Sie sind auf Seite 1von 19

sensors

Article
Bed-Embedded Heart and Respiration Rates
Detection by Longitudinal Ballistocardiography
and Pattern Recognition
Almothana Albukhari * , Frederico Lima and Ulrich Mescheder
Medical and Mechanical Engineering Faculty, and Institute for Microsystem Technology (iMST),
Furtwangen University, 78120 Furtwangen, Germany; lima@hs-furtwangen.de (F.L.);
mes@hs-furtwangen.de (U.M.)
* Correspondence: alb@hs-furtwangen.de; Tel.: +49-7723-920-2946

Received: 31 January 2019; Accepted: 19 March 2019; Published: 25 March 2019 

Abstract: In this work, a low-cost, off-the-shelf load cell is installed on a typical hospital bed and
implemented to measure the longitudinal ballistocardiogram (BCG) in order to evaluate its utility
for successful contactless monitoring of heart and respiration rates. The major focus is placed on the
beat-to-beat heart rate monitoring task, for which an unsupervised machine learning algorithm is
employed, while its performance is compared to an electrocardiogram (ECG) signal that serves as a
reference. The algorithm is a modified version of a previously published one, which had successfully
detected 49.2% of recorded heartbeats. However, the presented system was tested with seven volunteers
and four different lying positions, and obtained an improved overall detection rate of 83.9%.

Keywords: ballistocardiography; vital signs monitoring; contactless monitoring; heart rate; load cell;
pattern recognition; unsupervised machine learning; clustering

1. Introduction
In 2015 the number of people aged over 60 years reached almost 900 million worldwide,
and between 2015 and 2050 that number is forecast to increase by 58% in higher income countries and
by higher rates in lower income countries reaching 239% increase in countries at the lower end of the
global income spectrum. This demographic shift and the increased life expectancy is associated with
increased prevalence of chronic diseases like dementia, from which in 2015 around 46.8 million people
had suffered, with a projection of this number almost doubling every 20 years [1].
This epidemic has inspired some new research trends in the medical field, particularly regarding
sensory modalities that can be utilized in monitoring the status of dementia patients and their
vital signs, because the type of sensor will determine the level of intrusiveness on a patient’s life
and daily routines. This is not only true on the hospital bed but also in home care applications,
where incorporating monitoring systems could have major benefits for patients as well as caregivers;
financially, psychologically, and most importantly in terms of quality of provided healthcare [2].
The development of unobtrusive monitoring systems that cause the least amount of discomfort, while
being as ubiquitous in the patients’ environment as possible, is a goal of many recent research efforts,
which could accommodate the special conditions of this growing segment of society. Moreover, tracking
the cardiopulmonary activity of patients is paramount for determining their well-being and a key
indicator for detecting distress situations of patients with dementia.
One of the most promising unobtrusive techniques, which has recently regained considerable
attraction among researchers, is ballistocardiography, which is the measurement of the body recoil
forces resulting from the ejection of blood in the vasculature at each cardiac cycle. BCG was first

Sensors 2019, 19, 1451; doi:10.3390/s19061451 www.mdpi.com/journal/sensors


Sensors 2019, 19, 1451 2 of 19

Sensors 2018, 18, x FOR PEER REVIEW 2 of 19


observed by Gordon in 1877, when he noticed the index of a weighing scale having rhythmic movement
movement synchronously
synchronously with the pulse with
of athe pulsestanding
subject of a subject standing
vertically vertically
on it, on it, he then
he then reproduced reproduced
a similar motiona
similar motion in a horizontal hammock-wise suspension of the subject [3].
in a horizontal hammock-wise suspension of the subject [3]. The scientific research into BCG surged The scientific research
into BCG
later in thesurged later an
mid-1900s, in era
theknown
mid-1900s,as theangolden
era known
age ofasBCGthe[4,5].
goldenTheagemostof prominent
BCG [4,5]. pioneer
The most in
prominent
this regard pioneer
was Isaac inStarr,
this regard was Isaac what
who published Starr,iswho published
considered thewhat
first is considered
scientific theinfirst
works thisscientific
field [6]
works
and in headed
later this fieldthe[6] and later on
Committee headed the Committee on
Ballistocardiographic Ballistocardiographic
Terminology [7]; since hisTerminology
first publication, [7];
since
he his firsttopublication,
continued investigate he BCG continued to investigate
for decades [5]. Starr wasBCGtheforfirstdecades
to present[5].aStarr
model was the signal’s
of the first to
commonest morphology with a proposed terminology [6], which was later adopted and promotedwas
present a model of the signal’s commonest morphology with a proposed terminology [6], which by
lateraforementioned
the adopted and promoted
committeeby[8], the aforementioned
Figure 1 depicts thecommittee
signal with [8],
theFigure 1 depicts
terminology the signal
proposed with
by Starr
thethat
at terminology
time, which proposed
was laterbyelaborated
Starr at that time,
on by which
other was laternotably
researchers, elaborated on by other
by reporting researchers,
preceding F and
notably by reporting preceding F and G waves [9,10]. However, the most
G waves [9,10]. However, the most prominent segments of the signal are H–L, which represent theprominent segments of the
signal are
systole H–L,ofwhich
phase represent
the cardiac cyclethe systole
and should phase of the cardiac cycle
be distinguishable in a and should
healthy be distinguishable
subject under normal
in a healthy subject
circumstances, since theunder normalcomponents
non-systolic circumstances, since
are less the non-systolic
perceptible, especiallycomponents are rate,
with a fast heart less
perceptible, especially with a fast
when superimposition of waves is evident [4].heart rate, when superimposition of waves is evident [4].

Figure 1.1. AADiagram


Figure Diagram of of
thethe
commonest formform
commonest of theofnormal ballistocardiogram
the normal with thewith
ballistocardiogram lettersthe
denoting
letters
its various waves (tips), redrawn after Starr’s publication [6].
denoting its various waves (tips), redrawn after Starr’s publication [6].

The body of research at that time aimed at exploiting the technique’s potential for medical diagnosis;
The body of research at that time aimed at exploiting the technique’s potential for medical
however, the lack of consensus among researchers about the physiological genesis of the signal’s
diagnosis; however, the lack of consensus among researchers about the physiological genesis of the
morphology, the failure to adhere to standardized techniques for measurements, and the ambiguity
signal’s morphology, the failure to adhere to standardized techniques for measurements, and the
in interpretation of the results, given the inherent great level of scrutiny a medical diagnostic tool is
ambiguity in interpretation of the results, given the inherent great level of scrutiny a medical
naturally placed under, led to its effective dismissal from the medical community in the late 1970s [11].
diagnostic tool is naturally placed under, led to its effective dismissal from the medical community
In the last two decades, however, the research community has revived its interest in BCG with substantial
in the late 1970s [11]. In the last two decades, however, the research community has revived its
momentum. Several factors have contributed to this phenomenon. Most notably, the BCG has a great
interest in BCG with substantial momentum. Several factors have contributed to this phenomenon.
capacity to be realized in an unobtrusive sensory form and be embedded with different configurations
Most notably, the BCG has a great capacity to be realized in an unobtrusive sensory form and be
in the patient’s environment. Secondly, owing to its physical nature, it is able to convey key medical
embedded with different configurations in the patient’s environment. Secondly, owing to its physical
information about the cardiovascular system, which might be otherwise unattainable [10], e.g., the force
nature, it is able to convey key medical information about the cardiovascular system, which might be
of the heart’s contraction, which is a key indicator of the heart’s physiologic age and its decline could be
otherwise unattainable [10], e.g., the force of the heart’s contraction, which is a key indicator of the
an early warning of a heart disease [12]. Furthermore, the recent technological advancements in sensory
heart’s physiologic age and its decline could be an early warning of a heart disease [12]. Furthermore,
fabrication, signal processing, and computational powers facilitated the resurgence of BCG.
the recent technological advancements in sensory fabrication, signal processing, and computational
Considering the nature of BCG being the measurement of the body motion that comes about
powers facilitated the resurgence of BCG.
to oppose the motion of mass within it, primarily blood, as Newton’s third law of motion dictates,
Considering the nature of BCG being the measurement of the body motion that comes about to
and the fact that the motion of blood in the body is not confined in a single direction, making
oppose the motion of mass within it, primarily blood, as Newton’s third law of motion dictates, and
BCG practically omnidirectional, this made it possible to measure BCG along several directions.
the fact that the motion of blood in the body is not confined in a single direction, making BCG
Consequently, the Committee on Ballistocardiographic Terminology recommended three standard
practically omnidirectional, this made it possible to measure BCG along several directions.
spatial axes of measurement: longitudinal (head-foot), transverse (side to side), and dorsoventral
Consequently, the Committee on Ballistocardiographic Terminology recommended three standard
or anteroposterior (back to chest) [8]. However, it has been reported that a BCG measured in the
spatial axes of measurement: longitudinal (head-foot), transverse (side to side), and dorsoventral or
longitudinal direction, particularly of a bedridden subject, has stronger cardiac output components
anteroposterior (back to chest) [8]. However, it has been reported that a BCG measured in the
than those measured in the other directions [13,14]. Additionally, the signal’s morphology of a
longitudinal direction, particularly of a bedridden subject, has stronger cardiac output components
than those measured in the other directions [13,14]. Additionally, the signal’s morphology of a
transverse or a dorsoventral BCG is highly dependent on the subject and the subject’s position on the
Sensors 2019, 19, 1451 3 of 19

