IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS , APRIL 2016 1
Automated Detection of Atrial Fibrillation Based on
Time-Frequency Analysis of Seismocardiograms
Tero Hurnanen, Eero Lehtonen, Mojtaba Jafari Tadi, Tom Kuusela, Tuomas Kiviniemi, Antti Saraste, Tuija
Vasankari, Juhani Airaksinen, Tero Koivisto, Mikko Pa¨nka¨a¨la¨
Abstract—In this paper, a novel method to detect atrial fibrilla-
tion from a seismocardiogram (SCG) is presented. The proposed
method is based on linear classification of the spectral entropy
and a heart rate variability index computed from the SCG. The
performance of the developed algorithm is demonstrated on data
gathered from 13 patients in clinical setting. After motion artefact
removal, in total 119 minutes of AFib data and 126 minutes
of sinus rhythm data were considered for automated atrial
fibrillation detection. No other arrhythmias were considered in
this study. The proposed algorithm requires no direct heartbeat
peak detection from the SCG data, which makes it tolerant
against interpersonal variations in the SCG morphology, and
noise. Furthermore, the proposed method relies solely on SCG
and needs no complementary electrocardiography (ECG) to
be functional. For the considered data, the detection method
performs well even on relatively low quality SCG signals. Using a
majority voting scheme which takes 5 randomly selected segments
from a signal and classifies these segments using the proposed
algorithm, we obtained an average true positive rate of 99.9%
and an average true negative rate of 96.4% for detecting atrial
fibrillation in leave-one-out cross-validation. The presented work
facilitates adoption of MEMS-based heart monitoring devices for
arrhythmia detection.
Index Terms—Seismocardiography, MEMS, accelerometer,
atrial fibrillation
I. INTRODUCTION
ATRIAL fibrillation (AFib) is a very common cardiacanomaly, present in approximately two percent of all
people, i.e. in approximately 140 million people globally [1].
During AFib, the atria fail to contract in a coordinated manner,
instead vibrating approximately 400 to 600 times per minute.
In this case, contraction of the ventricles is irregular and
excessively infrequent, for example 120 to 180 times per
minute. AFib poses a major diagnostic challenge, as symptoms
may be sporadic and absent during medical examinations.
Identifying scarcely occurring silent AFib thus requires long-
term continuous monitoring, e.g. months or years at a time [2].
Today, the heart’s activity can be investigated and measured
by several methods, of which the most common are electrocar-
diography (ECG), ultrasound cardiography (or echocardiogra-
T. Hurnanen, E. Lehtonen, M. Jafari Tadi, T. Koivisto and M. Pa¨nka¨a¨la¨
are with Technology Research Center, University of Turku, Finland. M.
Jafari Tadi is also with the Department of cardiology and cardiovascular
medicine, Faculty of Medicine, University of Turku, Finland. T. Kuusela
is with Department of Physics and Astronomy, University of Turku, Turku,
Finland. T. Kiviniemi, A. Saraste, T. Vasankari and J. Airaksinen are with
Heart Center, Turku University Hospital, Turku, Finland.
This work was supported by Finnish funding agency for innovation
(TEKES) under grant 40254/12, by the Academy of Finland under grant
277383, and by the University of Turku Graduate School.
phy), phonocardiography and photoplethysmography. One al-
ternative to the above mentioned techniques is mechanocardio-
graphy (MCG), where the idea is to assess the condition of the
heart [3] by measuring the mechanical activity (cardiac muscle
motion) of the heart. MCG has multiple names in the literature,
for example, ballistocardiography (BCG) and seismocardiog-
raphy (SCG) [4]–[6]. Although ECG is currently the dominant
technique, using microelectromechanical sensors (MEMS) for
heart telemonitoring has recently become a popular research
topic, e.g [7]–[9]. Miniaturized low-power, three-axis, high-
resolution and low-noise accelerometers enable inexpensive
access to mechanical cardiac monitoring via many existing
devices such as different smart devices (e.g. smart watches and
phones) [10], [11], as well as via dedicated hardware attached
for example with chest straps or even indirectly from beds,
chairs or similar set ups [12]–[15].
As the ECG currently is the gold standard for diagnosing
arrhythmias, a lot of research efforts have been targeted in
the development of EPMs (ECG Patch Monitors) [16]. Main
challenges with EPMs are that a reliable analysis requires good
electrode-to-patient contact and that the electrode distance
should be large enough to provide sufficient amplitude of
the biosignal [7] — a typical distance between electrodes
is in the order of 5-10 cm [7], [8]. Similar minimum size
requirements are present in insertable ECG sensors, such as
Medtronic’s Reveal LINQ ICM, whose length is approxi-
mately 4.5 cm. Furthermore, the electrode interface required
by ECG must be provided with additional hardware to existing
smart devices which slows down the penetration to consumer
market. In contrast, there are over two billion smart phone
users which already have MCG-ready device in their pocket.
Of course it would also be possible to integrate a MEMS-
based AFib detector into a single package (System in Package,
SiP) measuring few millimeters and attach it with a skin-
friendly adhesive to patient’s chest [8], [9], [17]. That kind of
cheap, noninvasive (and possibly disposable) device capable to
frequent monitoring of the cardiac activity in an unnoticeable
manner would be a preferred choice for massive screening
purposes of silent AFib. The third alternative is to enhance
the performance of EPMs with concurrent MCG. Since all
EPMs already include an accelerometer for motion artefact
removal/compensation the price to be paid for this extra
analysis remains almost negligible.
The first step toward the implementation of the above men-
tioned MCG applications is to evaluate the feasibility of seis-
mocardiography for AFib detection. To this end, we propose
an algorithm based on computing the spectral entropy [18] and
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS , APRIL 2016 2
a heart rate variability index of the measured SCG signal. This
automated classification procedure is based on the temporal
randomness in the SCG signal rather than finding individual
heart beats. This approach relaxes the requirements for the
acquired signal quality and therefore results in more reliable
detection results. A flowchart of the presented automated AFib
detection algorithm is shown in Fig.1.
