IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO.

11, NOVEMBER 2012

3177

SoM: A Smart Sensor for Human Activity

Monitoring and Assisted Healthy Ageing
David Naranjo-Hernandez, Student Member, IEEE, Laura M. Roa, Fellow, IEEE,

Javier Reina-Tosina, Senior Member, IEEE, and Miguel Angel

Estudillo-Valderrama, Student Member, IEEE

AbstractThis paper presents the hardware and software design and implementation of a low-cost, wearable, and unobstructive
intelligent accelerometer sensor for the monitoring of human physical activities. In order to promote healthy lifestyles to elders for
an active, independent, and healthy ageing, as well as for the early
detection of psychomotor abnormalities, the activity monitoring is
performed in a holistic manner in the same device through different approaches: 1) a classification of the level of activity that allows
to establish patterns of behavior; 2) a daily activity living classifier
that is able to distinguish activities such as climbing or descending
stairs using a simple method to decouple the gravitational acceleration components of the motion components; and 3) an estimation
of metabolic expenditure independent of the activity performed
and the anthropometric characteristics of the user. Experimental
results have demonstrated the feasibility of the prototype and the
proposed algorithms.
Index TermsActive ageing, independent living, metabolic
expenditure, physical activity monitoring, smart sensor.

I. INTRODUCTION
EMOGRAPHIC changes associated with population ageing and the increasing number of elders living alone are
leading to a significant change in the social and economic structure of society [1]. In 2030, a third part of the European citizens
will have more than 65 years and 40% of them may not reach
a decent standard of independent living, needing the help of
family and professional or nonprofessional caregivers [2].

Manuscript received February 19, 2012; revised April 20, 2012 and June
4, 2012; accepted June 20, 2012. Date of publication July 6, 2012; date of
current version October 16, 2012. This work was supported in part by the
CIBER de Bioingeniera, Biomateriales y Nanomedicina (CIBER-BBN) and
the intramural Grant PERSONA, in part by the Instituto de Salud Carlos III
under Grants PI082023 and PI11/00111, and in part by the Direccion General de Investigacion, Tecnologa y Empresa, Government of Andaluca, under
Grants P08-TIC-04069 and TIC6214. CIBER-BBN is an initiative funded by
the 6th National R&D&i Plan 20082011, Iniciativa Ingenio 2010, Consolider
Program, CIBER Actions and financed by the Instituto de Salud Carlos III with
assistance from the European Regional Development Fund. Asterisk indicates
corresponding author.
D. Naranjo-Hern
andez is with CIBER de Bioingeniera, Biomateriales y
Nanomedicina (CIBER-BBN), Spain, and also with the Biomedical Engineering
Group, University of Seville, Seville 41092 (e-mail: davidazuaga@gmail.com).
L. M. Roa is with the Biomedical Engineering Group, University of Seville,
Seville 41092, Spain, and also with CIBER de Bioingeniera, Biomateriales y
Nanomedicina (CIBER-BBN) (e-mail: lroa@us.es).
J. Reina-Tosina is with the Department of Signal Theory and Communications, University of Seville, Seville 41092, Spain, and also with CIBER de Bioingeniera, Biomateriales y Nanomedicina (CIBERBBN) (e-mail: jreina@us.es).
Estudillo-Valderrama is with CIBER de Bioingeniera, Biomateriales
M. A.
y Nanomedicina (CIBER-BBN), Spain, and also with the Biomedical Engineering Group, University of Seville, Seville 41092 (e-mail: mestudillo@us.es).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TBME.2012.2206384

