Content uploaded by Marco Altini

Author content

All content in this area was uploaded by Marco Altini on Jun 01, 2015

Content may be subject to copyright.

1

Estimating Oxygen Uptake during Non-Steady-State

Activities and Transitions Using Wearable Sensors

Marco Altini

1

, Julien Penders

2

and Oliver Amft

3

Abstract—In this paper, we present a method to estimate

oxygen uptake (VO

2

) during daily life activities and transitions

between them. First, we automatically locate transitions between

activities and periods of non-steady-state VO

2

. Subsequently, we

propose and compare activity-speciﬁc linear functions to model

steady-state activities and transition-speciﬁc non-linear functions

to model non-steady-state activities and transitions. We evaluate

our approach in study data from 22 participants that wore a

combined accelerometer and heart rate (HR) sensor while per-

forming a wide range of activities (clustered into lying, sedentary,

dynamic/household, walking, biking and running), including many

transitions between intensities, thus resulting in non-steady-

state VO

2

. Indirect calorimetry was used in parallel to obtain

VO

2

reference. VO

2

estimation error during transitions between

sedentary, household and walking activities could be reduced by

16% on average using the proposed approach, compared to state

of the art methods.

Index Terms—Accelerometers, Energy Expenditure, Heart

Rate, Non-Steady-State, VO

2

I. INTRODUCTION

Ubiquitous sensing technologies that objectively and non-

invasively monitor human behavior, started to provide insights

into the relation between physical activity (PA) and health.

Among the parameters used to objectively quantify PA, energy

expenditure (EE) is the most commonly used single metric

[1], [2], [3]. The measurement of steady-state oxygen uptake

(VO

2

) is considered to be the gold standard for estimating

EE during light to moderate steady-state exercise [4], [5],

where aerobic pathways are predominant. In this context, VO

2

measurements are proportional to metabolic heat production

[4]. Due to the practical limitations of measuring VO

2

in free

living, different methods to estimate VO

2

using miniaturized

wearable sensors have been developed in the past.

VO

2

estimation methods are typically based on activity-

speciﬁc linear regressions developed using steady-state data

and therefore can describe the variations within each modeled

activity during steady-state. However, transitions to activities

with other VO

2

levels cannot be accurately estimated, since

VO

2

dynamics differ during steady-state and non-steady-

state periods. Proper modeling of non-steady-state transitions

including transitions of activity types (e.g. sitting to walking)

and activity intensities (e.g. walking at different speed) are

necessary in order to provide accurate VO

2

estimation in free

1

M. Altini is with Bloom Technologies, Agoralaan Building Abis 2.13,

3590 Diepenbeek, Belgium, and Eindhoven University of Technology, Eind-

hoven, The Netherlands (e-mail: marco.altini@imec-nl.nl)

2

J. Penders is with Holst Centre/imec The Netherlands, Eindhoven, Nether-

lands, (e-mail: julien.penders@imec-nl.nl)

3

O. Amft is with University of Passau, Germany, and Eindhoven University

of Technology, The Netherlands (e-mail: amft@ieee.org)

living conditions. Accurate identiﬁcation of non-steady-state

periods as well as quantiﬁcations of VO

2

during transitions

could improve EE estimation because during non-steady-state

periods total EE is composed of aerobic and anaerobic compo-

nents. [4]. Earlier work on VO

2

transitions analysis focused

mainly on Post-Exercise Oxygen Consumption (EPOC) [6],

or VO

2

estimation for single activities [7], [8]. However

non-steady-state VO

2

is very frequent during varying low

intensities activities of daily living (ADLs). Studies showed

that most activities performed in free living last shorter than

the time needed to reach steady state. For example, 60% of

all walking bouts last shorter than 30 seconds [9]. Identifying

non-steady-state VO

2

can provide more insights on the aerobic

and anaerobic dynamics during the onset of exercise as well as

during ADLs characterized by short duration. While solutions

have been proposed to model non-steady-state VO

2

for spe-

ciﬁc activities [7], [8] or transitions [6], a uniﬁed approach able

to continuously estimate VO

2

is missing. Thus, we propose

a novel VO

2

estimation method, which combines activity-

speciﬁc VO

2

estimation using linear regression, with non-

steady-state detection and transition-speciﬁc VO

2

estimation

using non-linear equations.

This paper provides the following contributions:

1) We introduce a method to automatically locate periods

of non-steady-state VO

2

by analyzing the coefﬁcient

of variation (CV) of the predicted VO

2

. Using the

CV allows for detection of transition both between and

within activities. Then, we compare linear, exponential

and logistic transfer functions to model non-steady-state

VO

2

during individual transition types.

2) We evaluate the proposed approach on a dataset ac-

quired from 22 participants performing a wide set of

physical activities, including many transitions between

activities and changes of intensity within activities. We

show that the transition-speciﬁc modeling could reduce

VO

2

estimation error by 16% during activity transitions,

compared to state of the art methods.

