Validated Caloric Expenditure Estimation using a Single
Body-Worn Sensor
Jonathan Lester, Carl Hartung, Laura Pina,
Ryan Libby, Gaetano Borriello
Computer Science and Engineering
University of Washington
Seattle, Washington, USA
{jlester,chartung,lpina,libby,
gaetano}@cs.washington.edu

Glen Duncan
Epidemiology and Nutritional Sciences
University of Washington
Seattle, Washington, USA
duncag@u.washington.edu


ABSTRACT
In 2007, approximately 30% of US adults were obese, with
related health care costs exceeding 100 billion dollars.
Clearly, the obesity epidemic represents a growing societal
concern. One challenge in weight control is the difficulty of
tracking food calories consumed and calories expended by
activity. This paper presents a system for automatic
monitoring of calories expended using a single body-worn
accelerometer. Our system uses activity inference combined
with signal analysis to estimate calories expended in real-
time using regression formulas developed by the American
College of Sports Medicine. To validate our system, we
have collected data from 51 subjects in a laboratory setting
using a treadmill and a more natural field test. Actual
caloric expenditure was determined using the medical “gold
standard” measurement, of oxygen consumption. We are
able to achieve 89% accuracy with lab data and 79% with
field data – both high enough to be useful in practice.
Author Keywords
pervasive health, long term health monitoring, wellness,
caloric balance
General Terms
Human Factors, Measurement, Verification
ACM Classification Keywords
H5.m. Information interfaces and presentation: Miscellaneous.
INTRODUCTION
The United States and many other industrialized countries
currently face an epidemic of obesity. An overabundance of
highly palatable, energy-dense foods, and a lack of regular
exercise has resulted in 35% of the US population being
overweight, and the percentage of obese adults has doubled
to 30% over the past decade [2]. Often viewed as a US
problem, the World Health Organization estimates that
there are more than one billion overweight adults in the
world; 300 million of whom are actually obese [1]. In
addition to social and lifestyle effects, obesity is associated
with increased risk of chronic and fatal diseases such as
heart disease, stroke, some forms of cancer, type 2 diabetes,
and hypertension. The costs for these diseases in terms of
human life and medical expenses are staggering. In 1998 it
was estimated that 9.1% of all US medical expenses, $78.5
billion dollars, was spent on overweight/obesity attributed
expenses.
While there is no single solution to the obesity epidemic,
research has shown that overweight/obese individuals tend
to underestimate the calories they consume [3,4] and
overestimate the calories they burn [5]. This is, of course,
the worst possible error; people consume more calories and
burn fewer calories than they believe, and remain
overweight. This misconception is far from the only reason
for the obesity epidemic; however, a better idea of the
calories consumed and expended by a person would be a
valuable piece of information for individuals to monitor
their energy balance, as well as for doctors and other
caregivers. With our busy lifestyle, it is difficult to imagine
that most people would, or are even capable of, accurately
and continuously tracking their eating and exercise habits.
Systems which can automatically track and provide a user
with feedback about their energy balance might have the
potential to help obese, overweight, and even normal
weight individuals.
To create such a system, we are working on two projects,
funded by the National Institutes of Health, to develop an
up-to-the-minute display of a user’s caloric balance: what
they've eaten versus what they've burned through physical
activity. There are two components to our energy balance
monitoring system: 1) an interface used to enter foods
consumed and 2) the system presented here, algorithms to
automatically compute caloric expenditure. For some
preliminary details regarding food entry and some of our
initial lab tests results please see [6]. In this paper we

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise,
or republish, to post on servers or to redistribute to lists, requires prior
specific permission and/or a fee.
UbiComp 2009, Sep 30 – Oct 3, 2009, Orlando, Florida, USA.
Copyright 2009 ACM 978-1-60558-431-7/09/09...$10.00.
225