This paper is based upon earlier publications [19], [20],
in which we represented a proof of concept and preliminary
results for the automatic detection of atrial fibrillation in
seismocardiograms. These methods are refined and explained
in detail in this manuscript, and a statistical analysis is in-
cluded. This paper is organized as follows. We review the most
significant publications related to this work in Subsection I-A.
In Section II the experimental study protocol used in this work
and the used data collection system are described. Section III
describes the signal processing methods used in this work.
Section IV concentrates on the evaluation of the developed
classification method with related experimental statistics. Sec-
tion V discusses the potential impact of this study and future
directions in this research, and Section VI concludes the paper.
Noise-removal filter: FFT-based brick
wall filter, bandpass 1-45 Hz. 
Motion artefacts removal filter: 
sliding window (500 ms) RMS filter, 
threshold-based artefact removal.
Subtraction of the estimated
respiration component given by
median filter (window size
125 ms). 
Rectify, multiply with Hamming
window and compute PSD for 
bandwidth of 2-8 Hz.
Divide signal into nonoverlapping
segments of 12.5 s. 
Estimate durations of 
cardiac cycles: divide
signal into overlapping
segments of 2.5 s 
(overlap 1.5 s) and 
compute cross-
correlation. Find the
location of the first side 
peak with threshold
opearation. 
Remove noise floor with
threshold operation and 
compute spectral entropy for 
normalized data.
Spectral entropyHRV index
Preprocessing
Estimate HRV: Compute
the median of absolute
differences between beat 
intervals.
Linear least squares classifier. 
Computation of spectral entropy and HRV index for each segment. 
Fig. 1. A flowchart of the proposed automated AFib detection algorithm, as
described in more detail in Sections III and IV.
A. Literature Review
Up to now, a number of techniques have been pro-
posed for an automatic detection of AFib. For example,
methods based upon Wavelet transform [21], Kolmogorov
entropy [22], sample entropy [23], time-frequency analysis
of heart rate variability [24], [25], Shannon entropy [26],
[27], approximate entropy [25], sequential hypothesis test-
ing [28], threshold-crossing intervals [29], auto-correlation
function [29], spectrum analysis [29] and algorithms based on
neural-networks [30], pattern recognition [31], and machine
learning techniques [32]. However, most of these techniques
are developed for the ECG signal and there are only a few
studies focusing on the detection of AFib from the mechanical
signals. In 2011, Bru¨ser et al. investigated the feasibility of
AFib detection in BCG signal using support vector machine
(SVM) classifiers and latter in 2013 reported results achieved
for analysis of ballistocardiogram signals using machine learn-
ing techniques [33], [34]. Accordingly, the best algorithm,
random forest, achieved a sensitivity (or true positive rate)
and specificity (or true negative rate) of 93.8% and 98.2%,
respectively. This result was achieved by analyzing 856 epochs
(each 30 s) of BCG signal recorded from 10 patients. A recent
study reported a pitch-tracking technique for cycle length
analysis of BCG signals in AFib and normal sinus rhythm
modes [35]. However, as no quantitative results were reported,
it is difficult to estimate the accuracy of the method in AFib
detection.
II. DATA ACQUISITION
A. Study Protocol
The clinical study was performed in the Heart Center,
Turku University Hospital, Finland, by using the measurement
system described in Subsection II-B. The research protocol
was approved by the Ethical Committee of the Hospital
District of the South-Western Finland. This study included 13
patients with AFib who were medically treated or to whom a
cardioversion was anticipated. Of these 13 patients, 12 were
male and one was female. Other demographics of the patients
were as follows (min-max, mean, standard deviation): age (38-
71, 60.7, 9.1 years), height (165-193, 180.4, 8.8 cm), mass
(77-123, 92.2, 11.8 kg), BMI (23.5-45.2, 28.6, 5.4 kg/m2).
Patients were enrolled from the outpatient clinic, from those
referred for cardioversion and from the ward. The main criteria
for inclusion were:
• Age at least 18 years old
• History of AFib requiring medical therapy or cardiover-
sion
• Patient is willing to comply with specified evaluations
• Patient or legally authorized representative has been in-
formed of the nature of the study, agrees to its provi-
sions and has been provided written informed consent,
approved by the appropriate review board.
The main exclusion criteria for this study were:
• Ages under 18 years old
• Any significant medical condition, which in the inves-
tigators opinion may interfere with the patients optimal
participation in the study
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS , APRIL 2016 3
0 1 2 3 4 5
−10
0
10
AFib
Time (s)
Ac
ce
le
ra
tio
n 
(m
g n
)
−5 0 5
0
0.5
1
Time (s)A
ut
oc
or
re
la
tio
n 
of
 S
CG
0 1 2 3 4 5
−100
0
100
Time (s)
EC
G
0 1 2 3 4 5
−10
0
10
sinus rhythm
Time (s)
Ac
ce
le
ra
tio
n 
(m
g n
)
−5 0 5
0
0.5
1
Time (s)A
ut
oc
or
re
la
tio
n 
of
 S
CG
0 1 2 3 4 5
−100
0
100
Time (s)
EC
G
Fig. 2. Five seconds of SCG data measured from the same patient. Left panels: AFib. Right panels: sinus rhythm (after successful cardioversion). Bandpass
filtered SCG signal is plotted in upper panels (gn is earth’s gravity) and corresponding ECG signal in middle panels. Autocorrelations of the SCG signals are
plotted in lower panels to demonstrate the change in the regularity of the heart rhythm.
Two sets of SCG data were collected from each patient in
supine position. The first data set was acquired during AFib
and the second in sinus rhythm after successful cardioversion.
The measurement took approximately 30 minutes per patient.
B. Data Collection
The custom-made data collection system used in this study
consisted of a seismocardiograph, electrocardiograph and a
data acquisition board for storing synchronized SCG and
ECG data on a memory card. SCG was obtained using a
three-axis, low-power, micro electromechanical accelerometer
(MMA8451Q from Freescale Semiconductor), with 14 bits of
resolution and the size of 3 mm × 3 mm × 1 mm. MMA-
sensor was carefully selected because it is low-power, small
size, has enough output bits and reasonable noise floor. There
are MCG studies conducted with more accurate accelerometers
e.g. [36], [37], but using superior sensors was avoided in
this study to maintain the compatibility of the presented
AFib detection method with the current and near future smart
phones. The sensor was attached to the body of sternum using
double-sided tape without hair removal in the chest area.