On the other hand, there is a close correlation between the

loss of mobility in the elderly with the loss of independence [1],
[3], [4], what can even lead to a higher rate of morbimortality [1],
[3], [4]. Therefore, it is evident that preventive interventions,
aimed at preserving or restoring mobility of the elderly, may
improve both their quantity and quality of life [5][7].
This scenario justifies the major effort currently undertaken to
develop new technologies which, mainly through exercise motivation [8] and diet control [9], [10], promote healthy lifestyles
and an active ageing for the elderly and chronic patients at their
homes. Many of the developed systems base their operation on
an unobtrusive monitoring of physical activity [11][13], either
through the classification and identification techniques [14] or
through the estimation of energy expenditure [15], [16].
Accelerometry is the most commonly used method to assess
physical activity in observational studies [17] as it allows to
record variations in orientation as well as movements are easily detected [7]. This way, many accelerometry-based wearable
systems for the pervasive monitoring of activities of daily living
(ADL) have been developed. Some authors propose measuring
the number of walked steps as an estimation of the developed
physical activity, whereas other studies suggest either using the
total duration of the activity in minutes [18] or measuring the
intensity of the activity [19]. Nevertheless, the majority of accelometric devices only store a measurement proportional to the
average acceleration over a specific time period [17], [20] that
may lead to an imprecise estimation when used for ADL classification compared to the gold standard [21]. Another approach
made by a remarkable number of proposals is based on the
frequency of postural transitions [22]. Estimation of metabolic
expenditure is also an important area of study [11], [15], [23].
In any case, several studies demonstrate that systems based
on the ADL classification are the most appropriate option to
study and analyze the behavior patterns and habits of monitored
users [24], [25]. These systems can be employed in the detection
of abnormalities related to the progression of neurodegenerative
diseases or in the early stages of health disorders [26]. However,
and to the best of our knowledge, there are no solutions able to
provide an accurate estimation of daily activities by means of
affordable wearable devices that use simple algorithms, but flexible enough to develop various types of classification.
Thus, the main objective of this paper is to present the implementation design of a low-cost device for human activity
monitoring that can be applied to the majority of scenarios
considered, such as subject homes or residential centers for
seniors, and which address the previous lacks. This device incorporates the following capabilities. 1) A classification system

0018-9294/$31.00 2012 IEEE

3178

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 11, NOVEMBER 2012

III. FUNCTIONAL DESCRIPTION

This section describes three processing modules applied to
human movement monitoring.
A. Processing Module for the Classification of Activity Level

Fig. 1.

Proposed distributed monitoring system and SoM prototype.

of the level of activity that can be used to quantify the thresholds of minimum required levels of activity, so as to improve the
prescription of physical exercise by the medical staff to obtain
optimal preventive results [7]. 2) A classification system of activity patterns, which is enabled to distinguish vertical displacements like climbing or descending stairs, which is a problem
that is usually addressed through complex schemes with multiple sensors or a difficult processing [11], [27], [28]. 3) A system
for the estimation of metabolic expenditure, for which a simple
equation has been proposed, valid for different users independently of their anthropometric characteristics, and which also
benefits from the data provided by the activity classification
system to make an estimation independently of the developed
activity.

The characteristics of this processing module were presented

in [29]. This classifier distinguishes among the following activity levels: very low (like sleeping), low (like eating), medium
(like walking), and high (like climbing stairs). Its operation can
be decomposed into the following phases.
1) The 40-Hz acceleration samples are filtered through a
second-order high-pass Butterworth filter (Ord|Filt = 2)
with a cutoff frequency of 0.9 Hz (FC = 0.9 Hz). This
value of FC provided the highest sensitivity and specificity
in a preliminary set of experiments. Filtered accelerations
are defined as AHP,x , AHP,y , and AHP,z ).
2) Afterward, a global energy estimation is calculated by
adding the square of NAS acceleration samples (NAS =
256), and multiplying the vertical component with a
weighting coefficient (W eCo = 0.5) to give a major importance to the horizontal components for the activity level
estimation. A value of 256 samples was chosen because
at 40 Hz, it corresponds to a time comparable to the average duration of an activity in accordance with other
authors [14], [16], [26]. Furthermore, this value facilitates
the division operation in (1). This value of W eCo provided
the highest sensitivity and specificity in a preliminary set
of experiments

II. SYSTEM DESCRIPTION

In the proposed paradigm (see Fig. 1), the sensor device is an
accelerometer smart sensor, referred to as sensor of movements
(SoM), which performs personalized and real-time monitoring
of physical activity. The SoM has been conceived to be integrated into a biocompatible and impermeable skin patch so
that it can be comfortably carried by the user on the back at
the height of the sacrum. It connects with a Remote Telecare
Center (RTC) [29] through a wearable device, the Decision and
Analysis Device (DAD).
The design applied to the SoM uses a modular and generic
scheme [29] supported by the following elements (see Fig. 1).
1) Sensor device: A triaxial LIS3LV02DQ accelerometer
from STMicroelectronics.
2) Communications module: A Chipcom CC2430 transceiver has been used to develop a simple low-power protocol
based on the IEEE 802.15.4 standard.
3) Processing unit: The intelligence of the sensor device is
provided by the processing modules that are executed in
real-time and in parallel within the SoM processing unit.
Each processing module is capable of transmitting the acquired information or the result of its processing. These
data are structured into information samples (abstract concept defined in [29]) generated with a given sampling frequency, which can be set via commands. A PIC18F2431
microcontroller of Microchip Technology Inc. has been
used for the purpose of data processing.