transverse or a dorsoventral BCG is highly dependent on the subject and the subject’s position on the
bed, more so than that of a longitudinal BCG, and, by the same token, the morphological variability of
longitudinal BCG is lesser than that of the others [14]. Hence, unlike most modern bed-embedded
BCG systems, the BCG in this work is set up to be measured along the longitudinal direction in an
attempt to obtain the best possible signal to work with.
Since there are different ways to take account of the reaction motion of the body; namely,
via displacement, velocity, or acceleration, and with each of these measurands being measurable through
various sensor modalities, ballistocardiographers have experimented with various BCG systems throughout
the decades. Furthermore, in earlier times, notwithstanding all setups being table-based, they differed with
regard to the implemented technology and the resulting setup’s natural frequency, but they were usually
labelled according to the latter. To name a few, the earliest BCGs, by Gordon [3] and Henderson [15],
were deemed ultra-low frequency BCGs, on the other hand, Starr’s spring-coupled undamped table
produced high-frequency BCG [6], and Nickerson’s apparatus, capable of continuously adjustable damping,
presented low-frequency BCG under critical damping [16]. Later, Dock introduced a different measurement
process that provided direct body BCG, however, he kept the table-based setup [17]. Each of these types
produced a different common BCG waveform [8] while attempting to address drawbacks or forms of
distortion that other types may have suffered [4].
As mentioned previously, when the potential of the BCG started to be recognized in the medical field,
it was subjected to great scrutiny to substantiate its value as a cardiological examination. This necessitated
the aforementioned classification of BCG systems, as they appeared, and the standardization of their
respective morphologies because medical diagnosis needs above all to establish a norm to contrast
anomalies against. Nevertheless, with the recent interest in BCG shifting its objective towards ubiquitous
unobtrusive monitoring, the recent novelties in the field manifested mainly in the implemented sensors,
the manner of embedding sensors in patients’ surroundings, and the techniques and algorithms employed
in signal processing and data extraction. In this respect, BCG measurements have been installed in
weighing scales [18–20], which have the benefit of pure longitudinal BCG, however, practical constraints
on long measurement times is a major disadvantage. Additionally, chair-based BCG is popular among
researchers, where BCG transducers can be embedded in the seat [21], backrest [22], or even headrest [23];
moreover, equipping wheelchairs with cardiac monitoring has also been explored [24]. At the same time,
the widespread utilization of BCG in sleep monitoring applications, made embedding BCG sensors in beds
quite prevalent [14,25–28], even in air mattresses [29] and pillows [30].
The BCG transducers found in recent literature are even more diverse. For example, we find
film sensors, like electromechanical film (EMFi) [22,24] and polyvinylidenefluorid film (PVDF) [21,25];
strain gauges, which are dominant in but not limited to weighing scales [14,18–20]; MEMS
accelerometers [24]; fiber optical sensors [23]; piezoelectric pressure sensors [27]; hydraulic pressure
sensors [26]; and pneumatic systems based on air pressure sensors employed in elaborated
configurations [28,29]. It is worth pointing out that this variety of transducers has further increased the
variability of the BCG waveform, which, for the most part, still relates to the conventional waveform
depicted in Figure 1, however, significant alterations can often be noticed.
Furthermore, given the morphological nature of the BCG signal distinguishing individual
heartbeats in it is not usually straightforward, unlike, for example, detecting the sharp QRS complex
in ECG. Thus, various methodologies are constantly being developed, some are as simple as an
adaptive threshold to detect the IJ segment (Figure 1) [19], or a basic peak detection algorithm applied
after back-and-forth transformation between time and frequency domains for signal screening and
cleaning [23]. On the other hand, BCG measurement setups with their resulting waveforms usually
required more sophisticated approaches, which have made use of, among others, unsupervised
machine learning techniques like hierarchical clustering [27], or k-means clustering [26], which in
Reference [14] was part of an elaborate detection algorithm based on three different methods, preceded
by a training phase: cross-correlation, modified k-means clustering, and heart valve function, which
was concluded by weighting each method’s outcome through reliability functions. Other useful tools
Sensors 2019, 19, 1451 4 of 19

often found in the literature are the continuous wavelet transform (CWT) [20] as well as the discrete
wavelet transform (DWT) [31].
In this paper, the longitudinal BCG that is measured through a highly reliable, and readily
available sensor, in the form of a load cell installed on a bed, is assessed, and the likelihood of
successful heart and respiration rates monitoring through it is evaluated. An unsupervised machine
learning method is employed for heartbeat monitoring, while its performance is referenced to an
ECG signal that serves as ground truth. The algorithm is a modified version of the one found in
Reference [27], which reported a successful detection rate of 49.2% of recorded heartbeats. The main
purpose of this project is to improve the heartbeat recognition rate and to investigate the influence of
lying positions on recognition probability.

2. Materials and Methods

2.1. Measurement Setup


The measurement system is comprised of a force sensor to measure body motions, namely, those
induced by cardiac and respiration activities; an ECG sensor, which establishes a reference for the
evaluation of the heart rate detection; and a data acquisition tool.

2.1.1. Force Sensor


The force sensor utilized to acquire the BCG signal is convenient in a number of respects. It is a
low-cost, off-the-shelf load cell that is usually utilized in weighing scales (3D-Mechatronics, Bremen,
Germany). It is practically very reliable with low demands on maintenance. It comprises an aluminum
beam with a hollowed section in the middle, making it less resistant to bending forces, and two strain
gauges placed on either side of the hollowed part, forming a two-element bridge circuit. This Wheatstone
bridge setup has a very reasonable linear response and sensitivity. The load cell is interfaced through an
electronic circuit based on the 24-bit HX711 analog-to-digital converter (AVIA Semiconductor, Xiamen,
China). The transducer is suited for up to around 50 N (5 kg) of load, which when calibrated was found to
have a theoretical resolution of about 24 µN per A/D count over its operational range. The accompanying
circuit with the built-in oscillator is capable of providing up to an 80 Hz sampling frequency.
The influence of the force sensor placement on the resulting signals was initially investigated. For this
reason, a mechanical structure was designed, which sandwiches the load cell between two parallel plates,
with the load cell beam being fixed to each plate at its opposing sides. Firstly, in order to obtain the
transverse/dorsoventral signal the sensor was placed horizontally between the mattress and the bed base,
below the upper part of a volunteer’s body, similar to the placement of most modern bed-embedded BCG
sensors. The upper plate, facing the mattress, is designated as the sensing surface, while the lower one,
resting on the bed base, as the stationary plate. Figure 2 (a) depicts the sensor’s structure diagram in this
configuration, while (b) shows the actual placement of the sensor in relation to the bed. Secondly, in order
to measure the BCG signal in the longitudinal direction, the sensor was set up vertically to sense the body
motions transmitted through the mattress at its top side. In this configuration, which is the one adopted
for this monitoring system, the stationary plate is made as such by a structure that is attached firmly to
the slotted bed base, while the sensing surface faces the top side of the mattress. Figure 3 (a) displays
the structure diagram of this configuration, (b) the complete mechanical structure for the longitudinal
measurement, and (c) shows the structure fitted on the bed.

2.1.2. ECG Sensor


The ECG signal is recorded by means of a single-lead heart rate monitor board (SparkFun
Electronics, Niwot, CO, USA), which has an integrated signal conditioning chip intended for ECG
measurement, named AD8232 (Analog Devices, Norwood, MA, USA). This tool provides the ECG
signal in 0–3.3 V analog form; therefore, its conversion to digital form is carried out by the data
acquisition system.
Sensors 2019, 19, 1451 5 of 19
Sensors
Sensors2018,
2018,18,
18,xxFOR
FORPEER
PEERREVIEW
REVIEW 5 5ofof19
19

(a)
(a) (b)
(b)
Figure2.2.2.The
Figure
Figure The Load
TheLoad cell
Loadcell setup
cellsetup
setupto totomeasure
measure
measurethethebody
the bodyforces
body forcesininthe
forces thetransverse/dorsoventral
transverse/dorsoventral
transverse/dorsoventral axis
axis for
forfor
axis
comparison
comparison
comparison with
with
with the
the signal
thesignal in the
signalininthe longitudinal
thelongitudinal axis:
longitudinal axis: (a)
axis: (a) The
(a) The load
The load cell
loadcell structure
cellstructure diagram;
structurediagram;
diagram; (b) the
(b)(b) actual
thethe
actual
actual
sensor
sensor
sensor placement
placement
placement relative totothe
relativeto
relative the bed;
bed;the
thebed; themattress
the mattressisis
mattress islifted
liftedtemporarily
lifted temporarilyfor
temporarily forthe
forthesake
thesakeofofof
sake the
the picture.
thepicture.
picture.

(a) (b) (c)


(a) (b) (c)
Figure 3. The load cell setup for the longitudinal BCG: (a) The load cell structure diagram; (b) the
Figure3.3.The
Figure Theload
load cell
cell setup
setup for the
the longitudinal
longitudinalBCG:BCG:(a) (a)The
Theload
load cell
cell structure
structure diagram;
diagram; (b) (b)
the the
complete mechanical structure; (c) the structure fitted at the top side of the bed base. The clamps fit
completemechanical
complete mechanicalstructure;
structure;(c)
(c)the
thestructure
structurefitted
fitted at
at the
the top
top side
side of
of the
the bed
bed base. The
The clamps
clamps fit
fit on
on the bed base from underneath.
onbed
the the base
bed base
fromfrom underneath.
underneath.
2.1.3. Data
2.1.3. Acquisition
2.1.3.Data
DataAcquisition
Acquisition
TheThe two sensors are interfaced through an Arduino Uno microcontroller (Arduino AG,
Thetwo twosensors
sensors are are
interfaced through
interfaced an Arduino
through an ArduinoUno microcontroller
Uno microcontroller (Arduino AG, Somerville,
(Arduino AG,
MA, Somerville,
USA); Figure MA, USA);
4 shows Figure
the44complete
shows the complete measurement system. The microcontroller reads
Somerville, MA, USA); Figure shows themeasurement system. The
complete measurement system.microcontroller
The microcontroller reads the
reads force
the
sensor force sensor
by communicatingby communicating
with thewith with
HX711 the HX711 chip
chip through through the
the Serial Serial Peripheral
Peripheral Interface
Interface (SPI) (SPI)
at the
the force sensor by communicating the HX711 chip through the Serial Peripheral Interface (SPI)
at the maximum available sampling frequency, dictated, as mentioned, by the HX711 chip. The
maximum available
at the maximum sampling
available frequency,
sampling dictated,
frequency, as mentioned,
dictated, by the HX711
as mentioned, chip. The
by the HX711 sampling
chip. The
sampling frequency in the case of the board at hand is around 84.8 Hz. On the other hand, the analog
sampling frequency
frequency in the case of in the case of the board at hand is around 84.8 Hz. On the other
board at hand is around 84.8 Hz. On the other hand, the analog ECG signal hand, the analog
ECG signal is sampled at three times that rate, i.e., 254.4 Hz. The microcontroller carries out the
is ECG
sampledsignalat is sampled
three times at three
that rate,times that rate,
i.e., 254.4 i.e., microcontroller
Hz. The 254.4 Hz. The microcontroller
carries out the carries
sampling outatthefixed
sampling at fixed time intervals with proper independent timestamping of both signals. Additionally,
sampling
time at fixed
intervals withtime
proper intervals with proper
independent independent
timestamping of timestamping of both signals.the
both signals. Additionally, Additionally,
force sensor is
the force sensor is scaled to a force unit of millinewton (mN), as per previously established calibration
the force
scaled to a sensor
force is scaled
unit of to a force unit
millinewton of millinewton
(mN), (mN), as per previously
as per previously established calibration
parameters, before forwarding both measurands serially to established
the PC alongcalibration
with theirparameters,
timestamps.beforeOn
parameters,
forwarding before
both forwarding
measurands both
serially measurands
to the PC serially
along with to the
their PC along
timestamps. withOn their
the timestamps.
PC side, On
a desktop
the PC side, a desktop application organizes and saves the transmitted data, namely, the Processing
the PC side,
application a desktop application theorganizes anddata,
savesnamely,
the transmitted data, namely, the Processing
Integratedorganizes
Development and saves
Environment transmitted
(IDE) (Processing the Processing
Foundation, Integrated
processingfoundation.org).Development
Integrated Development Environment (IDE) (Processing Foundation, processingfoundation.org).
Environment (IDE) (Processing Foundation, processingfoundation.org).
2.2. Signal Preprocessing
2.2.Signal
2.2. SignalPreprocessing
Preprocessing
The force sensor in its basic setup delivers a signal containing various components. Naturally, it
The force sensor in its basic setup delivers a signal containing various components. Naturally, it
The
senses force
all motionssensorinducedin its basic
by the body,setup delivers
whether caused a by signal
cardiac containing
or respiratory various components.
activities, or any
senses all motions induced by the body, whether caused by cardiac or respiratory activities, or any
Naturally,
voluntary it senses
muscleall motions induced
movements. by the depending
Additionally, body, whether on thecaused by cardiacvibrations
surroundings, or respiratory activities,
through the
voluntary muscle movements. Additionally, depending on the surroundings, vibrations through the
or any voluntary muscle movements. Additionally, depending on the surroundings,
floor might be picked up in the signal, not to mention the electromagnetic interference (EMI), and the vibrations through
floor might be picked up in the signal, not to mention the electromagnetic interference (EMI), and the
theDC floor might bepresent
component pickeddue uptointhethenecessary
signal, not to mention
preloading the electromagnetic
by initial interference
contact with the mattress. Figure(EMI),
5
DC component present due to the necessary preloading by initial contact with the mattress. Figure 5
and displays
the DCtwo frequency
component spectrum
present due diagrams of the sensor’s
to the necessary preloading rawby signal that
initial are truncated
contact with the for the
mattress.
displays two frequency spectrum diagrams of the sensor’s raw signal that are truncated for the
convenience
Figure 5 displays of two
display. These two
frequency sample signals were
spectrum taken from rawtwo volunteers lying in the supine
convenience of display. These two samplediagrams
signals wereof the
takensensor’s
from two signal thatlying
volunteers are truncated for the
in the supine
convenience of display. These two sample signals were taken from two volunteers lying in the supine
Sensors 2019, 19, 1451 6 of 19
Sensors 2018,
Sensors 2018, 18,
18, xx FOR
FOR PEER
PEER REVIEW
REVIEW 66 of
of 19
19