II. RELATED WORK

Accelerometer and HR monitors are the most commonly

used single sensor devices in epidemiological studies. Ac-

celerometers use features representative of whole body motion,

as independent variables in the linear regression model devel-

oped to predict EE. However there are limitations due to the

inability of a single linear model to ﬁt all activities, since the

slope and intercept of the regression model change based on

the activity performed while data is collected [3]. On the other

2

hand, the high correlation between HR and EE within one

individual changes substantially between individuals [1]. The

latest EE estimation algorithms extended approaches based

on simple linear regression models performing activity recog-

nition over a predeﬁned set of activities, and then applying

different methods - typically regression models - to predict

EE, based on the activity [2], [3]. The regression models use

accelerometer features and anthropometric characteristics as

independent variables. Some authors included HR features as

well in the activity-speciﬁc linear models, showing consistent

improvements in EE estimation accuracy compared to algo-

rithms using accelerometer only features [1]. However, none

of these models explicitly models non-steady-state VO

2

.

A. Non-Steady-State VO

2

Estimation

Non-steady-state EE is often deﬁned as periods of time

where VO

2

and carbon dioxide production vary by more

than 5-10% [10]. Previous research on non-steady-state VO

2

focused mainly on metabolic responses to exercise (EPOC,

[6]). However, VO

2

estimation is intrinsically a temporal

problem and non-steady-state VO

2

is very common during

low intensity ADLs as well. While VO

2

increases rapidly,

normally reaching steady-state within 1-4 minutes, most ac-

tivities performed in free living last less than the time required

to reach steady state [9]. Most EE models are developed

deriving EE from VO

2

, and discarding the ﬁrst 1 or 2

minutes of data [1], [3], to isolate steady-state. Thus, the

predictions of these models will be negatively affected by the

real life nature of ADL activities. Other models incorporate

non-steady-state VO

2

but without providing details on the

models accuracy during transitions [11], thus limiting our

understanding of the models performance in non-steady-state

conditions. Using physiological data to predict VO

2

involves

slower dynamics present in both physiological changes (e.g.

HR slowing increasing) and aerobic pathways (VO

2

reaching

steady-state). However, these dynamics are different, typically

with HR being much slower than VO

2

[12], [13].

Few attempts to model non-steady-state VO

2

are found in

literature [14], [15], [16], [7], [8]. In [14], [15], the proposed

system relies on HR data only, suffering from all limitations

of non-activity-speciﬁc models, while [7], [8] analyzed walk-

ing data only. In [16], the authors used mono-exponential

functions to better capture the relation between movement

and EE during transitions between activities. However, the

EE prediction ignores the fact that during non-steady-state

EE cannot be derived from VO

2

alone, since total EE is

composed of both aerobic (estimated via VO

2

) and anaerobic

components. Finally, one single mono-exponential equation

might not be sufﬁcient to model energy deﬁcit and energy debt

situations between different activities [5]. While solutions have

been proposed to model non-steady-state VO

2

for speciﬁc

activities [7], [8] or transitions [6], a uniﬁed approach able

to continuously estimate VO

2

is missing.

III. ANA LY S IS A N D ESTIMATION APPROACH

This section describes the problem of non-steady-state VO

2

estimation and our approach to such problem.

5.0

7.5

10.0

12.5

15.0

17.5

0 20 40 60

VO2 (ml/kg/min)

State

steady−state

non−steady−state

a)

5

10

15

0 20 40 60

time windows (4 seconds each)

VO2 (ml/kg/min)

Activity Type

sedentary

dynamic

walking

b)

Fig. 1. Example of non-steady-state VO

2

. a) Non-steady state reference

VO

2

. b) Predicted VO

2

as estimated by steady-state activity-speciﬁc models.

In this work, non-steady state modelling is used combined with classic steady-

state models to better estimate actual VO2.

Our approach combines detection of non-steady-state VO

2

,

activity-speciﬁc linear models and non-steady state transition-

speciﬁc non-linear functions. Fig. 1 shows typical non-steady

state dynamics in activities of daily living, such as transitions

between sedentary activities (e.g. sitting or standing), walking

and then sedentary again. Fig. 1.b shows color-coded activities

as detected by an activity recognition system, while Fig. 1.a

shows steady-state and non-steady-state-data. Non-steady-state

VO

2

is present when the participant starts walking and when

the participant stops walking. Fig. 1.a show measured VO

2

while Fig. 1.b shows VO

2

as predicted by state of the art

activity speciﬁc models. Prediction models (Fig. 1.b) jump

between one value to the other as soon as a new activity starts,

while transitions shown in Fig. 1.a are much slower. Thus, a

different modeling technique is required during non-steady-

state.

A. Estimation Architecture

VO

2

predictions are generated by a sequence of steady-state

and non-steady-state models (see, e.g. Fig. 2):

...W

1

,V

1

,W

2

,V

2

,W

1

,V

3

...

where W

i

are activity-speciﬁc models used when steady-

state is detected and V

i

are transition-speciﬁc non-steady-state

models used when non-steady-state is detected. Each state W

i

or V

i

comprises t predictions based on the time spent in a

speciﬁc activity or transition duration:

W

i

= {VO2

Wi

SS

1

,...,VO2

Wi

SS

t

}

V

i

= {VO2

Vi

NS

1

,...,VO2

Vi

NS

t

}

In the following sections, we describe W

i

, V

i

and model

selection.

