniques which could gather data for precision health. This chapter outlines the available and growing tracking and sensing techniques for precision health, focusing especially on the highly needed techniques like portable and mobile devices, implantable sensors, and wearables.
... Recent research studies on VO 2 max prediction have incorporated the use of technology and have gone as far as estimating VO 2 max without the need to perform predefined protocols. Cao et al. used objectively measured physical activity levels to predict VO 2 max, and Altini et al. used wearable sensors to identify context-specific intensity during 'freeliving' and used HR to predict VO 2 max [20,21]. Although such approaches are novel and offer easy means to predict VO 2 max, these methods are still being tested and require specialized equipment, making it difficult for the general population to use. ...
Article
The aim of the study was to develop a simple submaximal walk test protocol and equation using heart rate (HR) response variables to predict maximal oxygen consumption (VO2max). A total of 60 healthy adults were recruited to test the validity of 3 min walk tests (3MWT). VO2max and HR responses during the 3MWTs were measured. Multiple regression analysis was used to develop prediction equations. As a result, HR response variables including resting HR and HR during walking and recovery at two different cadences were significantly correlated with VO2max. The equations developed using multiple regression analyses were able to predict VO2max values (r = 0.75-0.84; r2 = 0.57-0.70; standard error of estimate (SEE) = 4.80-5.25 mL/kg/min). The equation that predicted VO2max the best was at the cadence of 120 steps per minute, which included sex; age; height; weight; body mass index; resting HR; HR at 1 min, 2 min and 3 min; HR recovery at 1 min and 2 min; and other HR variables calculated based on these measured HR variables (r = 0.84; r2 = 0.70; SEE = 4.80 mL/kg/min). In conclusion, the 3MWT developed in this study is a safe and practical submaximal exercise protocol for healthy adults to predict VO2max accurately, even compared to the well-established submaximal exercise protocols, and merits further investigation.
... 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. ...
Article
Background: Wearable technology has the potential to improve cardiovascular health monitoring by using machine learning. Such technology enables remote health monitoring and allows for the diagnosis and prevention of cardiovascular diseases. In addition to the detection of cardiovascular disease, it can exclude this diagnosis in symptomatic patients, thereby preventing unnecessary hospital visits. In addition, early warning systems can aid cardiologists in timely treatment and prevention. Objective: This study aims to systematically assess the literature on detecting and predicting outcomes of patients with cardiovascular diseases by using machine learning with data obtained from wearables to gain insights into the current state, challenges, and limitations of this technology. Methods: We searched PubMed, Scopus, and IEEE Xplore on September 26, 2020, with no restrictions on the publication date and by using keywords such as "wearables," "machine learning," and "cardiovascular disease." Methodologies were categorized and analyzed according to machine learning-based technology readiness levels (TRLs), which score studies on their potential to be deployed in an operational setting from 1 to 9 (most ready). Results: After the removal of duplicates, application of exclusion criteria, and full-text screening, 55 eligible studies were included in the analysis, covering a variety of cardiovascular diseases. We assessed the quality of the included studies and found that none of the studies were integrated into a health care system (TRL<6), prospective phase 2 and phase 3 trials were absent (TRL<7 and 8), and group cross-validation was rarely used. These issues limited these studies' ability to demonstrate the effectiveness of their methodologies. Furthermore, there seemed to be no agreement on the sample size needed to train these studies' models, the size of the observation window used to make predictions, how long participants should be observed, and the type of machine learning model that is suitable for predicting cardiovascular outcomes. Conclusions: Although current studies show the potential of wearables to monitor cardiovascular events, their deployment as a diagnostic or prognostic cardiovascular clinical tool is hampered by the lack of a realistic data set and proper systematic and prospective evaluation.
Article
Full-text available
We introduce an approach to personalize energy expenditure (EE) estimates in free living. First we use Topic Models (TM) to discover activity composites from recognized activity primitives and stay regions in daily living data. Subsequently, we determine activity composites that are relevant to contextualize heart rate (HR). Activity composites were ranked and analyzed to optimize the correlation to HR normalization parameters. Finally, individual-specific HR normalization parameters were used to normalize HR. Normalized HR was then included in activityspecific regression models to estimate EE. Our HR normalization minimizes the effect of individual fitness differences from entering in EE regression models. By estimating HR normalization parameters in free living, our approach avoids dedicated individual calibration or laboratory tests. In a combined free-living and laboratory study dataset, including 34 healthy volunteers, we show that HR normalization in 14-day free living data improves accuracy compared to no normalization and normalization based on activity primitives only (29:4% and 19:8% error reduction against lab reference). Based on acceleration and HR, both recorded from a necklace, and GPS acquired from a smartphone, EE estimation error was reduced by 10:7% in a leave-oneparticipant- out analysis.