position; both
position;
position; bothare
both depicted
are
are depictedtoto
depicted toshowcase
showcasethe
showcase thevariability
the variability of
variability the BCG
of the
the BCGwaveform
BCG waveformeven
waveform evenin
even inin the
the
the frequency
frequency
frequency
domain.
domain. In order to separate the sought-after components from the aforementioned artifacts as
domain.In order
In to
order separate
to separatethe sought-after
the sought-aftercomponents
components from
fromthe
theaforementioned
aforementioned artifacts
artifacts as well
as well as
well
as from
from
as from one another,
one another,
one another, signalsignal filtration
filtration
signal is implemented
implemented
is implemented
filtration is digitally
digitally by means
means
by means
digitally by of Matlab
Matlab
of Matlab
of (The
(The MathWorks,
MathWorks,
(The MathWorks, Inc.,
Inc.,
Natick, Natick,
MA, MA,
USA). USA).
Inc., Natick, MA, USA).

Figure4.4.
Figure
Figure 4.The
Thecomplete
The complete measurement
complete measurement system.
measurement system.
system.

Figure
Figure
Figure
Frequency
5. 5.
5. Frequencyspectrum
Frequency spectrumdiagrams
spectrum diagrams of
diagrams
ofraw
of rawBCG
raw BCGsignals
BCG signals
signals taken
taken
taken
from
from
from twotwo
two
volunteers,
volunteers,
volunteers, bothboth
both lyinglying
lying in
in
in the
the supine
supine position.
position. The
The major
major signal
signal components
components are
are annotated.
annotated. The
The upper
upper
the supine position. The major signal components are annotated. The upper parts of the diagrams parts
parts of of the
the diagrams
diagrams
have been
have
have truncated
been
been since
truncated
truncated the magnitude
since
since the magnitude
the of theof
magnitude ofDC component
the
the DC component
DC (the preload)
component extends extends
(the preload)
(the preload) to the thousands.
extends to the
to the
Thethousands. The clear difference in the distribution of the frequency components of the two signalsthe
clear
thousands.difference
The in
clear the distribution
difference in the of the frequency
distribution of thecomponents
frequency of the
componentstwo signals
of the reflects
two signals
reflects the
variability
reflects the variability
of variability of the
the BCG waveforms
of the BCG
BCG waveforms
waveforms
across people.acrossThe
across people. The significant
significant
people. The significant
variation variation in amplitudes
in amplitudes
variation in amplitudes
overall
overallthe
between
overall between
between the two
two signals
the two signals
will will be
be discussed
signals will be discussed
discussed
later. Givenlater.
later. Given
that that the
the Nyquist
Given that the Nyquist
Nyquist
frequency frequency (half
(half the(half
frequency the
sampling
the
samplingisfrequency)
frequency)
sampling frequency)
42.4 Hz, an is 42.4
is 42.4 Hz,
alias Hz, an alias
of the
an alias
EMI ofof the
(50the EMI
Hz)
EMI (50 Hz)
appears
(50 Hz)atappears
appears
34.8 Hz. at 34.8
at 34.8 Hz.
Hz.

TheThereported
The reportedfrequency
reported frequencybandwidth
frequency bandwidth of
bandwidth of interest
of interest for
interest the
for the cardiac
the cardiac component
cardiac component
componentvaries varies among
variesamongamong
publications,
publications, where
where thethe cutoff
cutoff frequencies
frequencies range
range between
between 0.1–1 Hz
Hz for
for
publications, where the cutoff frequencies range between 0.1–1 Hz for high-pass filters, and 10–25 Hz high-pass
high-pass filters, and
filters, and10–25
10–25HzHz
for
forfor low-pass
low-pass
low-pass filters
filters [32].
[32].
filters [32]. Previously,
Previously,
Previously, namely
namely
namely in
inin Reference
Reference
Reference [33],
[33], we
[33], we assumed
weassumed
assumedthe the bandwidth
thebandwidth of
bandwidthofof0.5–20 0.5–20
0.5–20Hz
andHz
Hz and carried
carried
and carried
out the out
out the filtration
the filtration
filtration accordingly;
accordingly;
accordingly;however,however,
however, since
sincesince then,
then,then, with
withwith further
further experimentation
experimentation
further experimentation andand
new
and
new measurements
new measurements
measurements taking
taking taking place,
place, place, the bandwidth
the bandwidth
the bandwidth 0.5–15
0.5–150.5–15
Hz, was Hz, was
Hz, found found
was found empirically
empirically
empirically to yield
to yield
to yield overall
overall
overall more
moreresults
more
fruitful fruitfulasresults
fruitful results
far as asas far
heartfarrate
as heart
as heart rate monitoring
rate
monitoring monitoring
is concerned. is concerned.
is concerned. Moreover,
Moreover,Moreover,
the BCG signal the BCG
the BCG signal is
signal
is obtained isby
obtained by implementing
obtained by implementing finite impulse response filters (FIR), which are advantageous in the sense
implementing finite impulsefinite responseimpulse response
filters filters (FIR),
(FIR), which which are advantageous
are advantageous in the senseinthat the sense
they do
that they
that they dodo not
not induce
induce any any phase
phase distortion
distortion in in the
the filtered
filtered signal
signal because
because theythey maintain
maintain aa perfectly
perfectly
not induce any phase distortion in the filtered signal because they maintain a perfectly linear phase
linear phase
linear phase response.
response. Specifically,
Specifically, the the fircls1
fircls1 filter
filter function
function is is used,
used, which
which is an an inbuilt
inbuilt function
function in
response. Specifically, the fircls1 filter function is used, which is an inbuiltisfunction in Matlab in that
Matlab that
Matlab that facilitates
facilitates additionally
additionally the the fine-tuning
fine-tuning of of the
the levels
levels of of ripple
ripple andand attenuation
attenuation in in the
the
facilitates additionally the fine-tuning of the levels of ripple and attenuation in the passband and
passband and
passband and stopband,
stopband, respectively.
respectively. The The lengths
lengths of of the
the low-
low- and
and high-pass
high-pass filters
filters were
were 80 80 and
and 540,
540,
stopband, respectively.
respectively; whereas
whereas the
Thelevels
lengths
the levels
of theand
of ripple
ripple and
low- and high-pass
attenuation in in both
filters
both filters
were
filters were
were set
80 toand
set to
540, respectively;
0.005% andand 0.01%,
0.01%,
respectively; of attenuation 0.005%
whereas the
respectively. levels
In of
the ripple
upper and
plot attenuation
of Figure 6 a in both
sample filters
raw loadwere
cellset to
signal 0.005%
is and
plotted, 0.01%,
while itsrespectively.
realigned
respectively. In the upper plot of Figure 6 a sample raw load cell signal is plotted, while its realigned
In filtered
the upper plot of Figure
counterpart is shown
shown 6 ainsample
the bottom rawplot.
bottom load cell signal is plotted, while its realigned filtered
filtered counterpart is in the plot.
counterpart is shown in the bottom plot.
Sensors 2019, 19, 1451 7 of 19
Sensors 2018, 18, x FOR PEER REVIEW 7 of 19

Figure
Figure 6. 6. A
A sample
sample load
load cell
cell signal. The upper
signal. The upper plot
plot shows
shows thethe raw
raw signal,
signal, while
while the
the lower
lower one
one shows
shows
its counterpart. The filtration is carried out by high- and low-pass fircls1
its filtered counterpart. The filtration is carried out by high- and low-pass fircls1 filters with 0.5 and
filtered filters with 0.5 and
20
20 Hz
Hz cutoff
cutofffrequencies,
frequencies,respectively.
respectively.TheThemagenta
magenta lines’ intersections
lines’ with
intersections thethe
with time axesaxes
time highlight the
highlight
delay caused by filtration; about 3.7
the delay caused by filtration; about 3.7 s.s.

On the other hand, the respiration component is obtained through a single 3rd order band-pass
On the other hand, the respiration component is obtained through a single 3rd order band-pass
Butterworth filter that belongs to the infinite impulse response filters (IIR) category, which is also an
Butterworth filter that belongs to the infinite impulse response filters (IIR) category, which is also an
inbuilt function in the Matlab environment called butter. This filter is designed to have a passband that
inbuilt function in the Matlab environment called butter. This filter is designed to have a passband
is as flat as possible, meaning minimum ripple, while retaining decent phase linearity throughout that
that is as flat as possible, meaning minimum ripple, while retaining decent phase linearity throughout
passband. Additionally, as an IIR filter, it requires less computational power and induces less delay
that passband. Additionally, as an IIR filter, it requires less computational power and induces less
in the signal. The adopted bandwidth for this component also varied in BCG-related publications;
delay in the signal. The adopted bandwidth for this component also varied in BCG-related
however, more medically specialized research has reported that the normal respiration rate of elderly
publications; however, more medically specialized research has reported that the normal respiration
people above the age of 65 is 8–30 respirations per minute (2.5–97.5 percentile), while younger people
rate of elderly people above the age of 65 is 8–30 respirations per minute (2.5–97.5 percentile), while
have a narrower distribution, however well within these boundaries [34]. Those rates correspond to a
younger people have a narrower distribution, however well within these boundaries [34]. Those rates
frequency bandwidth of 0.13–0.5 Hz; nevertheless, the employed band-pass filter has a slightly wider
correspond to a frequency bandwidth of 0.13–0.5 Hz; nevertheless, the employed band-pass filter has
passband with the cutoff frequencies 0.05 Hz and 0.5 Hz.
a slightly wider passband with the cutoff frequencies 0.05 Hz and 0.5 Hz.
Even though the ECG signal is more susceptible to noise, especially when a volunteer lays prone
over the electrodes, and considerable levels of noise are sometimes encountered, yet the QRS complex
Even though the ECG signal is more susceptible to noise, especially when a volunteer lays prone
remained distinct enough during recorded measurements, hence, the signal has not been filtered by
over the electrodes, and considerable levels of noise are sometimes encountered, yet the QRS complex
any means.distinct enough during recorded measurements, hence, the signal has not been filtered by
remained
any Heart
2.3. means.
and Respiration Rates Detection Algorithms
All the
2.3. Heart detection
and algorithms
Respiration and theAlgorithms
Rates Detection evaluation of their outcome reported here are implemented
through Matlab.
All the detection algorithms and the evaluation of their outcome reported here are implemented
through Matlab.
2.3.1. Heart Rate in BCG