1) Activity-Speciﬁc Steady-State Models - W

i

: W

i

are com-

posed of two parts: activity recognition and activity-speciﬁc

multiple linear regression equations. Assuming n clusters of

activities:

C = {c

1

,...,c

n

}, 8c

i

2 C, 9 W

i

3

Activity

Recognition

State

1

model

State

n

model

VO2

ss

Activity-specific VO

2

models

Transition

1

model

Transition

n

model

Transition-specific VO

2

models

Model

Selection

VO2

ns

VO2

Physiological

Features

Anthropometric

Features

activity

class

c

i

W

1

W

n

V

1

V

n

VO2

ss

Accelerometer

Features

Fig. 2. Block diagram of out approach using activity-speciﬁc linear models

and transition-speciﬁc non-linear functions to estimate VO2. Features are used

for activity recognition and activity-speciﬁc VO

2

models. Activity recognition

is used to select the proper activity-speciﬁc or transition-speciﬁc model.

where the parameters of the activity-speciﬁc model W

i

are

derived from VO

2

values for a speciﬁc cluster of activities c

i

:

VO2

Wi

SS

= X

act

i

act

i

+ ✏ (1)

is the vector of regression coefﬁcients, and X

act

i

is the

vector of input features. Features are grouped into accelerom-

eter, physiological and anthropometric features (see Fig. 2).

2) Transition-Speciﬁc Non-Steady-State models - V

i

: We

compare linear, exponential and logistic functions as V

i

.

Assuming r types of transitions:

Tr = {tr

1

,...,tr

r

}, 8tr

i

2 Tr, 9 V

i

Linear functions: for a transition type V

i

from steady-state

W

j

to steady-state W

k

, a linear model is derived as follows:

VO2

Vi

NS

= ↵ + t + VO2

Wj

SS

+ ✏ (2)

where VO2

Vi

NS

are target VO

2

values during non-steady-state

for the transition V

i

, deﬁned by transition-speciﬁc linear func-

tions. The predictor t is the time elapsed since the transition

started. VO2

Wj

SS

is the VO

2

of the steady-state before the

transition (W

j

). Parameters ↵ and are the slope and intercept

of the linear model.

Logistic functions: for a transition type V

i

, a non-linear

model is derived as follows:

VO2

Vi

NS

=

✓

1

1+exp

(✓

Vi

2

+✓

Vi

3

t)

+ VO2

Wj

SS

+ ✏ (3)

where VO2

Vi

NS

are target VO

2

values during non-steady-state

for the transition V

i

, deﬁned by transition-speciﬁc logistic

functions. The predictor t is the time elapsed since the tran-

sition started. VO2

Wj

SS

is the VO

2

of the steady-state before

the transition (W

j

). Transitions between activities of different

type and intensity will have different dynamics [5], [17], thus

require transition-speciﬁc functions. The parameter ✓

1

is the

asymptote of the logistic function, and can be automatically

derived as the difference between the steady-state VO

2

before

(W

j

) and after (W

k

) the transition:

✓

1

= VO2

Wk

SS

VO2

Wj

SS

. (4)

Parameters ✓

Vi

2

and ✓

Vi

3

control the shape of the logistic curve

(e.g the transition speed), and were determined by ﬁtting

Recognized*ac,vity*

sedentary* walking* sedentary*

0

25

50

75

0 20 40 60

CV (%)

State

steady−state

non−steady−state

Time*windows*(4*s)*

Fig. 3. Example waveform of the CV of the estimated VO

2

for a sequence

of sedentary, walking and sedentary activities. State is non-steady when

CV > 15%. The CV captures transitions between activities as well as within

activities, thus detecting non-steady-state VO

2

.

for each transition type. Thus, ✓

Vi

2

and ✓

Vi

3

are transition-

speciﬁc. The transition type (i.e. V

i

) depends on the activities

recognized by the activity recognition system during steady-

states W

j

and W

k

.

Exponential functions: a transition model for a transition

type V

i

is derived as follows:

VO2

Vi

NS

= ✓

1

exp

↵+t

+ VO2

Wj

SS

+ ✏ (5)

where VO2

Vi

NS

are target VO

2

values for the transition V

i

,

deﬁned by transition-speciﬁc exponential functions. The pre-

dictor t is the time elapsed since the transition started. VO2

Wj

SS

is the VO

2

of the steady-state before the transition (W

j

). The

parameter ✓

1

is the same as for the logistic function, indicating

the asymptote of the exponential, and can be automatically

derived as shown in Eq. 4.

3) Model Selection: Non-steady-state is detected using the

CV of the predicted VO

2

over the last minute of data,

according to a preselected threshold:

CV

VO2

SS

=

VO2

SS

µ

VO2

SS

(6)

where

VO2

SS

and µ

VO2

SS

are the standard deviation and

mean of a 60 s sliding window of predicted VO

2

values.