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.
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.
Article
Full-text available
There is growing interest in the role of sedentary behavior as a risk factor for poor health, independent of physical activity (PA). To guide the spectrum of descriptive, analytic, and intervention studies on sedentary behavior, the authors advocate a behavioral epidemiology framework. This 5-phase framework is useful because it outlines a series of sequential stages important for developing, evaluating, and diffusing interventions to reduce sedentary behavior and improve population health. Studies of sedentary behavior and health outcomes (phase I) have found consistent evidence that excessive use of screen-based media is linked to overweight and obesity in children, and there is some evidence among adults that overall sedentary time is associated with risk factors for cardiometabolic disease, some cancers, and mortality. Biological mechanisms to explain possible relationships have started to emerge but are mostly based on animal models. Obtaining valid and reliable measurements of sedentary behavior (phase II) remains a research priority because self-reports are prone to recall bias, and it appears that sedentary habits do not appear to be well represented by measures of individual behaviors such as TV viewing. Studies have identified few modifiable correlates of sedentary behavior (phase III), although research appears to be limited to studies of TV viewing or to scenarios in which sedentary behavior is defined as an absence of PA. Rigorous intervention research (phase IV) has focused almost exclusively on reducing self-reported TV viewing among children and adolescents, and there is consistent evidence that these interventions are efficacious. There appear to be no interventions focused exclusively on reducing sedentary behavior of adults. Translation studies (phase V) are absent because the underlying evidence is still emerging. Future research should focus on examining causal associations between sedentary behavior and health, developing objective measures of domain-specific sitting time, and identifying modifiable correlates of sedentary behavior that can be used as leverage points for behavioral interventions.
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
Accurate estimation of Energy Expenditure (EE) and cardiorespiratory fitness (CRF) is a key element in determining the causal relation between aspects of human behavior related to physical activity and health. In this paper we estimate CRF without requiring laboratory protocols and personalize energy expenditure (EE) estimation models that rely on heart rate data, using CRF. CRF influences the relation between heart rate and EE. Thus, EE estimation based on heart rate typically requires individual calibration. Our modeling technique relies on a hierarchical approach using Bayesian modeling for both CRF and EE estimation models. By including CRF level in a hierarchical Bayesian model, we avoid the need for individual calibration or explicit heart rate normalization since CRF accounts for the different relation between heart rate and EE in different individuals. Our method first estimates CRF level from heart rate during low intensity activities of daily living, showing that CRF can be determined without specific protocols. Reference VO2max and EE were collected on a sample of 32 participants with varying CRF level. CRF estimation error could be reduced up to 27.0% compared to other models. Secondly, we show that including CRF as a group level predictor in a hierarchical model for EE estimation accounts for the relation between CRF, heart rate and EE. Thus, reducing EE estimation error by 18.2% on average. Our results provide evidence that hierarchical modeling is a promising technique for generalized CRF estimation from activities of daily living and personalized EE estimation. Copyright © 2015. Published by Elsevier Inc.
Article
In this work we investigate unsupervised activity discovery approaches using three topic model (TM) approaches, based on Latent Dirichlet Allocation (LDA), n-gram TM (NTM), and correlated TM (CTM). While LDA structures activity primitives, NTM adds primitive sequence information, and CTM exploits co-occurring topics. We use an activity composite/primitive abstraction and analyze three public datasets with different properties that affect the discovery, including primitive rate, activity composite specificity, primitive sequence similarity, and composite-instance ratio. We compare the activity composite discovery performance among the TM approaches and against a baseline using kk-means clustering. We provide guidelines for method and optimal TM parameter selection, depending on data properties and activity primitive noise. Results indicate that TMs can outperform kk-means clustering up to 17%, when composite specificity is low. LDA-based TMs showed higher robustness against noise compared to other TMs and kk-means.
Article
We describe latent Dirichlet allocation (LDA), a generative probabilistic model for collections of discrete data such as text corpora. LDA is a three-level hierarchical Bayesian model, in which each item of a collection is modeled as a finite mixture over an underlying set of topics. Each topic is, in turn, modeled as an infinite mixture over an underlying set of topic probabilities. In the context of text modeling, the topic probabilities provide an explicit representation of a document. We present efficient approximate inference techniques based on variational methods and an EM algorithm for empirical Bayes parameter estimation. We report results in document modeling, text classification, and collaborative filtering, comparing to a mixture of unigrams model and the probabilistic LSI model.