ECG was recorded for reference purposes only. The ECG
electrodes were mounted on the anterior lateral regions to
the abdomen on the left and the right hypochondriac leading
to one-lead setup. Data acquisition was hosted by Freescale
FRDM-KL25Z board, which sampled both data at a sampling
frequency (FS) of 800 Hz and stored the collected data on
a memory card. Data processing was performed offline after
transferring the binary data from the memory card to the
computer [38].
A typical sample of the SCG signal recorded from a patient
within this study with the above described data collection
system is shown in Fig. 2. The data on the left panel of
Fig. 2 is recorded during AFib, while the data on the right
panel is from the same patient during sinus rhythm after
succesful cardioversion. The corresponding ECG signals are
also plotted for reference. Autocorrelation curves computed
from the plotted SCG samples (5 s) demonstrate the change
in the periodicity of heart rate between the sinus rhythm and
AFib.
III. SIGNAL PROCESSING METHODS
In the following a relatively simple and computationally
light algorithm for the detection of AFib from SCG signal
is presented. The motivation for this work is to investigate
methods which allow for automatic detection of AFib episodes
solely from SCG, while being as tolerant against interpersonal
variations in the SCG morphology and noise as possible.
A. Preprocessing of the SCG signal
The first preprocessing step is to apply a noise-removing
filter on the acquired SCG signal. To this end we use an FFT
based brick-wall filter with the bandpass frequency range of
[1 Hz, 45 Hz]. It should be noted that while the heart rate
can be lower than 60 bpm, the majority of heart-related SCG
spectrum is still above 1 Hz. Subsequently, motion artefacts
are removed from the signal by applying a sliding window
(500 ms) root mean square (RMS) filter and then determining
the median value of the filtered signal. Parts of the signal
exceeding twice this median value are considered to be motion
artefacts and are discarded. The remaining, artefactless signal
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS , APRIL 2016 4
is divided into non-overlapping segments of 10000 samples
— which equals 12.5 seconds. The selection criterion for this
experimentally determined segment size was that it yielded
sufficient accuracy in the subsequent steps of the algorithm.
Shorter segments would reduce the accuracy of detection
algorithm, while longer segments would reduce the number of
segments. Fig. 3 demonstrates the performance of the motion
artifact detection algorithm on a noisy SCG signal. The upper
panel of Fig. 3 shows the energy envelope of the signal and
the median value based noise removal threshold is drawn
with a dashed line. In the lower panel of Fig. 3, the motion
artefacts in the original (bandpass filtered) SCG signal are
detected, and hence can be omitted from further analysis. In
this example, the algorithm finds eleven motion artefact-free
signal segments, marked with numbers and dashed vertical
lines. In the following we denote the resulting motion artefact-
Time (s)
0 20 40 60 80 100 120 140 160
SC
G
 S
ig
na
l (m
g n
)
-20
0
20
1 2 3 4 5 6 7 8 9 10 11
Time (s)
0 20 40 60 80 100 120 140 160
Si
gn
al
 e
nv
el
op
e
(n
or
ma
liz
ed
)
0
5
10
15
Detection Threshold
Fig. 3. Detection of motion artefacts from mechanical signals using the RMS
filter. Upper panel: the RMS-filtered signal and the threshold for the motion
artefact detection. The signal is normalized so that the threshold value is 1.
Bottom panel: motion artefacts are detected and highlighted in light gray.
Signal segments which are determined to be artefact-free, are numbered and
marked with dashed vertical lines.
free SCG signal segments by s(t). These segments are used
for the computation of the spectral entropy and the heart rate
variability index.
B. Spectral entropy
Our hypothesis is that by analyzing the randomness of
the preprocessed SCG signal it is possible to determine if
it corresponds to sinus rhythm or AFib. A specific way to
measure the randomness of a signal segment is to compute its
spectral entropy [18] using FFT analysis and power spectral
density (PSD). To compute the spectral entropy of a cardiac
signal segment, the following steps are applied.
The first step is to remove the respiration component. Let
us assume that the considered signal segments s(t) consist of
three additive components as
s(t) = sh(t) + sb(t) + n(t), (1)
where sh(t) is the acceleration signal segment of interest
caused by the heart motion, sb(t) corresponds to the respi-
ration component, and n(t) includes all the other residual ac-
celeration signals and noise after removal of motion artefacts.
The above described preprocessing steps significantly reduce
the power of the third component n(t). Thus it can be assumed
to be small in comparison to sh(t) and sb(t) in the following.
The effect of breathing component sb(t) was reduced by
subtracting an estimate of the breathing from the signal
segment s(t). The estimated breathing signal was obtained
by applying a median filter to s(t) with a window length
of 100 samples — which equals 0.125 seconds with the
considered sampling frequency of 800 Hz. This window size
was found to be a good compromise for this filter. It is
short enough to capture sb(t), but long enough to effectively
filter out the cardiac-related part sh(t), thus resulting into
good estimate of breathing component. The final approximated
cardiac acceleration signal segment is then given by
sˆh(t) = s(t)−Median100(s(t)). (2)
The second step is to rectify sˆh(t) and consider only its
positive values. The resulting signal segment is denoted by
sˆh,rect.(t), where
sˆh,rect.(t) =
{
sˆh(t), if sˆh(t) ≥ 0,
0, otherwise.
This rectification step is adopted from [39], where it is ex-
plained to restore the missing periodicity information, which is
beneficial when trying to distinguish AFib from sinus rhythm
on the basis of the peridiocity of the signal.
The third step is to compute the PSD for the signal segment
using FFT. Prior to taking the FFT, the signal was multiplied
with the Hamming window in order to reduce the edge effects
of the FFT and smoothen the spectrum. Finally the FFT was
squared to get an approximation of the PSD. Let us denote
this approximation of the power spectral density of sˆh(t) by
p(f), where f is limited to the interval [2 Hz, 8 Hz].