2.3.1.Keeping in mind
Heart Rate in BCGthe variability of the BCG signal’s morphology across people, lying positions,
and even with time, a versatile algorithm that is independent of history, at least apart from the
Keeping
immediate inand
one, mind the variability
capable of thethe
of perceiving BCG signal’sslightly
short-term morphology across
changing people,
pattern lying
of the BCG positions,
cardiac
and even with time, a versatile algorithm that is independent of history, at least
cycle was required. Consequently, the choice of utilizing an unsupervised pattern recognition algorithm apart from the
immediate
that operates one, and capablewithout
immediately of perceiving the short-term
a training phase was slightly
made.changing
Hence, anpattern of the BCGover
improvement cardiac
the
cycle was required. Consequently, the choice of utilizing an unsupervised
agglomerative clustering algorithm put forward by Paalasmaa in [27] is attempted. In Appendix A pattern recognition
algorithm
further that operates
explanations immediately
for some without
of the terms a training
related phase wascan
to this algorithm made. Hence, an improvement
be found.
over Thethe BCG
agglomerative clustering algorithm put forward by Paalasmaa
signal is run through the algorithm in segments, each with a length in [27] is attempted.
of 10.66 In
s; effectively
Appendix
it A further
is 10 s since explanations
there is for some
0.66 s of overlap of thesuccessive
between terms related to this algorithm
operations. canderivative
Firstly, the be found. of the
The BCG signal is run through the algorithm in segments, each with a length of 10.66 s;
effectively it is 10 s since there is 0.66 s of overlap between successive operations. Firstly, the
Sensors 2019, 19, 1451 8 of 19
Sensors 2018, 18, x FOR PEER REVIEW 8 of 19

derivative of the signal


signal is calculated, is calculated,
through which the through which the
local maxima thatlocal maxima that
presumably presumably
represent represent
the upward wavesthe
in
upward waves
the signal (H, J,in
L, the
etc.,signal (H, J,1)L,are
see Figure etc., see Figure
identified, 1)step
this are identified,
is depicted this step is
in Figure 7a.depicted in Figure
Afterwards, 7
feature
(a). Afterwards, feature vectors are formulated. The feature selection here is straightforward;
vectors are formulated. The feature selection here is straightforward; each feature vector begins with a each
feature vector begins
local maximum pointwith
andacontains
local maximum pointofand
the segment thecontains thefollows.
signal that segment of the signal that follows.

(a) (b)

Figure
Figure7.7. A
A segment
segment of aa BCG
BCGsignal
signalbeing
beingprocessed
processed byby
thethe clustering
clustering algorithm,
algorithm, which
which is shorter
is shorter than
than the usual
the usual casecase for convenience
for convenience of display:
of display: (a) (a) local
local maxima
maxima areare identified;
identified; (b)(b) two
two of of
thethe twenty-
twenty-six
six feature
feature vectors
vectors contained
contained ininthe
thesegment
segmentarearehighlighted,
highlighted,namely,
namely,vectors
vectors 99 and 20, with
with each
eachone
one
comprising 28 features, i.e., signal samples.
comprising 28 features, i.e., signal samples.

Thetemporal
The temporallength lengthof ofthe
thefeature
featurevector,
vector,i.e.,
i.e.,the
thetime
timethethesignal
signalsegment
segmentcontained
containedin inthe
thevector
vector
covers, is 0.66 s, which corresponds to the time interval of a 90-bpm
covers, is 0.66 s, which corresponds to the time interval of a 90-bpm heart cycle. Reducing this heart cycle. Reducing this
parameter risks
parameter risks losing
losing important
important information,
information, making making the the wave
wave represented
represented in in the
the vector
vector less
less
distinctive from other types of waves; however, increasing it might unnecessarily
distinctive from other types of waves; however, increasing it might unnecessarily increase algorithm increase algorithm
computations.Furthermore,
computations. Furthermore,totokeep keepaaminimum
minimuminformation
informationredundancy
redundancywithin withinthe thefeature
featurevector,
vector,
aa down-sampling
down-samplingby byaa factor
factor of of 22 isis employed,
employed, meaning,
meaning, the the feature
feature vector
vector contains
contains everyevery other
other
sample of the signal segment. Figure 7b highlights two of the feature
sample of the signal segment. Figure 7 (b) highlights two of the feature vectors being processed, as vectors being processed,
as examples.
examples. TheThe downsampling
downsampling reduces
reduces thethe computational
computational timewithout
time withoutcompromising
compromisingthe theoverall
overall
process. Since the sampling frequency at hand is high enough to
process. Since the sampling frequency at hand is high enough to contain frequency components up contain frequency components
up42.4
to to 42.4
Hz, as Hz,perasthe
persampling
the samplingtheorem, theorem, and given
and given that thethat
BCG thesignal
BCG is signal is low-pass
low-pass filtered filtered
at 15 Hz,at
a15downsampling
Hz, a downsampling to halftothe halforiginal
the original sampling
sampling frequency
frequency willwillstillstill preserve
preserve allallthe
thefrequency
frequency
components of interest in the signal. As a result, each feature vector
components of interest in the signal. As a result, each feature vector is comprised of 28 features to is comprised of 28 features to
cover the aforementioned temporal
cover the aforementioned temporal length. length.
Theproximity
The proximitymeasure,
measure,according
accordingtoto which
which thethe feature
feature vectors
vectors are are clustered,
clustered, has has to handle
to handle the
the varying amplitudes of the BCG waves across different heartbeats
varying amplitudes of the BCG waves across different heartbeats with minimal differentiation. with minimal differentiation.
Therefore,the
Therefore, theinverse
inversecosine
cosinedissimilarity
dissimilaritymeasure
measure is isquite
quiteapt
aptfrom
fromthisthisperspective,
perspective,since sinceititisisaapure
pure
angular measure. Nevertheless, some boundary conditions are compounded
angular measure. Nevertheless, some boundary conditions are compounded with the dissimilarity with the dissimilarity
measure,which,
measure, which,on onthe
theone
onehand,hand,avoid
avoidcomputations
computationsdeemed deemedpractically
practicallyunnecessary,
unnecessary,and andon onthe
the
other hand,
other hand, can canbebevery vital
very for proper
vital for proper clustering underunder
clustering certaincertain
circumstances. If any ofIftheany
circumstances. conditions
of the
is fulfilled, the angular dissimilarity between the pair of vectors in question
conditions is fulfilled, the angular dissimilarity between the pair of vectors in question is given the is given the maximum
value of π value
maximum (rad). ofThose conditions
π (rad). are:
Those conditions are:
1.1. IfIfthe
theratio
ratioofofthe
thefeature
featurevectors’
vectors’magnitudes
magnitudestotooneoneanother
anotherisislarger
largerthan
thanα,α,or,
or,as
asthethecase
casemay
may
be, smaller than 1/α, where α is
be, smaller than 1/α, where α is set to 3.set to 3.
2.2. IfIfthe
thedifference
differencebetween
betweenthethetimestamps
timestampsof ofthe
therespective
respectivevectors’
vectors’first
first elements
elements is is less than ttmin
less than min,,
which is set to 0.33 s, making the algorithm only suitable for heart rates lower
which is set to 0.33 s, making the algorithm only suitable for heart rates lower than 180 bpm. than 180 bpm.
3.3. IfIfthe
theabsolute
absolutedifference
differenceininthe
thechronological
chronologicalorder
orderofofthe
thevectors
vectorsisisless thanii(index).
lessthan (index).
Thethird
The third condition
condition relies
relies on
on the
thefact
factthat,
that,asasmentioned
mentionedearlier,
earlier,the BCG
the BCG waves
waves H–L dominate
H–L the
dominate
signal
the andand
signal should usually
should be distinctively
usually observed
be distinctively forming
observed a W shape
forming [4]. Consequently,
a W shape it should
[4]. Consequently, it
should be reasonable to assume that a single cardiac cycle must contain a minimum of three upward
Sensors 2019, 19, 1451 9 of 19

Sensors 2018, 18, x FOR PEER REVIEW 9 of 19

be reasonable to assume that a single cardiac cycle must contain a minimum of three upward waves,
waves, meaning i = 3, which is valid most often; however, when the heart rate is high enough,
meaning i = 3, which is valid most often; however, when the heart rate is high enough, sometimes
sometimes superimposition of waves is strong, and the BCG waveform does not assume more than
superimposition of waves is strong, and the BCG waveform does not assume more than two local
two local maxima. When that is the case, having this condition with i = 3 is detrimental to the
maxima. When that is the case, having this condition with i = 3 is detrimental to the clustering because
clustering because it will prevent waves of the same type from being clustered together. One solution
it will prevent waves of the same type from being clustered together. One solution is to perform the
is to perform the whole clustering procedure twice, first with i = 3, then i = 2, afterwards the
whole clustering procedure twice, first with i = 3, then i = 2, afterwards the consistency of the detected
consistency of the detected beat-to-beat heart rate is assessed for both runs via a standard deviation
beat-to-beat heart rate is assessed for both runs via a standard deviation function, and if σ when i =
function, and if σ when i = 2 is less, then the result of that clustering is adopted, otherwise the result
2 is less, then the result of that clustering is adopted, otherwise the result of i = 3 remains.
of i = 3 remains.
After combining the inverse cosine dissimilarity measure with the aforementioned boundary
After combining the inverse cosine dissimilarity measure with the aforementioned boundary
conditions, the overall dissimilarity function becomes:
conditions, the overall dissimilarity function becomes:
(  T 
cos−x1 ykxk kyk , 1if α1 ‖x‖ ≤ kkyxkk ≤ α and tx1 − ty1 ≥ tmin and ix − iy ≥ i
x y
d(x, y) =cos , if α and t t t and i i i (1)
d x, y ‖x‖ ‖y‖ π, α ‖y‖ otherwise (1)

where x and y are the π , otherwise


pair of feature vectors, the dissimilarity between which is being quantified.
whereThex and y are the pair
complete-link of feature
hierarchical vectors, scheme
clustering the dissimilarity
begins with between
each vector which is being
forming quantified.
a single-element
Thethen
cluster, complete-link
merges these hierarchical
clusters together, clustering twoscheme
at a time, begins
until with each vector
all vectors are contained forminginaa single-
single
elementFigure
cluster. cluster, then merges
8 depicts these clusters
a dendrogram together,the
that displays twocomplete
at a time,clustering
until all vectors hierarchy areofcontained
the sample in
a single cluster. Figure 8 depicts a dendrogram that displays the complete
signal shown in Figure 7. When the dissimilarity between a pair of clusters, at least one of which clustering hierarchy of the
sample signal
contains more shown
than ainsingle
Figure 7. Whenisthe
element, dissimilarity
evaluated, between a pairalgorithm
the complete-link of clusters, at least
adopts theone
max of
which contains
proximity more
function than
[35] (p. a620),
single thatelement,
is: is evaluated, the complete-link algorithm adopts the max
proximity function [35] (p. 620), that is: 
dmax Ci , Cj = max d(x, y) (2)
d C , C max x∈Ci , y∈dCj x, y (2)
∈ , ∈

where
whereCCi iand
andCCj jare
arethe
thepair
pairofof
clusters in question.
clusters Therefore,
in question. the the
Therefore, dissimilarity between
dissimilarity clusters
between is fully
clusters is
determined by theirbyrespective
fully determined most dissimilar
their respective vectors. This
most dissimilar is crucial
vectors. Thisfor
is the conditional
crucial for thedissimilarity
conditional
function implemented
dissimilarity here, otherwise
function implemented a min
here, proximity
otherwise function,
a min for instance,
proximity function, would renderwould
for instance, those
boundary conditions practically void.
render those boundary conditions practically void.