Non-steady-state is detected if CV

VO2

SS

>CV

T hres

, where

CV

T hres

was set to 15%. Fig. 3 shows the CV for the predicted

VO

2

data. CV can be used to locate non-steady-states, such

as when activity intensity changes, thus even when there is no

transition between two activities.

IV. EVA L UAT I O N STUDY

A. Participants and Data Acquisition

Participants were 22 (17 male, 5 female), mean age

30.4 ± 5.7 years, mean weight 71.2 ± 12. 4 kg, mean height

1.76 ± 0.09 m, mean BMI 23.0 ± 2.8 kg/m

2

. Imec’s IRB

approved the study. Each participant signed an informed con-

sent form. The sensor platform used was the ECG Necklace, a

wearable sensor acquiring one lead ECG data at 256 Hz, and

three-axial accelerometer data at 32 Hz. Activity type was

annotated manually by experimenter. Breath-by-breath data

were collected using the Cosmed K4b

2

indirect calorimeter.

We interpolated calorimeter data at 0.25 Hz to align it with

the ECG Necklace data, and applied a moving average with

window size 4 elements to reduce high frequency noise.

4

B. Experiment Design

Participants reported at the lab after refraining from drink-

ing, eating and smoking in the two hours before the experi-

ment. Activities were grouped into six clusters to be used for

activity classiﬁcation. The six clusters were lying (lying down),

sedentary (sitting, standing, desk work, reading, writing, PC

work), dynamic (stacking groceries, washing dishes, cleaning,

sweeping, vacuuming), walking (treadmill ﬂat at 3,4,5,6 km/h,

inclined 3-5%, 3-5 km/h), biking (cycle ergometer, low medium

and high resistance level), running (treadmill 7,8,9, and 10

km/h). Activities were carried out for a period of at least 4

minutes, with the exception of running (1 to 4 minutes). All

transitions were manually annotated.

C. Statistics and Performance Measure

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 perform feature selec-

tion, activity recognition, activity-speciﬁc VO

2

estimation and

transition-speciﬁc VO

2

estimation models. The data from the

remaining participant was used for validation. This procedure

was repeated n (n = number of participants) times, and re-

sults were averaged. All parameters used in transition-speciﬁc

functions were determined in the same way, no data used for

model building was used for model evaluation. Performance

of the activity recognition models was evaluated using the

class-normalized accuracy. Results for VO

2

estimates are

reported in terms of Root-mean-square error (RMSE), where

the outcome variable was VO

2

in ml/kg/min. Paired t-tests

were used to compare RMSE between models.

V. I MPLEMENTATION

A. Features Extraction and Selection

Features extracted from the sensors’ raw data were used

to derive activity recognition and VO

2

estimation models.

Activity recognition was performed to classify the six activity

clusters introduced in Section IV-B. Accelerometer data were

segmented in 4 s windows, band-pass ﬁltered between 0.1

and 10 Hz, to isolate the dynamic component caused by

body motion, and low-pass ﬁltered at 1 Hz, to isolate the

static component, due to gravity. Feature selection for activity

type recognition was based on mutual information. The ﬁnal

feature set included: mean of the absolute signal, inter-quartile

range, median, variance, main frequency peak, low and high

frequency band signal power. Feature selection for VO

2

estimation was based on how much variation in VO

2

each

feature could explain within one cluster. The process was

automated using linear forward selection.

B. Activity Recognition

We selected a time window of 4 s, which is short enough

to detect short breaks in sedentary time, and long enough to

capture the repetitive patterns of some activities (e.g. walking

or running). Given the positive results in past research on

activity recognition, we selected Support Vector Machines

(SVMs) as classiﬁers. For the SVMs, we used a polynomial

kernel with degree 5 ( = 10, C = 1).

C. Activity-Speciﬁc Steady-State Models - W

i

Within one activity cluster, VO

2

can be estimated using

features representative of VO

2

changes within the activity

cluster [3], [1]. We used the mean of the absolute signal

to model changes in intensity within an activity, together

with HR. Anthropometrics features (body weight and resting

metabolic rate (RMR), estimated with the Harris-Benedict

formula) were added depending on the activity cluster.

D. Non-Steady-State Detection

The CV over one minute windows was computed to locate

non-steady-state segments of data. When the CV was higher

than CV

T hres

, a new non-steady-state transition was detected.

CV

Thres

was derived from previous literature on detection

of non-steady-state VO

2

, and empirically cross-validated on

our dataset, since no deﬁnition of non-steady-state is widely

accepted [10]. Once a non-steady-state transition V

i

was

detected between two steady-states W

j

and W

k

, the system

compared the steady-state VO

2

levels and enables the non-

steady-state models only if the difference in VO

2

was greater

than a threshold. This ﬁlter avoids activating the non-steady-

state models for transitions that are too small or short.

E. Transition-Speciﬁc Non-Steady-State models - V

i

Transition-speciﬁc non-steady-state models V

i

were devel-

oped for the most common transitions. More speciﬁcally,

logistic functions were derived by ﬁtting the parameters ✓

2

and ✓

3

while linear and exponential functions were derived by

ﬁtting the parameters ↵ and of the respective models, for

the following transitions: sedentary to walking (SW), walking

to sedentary (WS), sedentary to dynamic/household (SD),

dynamic/household to sedentary (DS), walking fast to walking

slow (WWDOWN), walking slow to walking fast (WWUP).