EA =
256
+

255


 255


(AHP,x (n i))2

i=0

W eCo (AHP,y (n i))2

i=0

255



(AHP,z (n i))

(1)

i=0

3) The activity level variable is determined depending on

the situation of the EA variable value in the four areas
established by three preset thresholds (ThVeryLow , ThLow ,
ThM edium ) and recalculated with each new 256 samples.
The possible values of activity level variable are very low,
low, medium, or high.
4) Each information sample takes the value (very low, low,
medium, or high) of the activity level variable that has
been assigned in a higher number of occasions during the
corresponding intersample period.
In the processing module, the alarm events correspond to
events related with the presence or absence of activity in certain
time observation periods in the daily life of the end-user. It
has been programmed to trigger a warning event if no physical
activity is detected during the observation period (warn if the
user has not woken up at his/her regular schedule).

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 11, NOVEMBER 2012

Fig. 2.

Characteristics of accelerations observed during the period of a step.

B. Processing Module for ADL Classification

Each information sample of this processing module corresponds to a variable that classifies the physical activity developed by the user, which is estimated during the period
between the current sample and the previous one. The classifier
distinguishes among the following activities: 1) walk (normal
pace), 2) walk (fast pace), 3) stairs-up, 4) stairs-down, 5) lying,
and 6) none of the above. The information sample includes a
variable that represents the number of steps performed for the
estimated activity, which takes zero value in the activities 5 and
6. The traveled distance can be approximated from the number
of steps by multiplying by a customized mean step length (both
for normal and fast pace). The purpose of this processing module
is to provide an additional level of information about the physical activity developed by the user to enhance the information
provided by the previous processing module.
The main novelty of this processing second layer is the
method used to decouple the gravitational acceleration components from those caused by the movement itself, in order
to discriminate between horizontal activities such as walking
and sloping ones like climbing or descending stairs, which is
an issue rarely studied in the related literature. Many authors
solve this problem in a simple manner by filtering the accelerations [30], so that low-frequency components correspond to
the gravitational acceleration and high-frequency components
correspond to the motion. Starting with experimental data, a preliminary study was conducted to analyze the feasibility of these
methods; however, we were unsuccessful when trying to get a
robust discrimination of walk, stairs-up/-down activities by applying this filtering and integrating the acceleration components
associated with vertical movement.
To solve this issue, other authors add information from additional sensors by means of gyroscopes and magnetometers [16],
[27], or use algorithms with a high computational load [28] to
reliably decouple the gravitational components of the motion
components. However, the mentioned techniques are not applicable in our work due to need of using algorithms with a low
computational load because of the implications of the compact,
low-cost sensors with limited processing capabilities, required
for the desired product. To tackle this issue, a simple algorithm
that is capable of being processed in real time is presented.

3179

A preliminary experimental analysis of the accelerations

served to establish a set of assumptions following an empirical rationale. These were subsequently applied to simplify the
classification methodology (see Fig. 2).
1) First, to consider the acceleration component Ay as the
vertical acceleration component. This assumption would
imply that when the subjects back is not completely vertical, an error would be introduced in the exact calculation of the traveled vertical distance; however, the error
introduced does not affect the estimation of movement
direction.
2) Second, to consider that the vertical acceleration component Ay can be decomposed as the sum of the acceleration associated with the gravitational force in the vertical
direction Ay |g and the acceleration associated with the
movement in the vertical direction Ay |m ov .
3) Third, during the walking process, either in horizontal
direction or stairs-up/-down, the vertical acceleration associated with movement is zero (Ay |m ov = 0) when the
inclination of the subjects back is aligned with the vertical direction. This assumption is supported by additional
reasons. On the one hand, during the walking process,
we observed a short unbalance period where the body resembles undergoing a free fall; it is the instant when
the movement losses the inertia in the vertical direction
and triggers a short fall on a foot, which is ready for the
short impact damping. If the free fall occurred over a
larger time, the vertical acceleration component Ay acquired by the accelerometer would tend to zero, but as
this event is very short, there is a negligible variation and
its value is kept around the nominal gravitation (9.8 m/s2
if the body is completely vertical). At this instant, it was
also observed that the subjects back laid on a vertical
position. On the other hand, the assumption implies an
additional simplification of the acting forces by which the
biomechanical forces are neglected to consider only the
gravitational force at that instant of free fall.
If the activity corresponds to walking, climbing, or descending stairs, the displacement obtained by integrating the movement components can be used to discriminate its movement.
However, it is not trivial to decouple the acceleration components to obtain the motion components, taking into account that
the gravitational components also vary with the position and,
therefore, with the movement.
The proposed method is based on setting a time-varying gravitational reference level (Ay |g ,ref ) for the vertical acceleration
component (Ay , according to assumption 1), which tries to approximate the gravitational acceleration component at the vertical direction (Ay |g ). This way, the vertical acceleration component associated with movement (Ay |m ov ) is estimated as the
difference between the vertical acceleration Ay and the gravitational reference level (Ay |g ,ref ) (assumption 2).
Given the complexity to set this gravitational reference level
(Ay |g ,ref ) for each time instant, it is approximated as an accelerometric curve that matches a series of control points. The
problem is thus reduced to find a number of points in the space
of accelerations that can be approximated by the gravitational