Figure 8. A dendrogram showing the complete hierarchical clustering of the sample signal presented
in the previous figure.
figure. The magenta line represents
represents the
the dissimilarity
dissimilarity threshold
threshold (θ
(θ== 1.2)
1.2) that
that is used to
determine the final
final clustering.
clustering.

To
To conclude
concludethe
thehierarchical
hierarchicalclustering
clusteringmeaningfully
meaningfullya athreshold
thresholdfor
forthe
theangular dissimilarity,
angular θ,
dissimilarity,
is employed. The clustering that occurs up to this level of dissimilarity determines the final clusters,
θ, is employed. The clustering that occurs up to this level of dissimilarity determines the final clusters,
and
and beyond
beyond which
which any
any clustering
clustering is
is discarded.
discarded. This
This threshold
threshold was
was empirically
empirically tuned
tuned to
to aa value
value of
of 1.2
1.2
(rad); the magenta line in Figure 8 depicts this parameter. The final step of the algorithm is choosing
Sensors 2019, 19, 1451 10 of 19
Sensors 2018, 18, x FOR PEER REVIEW 10 of 19

a cluster
(rad); the that reflects
magenta linethe
in cardiac
Figure 8cycles best.
depicts thisTwo criteria are
parameter. Theutilized forofthis
final step thepurpose,
algorithm the
is second
choosingof
awhich comes
cluster into effect,
that reflects the only if the
cardiac firstbest.
cycles was indecisive:
Two criteria are utilized for this purpose, the second of
which comes into effect, only if the first was indecisive:
1. The size of the cluster. The cluster containing the largest number of vectors is chosen.
2. The
1. mean
The size of of
thedissimilarity. If the containing
cluster. The cluster largest number of vectors
the largest number contained
of vectorsin isa chosen.
cluster is equally
2. matched by more than one cluster, the one with the lowest mean of
The mean of dissimilarity. If the largest number of vectors contained in a cluster isdissimilarity among its
equally
vectors is picked.
matched by more than one cluster, the one with the lowest mean of dissimilarity among its
vectors
In the is picked.
dendrogram of Figure 8, considering the clusters formed below the dissimilarity
threshold, it can be clearly observed that the largest cluster, which is the only one containing four
In the dendrogram of Figure 8, considering the clusters formed below the dissimilarity threshold,
vectors, is the first one to the left; therefore waves 1, 9, 18 and 23, which happen to be all the J waves
it can be clearly observed that the largest cluster, which is the only one containing four vectors, is the
present in the segment are chosen to represent the cardiac cycles.
first one to the left; therefore waves 1, 9, 18 and 23, which happen to be all the J waves present in the
segment are chosen to represent the cardiac cycles.
2.3.2. Respiration Rate in BCG
2.3.2.The
Respiration
scope ofRate thisinpaper
BCG with regard to respiration is primarily to determine whether the
respiration component
The scope in the
of this paper withload
regardcelltosignal is present
respiration with sufficient
is primarily to determinemagnitude thatrespiration
whether the facilitates
monitoring it.
component in the load cell signal is present with sufficient magnitude that facilitates monitoring it.
Giventhat
Given thatthethe narrow
narrow bandwidth
bandwidth of theofrespiration
the respiration component
component lies at thelies at end
lower the oflower end of
frequencies,
frequencies, where the motions of respiration dominate, unchallenged by the usual
where the motions of respiration dominate, unchallenged by the usual noise sources, this simplified noise sources,
thisalgorithm
the simplifiedsignificantly.
the algorithmConsequently,
significantly. Consequently, using its
using its derivative, thederivative,
local maximathe local maxima
of the of the
respiration
respiration
signal signal are
are identified in identified in order
order to mark to mark the
the boundaries of boundaries of respiration
respiration cycles cyclesthe
and calculate and calculate
respiration
the respiration rate accordingly, Figure 9 depicts a sample
rate accordingly, Figure 9 depicts a sample output of this procedure. output of this procedure.

Figure 9. Sample of the workings of the respiration rate calculation method. The upper plot depicts
Figure 9. Sample of the workings of the respiration rate calculation method. The upper plot depicts a
a respiration signal with red asterisks marking the detected local maxima, i.e., the inhalation peaks,
respiration signal with red asterisks marking the detected local maxima, i.e., the inhalation peaks,
while the corresponding calculated breath-to-breath respiration rate is shown in the lower plot.
while the corresponding calculated breath-to-breath respiration rate is shown in the lower plot.
2.3.3. Heart Rate in ECG
2.3.3. Heart Rate in ECG
For the most part, a simple peak detection with an absolute threshold works well to detect the
For the most
QRS complex in ECGpart, a simple
signals; Figurepeak
10adetection with
depicts the an absolute
common form ofthreshold
the signal.works well to
However, twodetect
typesthe
of
QRS complex
noises have been in encountered,
ECG signals;usually
Figure due
10 (a) depicts
to poor the common
electrode form which
connection, of the require
signal. more
However, two
elaborate
types of noises
algorithms: thehave
first been
one isencountered,
a deviation usually due to poor
of the baseline electrode
of the signal, connection, which10b,
shown in Figure require more
which is
elaborate algorithms:
addressed by calculatingthe first one is a deviation
the immediate relativeofincrease
the baseline of the signal,
in amplitude shown
at each peak in and
Figure 10 (b),
setting a
which is addressed by calculating the immediate relative increase in amplitude
threshold for that measure; and the second one is a considerable amount of EMI, shown in Figure 10c, at each peak and
settingisararely
which threshold for that measure;
encountered and the
in the depicted second one
intensity; is a considerable
it is handled amount
by comparing theofrelative
EMI, shown in
increase
Figure
in 10 (c), at
amplitude which
a peakis with
rarelya encountered
moving average in the
of depicted
the same intensity;
measure of it other
is handled
peaksby comparing the
surrounding the
relative increase in amplitude at a peak with a moving average of the
peak in question, fine-tuned by a coefficient that is multiplied by the moving average.same measure of other peaks
surrounding the peak in question, fine-tuned by a coefficient that is multiplied by the moving
average.
Sensors 2018, 18, x FOR PEER REVIEW 11 of 19

Figure 10 (c), which is rarely encountered in the depicted intensity; it is handled by comparing the
relative increase in amplitude at a peak with a moving average of the same measure of other peaks
surrounding the peak in question, fine-tuned by a coefficient that is multiplied by the moving
Sensors 2019, 19, 1451 11 of 19
average.

(a) (b) (c)

Figure
Figure 10.
10. ECG
ECGsignal
signalsamples
samples with
withQRS
QRScomplexes
complexes marked
marked with
with red
red asterisks:
asterisks: (a)
(a)the
thecommon
commonform;
form;
(b) baseline fluctuation; (c) large EMI.

2.4. Measurements
2.4. Measurements
The measurements
The measurements involved involved sevenseven volunteers,
volunteers, whowho have
have approved
approved of of and
and signed
signed aa consent
consent form
form
permitting the publishing of the results based on their
permitting the publishing of the results based on their measurements. measurements.
Each of
Each of the
the volunteers
volunteers lay lay inin the
the four
four commonest
commonest lying lying positions:
positions: thethe supine,
supine,prone,
prone,left
leftlateral,
lateral,
and right
and right lateral.
lateral.In Inthethesupine
supineand andprone
pronepositions,
positions,the thevolunteers
volunteerslay layflat
flatononthe
the mattress
mattress without
without a
a pillow, whereas a pillow was provided when they lay on their sides.
pillow, whereas a pillow was provided when they lay on their sides. During the measurements the During the measurements
the volunteers
volunteers werewereasked asked
to lietostill;
lie still; however,
however, smallsmall movements
movements takingtaking
placeplace
everyevery
now andnowthenandwere
then
were inevitable. In each measurement, BCG and ECG signals were
inevitable. In each measurement, BCG and ECG signals were recorded continuously for five minutesrecorded continuously for five
in a certain position, then repeated twice in the same position. This division was meant to allow to
minutes in a certain position, then repeated twice in the same position. This division was meant a
allow a short pause for the volunteer to move between the measurements
short pause for the volunteer to move between the measurements if they wished to do so. Therefore, if they wished to do so.
Therefore,
the the measurements
measurements added up added to oneup hourto one hour of measurement
of measurement time for time
eachfor each volunteer,
volunteer, and sevenand hours
seven
hours in
in total. total.
At the
At the beginning
beginning of of the
the measurements,
measurements, aa good good coupling
coupling of of the
the sensing
sensing surface
surface with
with the
the mattress
mattress
after a volunteer lay down on it was ensured by checking the amount
after a volunteer lay down on it was ensured by checking the amount of preload the load cell of preload the load cell was
was
reading. It
reading. It is
is worth
worthmentioning
mentioning that thatthe
themeasurements
measurements presented
presented herehere were
were carried
carried out
outin
intwo
twostages,
stages,
which involved two and five volunteers, respectively. In the first stage, the
which involved two and five volunteers, respectively. In the first stage, the preload was kept between preload was kept between
4–5 N,
4–5 N, whereas
whereasininthe thesecond
secondone, one,it itwas
was between
between 12–21.5
12–21.5 N. N. This
This variation
variation hadhad an influence
an influence on
on the
the magnitudes
magnitudes of the
of the transmitted
transmitted forces,
forces, which
which will
will bebe discussed
discussed later.OfOfcourse,
later. course,thethefact
fact that
that the
the
operating range of the load cell is up to 50 N provides a considerable
operating range of the load cell is up to 50 N provides a considerable margin for maneuvering. margin for maneuvering.