Fig. 4 presents samples of the bandwidth limited power
spectral density curves p(f) for the sinus rhythm and AFib. It
can be observed that the power spectrum of the sinus rhythm
signal is divided between a smaller number of frequencies in
comparison to the AFib signal, which has a more wide-band
signal structure, as can be expected from the random nature
of AFib.
Before computing the spectral entropies, an extra threshold
operation given by
pˆ(f) =
{
0, if p(f) < Θ,
p(f), otherwise,
is applied in order to remove the noise floor from the PSD. The
sufficient threshold value Θ depends mainly on the amount of
noise in the sensor, and was once again determined experimen-
tally in this work. It was found out that Θ = max(p(f))/6 is
a good choice with the available data. Next pˆ(f) is normalized
sum up to 1 by setting
pˆnorm(f) = pˆ(f)/
∑
f
pˆ(f).
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS , APRIL 2016 5
This normalization is done in order to compute the spectral
entropy, as it essentially considers the frequency spectrum as
a probability distribution. The spectral entropy S of pˆnorm(f)
is then computed as
S = −
∑
f
pˆnorm(f) log(pˆnorm(f)). (3)
The computed spectral entropy S is higher for a SCG signal
segment containing AFib than for sample containing sinus
rhythm only. The computed spectral entropies for the AFib
signal segment and a sinus rhythm signal segment of example
of Fig. 4 are approximately 5.65 and 2.33, respectively, where
natural logarithm is used in (3). In Section IV we describe
how the discrimination between AFib and sinus rhythm can
be automated by using the computed spectral entropy values.
0 2 4 6 8 10 12
0
0.02
0.04
0.06
0.08
0.1
0.12
Frequency (Hz)
Po
w
er
 (n
orm
ali
ze
d)
Power spectral density
0 2 4 6 8 10 12
0
0.02
0.04
0.06
0.08
0.1
0.12
Frequency (Hz)
Po
w
er
 (n
orm
ali
ze
d)
Power spectral density
Fig. 4. Power spectral density curves obtained by using the method described
in the text part. Upper panel: power spectral density of an SCG signal
corresponding to sinus rhythm. Lower panel: power spectral density of an
SCG signal corresponding to AFib. The power spectral densities have been
normalized so that the summed PSD between 2 Hz and 8 Hz equals 1. The
spectral entropies for these PSD curves are 2.78 and 5.74, for the sinus rhythm
and the AFib, respectively.
C. Heart rate variability
In addition to spectral entropy, the variations between
consecutive interbeat time intervals, defined as the heart rate
variability (HRV) [40] is used as a second indicator of the
randomness of the cardiac signal in this study. The HRV is
considered to be a key indicator of the cardiovascular condition
of an individual [40], [41].
To compute the amount of HRV in a given 12.5 s segment
of SCG signal, the durations of the individual cardiac cycles
in this signal must first be estimated. Due to the signal quality
we are unable to detect individual heart beats from the signal.
Autocorrelation based methods have been proposed to find
heart rate in ultrasound signal [42] and in ballistocardiography
signal [43]. Correlation based methods are beneficial, as no
explicit peak detection algorithms are needed, and hence the
methods tolerate noise and interpersonal variations in the sig-
nal morphology. In this work the following cross-correlation
based method is applied to get this estimation.
1) Determining the cardiac cycle duration: To begin with,
the acquired 12.5 s segment of SCG signal is divided into
smaller subsegments with the duration of 2.5 seconds. The
subsegment size is chosen so that each subsegment should
contain at least two heartbeats. This assumption is true as
long as the instantaneous heart rate is at least 48 beats per
minute. The segmentation is performed so that consecutive
subsegments overlap by 1.5 seconds. Therefore, each 12.5 s
segment s(t) consists of eleven 2.5 s subsegments, which are
all unique but should share at least one heart beat with the
neighboring subsegment.
Let us consider a 2.5 second subsegment u(t) ⊂ s(t). If
u(t) contains exactly two heartbeats, then the time interval
between these heartbeats (or duration of the single cardiac
cycle) can be accurately estimated by computing the period of
u(t). To achieve this, the first 1.5 seconds of u(t) — denoted
by u1.5(t) in the following — is cross-correlated with u(t).
This yields
R(u(t), i) =
∑
j
u(j)u1.5(j + i), (4)
where j is a discrete variable denoting the time indices, and
only positive indices j + i up to the number of samples in
u1.5(j) are taken into account.
The duration of the corresponding cardiac cycle can now
be estimated by locating the first side peak of R(i). Formally,
this is performed by computing the index of the first side peak
ifirst peak = arg|i|>i0 max(R(u(t), i), (5)
where i0 is a threshold, which is chosen to be i0 = 1/3Fs.
The justification for this choice is that io corresponds to
a period of at least 1/3 seconds in the signal, which is
sufficient threshold for heart rates below 180 bpm. Now the
corresponding estimated cardiac cycle duration is obtained as
d = ifirst peak/Fs. (6)
Although the atria contract fast at rate of 400 to 600 beat per
minute in AF, the AV node allows only occasional impulses to
pass through. Therefore, the ventricular rate registered by SCG
during AFib is typically from 160 to 180 beats per minute [44].
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS , APRIL 2016 6
Time (s)
1 2 3 4 5 6 7 8
Es
tim
at
ed
 c
yc
le
 d
ur
at
io
n 
d i
0.5
1
1.5
AFib
Time (s)
1 2 3 4 5 6 7 8
|d i
+1
-
d i
|
0
0.1
0.2
0.3
0.4
0.5
0.6
Time (s)
1 2 3 4 5 6 7 8
Es
tim
at
ed
 c
yc
le
 d
ur
at
io
n 
d i
0.5
1
1.5
Sinus rhythm
Time (s)
1 2 3 4 5 6 7 8
|d i
+1
-
d i
|
0
0.1
0.2
0.3
0.4
0.5
0.6
Fig. 5. HRV determination. Left side panels correspond to a signal acquired
during AFib, and the right side panels to signal acquired during sinus rhythm.