3180

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 11, NOVEMBER 2012

TABLE I
PARAMETERS OF THE PROCESSING MODULE FOR ADL CLASSIFICATION

Fig. 3.

Flow diagram of the processing module for ADL classification.

acceleration components in the vertical direction (Ay |g ). Due to

the fact that in the absence of acceleration components associated with the movement, the square root of the sum of the square
of the accelerations must take the value g, these points could be
established as the values of the acceleration component Ay , in
which the sum of the squared values of the accelerations is g 2 or
approximates to it. However, our results with this approximation
for a set of experiments showed a significant error.
To solve this problem, an indirect method has been established that involves the assumption of the third assumption.
According to it, at the time instants when the inclination of the
subjects back is aligned with the vertical direction, the acceleration of the vertical component associated with movement is
zero (Ay |m ov = 0), and in agreement with the second assumption, the vertical accelerations acquired by the accelerometer
would match those associated with the gravitational force at
the vertical direction (Ay |g ). This way, the vertical acceleration
components at these time instants (Ay ) can be used to set the
control points of the reference gravitational level (Ay |g ,ref ).
In agreement with the proposed methodology, the solution of
the problem is simplified just to find the time instants of the
control points. Experiments conducted in laboratory suggested
two alternatives (see Fig. 2).
1) If the frequency of the step is low, the lateral acceleration
(which can be approximated by Ax regarding the first assumption) is maximum (either positive or negative) when
the inclination of the subjects back is aligned to the vertical position. Thus, the maximum and minimum values
of Ax correspond to the instants of the control points.
Upstream and downstream crossing points of the smooth
lateral acceleration (lateral acceleration Ax filtered with a
bandpass filter around the gait frequency) with the gravitational lateral acceleration (lateral acceleration Ax filtered
with a low-pass filter) allow to establish the search intervals of this points.
2) If the frequency of the step is high, lateral acceleration is
not adequate to define the control points. However, when
the same procedure was applied to the acceleration in the
front direction Az , it was proven to be completely valid.
The flow diagram described in Fig. 3 summarizes the most relevant aspects of the processing module classification algorithm.
The parameters shown in the same are described in Table I.

C. Processing Module for Metabolic Expenditure Estimation

Each information sample of this processing module corresponds to a variable that estimates the metabolic expenditure of
the subject, which is estimated during the period between the
current sample and the previous one. The variable EA defined
in the processing module for the classification of the level of
activity is used as a basis for metabolic expenditure estimation. Using the generic standard proposed in the Compendium
of Physical Activities [31], a series of correspondences have
been established between ADL and the associated metabolic
expenditure. The Compendium of Physical Activities has been
developed to facilitate the comparison between intensity values
of physical activity in observational studies. The Compendium
provides a classification that assigns to each activity a five-digit
code and the corresponding metabolic equivalent level of intensity (MET). One MET is defined as the standard of metabolic
expenditure associated with a rest activity (4.184 kJkg1 h1 ).
The metabolic cost in METs assigned to other activities is then
defined as the ratio between the metabolic expenditure associated with the activity and the expenditure of the 1-unit MET
activity pattern. Table II shows the correspondence between a
set of observational activities, the associated code according to
the Compendium and their equivalent in METs.
A validation study of the aforementioned activities in a preliminary set of experiments allowed us to establish a linear correspondence between the logarithmic value of the variable EA ,
for each activity, with the logarithm of its equivalent in METs.
This way, the following equation can be set as an approximation
to estimate the metabolic cost associated with the development
of an activity:
log10 (METs) = CoefB log10 (EA ) + CoefA