Data used for model development was not used for validation.

VI. RESULTS

We report results for activity recognition, activity-speciﬁc

steady-state models, non-steady-state detection and transition-

speciﬁc non-steady-state models, together with summary

statistics on the speciﬁc transitions considered in this work.

An example of the proposed method applied to a transition

between sedentary behavior and walking is shown in Fig. 4.

5

10

15

0 20 40 60 80

Samples

VO2 − ml/kg/min

type

Act−Spec ACC−HR

Act−Spec ACC−HR + Logistic

referenceVO2

ac#vity(

sedentary(

walking(

sedentary(

Fig. 4. Example of the results obtained when combining activity-speciﬁc

steady state models and transition-speciﬁc non-stead-state models (Act-Spec

ACC-HR + Logistic), as proposed by our method, for a transition between

sedentary to walking and walking to sedentary. Act-Spec ACC-HR show the

inability of steady-state models to predict VO

2

accurately during transitions.

5

Fig. 5. RMSE for VO

2

estimation using different sensing modalities and

methods for steady-state data. Activity-speciﬁc (Act-Spec) models outperform

all others, combined accelerometer and heart rate (ACC-HR) data improves

performance compared to accelerometer (ACC) or heart rate (HR) only.

A. Activity Recognition

Activity recognition accuracy was 94% on average across

all participants using the validation procedure detailed in Sec.

IV-C. More speciﬁcally, accuracy was 100% for lying down,

96% for sedentary, 81% for dynamic/household, 99% for

walking, 91% for biking and 98% for running.

B. Activity-Speciﬁc Steady-State Models

RMSE for activity-speciﬁc steady-state VO

2

estimation

models using combined accelerometer and HR data was 3.5

ml/kg/min. More speciﬁcally, RMSE was 1.16 ml/kg/min for

lying down, 1.75 ml/kg/min for sedentary, 3.93 ml/kg/min

for walking, 4.55 ml/kg/min for dynamic/household, 4.24

ml/kg/min for biking and 5.81 ml/kg/min for running. Sim-

ilarly to what was reported in literature for EE estimation

models, activity-speciﬁc VO

2

estimation models combining

accelerometer and HR data outperformed single regression

models relying on accelerometer only (46% RMSE reduc-

tion, p =4e

11

<↵), HR only (23% RMSE reduction,

p =0.0002 <↵), combined accelerometer and HR data

(17% RMSE reduction, p =0.001 <↵) and activity-speciﬁc

estimation models relying on accelerometer data only (10%

RMSE reduction, p =0.002 <↵). Fig. 5 provides an

overview.

C. Non-Steady-State Detection

221 transitions were analyzed in total (45 SW, 28 WS, 42

SD, 53 DS, 39 WWUP, 14 WWDOWN). 85% of all transitions

were correctly identiﬁed by the transition detection system.

Transition detection was more accurate for transitions between

sedentary to walking activities (98%) and for transitions be-

tween walking to sedentary activities (100%). Transition detec-

tion accuracy dropped to 78% and 84% for transitions between

sedentary to dynamic/household and dynamic/household to

sedentary respectively. Transitions in VO

2

within activities

(e.g. following changes in walking speed or inclination), that

were smaller than 1 ml/kg/min were not considered. All other

within activities transitions were correctly identiﬁed (100% ).

Fig. 6. RMSE for VO

2

estimation using different methods for non-steady-

state data. While standard activity-speciﬁc models that combine acc and

HR (Act-Spec ACC+HR) perform better than other models during steady-

state (see Fig .5), they offer poor performance during transitions. Linear,

exponential and logistic functions progressively reduce RMSE, with logistic

functions providing the lowest errors. Logistic Generic refers to logistic

functions were parameters did not vary depending on the transition type.

D. Transition Types

There was a signiﬁcant relation between transition time

and VO

2

change between the starting and target activities

during a transition (correlation coefﬁcient r =0.68, p =

0.0015 <↵). VO

2

difference between the starting and target

activities during a transition was 6.40 ml/kg/min for SW, 9.25

ml/kg/min for WS, 7.47 ml/kg/min for SD, 7.46 ml/kg/min

for DS , 4.38 ml/kg/min for WWUP and 5.80 ml/kg/min

for WWDOWN. Average transition duration was 97 s for

SW , 108 s for WS, 88 s for SD, 98 s for DS, 87 s for

WWUP and 125 s for WWDOWN. VO

2

estimation error

for steady-state models during transitions involving sedentary

and household activities was 4.62 ml/kg/min (SD) and 4.50

ml/kg/min (DS ), even though error during steady-state was

only 1.75 ml/kg/min for sedentary and 4.55 ml/kg/min for

household. VO

2

estimation error for steady-state models

during transitions involving sedentary and walking activities

was 5.63 ml/kg/min (SW) and 4.23 ml/kg/min (WS), even

though error during steady-state was only 1.75 ml/kg/min for

sedentary and 3.93 ml/kg/min for walking.