The estimated cardiac cycle durations di are presented in the upper panels,
and their absolute differences in the bottom panels. Median of the absolute
change, which is the used as the estimate for the HRV value, is plotted with
dashed line. The calculated HRV values are 0.30 and 0.03 for AFib and sinus
rhythm, respectively.
However, the considered 2.5 second subsegment may con-
tain more than two heartbeats. In that case, the resulting cross-
correlation can contain strong side peaks which correspond to
shifts by other than the desired cardiac cycle. In such situation
we may either detect wrong cardiac cycle or two consecutive
cardiac cycles as one cycle. The first one is not harmful, but
the latter can be considered as an outlier that will be dealt
with by using median function, as explained in the following.
2) Estimating the heart rate variability: Let U be a set
of above-defined subsegments ui(t) from a given SCG signal
segment, where i = 1, . . . n. Each ui(t) is associated with
the corresponding estimated duration of the cardiac cycle di.
In this paper, the estimated heart rate variability (HRV) is
given by the median of absolute differences in subsequent
cycle durations:
HRV = Median({|di+1 − di| : i = 2, 3, . . . , n}). (7)
This HRV index measures the amount of change in the
durations between subsequent cardiac cycles. The median
function is used here to protect this index against outliers and
incorrectly determined cardiac cycle durations. For the clas-
sification described in the following Section, the logarithmic
HRV is considered
HRVlog = log(1 + HRV). (8)
The justification for using the logarithmic HRV is that the
dynamic range of HRV is large in this study, and therefore
the logarithmic HRV makes visualization easier and also
results in a minor improvement in the classification. Examples
of obtained HRV curves, and their absolute differences and
median values, are presented in Fig. 5.
Spectral entropy
2 2.5 3 3.5 4 4.5 5 5.5 6 6.5
N
um
be
r o
f s
am
pl
es
0
5
10
15
20
25
30
35
40
Sinus
AFib
HRVlog
0 1 2 3 4 5 6 7 8
N
um
be
r o
f s
am
pl
es
0
5
10
15
20
25
30
35
40
45
Sinus
AFib
Fig. 6. Detection of atrial fibrillation using a single variable. Bars on the left
side of the boundary are classified as segments which contain sinus rhythm and
points on the right side the boundary are classified as segments which contain
atrial fibrillation. Top panel: classification using the spectral entropy. The
classification boundary obtained from a linear least-squares classifier equals
SE = 4.6. Bottom panel: classification using the logarithmic HRV, where the
linear boundary equals HRVlog = 4.0.
IV. AUTOMATED DETECTION OF ATRIAL FIBRILLATION
USING SPECTRAL ENTROPY AND HRV
In the following we use the spectral entropy and the HRV
index to determine whether a given SCG signal has been
recorded during atrial fibrillation or not. For this, all the data
acquired from the 13 patients was preprocessed according to
steps defined in Section III to obtain cardiac signal segments
s(t). After applying the spectral entropy and HRV estimation
algorithms to each signal segment, the histograms shown in
Fig. 6 are obtained. In Fig. 7 the results of both analyses are
combined so that the y-axis shows the location of the assessed
sample with respect to spectral entropy and x-axis with respect
to the logarithmic HRV. Both of presented figures also contain
linear decision boundaries to separate the clouds of points to
belonging to sinus rhythm and atrial fibrillation to demonstrate
that visually the linear classifier works very well.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS , APRIL 2016 7
HRVlog
1 2 3 4 5 6
Sp
ec
tra
l e
nt
ro
py
2
3
4
5
6
7 Sinus rhythmAtrial fibrillation
Fig. 7. Detection of atrial fibrillation using both the spectral entropy and the
logarithmic HRV. For visualization purposes we have restricted the number of
points in this scatter plot to 200. The classification boundary obtained from
a linear least-squares classifier equals SE = −1.1HRVlog + 8.8; a segment
corresponding to a point above this boundary is classified as atrial fibrillation,
whereas a segment corresponding to a point below this boundary is classified
to contain no atrial fibrillation.
A. Linear least-squares classifier with majority voting
A linear least-squares classifier — using only spectral
entropy, only logarithmic HRV, or both — was trained using
a training set of 12 patients’ data, and tested using a data
from the remaining patient. The test was performed for both
AFibPos signal and AFibNeg signal recorded from that patient.
The classification procedure was performed multiple times,
each time considering the data from a different patient as a
test set, and the data from the rest of the patients as the training
set. For a given patient, the classification was performed 1000
times, and the average classification performance was recorded
as presented in Table I. This leave-one-out cross-validation
was applied to reduce the effect of low amount of data.
In the first trial, only one signal segment from a random
temporal location for each case — AFibPos and AFibNeg
— per patient was designated to the test set, and that signal
segment was then classified with the trained linear classifier.
Hence this test required 12.5 seconds of data per patient.
To improve the classification performance, in the subsequent
trials an uneven number m = 3 or m = 5 of segments —
were randomly selected from the same patient, all segments
either AFibPos or AFibNeg — and the these signal segments
were classified using the trained classifier, and the final result
was determined using the majority rule. In other words, this
set of m segments was classified as AFibPos if more than
m/2 of the segments were classified by the linear classifier as
AFibPos, while otherwise the set was classified as AFibNeg.