(2)

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 11, NOVEMBER 2012

TABLE II
CORRESPONDENCE BETWEEN ACTIVITIES AND THEIR EQUIVALENT IN METS

3181

TABLE IV
GLOBAL SENSITIVITY ANALYSIS OF THE PROCESSING MODULE FOR THE
CLASSIFICATION OF ACTIVITY LEVEL (SENSIVILITY IN %)

TABLE V
GLOBAL SPECIFICITY ANALYSIS OF THE PROCESSING MODULE FOR THE
CLASSIFICATION OF ACTIVITY LEVEL (SPECIFICITY IN %)
TABLE III
CHARACTERISTICS OF THE EXPERIMENTS PERFORMED IN THE
VALIDATION OF RESULTS

and taking into account the definitions made in the Compendium

ME = CoefC (EA )Co ef B m t

(3)

10Co ef A 4.184
(4)
60
where ME is the metabolic expenditure associated with the activity in kJ, m is the mass of the subject in kg, and t is the
elapsed time of the activity in minutes.
CoefC =

IV. RESULTS
A first validation in a controlled environment has been developed to evaluate the performance of the proposed methods.
For this reason, it was considered a large number of activities (see Table II) developed by volunteers (90 experiments),
so that sex and anthropometric characteristics variations are
sufficiently covered in the captured data. The number of subjects and the age and number of activities used in this first
validation (see Table III) is similar to the ones employed in
comparable studies, both for the classification of the vertical movement and the estimation of the metabolic cost [11],
[15], [16], [28], [30].
A. Validation of the Processing Module for Classification of
Activity Level
The preset thresholds for the activity level variable in the
ADL classifier were previously fixed in a preliminary set of experiments (ThVeryLow = 0.0306, ThLow = 0.1490, ThM edium =
1.6026 [m2 /s4 ]). The following classification for the level of activity was established in the performed activities.

1) Very low level activities: sleeping (A1), sitting (A2).

2) Low level activities: eating (A3), hand craft (like sewing)
(A4), stand up and speaking (A5), writing seated (A6).
3) Medium level activities: walking (A7), washing dishes
(A8), ironing (A9), sweeping (A10), but not in an intense
way.
4) High level activities: descending stairs (A11), physical exercise (repeated rotations of the trunk, A12), walking fast
(A13), intensive homework (A14), climbing stairs (A15).
The motivation to classify the developed activity between the
four levels referred was primarily to gather activities with similar behavioral patterns and check the capability of the SoM to
distinguish them. This activity multilevel classification, used in
the literature for the postoperative home recovery [32], paves
the way to prescribe remote procedures for personalized medical
treatments.
Tables IV and V show the results of an analysis of sensitivity
and specificity of the classification algorithm performed with
each of the parameters. The best results (100% sensitivity and
specificity) were obtained with the nominal values of the parameters. Variations around their nominal values did not result
in significant errors in the classification, because the sensitivity
was higher than 93% in most cases. The most sensitive parameter was FC |,Filt0 , but only for parameter variations around 50%.
However, best results were obtained regarding specificity. Due
to the fact that this processing module operates isolated, there
is no propagation of errors from the other modules in this case.

3182

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 11, NOVEMBER 2012

TABLE VIII
GLOBAL SPECIFICITY ANALYSIS OF THE PROCESSING MODULE FOR THE
CLASSIFICATION OF ACTIVITY LEVEL (SPECIFICITY IN %)

C. Validation of the Processing Module for Metabolic

Expenditure Estimation
Fig. 4. Characteristics of accelerations corresponding to up/down-stairs
activities.
TABLE VI
PARAMETERS OF THE PROCESSING MODULE FOR ADL CLASSIFICATION

TABLE VII
GLOBAL SENSITIVITY ANALYSIS OF THE PROCESSING MODULE FOR ADL
CLASSIFICATION (SENSIVILITY IN %)

B. Validation of the Processing Module for ADL Classification

Using the acceleration values corresponding to the activities
performed by the volunteers of Table III and a linear approximation between each pair of control points (see Fig. 4) to approximate the accelerometric curve that describes the reference
level for the vertical acceleration, an iterative approximation
process was carried out to find robust values in the parameters of the activity classification algorithm. Table VI shows the
values obtained for the different parameters.
Tables VII and VIII show the results of an analysis of sensitivity and specificity of the classification algorithm performed
with each of the parameters. The best results (100% sensitivity
and specificity) were obtained with nominal parameter values,
although some thresholds like Hgait |up or Asum |th were robust
over a wide range of values. Significant variations in the parameters did not significantly decrease the sensitivity values, keeping
above 95% in the majority of cases. Specificity analysis showed
better results. This processing module operates independently
of the other processing modules, so that any errors that may
appear in them do not affect its results.