2.5. Evaluation
2.5. Evaluation
The effectiveness of the heart rate detection algorithm was evaluated based on the seven hours of
The effectiveness of the heart rate detection algorithm was evaluated based on the seven hours
measurements time. In this respect, few things are worth pointing out:
of measurements time. In this respect, few things are worth pointing out:
•• The correct
The correct detection
detectionwas wasevaluated
evaluatedvisually,
visually, which
which was carried
was outout
carried by plotting the BCG
by plotting and ECG
the BCG and
signals with their respective heart cycles’ detection markings and corresponding
ECG signals with their respective heart cycles’ detection markings and corresponding calculated calculated heart
rates. rates.
heart FigureFigure
11 displays part ofpart
11 displays a signal undergoing
of a signal visualvisual
undergoing assessment.
assessment.
•• The calculated heart rates were rounded to integers. Furthermore,
The calculated heart rates were rounded to integers. Furthermore, a discrepancy a discrepancy of 5up
of up to bpmto
between the ECG and corresponding BCG beat-to-beat heart rates was considered a correcta
5 bpm between the ECG and corresponding BCG beat-to-beat heart rates was considered
correct detection,
detection, as long as asitlong
was as it wasdetermined
visually visually determined that the
that the detected BCGdetected
wavesBCGof thewaves of the
consecutive
consecutive heart cycles are of the same type, which meant that the discrepancy
heart cycles are of the same type, which meant that the discrepancy was not a result of wrong was not a result
of wrong detection, but rather a time shift of the wave due to an artifact,
detection, but rather a time shift of the wave due to an artifact, for instance. Such high for instance. Such high
discrepancies happen
discrepancies happen infrequently,
infrequently, andand mostly
mostly with with higher
higher heart
heart rates.
rates.
•• Given that the algorithm does not single out any specific wave
the algorithm does not single out any specific wave to represent to represent the the
cardiac cycle,
cardiac like
cycle,
the Jthe
like wave for for
J wave instance,
instance,butbut
rather
rather looks
looksforforthethewaveform
waveformmost mostconsistently
consistently occurring
occurring in thethe
10-s segment being processed. It It follows
follows that
that thethe calculation
calculation of
of the
the beat-to-beat
beat-to-beat heart
heart rate
rate of
of the
the
cardiac cycle that lies between two consecutive segments based on potentially different waves is
not warranted. In Figure 11, which depicts 5 transitions between 10-s segments, it can be noticed
that the algorithm was first detecting the H wave, then in the next segment the J wave, and later
Sensors 2019, 19, 1451 12 of 19

Sensors 2018, 18, x FOR PEER REVIEW 12 of 19


the L wave. Consequently, the 29 intermediate cardiac cycles, which lie between the 30 segments
of a 5-min measurement are subtracted from the number of cycles assumed to be detectable by the
detectable by the algorithm, and the correct detection percentage is calculated accordingly. This
algorithm, and the correct detection percentage is calculated accordingly. This can be observed
can be observed in the plot of Figure 11 by the cuts in the BCG heart rate step diagram. It is
in the plot of Figure 11 by the cuts in the BCG heart rate step diagram. It is worth pointing out
worththat
pointing out that
this exclusion this
was not exclusion
entertained wasprevious
in the not entertained inReference
paper, namely the previous paper,
[33], and namely
all heart
Reference
cycles[33], and all
contained in aheart cycles contained
measurement in a measurement
were considered detectable. were considered detectable.
• Despite
• the fact
Despite thatthat
the fact artifacts caused
artifacts causedby
by volunteers’ movements
volunteers’ movements occurred,
occurred, sincesince
it wasitnot
was not that
that
frequent, no exclusion
frequent, no exclusion of of
heartbeats
heartbeatsthat
that might havebeen
might have been undetected
undetected because
because of artifacts
of such such artifacts
was made in this evaluation.
was made in this evaluation.

Figure 11. Visual evaluation diagram that contains both BCG and ECG signals, which have
Figure 11. Visual evaluation diagram that contains both BCG and ECG signals, which have been
been rescaled and shifted for the convenience of display; their respective detected heartbeats and
rescaled and shifted for the convenience of display; their respective detected heartbeats and
corresponding beat-to-beat heart rates; and the signal epochs’ markers that highlight the segment’s
corresponding beat-to-beat
divisions, by heartwas
which the signal rates; and This
handled. the signal epochs’
volunteer markers
has been that
diagnosed highlight
with the segment’s
heart arrhythmia.
divisions, by which the signal was handled. This volunteer has been diagnosed with heart arrhythmia.
3. Results and Discussion
3. Results The
and preliminary
Discussionmeasurements that were carried out to assess the influence of the force
sensor placement on the signals revealed that the respiration component generally dominates the
The preliminary measurements
transverse/dorsoventral that
signal, i.e., whenwerethecarried
sensor isout to assess
placed underthe theinfluence of the
mattress (see force2).sensor
Figure
placement onbethe
This can seensignals
in Figure revealed
12a by the that the respiration
magnitudes component
of the respiration wavesgenerally
marked with dominates
circles the
transverse/dorsoventral
when compared to the signal,
muchi.e., when
smaller BCGthewaves
sensor is placed
marked withunder the mattress
“x”. Whereas (see12b,
in Figure Figure
which 2). This
can becorresponds
seen in Figure
to the 12 (a) by the
longitudinal magnitudes
BCG of the3),respiration
setup (see Figure the magnitudes waves marked
of the withand
respiration circles
BCG when
components
compared to theare muchsmaller
much more comparable.
BCG waves This improves
marked the withquality
“x”.ofWhereas
the BCG signal by facilitating
in Figure 12 (b),a which
more optimized signal amplification and more effective filtration of the respiration
corresponds to the longitudinal BCG setup (see Figure 3), the magnitudes of the respiration and BCG component.
The results of the evaluation of the heart rate detection algorithm are presented in Table 1,
components are much more comparable. This improves the quality of the BCG signal by facilitating
which relates some of the volunteers’ biometrics and the measurements’ details to the rates of correct
a moredetection.
optimized signal amplification and more effective filtration of the respiration component.
The measurements of the different lying positions for each volunteer are grouped together;
The
the results of the evaluation
rate of detection of the heart
for each position rateasdetection
is given well as the algorithm
average rateare for
presented in Table
all positions 1, which
of each
relatesvolunteer.
some ofThe theoverall
volunteers’ biometrics
correct detection rate and the measurements’
was 83.9%, which is the average details
of thetodetection
the rates ofofcorrect
rates
detection. The measurements
all measurements, given thatofthey
the different
all have the lying
samepositions
weight, thatforis,each
theirvolunteer
equal length areofgrouped
time. together;
the rate of detection for each position is given as well as the average rate for all positions of each
volunteer. The overall correct detection rate was 83.9%, which is the average of the detection rates of
all measurements, given that they all have the same weight, that is, their equal length of time.
Sensors 2019, 19, 1451 13 of 19

Table 1. BCG measurements’ details and their corresponding heart rate detection evaluation.

Volunteer Details Average Heart No. Theoretically Detectable No. Cycles Not No. False Positive Correct Avg. Corr.
Lying Position
Sex Age [yr] Wt. [kg] Ht. [cm] Rate [bpm] Beat-to-Beat Cycles Correctly Detected 3 Detections 4 Detection (%) Detection (%)
Supine 76 1050 19 0 98.2
Prone 64 872 100 0 88.6
M 31 83 180 95.4
Left side 69 944 32 0 96.6
Right side 68 928 18 0 98.0
Supine 39 502 29 8 94.3
1
Prone 41 524 19 4 96.4
M 35 78 180 93.0
Left side 38 486 43 8 91.2
Right side 43 551 55 7 90.0
Supine 53 700 65 1 90.7
Prone 56 756 149 5 80.5
M 28 75 170 85.3
Left side 57 758 37 2 95.0
Right side 61 822 204 3 74.8
Supine 87 1211 200 0 83.5
Prone 83 1151 285 0 75.2
M 30 78 170 86.9
Left side 79 1089 67 1 93.9
Right side 82 1142 59 0 94.9
Supine 59 802 75 1 90.7
Prone 61 828 29 0 96.5
M 26 65 170 86.4
Left side 63 853 94 1 88.9
Right side 61 822 252 5 69.3
Supine 84 1177 383 0 67.4
Prone 88 1235 500 1 59.8
M 30 95 173 68.2
Left side 83 1150 173 0 84.9
Right side 86 1206 473 0 60.8
Supine 68 926 615 9 33.6
Prone 70 958 186 4 80.6
F2 70 67 158
Left side 65 879 117 36 86.7
72.2
Right side 67 921 109 12 88.1
1 The volunteer has been diagnosed with Heart arrhythmia; 2 The volunteer has four bypasses in her heart; 3 As a result of all three types of false detection mentioned in the text; 4 A false

positive is the detection of more than a single wave within a single cardiac cycle.
Sensors 2019, 19, 1451 14 of 19
Sensors 2018, 18, x FOR PEER REVIEW 13 of 19

(a) (b)

12. Two
Figure 12. Twosamples
samplesofofraw
rawload
loadcell
cellsignals
signals measured
measured forfor a volunteer
a volunteer in ainsupine
a supine position
position withwith
the
the sensor positioned (a) below the mattress (see Figure 2); (b) at the top side of the mattress
sensor positioned (a) below the mattress (see Figure 2); (b) at the top side of the mattress (see Figure 3).(see
Figure 3). Respiration
Respiration and
and cardiac cardiac
cycles cycles are
are marked bymarked by circles
circles and and “x”, respectively.
“x”, respectively.

TheTable
first two volunteers
1. BCG in thedetails
measurements’ table and
hadtheir
excellent detection
corresponding rates
heart rateindetection
all positions, whereas the
evaluation.
other five usually had at least one position with a relatively worse rate of detection. This shows that
Volunteer
there details ofLying
is an influence the lyingAverage No.on
position theoretically
the BCG signalNo. cycles
quality. No.
Forfalse Correct ofAvg.
an overview how Corr.
the
Sex Age Wt. Ht. position heart rate detectable beat- not correctly positive detection
algorithm performed with the four lying positions Table 2 presents the average rate of detection for detection
[yr] position
each lying [kg] [cm]across all volunteers.
[bpm] to-beat cycles
It can be detected
said that
3 detections 4
the left lateral position(%) (%)
had a distinctly
higher detection rate Supine
than the other76positions, which
1050 fared rather 19 the same.0 98.2
Prone 64 872 100 0 88.6
M 31 83 180
Table 2. The average correct detection rates for lying positions across all volunteers. 95.4
Left side 69 944 32 0 96.6
Right side Lying
68 Position 928Correct Detection
18 (%) 0 98.0
Supine Supine
39 502 79.8 29 8 94.3
Prone 41Prone 524 82.5 19 4 96.4
M1 35 78 180 Left side 91.0 93.0
Left side 38 side
Right 486 82.3 43 8 91.2
Right side 43 551 55 7 90.0
Supinevolunteer,
Moreover, the second 53 who is diagnosed
700 65 arrhythmia,
with heart 1 had some 90.7of the highest
detection Prone 56 algorithm 756 149 5 in the time
80.5 interval of a
M 28 rates,
75 which
170 shows that the handles the sudden change 85.3
heartbeat very well. A sample of his arrhythmic pattern is depicted in Figure 11. On95.0
Left side 57 758 37 2 the other hand,
the algorithm was notRight
as successful
side when dealing
61 822with the signal
204of the last male
3 volunteer
74.8 in the table
and produced overallSupine
the worst detection
87 results. This
1211 was anticipated,
200 given 0 his high body mass index
83.5
that seems to have a detrimental
Prone effect
83 on the BCG1151 signal. Interestingly,
285 the0last volunteer
75.2 in the table,
M is 30
who 78 170
an elderly female with four bypasses in her heart, gave reasonably good results, 86.9
Left side 79 1089 67 1 93.9 except in the
supine position, which fared the worst among all measurements, with one-third of the heartbeats being
Right side 82 1142 59 0 94.9
correctly detected. Perhaps the bypasses interfered most with the signal in that particular position.
Supine 59 802 75 1 90.7
The longitudinal sensor setup presented here uses the mattress as a medium for the transmission
Prone 61 828 29 0 96.5
ofMbody26forces
65 in170
a manner that introduces the preload on the sensor as an independent parameter, 86.4
Left side
which evidently has some influence63 on the transmitted
853 94 As mentioned
signals. 1 88.9
in the measurements
Right sidethe magnitudes
section, a correlation between 61 of822
the preload and252the measured5 force, and
69.3 by extension,
Supine components,
its cardiac and respiratory 84 was 1177
revealed; Figure38313a depicts 0 this correlation.
67.4 In the
figure, if grouped according
Prone to the 88 amount of preload,
1235 two groups
500 of measurements
1 can be observed.
59.8
M 30 95 173 68.2
The small group to the Left left,
side which 83 had much smaller preloads,
1150 173 resulted0 in the whole84.9 in smaller
magnitudes of forces.RightHowever,
side in
86the group1206
to the right, notwithstanding
473 0 the greater
60.8 dispersion,
the higher preloads resulted
Supine in overall
68 higher magnitudes.
926 615 9 33.6
Prone 70 958 186 4 80.6
F2 70 67 158 72.2
Left side 65 879 117 36 86.7
Right side 67 921 109 12 88.1
1 The volunteer has been diagnosed with Heart arrhythmia
2 The volunteer has four bypasses in her heart