E. Transition-Speciﬁc Non-Steady-State models

Fig. 6 shows VO

2

RMSE during transitions for the best

performing steady-state models (i.e. activity-speciﬁc models

combining accelerometer and HR data) and non-steady-state

models. VO

2

RMSE during transitions for steady-state models

was 4.38 ml/kg/min. Linear models reduced RMSE to 4.27

ml/kg/min (3% RMSE reduction, p =0.5 >↵), expo-

nential functions to 4.07 ml/kg/min (7% RMSE reduction,

p =0.2 >↵), and logistic functions to 3.68 ml/kg/min (16%

RMSE reduction, p =0.0007 <↵). Thus, logistic functions

were the best performing non-steady-state models for activities

transitions. For comparison, we evaluated logistic functions

were parameters were not ﬁtted to speciﬁc transitions, but were

held constant for all transitions, denoted as Logistic Generic in

Fig. 6 and Table I. Logistic generic results were VO

2

RMSE

of 3.77 ml/kg/min (14% RMSE reduction, p =0.002 <↵).

6

TABLE I

VO

2

RMSE FOR TRANSITIONS MODELS CONSIDERING PERFECT TRANSITION DETECTION (ML/KG/MIN)

Transition Activity-Speciﬁc Linear Exponential Logistic Logistic Generic

Sedentary to walking (SW) 5.42 ± 1.46 5.79 ± 1.86 4.67 ± 3.18 4.27 ± 2.20 4.94 ± 2.40

Walking to sedentary (WS) 3.84 ± 1.12 4.10 ± 1.62 3.26 ± 2.60 2.43 ± 1.59 2.48 ± 1.33

Sedentary to dynamic (SD) 4.28 ± 1.25 3.81 ± 1.54 4.07 ± 1.99 3.30 ± 1.29 3.29 ± 1.36

Dynamic to sedentary (DS) 4.20 ± 1.48 3.17 ± 1.43 3.32 ± 1.28 3.57 ± 1.31 3.42 ± 1.23

Walking slow to fast (WWUP) 4.10 ± 2.06 4.97 ± 2.43 3.15 ± 1.79 4.05 ± 2.10 4.06 ± 2.10

Walking fast to slow (WWDOWN) 4.03 ± 2.17 3.69 ± 2.32 4.87 ± 3.01 4.4 ± 2.28 3.95 ± 2.14

Mean 4.38 ± 0.80 4.27 ± 1.05 4.07 ± 1.28 3.68 ± 1.10 3.77 ± 1.03

VII. DISCUSSION

In this paper we introduced a method to estimate VO

2

dur-

ing ADLs. Our method combines non-steady-state detection

and transition-speciﬁc VO

2

estimation non-linear equations.

It is misleading to assume that EE can be derived from

VO

2

alone. Proper modeling of non-steady-state transitions

is important since both anaerobic and aerobic EE result in

heat production, which is the true EE. However, only aerobic

EE results in VO

2

consumption. To tackle non-steady-state

VO

2

estimation, we proposed a method to locate periods of

non-steady-state VO

2

using the CV over the predicted VO

2

level. Our results showed that the CV can be used for detecting

transitions between and within activities. Then, we showed that

modeling non-steady-state transitions using non-linear logistic

functions for individual transition types could reduce VO

2

estimation error by 16%.

A. Activity Recognition and Transition Detection

In previous research both unsupervised and supervised

methods for activity recognition were proposed in the context

of VO

2

estimation, with supervised methods providing better

results [18]. For this reason, we opted for supervised methods

typically able to provide good accuracy on big clusters of

activities as conﬁrmed by our results, where 94% average

accuracy was obtained in leave-one-participant-out validation

design. Transition detection was highly accurate for all tran-

sitions including the walking activity (98 100%). However,

transition detection was less accurate for transitions involving

dynamic/household activities (78 84%), since these activities

are both more difﬁcult to recognize and highly variable in

VO

2

. In literature, several approaches have been proposed

to detect transitions between activities [19], [20]. Detecting

transitions based on the CV of the predicted VO

2

provides

the advantage to detect not only transitions between activities,

but also changes in intensity within the same activity (e.g.

walking at a lower speed). Various deﬁnitions of steady-state

have been reported [10]. Some researchers deﬁned steady state

as 5 consecutive 1-minutes intervals where O

2

consumption

changes by less than 10%. However this deﬁnition resulted

from assessments of basal metabolic rate and is not directly

applicable to free living conditions, where we are interested in

detecting much shorter transitions. Our analysis shows that one

minute windows and a slightly less strict CV (i.e. 15%), are a

good trade-off for non-steady-state and transitions detection.

B. Transition-Speciﬁc Non-Steady-State Models

Our analysis focused on sedentary, dynamic/household and

walking activities, since our objective is to improve VO

2

estimation in common transitions during ADLs, where current

methods lack proper modeling.