Using the EA variables values that correspond to the activities

undertaken by the volunteers in Table III, a linear correlation
between the average values of the logarithm of the variable EA
with the logarithm of their equivalent in METs was established
by means of the least-squares method. This approximation was
valid for all the activities except for descending stairs, where
the associated metabolic expenditure was overestimated, and
stairs-up, where it was underestimated. Such results were predictable, since climbing stairs is a much bigger effort compared
to a horizontal movement, due to the opposition of the gravity force. By contrast, descending stairs takes advantage of the
contrary effect. In order to take into account this effect, a new
correspondence between the EA values and their approximate
equivalent in METs was established, now without considering
the activities of descending and climbing stairs. Specifically, for
such activities, the scaling factor SF that allowed the deviation
to be corrected was obtained. The obtained results allowed us
to establish the following equation for metabolic expenditure
estimation:
ME = 0.2175 (SF EA )0.237 m t

0.0679, for descending stairs

SF 9.397, for climbing stairs

1, otherwise

(5)

where ME is the metabolic expenditure associated with the activity in kJ, m is the mass of the subject in kg, and t is the
elapsed time of the activity in minutes.
Fig. 5 shows graphically the correspondence between the
logarithm of the corrected EA average value for each of the
activities with the logarithm of its equivalent in METs. In it, the
line represents the values predicted by the proposed equation.
The mean square error for estimating the metabolic rate was
18 J/s in the activities developed in the experiments (according
to [31], sleeping metabolic rate would be around 89 J, and
climbing stairs around 790 J for a 85-kg person). This processing
module, which, given its simplicity, can be executed in real time
within the SoM, will provide an estimation of the metabolic cost
in conjunction with the other two processing modules. This way,
errors in estimating EA or in the classification of the activity can
lead to errors to estimate the metabolic rate. The results shown
in Table IX corroborate this statement.

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 11, NOVEMBER 2012

Fig. 5. Correspondence between log 1 0 (E A ) and log 1 0 (METs) in the performed activities.
TABLE IX
MEAN SQUARE ERROR (IN JOULES PER SECOND) OF THE PROCESSING MODULE
FOR THE CLASSIFICATION OF ACTIVITY LEVEL

3183

ent users independently of their anthropometric characteristics,

customizable through the mass of the user, which takes into account the effects of gravity when climbing or descending stairs,
and that provides a realistic value independently of the activity
performed.
A set of experiments developed by a cohort of young volunteers has demonstrated the feasibility of the prototype and the
proposed classification and estimation algorithms for human
activity monitoring. Currently, we are preparing the monitoring
system for its validation with elderly persons. The information provided by the device will ease the promotion of healthy
lifestyles in the elderly and chronic patients through the motivation of exercise and diet control for an active and healthy
ageing in their own homes. In addition, the monitoring proposed system could be used by the medical staff to track and
prescribe physical exercises for an optimal preventive control,
as well as for the detection of alarm conditions originated by
unusual behaviors of the subject (not getting up from bed at
regular hours, no detected activity during a defined time period)
or other changes in routinely activities related to psychomotor
pathologies (Alzheimers, dementia, multiple sclerosis, etc.).
The low cost of the SoM can be analyzed from two related
perspectives. First of all, and taking into account the distributed
processing performed by the monitoring system, the embedded
algorithms are simpler and thus do not require complex devices.
Second, a convenient design methodology, like the followed to
develop the SoM, permits the seek for sustainable devices, in
terms of power consumption and costs, which optimize the bill
of materials to be used in SoM implementation. Besides, this design approach eases the multimodal operation of the SoM thanks
to its modular general scheme, which allows the integration of
other sensor devices within the SoM, avoiding a price rise.
REFERENCES

V. CONCLUSION
This paper has presented the hardware and software design of
a wearable and low-cost smart sensor of accelerometry for the
monitoring of physical activity, whose specifications have been
preset for it to be applied to the majority of possible scenarios,
such as subject homes or residential centers for seniors, both for
remote monitoring in real time as a local application executed
offline.
The monitoring of human movement is approached in a holistic manner in the device through a series of processing modules:
1) Processing module for the classification of activity level,
which can be used to know the distribution and intensity of activity and inactivity patterns to help to quantify the thresholds
of minimum required levels of activity, so as to improve the
prescription of physical exercise by the medical staff to obtain
optimal preventive results. 2) Processing module for ADL classification, able to distinguish the performing of activities such
as climbing or descending stairs, for which it has been proposed
a simple method that decouples the components of acceleration
associated with the movement of the gravitational components.
3) Processing module for metabolic expenditure estimation, for
which a simple equation has been proposed, valid for differ-