3 As a result of all three types of false detection mentioned in the text

4 A false positive is the detection of more than a single wave within a single cardiac cycle
section, a correlation between the magnitudes of the preload and the measured force, and by
extension, its cardiac and respiratory components, was revealed; Figure 13 (a) depicts this correlation.
In the figure, if grouped according to the amount of preload, two groups of measurements can be
observed. The small group to the left, which had much smaller preloads, resulted in the whole in
smaller
Sensors magnitudes
2019, 19, 1451 of forces. However, in the group to the right, notwithstanding the greater15 of 19
dispersion, the higher preloads resulted in overall higher magnitudes.

(a) (b)

Figure 13. (a) The relation between the sensor preload and the magnitude of the corresponding cardiac
and respiration forces. The respiration force magnitude is evaluated by its average peak-to-peak
amplitude, whereas the cardiac force is assessed based on the standard deviation of the signal sample
points since the signal generally assumes a Gaussian distribution; its magnitude is determined as 6 σ.
(b) The relation between the preload and the resulting rate of correct detection.

It is worth pointing out that the proportionality of the cardiac and respiration magnitudes to one
another, in the same measurement, differed across volunteers, as well as lying positions of the same
volunteer. From Figure 13a, it can be discerned that with smaller preloads the cardiac component
was generally stronger, which was not usually the case with higher preloads, especially since the
respiration forces, marked with blue triangles, showed greater dispersion than the cardiac forces. As a
consequence of this observation it might be said that the more the mattress is pressed against the
sensing surface, the better the coupling between the two is. On the other hand, however, when the
relation between the preload and the rate of detection is considered, which is portrayed in Figure 13b,
it seems that lower preloads gave, as a whole, higher detection rates. Nevertheless, since the lower
preloads involved only two volunteers, such a conclusion might be viewed statistically as premature.
The types of false detections that occur in this clustering algorithm can generally be classified
into three categories: false positive, which is the detection of more than a single wave within a single
cardiac cycle, i.e., multiplicate detection; false negative, when not a single wave is detected within a
cardiac cycle; and misplaced detection, which is the detection of a single wave in a cardiac cycle that
is different from the majority of detected waves in the other cardiac cycles within the segment being
processed. It is worth pointing out, that the dissimilarity threshold, θ, does not have a direct influence
on the third class of false detections the way it has on the other two classes. As may be discerned
from the dendrogram shown in Figure 8 the value of θ determines the size of the final clusters, and by
extension, the cluster chosen to represent the heartbeats. Therefore, increasing θ may lessen the false
negative detection on the one hand, but it may increase the false positive one on the other hand.
In the work we presented in Reference [33], the third boundary condition of the dissimilarity
function had not yet been implemented. It may be argued that there is an overlap of the effects the
second and third conditions have in the function, which is generally true; nevertheless, the need for an
improved algorithm became obvious, when it made a large amount of false positive detections in a
BCG signal of a volunteer lying in a certain position, even when lower values of θ where attempted.
The benefit of adding the third condition was two-fold. Firstly, it suppressed most of the false
positive detections happening at the previously employed θ (in Reference [33] it was 1 rad) across all
measurements. Secondly, as a result, it facilitated raising θ (to 1.2 rad), which decreased the probability
of false negative detection as well, improving the overall detection rates. This can be recognized to
a certain extent by comparing the results of the two volunteers, whose same measurements were
handled by the two versions of the algorithm; the first two volunteers in Table 1 of both publications,
Sensors 2019, 19, 1451 16 of 19

respectively. Furthermore, the effect of adding this new boundary condition can be directly observed
in Figure 14, which depicts the algorithm’s detections with a same sample signal before and after this
addition, all other parameters, including θ, are kept the same. In plot (a) almost all heart cycles shown
had duplicate
Sensors 2018, 18, xdetections, while in (b) each cycle is detected only once.
FOR PEER REVIEW 16 of 19

(a) (b)

14.The
Figure 14. Theperformance
performance of the
of detection algorithm
the detection with awith
algorithm peculiar sample signal
a peculiar sample(a)signal
when processed
(a) when
without
processedthewithout
lately added boundary
the lately addedcondition
boundary resulted in excessive
condition resultedpositive detections,
in excessive whereas
positive in (b)
detections,
the issue is resolved solely by this addition; all other parameters are kept the same.
whereas in (b) the issue is resolved solely by this addition; all other parameters are kept the same.

Finally, it is
Finally, it is worth
worthpointing
pointingout outthat,
that,
asas mentioned
mentioned in the
in the algorithms
algorithms section,
section, at higher
at higher heartheart
rates
rates
superimposition of waves takes place and the BCG waveform might occasionally include onlyonly
superimposition of waves takes place and the BCG waveform might occasionally include two
two upward waves in a cardiac cycle. Consequently, if the minimum index
upward waves in a cardiac cycle. Consequently, if the minimum index difference was set to three, difference was set to
three, the presence
the presence of even ofaeven
single a single such cardiac
such cardiac cycle within
cycle within the segment
the segment being being processed
processed wouldwould
cause
cause incorrect clustering for the majority of waves in the segment, hence
incorrect clustering for the majority of waves in the segment, hence the solution of carrying the solution of carrying
out out
the
the clustering with two values of the minimum index difference was
clustering with two values of the minimum index difference was implemented. However, this implemented. However, this
solution
solution requires
requires further
further improvement,
improvement, and and toto minimize
minimize the the computations
computations of of the
the algorithm
algorithm inin the
the
future,
future, a better solution would be to devise a proper measure for assessing the outcome of the first
a better solution would be to devise a proper measure for assessing the outcome of the first
clustering,
clustering, then
then carrying
carrying itit out
out aa second
second time
time only
only ifif the
the first
first outcome
outcome isis not
not satisfactory.
satisfactory.
4. Conclusions
4. Conclusions
In this work, the implementation of a low-cost force sensor as a BCG and respiration
In this work, the implementation of a low-cost force sensor as a BCG and respiration sensory
sensory element is investigated. A preliminary setup and integration scheme, measuring the
element is investigated. A preliminary setup and integration scheme, measuring the longitudinal
longitudinal BCG of patients on a bed, is proposed and its raw signal is compared to the BCG
BCG of patients on a bed, is proposed and its raw signal is compared to the BCG setup
setup (transverse/dorsoventral) most commonly found in today’s bed-embedded BCG systems.
(transverse/dorsoventral) most commonly found in today’s bed-embedded BCG systems. To evaluate
To evaluate the usefulness of the proposed setup for heart rate monitoring, the ability to correctly
the usefulness of the proposed setup for heart rate monitoring, the ability to correctly recognize the
recognize the individual heartbeats contained in the filtered BCG signal is tested by an unsupervised
individual heartbeats contained in the filtered BCG signal is tested by an unsupervised machine
machine learning algorithm and referenced to the ECG. The algorithm, which was originally proposed
learning algorithm and referenced to the ECG. The algorithm, which was originally proposed by
by another research group [27], is further improved, thoroughly discussed, and tested with seven
another research group [27], is further improved, thoroughly discussed, and tested with seven
volunteers and four lying positions. The overall detection rate achieved was 83.9%.
volunteers and four lying positions. The overall detection rate achieved was 83.9%.
Since the setup depends on a mattress as a medium for the transmission of forces, the preload
Since the setup depends on a mattress as a medium for the transmission of forces, the preload
aspect of the initial sensor-mattress contact is briefly discussed and some preliminary correlations are
aspect of the initial sensor-mattress contact is briefly discussed and some preliminary correlations are
presented; admittedly, however, more tests and statistical analyses are recommended for the future
presented; admittedly, however, more tests and statistical analyses are recommended for the future
in order to thoroughly understand this interaction. By the same token, as suggested, there is always
in order to thoroughly understand this interaction. By the same token, as suggested, there is always
room for further refinement of the algorithm itself.
room for further refinement of the algorithm itself.
Author Contributions: Methodology, A.A. and F.L.; Software, A.A.; Validation, A.A.; Formal Analysis, A.A.;
Author Contributions:
Investigation, Methodology,
A.A. and F.L.; Resources, A.A. and F.L.; Software,
F.L.; Writing-Original A.A.;
Draft Validation,
Preparation, A.A.;
A.A.; Formal Analysis,
Writing-Review A.A.;
& Editing,
A.A., F.L. and U.M.;
Investigation, A.A. Supervision, U.M.; Project
and F.L.; Resources, F.L.;Administration, F.L.;
Writing-Original Funding
Draft Acquisition,
Preparation, U.M.
A.A.; Writing-Review &
Editing, A.A., F.L. and U.M.; Supervision, U.M.; Project Administration, F.L.; Funding Acquisition, U.M.

Funding: This research was funded by the Federal Ministry of Education and Research (BMBF) in Germany
under the IKT2020—Innovations in Care for People with Dementia funding program (iSenDi, reference code:
16SV7546).
Sensors 2019, 19, 1451 17 of 19

Funding: This research was funded by the Federal Ministry of Education and Research (BMBF) in Germany under
the IKT2020—Innovations in Care for People with Dementia funding program (iSenDi, reference code: 16SV7546).
Acknowledgments: The authors are grateful to Rui Zhu for the mechanical design of the load cell structure.
Special thanks to Alexander Bejan and the Future Care Lab staff at Furtwangen University for their permission and
accommodation to our use of their laboratory. We are also thankful to our volunteers for their time and patience.
Conflicts of Interest: The authors declare no conflict of interest. Moreover, the funders had no role in the design
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the
decision to publish the results.