present a system for computing an estimate of the calories
expended using a wearable system and its validation against
the medical gold standard measurement. While the eventual
end users of our system are obese and overweight
individuals, the current target audience is primarily medical
researchers and practitioners. These researchers are not
amenable to overly complex algorithms or “black box”
systems. Rather than use more complex models, which may
perform better, we use more standardized algorithms and
techniques to make our system more understandable,
useful, and more likely to be adopted by this audience.
RELATED WORK
The area of fitness and exercise is filled with a variety of
research projects [19] and commercial systems. Several
commercial products exist which can measure calories
expended during exercise. The most basic systems are those
found on gym equipment. There have also been several
portable systems for measuring more free living conditions.
One popular device among researchers has been the RT3
accelerometer, a wearable accelerometer which computes
activity counts that can approximate the energy level of the
activity being performed by the user based on established
cut-points [18]. However, the coarse grain output (often
minute long averages of acceleration) of the RT3 means
that it is impossible to gain detailed insight into the
behavior of individuals, such as activity type and location,
among others. The Philips Respironics’ Actical is a similar
device which can compute activity counts as well as an
estimate of total energy expenditure [1].
In addition to more clinical devices; several casual systems
have been released, including the Nike+ system, a shoe
worn piezoelectric sensor that communicates wirelessly
with a receiver plugged into an Apple iPod [8]. The Nike+
system tracks the distance, pace, estimates the calories
burned during a run, and provides users with audio
feedback while running.
Nokia provides two systems for activity monitoring:
Activity Monitor and Sports Tracker [9, 10]. Sports Tracker
is a GPS based system that uses either the phones built-in
GPS or a Bluetooth GPS device to track users’ trips. Users
are provided with maps overlaid with their GPS tracks;
information about the distance traveled, average speed, and
duration is provided in real time. Activity Monitor uses the
built-in accelerometer on Nokia’s N95 and N82 phones to
count the number of steps taken by a user. It provides
tracking information for users recording the number of
steps taken per day, providing an estimate of the calories
consumed and the distance traveled by a user. Both the
Nike+ and Nokia systems are closed platforms which use
fairly simplistic methods for estimating calories without
any extensive validation.
The SenseWear/BodyBugg system by SenseMedia is a
wearable armband-based system equipped with a dual-axis
accelerometer, skin temperature, Galvanic Skin Response,
and heat flux sensors [7]. While the SenseWear system has
been used in several experiments its cost and closed black
box nature prevents it from widespread use or integration
into other systems.
Choi et al. [11] conducted a similar laboratory study to
ours, in which they collected wireless accelerometer data
from 94 subjects walking on a treadmill hooked up to a
portable V
ሶ
O
2
gas analyzer to measure total energy
expenditure. Accelerometer data was then filtered and
integrated and used as an input into a regression analysis to
match their accelerometer output with the energy
expenditure computed by the gas analyzer. The main
purpose of their study was to validate accelerometer
readings, based on correlation with energy expenditure
measured by V
ሶ
O
2
. The study consisted of a controlled
laboratory experiment tested on primarily fit college-age
students. The main differences between Choi et al. and the
results presented here are that we used a wider variety of
test subjects (ages/body types), are able to classify more
natural activities using a field experiment, and that we use
accepted algorithms to estimate energy expenditure instead
of simply modeling the data we’ve collected.
All of these devices are closed systems, they cannot be
modified or scrutinized, and do not provide extensive proof
by validation. While many of these systems are appropriate
for the casual user, we are focused on developing a system
useful both to researchers, medical doctors, and
practitioners alike, many of whom have qualms—perceived
and real—about the utility, operation, and reliability of
these commercial systems.
OVERVIEW
In this paper we first discuss the experimental design,
algorithms used for computing caloric expenditure, step
counting, and approach for estimating walking grade. We
then present three sets of results, results from our laboratory
experiments, field data collections, and a synthetic average
day’s worth of data. We are able to achieve 89% accuracy
with lab data and 79% with field data.
EXPERIMENT DESIGN
While we could take one of the systems described in the
related work section, we do not have a quantifiable
understanding of how accurate these systems are or have
access to their methodology. Researchers who wish to use
the system need to have confidence that the system uses a
reasonable approach and works well enough for them to
conduct their research.
System Setup
Our system consists of a body-worn 3-axis accelerometer.
The primary source of data for our experiments is the
University of Washington/Intel Mobile Sensing Platform
(MSP) [15], which is worn on the waist of subjects during
data collection. As one of the goals of our project is to
create a system which can be carried around by users
throughout the day, we would eventually like to be able to
use a cell phone as sole platform. To this end, we also
collected a subset of accelerometer data from the Apple
226