[1] M. Ferreira, S. Matsudo, M. Ribeiro, and L. Ramos, Health-related factors correlate with behavior trends in physical activity level in old age:
Longitudinal results from a population in Sao Paulo, Brazil, BMC Public
Health, vol. 10, 2010.
[2] N. Zouba, F. Bremond, and M. Thonnat, An activity monitoring system
for real elderly at home: Validation study, in Proc. IEEE Int. Conf. Adv.
Video Signal Based Surveillance, 2010, pp. 278285.
[3] (2012). World health organization web site, global health risks: Mortality and burden of disease attributable to selected major risks. [Online].
Available: http://www.who.int/healthinfo/global_burden_disease/GlobalHealthRisks_report_full.pdf
[4] E. Groessl, R. Kaplan, W. Rejeski, J. Katula, A. King, G. Frierson,
N. Glynn, F.-C. Hsu, M. Walkup, and M. Pahor, Health-related quality of life in older adults at risk for disability, Amer. J. Preventive Med.,
vol. 33, no. 3, pp. 214218, 2007.
[5] M. Hirvensalo, T. Rantanen, and E. Heikkinen, Mobility difficulties and
physical activity as predictors of mortality and loss of independence in the
community-living older population, J. Amer. Geriatrics Soc., vol. 48,
no. 5, pp. 493498, 2000.
[6] J. Stessman, R. Hammerman-Rozenberg, A. Cohen, E. Ein-Mor, and
J. Jacobs, Physical activity, function, and longevity among the very old,
Archives Internal Med., vol. 169, no. 16, pp. 14761483, 2009.
[7] S. Lord, S. Chastin, L. McInnes, L. Little, P. Briggs, and L. Rochester, Exploring patterns of daily physical and sedentary behaviour in communitydwelling older adults, Age Ageing, vol. 40, no. 2, pp. 205210, 2011.
[8] J. H. Lim, A. Zhan, J. Ko, A. Terzis, S. Szanton, and L. Gitlin, A closedloop approach for improving the wellness of low-income elders at home
using game consoles, IEEE Commun. Mag., vol. 50, no. 1, pp. 44 51,
Jan. 2012.

3184

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 11, NOVEMBER 2012

[9] B. Silva, I. Lopes, J. Rodrigues, and P. Ray, Sapofitness: A mobile health

application for dietary evaluation, in Proc. 13th IEEE Int. Conf. e-Health
Netw. Appl. Serv., Jun. 2011, pp. 375380.
[10] L. Yan, L. Yue-e, and H. Jian, Gait recognition based on MEMS accelerometer, in Proc. 10th IEEE Int. Conf. Signal Process., Oct. 2010,
pp. 16791681.
[11] T. Shany, S. J. Redmond, M. R. Narayanan, and N. H. Lovell, Sensorsbased wearable systems for monitoring of human movement and falls,
IEEE Sens. J., vol. 12, no. 3, pp. 658670, Mar. 2012.
[12] (2012). Actigraph company web site. [Online]. Available: http://www.
theactigraph.com
[13] (2012). Aipermon company web site. [Online]. Available: http://www.
aipermon.com
[14] I. Gyllensten and A. Bonomi, Identifying types of physical activity with
a single accelerometer: Evaluating laboratory-trained algorithms in daily
life, IEEE Trans. Biomed. Eng., vol. 58, no. 9, pp. 26562663, Sep.
2011.
[15] N. Twomey, S. Faul, and W. Marnane, Comparison of accelerometerbased energy expenditure estimation algorithms, in Proc. 4th Int. Conf.
Pervasive Comput. Technol. Healthcare, Mar. 2010, pp. 18.
[16] H. Vathsangam, A. Emken, E. Schroeder, D. Spruijt-Metz, and
G. Sukhatme, Determining energy expenditure from treadmill walking
using hip-worn inertial sensors: An experimental study, IEEE Trans.
Biomed. Eng., vol. 58, no. 10, pp. 28042815, Oct. 2011.
[17] V. Van Hees, M. Pias, S. Taherian, U. Ekelund, and S. Brage, A method
to compare new and traditional accelerometry data in physical activity
monitoring, in IEEE Int. Symp. A World Wireless, Mobile Multimedia
Netw., 2010.
[18] P. Gerdhem, M. Dencker, K. Ringsberg, and K. kesson, Accelerometermeasured daily physical activity among octogenerians: Results and associations to other indices of physical performance and bone density, Eur.
J. Appl. Physiol., vol. 102, no. 2, pp. 173180, 2008.
[19] A. Dinh, D. Teng, L. Chen, Y. Shi, C. McCrosky, J. Basran, and V.
Del Bello-Hass, Implementation of a physical activity monitoring system
for the elderly people with built-in vital sign and fall detection, in Proc.
6th Int. Conf. Inf. Technol.: New Generations, 2009, pp. 12261231.
[20] M. Makikawa, S. Asajima, K. Shibuya, R. Tokue, and H. Shinohara,
Portable physical activity monitoring system for the evaluation of activity
of the aged in daily life, in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol.,
2002, vol. 3, pp. 19081909.
[21] G. Plasqui and K. Westerterp, Physical activity assessment with accelerometers: An evaluation against doubly labeled water, Obesity,
vol. 15, no. 10, pp. 23712379, 2007.