The performance of the classifier was assessed by computing
its true positive and true negative rates. By true positive we
mean classification of an AFibPos signal as atrial fibrilla-
tion, and by true negative the classification of an AFibNeg
signal as sinus rhythm. False positives and false negatives
are defined accordingly as falsely classified atrial fibrillation
SE vote size 1 3 5
patient TPR TNR TPR TNR TPR TNR
1 0.823 0.801 0.909 0.908 0.947 0.964
2 0.954 0.955 0.998 0.997 1.000 0.999
3 0.850 0.797 0.933 0.870 0.962 0.932
4 0.970 0.780 0.999 0.868 1.000 0.922
5 0.818 0.893 0.909 0.971 0.965 0.988
6 0.963 0.401 0.999 0.375 1.000 0.331
7 0.959 0.917 0.992 0.979 1.000 0.993
8 0.810 0.915 0.895 0.967 0.950 0.984
9 0.827 0.952 0.901 0.987 0.964 0.995
10 0.886 0.964 0.972 0.998 0.987 1.000
11 0.813 0.802 0.915 0.902 0.970 0.937
12 1.000 0.612 1.000 0.685 1.000 0.728
13 0.885 1.000 0.954 1.000 0.985 1.000
mean 0.889 0.830 0.952 0.885 0.979 0.906
std 0.071 0.167 0.043 0.176 0.020 0.188
HRV vote size 1 3 5
patient TPR TNR TPR TNR TPR TNR
1 1.000 0.828 1.000 0.906 1.000 0.953
2 0.979 0.986 1.000 1.000 1.000 1.000
3 0.974 0.867 0.998 0.960 1.000 0.981
4 1.000 0.750 1.000 0.818 1.000 0.863
5 1.000 0.673 1.000 0.778 1.000 0.811
6 1.000 0.769 1.000 0.835 1.000 0.907
7 0.987 0.976 1.000 0.998 1.000 0.999
8 0.981 0.797 0.997 0.901 1.000 0.931
9 1.000 0.966 1.000 0.997 1.000 0.999
10 0.960 1.000 0.996 1.000 0.999 1.000
11 0.815 1.000 0.909 1.000 0.964 1.000
12 1.000 0.961 1.000 0.992 1.000 0.998
13 1.000 1.000 1.000 1.000 1.000 1.000
mean 0.977 0.890 0.992 0.937 0.997 0.957
std 0.050 0.115 0.025 0.081 0.010 0.062
SE+HRV vote size 1 3 5
patient TPR TNR TPR TNR TPR TNR
1 1.000 0.815 1.000 0.917 1.000 0.949
2 0.979 0.983 0.998 1.000 1.000 1.000
3 1.000 0.834 1.000 0.944 1.000 0.981
4 1.000 0.756 1.000 0.829 1.000 0.902
5 0.985 0.738 1.000 0.877 1.000 0.910
6 0.959 0.741 0.997 0.839 0.998 0.895
7 1.000 0.947 1.000 0.991 1.000 1.000
8 1.000 0.824 1.000 0.905 1.000 0.950
9 0.967 0.971 0.992 0.999 1.000 1.000
10 0.982 1.000 0.999 1.000 1.000 1.000
11 0.869 0.984 0.957 0.999 0.990 1.000
12 1.000 0.800 1.000 0.904 1.000 0.951
13 1.000 1.000 1.000 1.000 1.000 1.000
mean 0.980 0.876 0.996 0.939 0.999 0.964
std 0.036 0.106 0.012 0.064 0.003 0.041
TABLE I
AVERAGE CLASSIFICATION PERFORMANCES USING LEAVE-ONE-OUT
CROSS-VALIDATION. EACH VALUE IS AVERAGED OVER 1000 RANDOM
TRIALS WITH THE CORRESPONDING PATIENT’S DATA AS THE TEST SET.
TOP: PERFORMANCE USING ONLY SPECTRAL ENTROPY. MIDDLE:
PERFORMANCE USING ONLY HRV. BOTTOM: PERFORMANCE USING BOTH
SPECTRAL ENTROPY AND HRV.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS , APRIL 2016 8
and sinus rhythm, respectively. As expected, the true positive
rate (TPR) and the true negative rate (TNR) of the detection
were improved by taking more segments into account. Table I
summarizes the results of different classification methods. In
the table, each row represents results of one patient being used
as the test group. The classifier is trained with the remaining
12 patients.
All three choices of variables for the detection of atrial
fibrillation perform well for the considered data. When the
number of segments is increased, the true positive rate and
the true negative rate of the classification improves, and for
five segments, both of them are above 0.95 when using
only HRV, or when using both spectral entropy and HRV.
With the considered data, using only spectral entropy for the
classification seems to perform slightly worse as compared to
using only HRV or both HRV and spectral entropy.
V. DISCUSSION
It is practically significant to develop the automated detec-
tion of atrial fibrillation from an SCG signal obtained using
a MEMS accelerometer, as this facilitates the development of
miniaturized smart sensors for screening of atrial fibrillation
from masses. This is an important task, as atrial fibrillation
is a common cardiac anomaly that appers to cause of up
to 25 percent of all strokes. Therefore an inexpensive and
unobtrusive way to constantly monitor the rhythm of the heart
for scarcely occurring silent atrial fibrillation would be useful
from both personal health’s and economics’ perspective. For
example, patients with unexplained syncope could benefit from
such a long-term monitoring in order to find an explanation
for the symptoms.
This work shares some similarities with [34], but there are
also many differences. These two studies are briefly compared
in the following. The apparent similarity between these two
studies is that both try to detect AFib noninvasively with me-
chanical sensors instead of conventional ECG approach. Also,
the measurement scenario is the same: patients are measured
before and after cardioversion. Furthermore, the amount of
patients and the mean length of the individual recordings are
the same order of magnitude. In [34] the recordings were
performed with bed-mounted sensor (BCG), whereas in this
study the sensor was attached directly to chest (SCG). That
is, the former method is contactless the latter is not. Both
approaches have their pros and cons. For example, the bed-
mounted sensor allows recordings lasting the whole nighttime
and powering the measurement equipment is easy. Instead,
the wireless chest-mounted sensor (e.g. patch or smartphone)
needs to be battery-powered, but can also take measurements
during daytime when the person is staying still. There are
also significant differences in the computational approaches
between these studies. Specifically, [34] focuses on comparing
the performance of seven popular machine learning algorithms
(with automatically selected 17 general features) in detecting
AFib. The results are classified into three categories: normal,
AFib and motion artifact. Instead, the focus of the presented
work is to show that the used spectral entropy and HRV
measures contain sufficient information to classify between
SCG signals corresponding to AFib and sinus rhythm after
noise artefacts have been removed. In the light of the results,
these methods have comparable detection performance, but
as they are based on different features, it is possible that
an appropriate combination of the methods could yield even
higher detection rates of atrial fibrillation.