Thus, we have not considered transitions during intense

activities (e.g. biking or running), which are less frequent in

daily life and well studied in literature. The ADLs considered

are characterized by similar levels of VO

2

for both the starting

and target activity of a transition. Thus, the VO

2

change

during a transition and the time required to reach steady

state are similar. While parameters differ between transitions,

and transition-speciﬁc logistic functions provided the best

results, generic logistic functions which do not vary parameters

based on the transition were sufﬁcient to signiﬁcantly reduce

RMSE compared to steady-state models (14% RMSE reduc-

tion, p =0.002 <↵). Logistic functions were proposed in

literature to model HR recovery after intense exercise [21], due

to the ability to model different dose-response relationships

between two relatively stable levels. While we applied this

technique to model transitory oxygen uptake during ADLs,

the methodology could be extended to other physiological

parameters undergoing similar changes. For comparison, we

also ﬁtted linear and exponential functions, as previously

reported in literature [16]. However we found suboptimal

results compared to logistic functions.

Another advantage of the proposed technique is that at

runtime VO

2

estimations during transitions are only dependent

on the previously ﬁtted parameters and on the parameter ✓

1

,

representing the difference in VO

2

between the starting and

the target activity. Thus, these models could be applied to

systems based on accelerometers only, and do not depend on

other predictors or sensing modalities.

Performance of non-steady-state VO

2

models should be

assessed outside of the lab [22]. However, there is currently no

unobtrusive reference system that is able to measure breath-by-

breath VO

2

in unconstrained settings. Only by using indirect

calorimetry and supervised settings we can record data, which

allows us to analyze how non-steady-state models can improve

VO

2

estimations during transitions. In our evaluation protocol,

we could not elicitate abrupt VO

2

changes within one activity,

e.g. when changing the treadmill speed and inclination level,

there was a gradual change, which could last up to 30

seconds. Hence, it was difﬁcult to even collect representative

data that could beneﬁt from non-steady-state modeling within

an activity (e.g. walking at different speeds) and the VO

2

7

estimation did not provide improvements over the steady-state.

A set of thirty activities of varying intensities was used in

this study and further work could extend the set, for example

including activities involving carrying loads. However, we

believe that the lifestyle activities used here are representative

and relevant to assess VO

2

estimation in non-stationary daily

life conditions. The generalized method proposed in this work

is able to model abrupt within-activity changes in VO

2

without

modiﬁcations to the model, since it detects transitions relying

on the CV, instead of the activity type. While it is not feasible

for us to robustly detect activities such as carrying a weight

using a wearable sensor on the chest, the use of HR partially

solves the problem. Carrying a load would create higher

strain, thus raise HR, which is used in the estimation models,

and therefore result in higher predicted VO

2

. In this work,

CV changes were shown to even indicate intensity changes

within the same activity conﬁrming the method’s sensitivity.

By properly modeling VO

2

during transitions and locating

periods of non-steady-state, we quantiﬁed VO

2

for aerobic

dynamics during the onset of exercise, as well as during ADLs

characterized by short duration, with 16% error reduction

compared to current methods.

REFERENCES

[1] M. Altini, J. Penders, and O. Amft, “Personalizing energy expenditure

estimation using a cardiorespiratory ﬁtness predicate,” in Pervasive

Computing Technologies for Healthcare (PervasiveHealth), 2013 7th

International Conference on. IEEE, 2013, pp. 65–72.

[2] A. G. Bonomi, “Improving assessment of daily energy expenditure

by identifying types of physical activity with a single accelerometer.”

Journal of Applied Physiology, vol. 107, no. 3, pp. 655–661, 2009.

[Online]. Available: http://www.ncbi.nlm.nih.gov/pubmed/19556460

[3] E. Tapia, “Using machine learning for real-time activity recognition and

estimation of energy expenditure,” in PhD thesis, MIT, 2008.

[4] C. B. Scott, “Contribution of anaerobic energy expenditure to whole

body thermogenesis,” Nutr Metab, vol. 2, pp. 14–23, 2005.

[5] K. Wasserman, A. L. Van Kessel, and G. G. Burton, “Interaction of

physiological mechanisms during exercise,” J Appl Physiol, vol. 22,

no. 1, pp. 71–85, 1967.

[6] K. R. Short and D. A. Sedlock, “Excess postexercise oxygen consump-

tion and recovery rate in trained and untrained subjects,” Journal of

Applied Physiology, vol. 83, no. 1, pp. 153–159, 1997.

[7] S. W. Su, L. Wang, B. G. Celler, and A. V. Savkin, “Oxygen uptake

estimation in humans during exercise using a hammerstein model,”

Annals of biomedical engineering, vol. 35, no. 11, pp. 1898–1906, 2007.

[8] S. W. Su, B. G. Celler, A. V. Savkin, H. T. Nguyen, T. M. Cheng, Y. Guo,

and L. Wang, “Transient and steady state estimation of human oxygen

uptake based on noninvasive portable sensor measurements,” Medical

& biological engineering & computing, vol. 47, no. 10, pp. 1111–1117,

2009.

[9] M. S. Orendurff, J. A. Schoen, G. C. Bernatz, A. D. Segal, and G. K.

Klute, “How humans walk: bout duration, steps per bout, and rest

duration,” J Rehabil Res Dev, vol. 45, no. 7, pp. 1077–89, 2008.