[22] J. Copeland and D. Esliger, Accelerometer assessment of physical activity in active, healthy older adults, J. Aging Phys. Activity, vol. 17, no. 1,
pp. 1730, 2009.
[23] H. Jee, L. Jeongwhan, T. Hyun, P. Jong, C. Jae, and S. Kunsoo, Estimation of activity energy expenditure: Accelerometer approach, in
Proc. Annu. Int. Conf. IEEE Eng. Med. Biol., 2005, vol. 7, pp. 3830
3833.
[24] B. Najafi, K. Aminian, A. Paraschiv-Ionescu, F. Loew, C. Bula, and
P. Robert, Ambulatory system for human motion analysis using a
kinematic sensor: Monitoring of daily physical activity in the elderly, IEEE Trans. Biomed. Eng., vol. 50, no. 6, pp. 711723, Jun.
2003.
[25] C.-C. Yang and Y.-L. Hsu, Algorithm design for real-time physical activity identification with accelerometry measurement, in Proc. Ind. Electron.
Conf., 2007, pp. 29963000.
[26] M. ElSayed, A. Alsebai, A. Salaheldin, N. El Gayar, and M. ElHelw,
Ambient and wearable sensing for gait classification in pervasive healthcare environments, in Proc. 12th IEEE Int. Conf. e-Health Netw. Appl.
Services, Jul. 2010, pp. 240245.
[27] A. Sabatini, Estimating three-dimensional orientation of human body
parts by inertial/magnetic sensing, Sensors, vol. 11, no. 2, pp. 1489
1525, 2011.
[28] N. Wang, S. Redmond, E. Ambikairajah, B. Celler, and N. Lovell,
Can triaxial accelerometry accurately recognize inclined walking terrains? IEEE Trans. Biomed. Eng., vol. 57, no. 10, pp. 25062516, Oct.
2010.
[29] D. Naranjo, L. Roa, J. Reina-Tosina, and M. Estudillo-Valderrama, Personalization and adaptation to the medium and context in a fall detection
system, IEEE Trans. Inf. Technol. Biomed., vol. 16, no. 2, pp. 264271,
Mar. 2012.
[30] D. Curone, G. Bertolotti, A. Cristiani, E. Secco, and G. Magenes, A
real-time and self-calibrating algorithm based on triaxial accelerometer
signals for the detection of human posture and activity, IEEE Trans. Inf.
Technol. Biomed., vol. 14, no. 4, pp. 10981105, Jul. 2010.
[31] B. Ainsworth, W. Haskell, M. Whitt, M. Irwin, A. Swartz, S. Strath,
W. OBrien, D. Bassett, K. Schmitz, P. Emplaincourt, D. Jacobs, and
A. Leon, Compendium of physical activities: An update of activity codes
and met intensities, Med. Sci. Sports Exercise, vol. 32, no. 9 Suppl.,
pp. S498S504, 2000.
[32] O. Aziz, L. Atallah, B. Lo, E. Gray, T. Athanasiou, A. Darzi, and G. Yang,
Ear-worn body sensor network device: An objective tool for functional
postoperative home recovery monitoring, J. Amer. Med. Informat. Assoc.,
vol. 18, no. 2, pp. 156159, 2011.