Currently, ECG is the state-of-the-art method for moni-
toring heart function, and the Holter monitor can be used
to determine arrhythmias in longer measurements. However,
wearable ECG devices require a good electrical contact, which
in practice entails attaching adhesive electrodes firmly onto the
skin, from which body hair and the surface layer have been
removed. Electrical contact is further enhanced by means of
a special gel between the electrode and the skin. Electrodes
attached in this manner may irritate the skin and for this
reason, ECG-based techniques are infeasible for very long-
term monitoring; typically, Holter monitors are used up to a
period of two weeks [16]. The ECG devices equipped with
electrodes to be touched with fingers do not suffer from such
problems but are only good for occasional short term check
ups and are therefore inferior as compared to Holter monitors
when trying to catch silent AFib.
Besides miniaturized heart monitoring patches, the pre-
sented algorithms can also be implemented in a smart phone.
Here the idea would be to measure the rhythm of the heart for
example once per day by lying in a supine position with the
smart phone placed on the chest. A straightforward advantage
in comparison to ECG is that no additional devices are
required; the internal sensors of the smart phone are sufficient
for such measurements. Another possible use for (MEMS)
accelerometer -based atrial fibrillation detection would be to
use the proposed algorithm with bed sensors, where the data
would be acquired while the patient is sleeping as suggested
by e.g. [13].
VI. CONCLUSION
In this paper a novel method for detecting atrial fibrillation
from an SCG signal measured by using a MEMS accelerome-
ter was presented. For automated detection of atrial fibrillation
a linear least-squares classifier was used with spectral entropy
and heart rate variability measure as input variables. The
proposed method tolerates well interpersonal variations in the
signal morphology and also noise, as it does not require
accurate beat detection from the SCG signal. The obtained
results are promising and warrant further investigation into this
topic. It should be noted that in this work the focus was on
classification between atrial fibrillation and sinus rhythm; the
effect of other arrhythmias to the quality of the classification
was left for future work.
REFERENCES
[1] J. S. Healey, S. J. Connolly, M. R. Gold, C. W. Israel, I. C. Van Gelder,
A. Capucci, C. Lau, E. Fain, S. Yang, C. Bailleul, C. A. Morillo,
M. Carlson, E. Themeles, E. S. Kaufman, and S. H. Hohnloser, “Sub-
clinical atrial fibrillation and the risk of stroke,” New England Journal
of Medicine, vol. 366, no. 2, pp. 120–129, 2012, pMID: 22236222.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS , APRIL 2016 9
[2] E. I. Charitos, U. Stierle, P. D. Ziegler, M. Baldewig, D. R. Robinson,
H.-H. Sievers, and T. Hanke, “A comprehensive evaluation of rhythm
monitoring strategies for the detection of atrial fibrillation recurrence:
Insights from 647 continuously monitored patients and implications for
monitoring after therapeutic interventions,” Circulation, vol. 126, no. 7,
pp. 806–814, 2012.
[3] I. Starr, A. J. Rawson, H. A. Schroeder, and N. R. Joseph, “Studies on
the estimation of cardiac output in man, and of abnormalities in cardiac
function,” American Heart Journal, vol. 18, no. 4, pp. 506 –, 1939.
[4] J. W. Gordon, “Certain molar movements of the human body produced
by the circulation of the blood,” Journal of Anatomy and Physiology,
vol. 11, pp. 533–536, 1877.
[5] B. S. Bozhenko, “Seismocardiography– a new method in the study of
functional conditions of the heart,” Terapevticheskii Arkhiv, vol. 33, pp.
55–64, 1961.
[6] O. Inan, P.-F. Migeotte, K.-S. Park, M. Etemadi, K. Tavakolian,
R. Casanella, J. Zanetti, J. Tank, I. Funtova, G. Prisk, and M. Di Rienzo,
“Ballistocardiography and seismocardiography: A review of recent ad-
vances,” Biomedical and Health Informatics, IEEE Journal of, vol. 19,
no. 4, pp. 1414–1427, July 2015.
[7] Y. Chuo, M. Marzencki, B. Hung, C. Jaggernauth, K. Tavakolian, P. Lin,
and B. Kaminska, “Mechanically flexible wireless multisensor platform
for human physical activity and vitals monitoring,” IEEE transactions
on biomedical circuits and systems, vol. 4, no. 5, pp. 281–294, 2010.
[8] M. Etemadi, O. T. Inan, J. A. Heller, S. Hersek, L. Klein, and S. Roy, “A
wearable patch to enable long-term monitoring of environmental, activity
and hemodynamics variables,” IEEE transactions on biomedical circuits
and systems, vol. 10, no. 2, pp. 280–288, 2016.
[9] P. Castiglioni, A. Faini, G. Parati, and M. Di Rienzo, “Wearable
seismocardiography,” in 2007 29th Annual International Conference of
the IEEE Engineering in Medicine and Biology Society. IEEE, 2007,
pp. 3954–3957.
[10] J. A. Heathers, “Smartphone-enabled pulse rate variability: An alter-
native methodology for the collection of heart rate variability in psy-
chophysiological research,” International Journal of Psychophysiology,
vol. 89, no. 3, pp. 297 – 304, November 2013, psychophysiology in
Australasia conference.
[11] J. Hernandez, D. McDuff, and R. Picard, “Bioinsights: Extracting
personal data from still wearable motion sensors,” in Wearable and
Implantable Body Sensor Networks (BSN), 2015 IEEE 12th International
Conference on, June 2015, pp. 1–6.
[12] W. Jia, Y. Li, Y. Bai, Z.-H. Mao, M. Sun, and Q. Zhao, “Estimation of
heart rate from a chest-worn inertial measurement unit,” in Bioelectron-
ics and Bioinformatics (ISBB), 2015 International Symposium on, Oct
2015, pp. 148–151.
[13] C. Bru¨ser, K. Stadlthanner, S. de Waele, and S. Leonhardt, “Adaptive
beat-to-beat heart rate estimation in ballistocardiograms,” Information
Technology in Biomedicine, IEEE Transactions on, vol. 15, no. 5, pp.
778–786, Sept 2011.