Appendix A
A clarification of some of the terms found in the machine learning algorithm section is provided
in this appendix.

• Unsupervised pattern recognition:

Pattern recognition is the scientific discipline that aims to classify objects into a number of
categories or classes. These objects, which are referred to by the generic term patterns, can be images or
signal waveforms or any type of measurements that require classification. The unsupervised approach
deals with problems, where the categories or classes, into which the objects are to be classified, are not
known before the pattern recognition process takes place. Thus, the aim here would be to unravel
the underlying similarities (or dissimilarities) among the objects and group (cluster) similar objects
together, based on certain measures that quantify the similarities [35] (pp. 1–7).

• Agglomerative clustering:

A type of hierarchical clustering algorithm that begins with each object, i.e., its representative feature
vector, forming a cluster on its own, and produces a sequence of clusterings of decreasing number of
clusters at each step. That is, at each step a clustering is produced by merging two clusters from the
previous set of clusters. The algorithm ends with all objects contained in one cluster [35] (p. 629).

• Features and feature vectors:

Features are a number of variables or measurable properties of an object (pattern). A feature vector
is a convenient formation that contains these variables and should uniquely identify a single object;
therefore, features must be properly selected so as to encode as much information about the object as
possible. However, information redundancy among the features should be avoided. The representation
of objects by feature vectors facilitates their classification based on quantifiable proximity measures [35]
(pp. 5, 596).

• Proximity measure:

A proximity measure quantifies how “similar” or “dissimilar” two feature vectors are.
Thus, a proximity measure is identified either as a similarity measure or a dissimilarity measure.
The choice to implement one or the other depends on the application and the nature of the feature
vectors [35] (p. 597).

References
1. Prince, M.; Wimo, A.; Guerchet, M.; Ali, G.-C.; Wu, Y.-T.; Prina, M. World Alzheimer Report 2015: The Global
Impact of Dementia; Alzheimer’s Disease International: London, UK, 2015.
2. Nijhof, N.; van Gemert-Pijnen, L.J.E.W.C.; Woolrych, R.; Sixsmith, A. An evaluation of preventive sensor
technology for dementia care. J. Telemed. Telecare 2013, 19, 95–100. [CrossRef]
3. Gordon, J.W. Certain Molar Movements of the Human Body produced by the Circulation of the Blood.
J. Anat. Physiol. 1877, 11, 533–536. [PubMed]
Sensors 2019, 19, 1451 18 of 19

4. Pinheiro, E.; Postolache, O.; Girão, P. Theory and developments in an unobtrusive cardiovascular system
representation: Ballistocardiography. Open Biomed. Eng. J. 2010, 4, 201–216. [CrossRef] [PubMed]
5. Ballistocardiography, a Bibliography; National Aeronautics and Space Administration: Washington, DC, USA, 1965.
6. Starr, I.; Rawson, A.J.; Schroeder, H.A.; Joseph, N.R. Studies on the estimation of cardiac output in man,
and of abnormalities in cardiac function from the heart’s recoil and the blood’s impact; the ballistocardiogram.
Am. J. Physiol. 1939, 127, 1–28. [CrossRef]
7. Starr, I.; Braunstein, J.R.; Dock, W.; Gubner, R.; Hamilton, W.F.; Nickerson, J.L.; Rappaport, M.B.;
Scarborough, W.R.; Smith, J.E. First Report of the Committee on Ballistocardiographic Terminology.
Circulation 1953, 7, 929–931. [CrossRef]
8. Scarborough, W.R.; Talbot, S.A. Proposals for ballistocardiographic nomenclature and conventions: Revised
and extended report of Committee on Ballistocardiographic Terminology. Circulation 1956, 14, 435–450.
[CrossRef] [PubMed]
9. Thompson, W.B.; Rappaport, M.B.; Sprague, H.B. Ballistocardiography. II. The normal ballistocardiogram.
Circulation 1953, 7, 321–328. [CrossRef] [PubMed]
10. Gubner, R.S.; Rodstein, M.; Ungerleider, H.E. Ballistocardiography; an appraisal of technic, physiologic
principles, and clinical value. Circulation 1953, 7, 268–286. [CrossRef] [PubMed]
11. Giovangrandi, L.; Inan, O.T.; Wiard, R.M.; Etemadi, M.; Kovacs, G.T.A. Ballistocardiography—A Method
Worth Revisiting. In Proceedings of the 2011 Annual International Conference of the IEEE Engineering in
Medicine and Biology Society, Boston, MA, USA, 30 August–3 September 2011; pp. 4279–4282.
12. Starr, I.; Wood, F.C. Twenty-Year Studies with the Ballistocardiograph. Circulation 1961, 23, 714–732.
[CrossRef]
13. Soames, R.W.; Atha, J. Three-dimensional ballistocardiographic responses to changes of posture. Clin. Phys.
Physiol. Meas. 1982, 3, 169–177. [CrossRef] [PubMed]
14. Brüser, C.; Stadlthanner, K.; de Waele, S.; Leonhardt, S. Adaptive beat-to-beat heart rate estimation in
ballistocardiograms. IEEE Trans. Inf. Technol. Biomed. 2011, 15, 778–786. [CrossRef] [PubMed]
15. Henderson, Y. The mass-movements of the Circulation as shown by a Recoil Curve. Am. J.
Physiol.-Leg. Content 1905, 14, 287–298. [CrossRef]
16. Nickerson, J.L.; Curtis, H.J. The design of the Ballistocardiograph. Am. J. Physiol.-Leg. Content 1944, 142, 1–11.
[CrossRef]
17. Dock, W.; Taubman, F. Some technics for recording the ballistocardiogram directly from the body. Am. J. Med.
1949, 7, 751–755. [CrossRef]
18. Inan, O.T.; Etemadi, M.; Widrow, B.; Kovacs, G.T.A. Adaptive cancellation of floor vibrations in standing
ballistocardiogram measurements using a seismic sensor as a noise reference. IEEE Trans. Bio-Med. Eng.
2010, 57, 722–727. [CrossRef] [PubMed]
19. González-Landaeta, R.; Casas, O.; Pallàs-Areny, R. Heart rate detection from an electronic weighing scale.
Physiol. Meas. 2008, 29, 979–988. [CrossRef] [PubMed]
20. Gilaberte, S.; Gómez-Clapers, J.; Casanella, R.; Pallas-Areny, R. Heart and respiratory rate detection on a
bathroom scale based on the ballistocardiogram and the continuous wavelet transform. In Proceedings of
the 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology, Buenos Aires,
Argentina, 31 August–4 September 2010; pp. 2557–2560.
21. Baek, H.J.; Chung, G.S.; Kim, K.K.; Park, K.S. A smart health monitoring chair for nonintrusive measurement
of biological signals. IEEE Trans. Inf. Technol. Biomed. 2012, 16, 150–158. [CrossRef] [PubMed]
22. Koivistoinen, T.; Junnila, S.; Värri, A.; Kööbi, T. A new method for measuring the ballistocardiogram using
EMFi sensors in a normal chair. In Proceedings of the 26th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, San Francisco, CA, USA, 1–4 September 2004; pp. 2026–2029.
23. Zhu, Y.; Zhang, H.; Jayachandran, M.; Ng, A.K.; Biswas, J.; Chen, Z. Ballistocardiography with fiber optic
sensor in headrest position: A feasibility study and a new processing algorithm. In Proceedings of the 2013
35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC),
Osaka, Japan, 3–7 July 2013; pp. 5203–5206.
24. Pinheiro, E.; Postolache, O.; Girão, P. Study on Ballistocardiogram Acquisition in a Moving Wheelchair with
Embedded Sensors. Metrol. Meas. Syst. 2012, 19, 739–750. [CrossRef]
Sensors 2019, 19, 1451 19 of 19

25. Jung, D.W.; Hwang, S.H.; Yoon, H.N.; Lee, Y.-J.G.; Jeong, D.-U.; Park, K.S. Nocturnal awakening and sleep
efficiency estimation using unobtrusively measured ballistocardiogram. IEEE Trans. Bio-Med. Eng. 2014, 61,
131–138. [CrossRef] [PubMed]
26. Rosales, L.; Skubic, M.; Heise, D.; Devaney, M.J.; Schaumburg, M. Heartbeat detection from a hydraulic
bed sensor using a clustering approach. In Proceedings of the 2012 Annual International Conference of
the IEEE Engineering in Medicine and Biology Society, San Diego, CA, USA, 28 August–1 September 2012;
pp. 2383–2387.
27. Paalasmaa, J.; Ranta, M. Detecting Heartbeats in the Ballistocardiogram with Clustering. In Proceedings
of the ICML/UAI/COLT 2008 Workshop on Machine Learning for Health-Care Applications, Helsinki,
Finland, 9 July 2008.
28. Watanabe, K.; Watanabe, T.; Watanabe, H.; Ando, H.; Ishikawa, T.; Kobayashi, K. Noninvasive measurement
of heartbeat, respiration, snoring and body movements of a subject in bed via a pneumatic method.
IEEE Trans. Bio-Med. Eng. 2005, 52, 2100–2107. [CrossRef]
29. Chee, Y.; Han, J.; Youn, J.; Park, K. Air mattress sensor system with balancing tube for unconstrained
measurement of respiration and heart beat movements. Physiol. Meas. 2005, 26, 413–422. [CrossRef]
[PubMed]
30. Zhu, X.; Chen, W.; Nemoto, T.; Kanemitsu, Y.; Kitamura, K.-i.; Yamakoshi, K.-i.; Wei, D. Real-time monitoring
of respiration rhythm and pulse rate during sleep. IEEE Trans. Bio-Med. Eng. 2006, 53, 2553–2563.
31. Jin, J.; Wang, X.; Li, S.; Wu, Y. A Novel Heart Rate Detection Algorithm in Ballistocardiogram Based on
Wavelet Transform. In Proceedings of the 2009 Second International Workshop on Knowledge Discovery and
Data Mining, Moscow, Russia, 23–25 January 2009; Luo, Q., Gong, M., Eds.; Conference Pub. Services/IEEE
Computer Society: Los Alamitos, CA, USA, 2009; pp. 76–79.
32. Gomez-Clapers, J.; Serra-Rocamora, A.; Casanella, R.; Pallas-Areny, R. Towards the standardization of
ballistocardiography systems for J-peak timing measurement. Measurement 2014, 58, 310–316. [CrossRef]
33. Lima, F.; Albukhari, A.; Zhu, R.; Mescheder, U. Contactless Sleep Monitoring Measurement Setup. Proceedings
2018, 2, 1083. [CrossRef]
34. Rodríguez-Molinero, A.; Narvaiza, L.; Ruiz, J.; Gálvez-Barrón, C. Normal respiratory rate and peripheral
blood oxygen saturation in the elderly population. J. Am. Geriatr. Soc. 2013, 61, 2238–2240. [CrossRef]
[PubMed]
35. Theodoridis, S.; Koutroumbas, K. Pattern Recognition, 4th ed.; Elsevier/Academic Press: Amsterdam,
The Netherlands; London, UK, 2009.

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).