[10] M. M. Reeves, P. S. Davies, J. Bauer, and D. Battistutta, “Reducing

the time period of steady state does not affect the accuracy of energy

expenditure measurements by indirect calorimetry,” Journal of Applied

Physiology, vol. 97, no. 1, pp. 130–134, 2004.

[11] J. Wang, S. J. Redmond, M. Voleno, M. R. Narayanan, N. Wang,

S. Cerutti, and N. H. Lovell, “Energy expenditure estimation during nor-

mal ambulation using triaxial accelerometry and barometric pressure,”

Physiological measurement, vol. 33, no. 11, pp. 1811–1830, 2012.

[12] T. Yoshida and B. J. Whipp, “Dynamic asymmetries of cardiac output

transients in response to muscular exercise in man.” The Journal of

Physiology, vol. 480, no. Pt 2, pp. 355–359, 1994.

[13] J. Smolander, T. Juuti, M.-L. Kinnunen, K. Laine, V. Louhevaara,

K. M

¨

annikk

¨

o, and H. Rusko, “A new heart rate variability-based

method for the estimation of oxygen consumption without individual

laboratory calibration: application example on postal workers,” Applied

Ergonomics, vol. 39, no. 3, pp. 325–331, 2008.

[14] A. Pulkkinen, J. Kettunen, K. Martinm

¨

aki, S. Saalasti, and H. Rusko,

“On–and off dynamics and respiration rate enhance the accuracy of

heart rate based vo2 estimation,” in Proceedings of the ACSM Congress,

Indianapolis, Abstract, vol. 36, 2004.

[15] L. R. Dugas, L. van der Merwe, H. Odendaal, T. D. Noakes, and

E. V. Lambert, “A novel energy expenditure prediction equation for

intermittent physical activity.” Medicine and Science in Sports and

Exercise, vol. 37, no. 12, pp. 2154–2161, 2005.

[16] F.-C. Chuang, Y.-T. C. Yang, and J.-S. Wang, “Accelerometer-based

energy expenditure estimation methods and performance comparison,”

in 2nd International Conference on Advances in Computer Science and

Engineering (CSE 2013). Atlantis Press, 2013, pp. 99–103.

[17] E. Børsheim and R. Bahr, “Effect of exercise intensity, duration and

mode on post-exercise oxygen consumption,” Sports Medicine, vol. 33,

no. 14, pp. 1037–1060, 2003.

[18] S. Chen, J. Lach, O. Amft, M. Altini, and J. Penders, “Unsupervised

activity clustering to estimate energy expenditure with a single body

sensor,” in Body Sensor Networks (BSN), 2013 IEEE International

Conference on. IEEE, 2013, pp. 1–6.

[19] R. Ali, L. Atallah, B. Lo, and G.-Z. Yang, “Detection and analysis

of transitional activity in manifold space,” Information Technology in

Biomedicine, IEEE Transactions on, vol. 16, no. 1, pp. 119–128, 2012.

[20] C. Zhu and W. Sheng, “Recognizing human daily activity using a single

inertial sensor,” in Intelligent Control and Automation (WCICA), 2010

8th World Congress on. IEEE, 2010, pp. 282–287.

[21] N. Christenfeld, L. M. Glynn, and W. Gerin, “On the reliable assessment

of cardiovascular recovery: An application of curve-ﬁtting techniques,”

Psychophysiology, vol. 37, no. 4, pp. 543–550, 2000.

[22] L. Bouarfa, L. Atallah, R. M. Kwasnicki, C. Pettitt, G. Frost, and G.-Z.

Yang, “Predicting free-living energy expenditure using a miniaturized

ear-worn sensor: An evaluation against doubly labelled water,” IEEE

Transactions on Biomedical Engineering, vol. 61, no. 2, pp. 566–575,

2014.

Marco Altini received the M.Sc. degree in engi-

neering and computer science from the University of

Bologna in 2010. He is currently pursuing the Ph.D.

degree with the Technical University of Eindhoven,

Eindhoven, The Netherlands. His current research

interests include machine learning techniques for

the development and implementation of biomedical

applications using wearable sensors data.

Julien Penders received the M.Sc. degree in systems

engineering from the University of Liege, Belgium,

and the M.Sc. degree in biomedical engineering

from Boston University, in 2004 and 2006, respec-

tively. He is currently the Program Manager with the

Holst Centre/IMEC, Eindhoven, The Netherlands.

He has authored or co-authored over 30 papers in

journals and conference proceedings on body area

networks and autonomous wireless sensor networks.

Oliver Amft received the M.Sc. from TU Chem-

nitz in 1999 and the Ph.D. from ETH Zurich in

2008, both in Electrical Engineering and Information

Technology. Oliver is a Full Professor (W3) and

Chair of Sensor Technology at University of Passau.

He is also afﬁliated with the Wearable Computing

Lab, ETH Zurich (CH) and the Signal Processing

Systems section at TU Eindhoven (NL). Oliver is

interested in multi-modal activity recognition and

human behaviour inference algorithms and has co-

authored more than 90 publications in this ﬁeld.