[14] O. Postolache, P. Girao, G. Postolache, and M. Pereira, “Vital signs
monitoring system based on emfi sensors and wavelet analysis,” in
Instrumentation and Measurement Technology Conference Proceedings,
2007. IMTC 2007. IEEE, May 2007, pp. 1–4.
[15] J. Ramos-Castro, J. Moreno, H. Miranda-Vidal, M. Garcia-Gonzalez,
M. Fernandez-Chimeno, G. Rodas, and L. Capdevila, “Heart rate
variability analysis using a seismocardiogram signal,” in Engineering
in Medicine and Biology Society (EMBC), 2012 Annual International
Conference of the IEEE, Aug 2012, pp. 5642–5645.
[16] J. A. Walsh, E. J. Topol, and S. R. Steinhubl, “Novel wireless devices
for cardiac monitoring,” Circulation, vol. 130, no. 7, pp. 573–581, 2014.
[Online]. Available: http://circ.ahajournals.org/content/130/7/573.short
[17] S. CI., “Bioelectronics: The way to discover the world of arrhythmias?”
Hellenic journal of cardiology, vol. 55, no. 3, pp. 267–8, 2014.
[18] I. A. Rezek and S. J. Roberts, “Stochastic complexity measures for
physiological signal analysis,” IEEE Transactions on Biomedical Engi-
neering, vol. 45, no. 9, pp. 1186–1191, Sept 1998.
[19] “European Congress on e-Cardiology and e-Health October 2014, Se-
lected Abstracts,” European Journal of Preventive Cardiology, vol. 21,
no. 2 suppl, pp. 38–44, 2014.
[20] T. Koivisto, M. Pa¨nka¨a¨la¨, T. Hurnanen, T. Vasankari, T. Kiviniemi,
A. Saraste, and J. Airaksinen, “Automatic detection of atrial fibrillation
using mems accelerometer,” in Computes in cardiology 2015, 42nd
annual conferene, vol. 42, September 2015, pp. 829–832.
[21] J. Slocum, A. Sahakian, and S. Swiryn, “Diagnosis of atrial fibrilla-
tion from surface electrocardiograms based on computer-detected atrial
activity,” Journal of Electrocardiology, vol. 25, no. 1, pp. 1 – 8, 1992.
[22] K. Tateno and L. Glass, “Automatic detection of atrial fibrillation using
the coefficient of variation and density histograms of rr and δrr intervals,”
Medical and Biological Engineering and Computing, vol. 39, no. 6, pp.
664–671.
[23] D. E. Lake and J. R. Moorman, “Accurate estimation of entropy
in very short physiological time series: the problem of atrial
fibrillation detection in implanted ventricular devices,” American
Journal of Physiology - Heart and Circulatory Physiology,
vol. 300, no. 1, pp. H319–H325, 2011. [Online]. Available:
http://ajpheart.physiology.org/content/300/1/H319
[24] D. Andresen and T. Brggemann, “Heart rate variability preceding
onset of atrial fibrillation,” Journal of cardiovascular electrophysiology,
vol. 9, no. 8 Suppl, p. S269, August 1998. [Online]. Available:
http://europepmc.org/abstract/MED/9727672
[25] C. W. Hogue, P. P. Domitrovich, P. K. Stein, G. D. Despotis, L. Re, R. B.
Schuessler, R. E. Kleiger, and J. N. Rottman, “Rr interval dynamics
before atrial fibrillation in patients after coronary artery bypass graft
surgery,” Circulation, vol. 98, no. 5, pp. 429–434, 1998.
[26] D. D. McManus, J. Lee, O. Maitas, N. Esa, R. Pidikiti, A. Carlucci,
J. Harrington, E. Mick, and K. H. Chon, “A novel application for the
detection of an irregular pulse using an iphone 4s in patients with atrial
fibrillation,” Heart Rhythm, vol. 10, no. 3, pp. 315 – 319, 2013.
[27] J. Lee, B. Reyes, D. McManus, O. Mathias, and K. Chon, “Atrial
fibrillation detection using an iphone 4s,” Biomedical Engineering, IEEE
Transactions on, vol. 60, no. 1, pp. 203–206, Jan 2013.
[28] N. V. Thakor, Y. S. Zhu, and K. Y. Pan, “Ventricular tachycardia and
fibrillation detection by a sequential hypothesis testing algorithm,” IEEE
Transactions on Biomedical Engineering, vol. 37, no. 9, pp. 837–843,
Sept 1990.
[29] R. H. Clayton, A. Murray, and R. W. F. Campbell, “Comparison of four
techniques for recognition of ventricular fibrillation from the surface
ecg,” Medical and Biological Engineering and Computing, vol. 31, no. 2,
pp. 111–117. [Online]. Available: http://dx.doi.org/10.1007/BF02446668
[30] T. F. Yang, B. Devine, and P. W. Macfarlane, “Artificial neural net-
works for the diagnosis of atrial fibrillation,” Medical and Biological
Engineering and Computing, vol. 32, no. 6, pp. 615–619.
[31] K. Sternickel, “Automatic pattern recognition in ECG time series,”
Computer Methods and Programs in Biomedicine, vol. 68, no. 2, pp.
109 – 115, 2002.
[32] M. G. Tsipouras and D. I. Fotiadis, “Automatic arrhythmia detection
based on time and timefrequency analysis of heart rate variability,”
Computer Methods and Programs in Biomedicine, vol. 74, no. 2, pp.
95 – 108, 2004.
[33] C. Bru¨ser, M. D. H. Zink, S. Winter, P. Schauerte, and S. Leonhardt,
“A feasibility study on the automatic detection of atrial fibrillation using
an unobtrusive bed-mounted sensor,” in Computing in Cardiology, 2011,
Sept 2011, pp. 13–16.
[34] C. Bru¨ser, J. Diesel, M. D. H. Zink, S. Winter, P. Schauerte, and
S. Leonhardt, “Automatic detection of atrial fibrillation in cardiac
vibration signals,” IEEE Journal of Biomedical and Health Informatics,
vol. 17, no. 1, pp. 162–171, Jan 2013.
[35] M. D. H. Zink, C. Bru¨ser, P. Winnersbach, A. Napp, S. Leonhardt,
N. Marx, P. Schauerte, and K. Mischke, “Heartbeat cycle length de-
tection by a ballistocardiographic sensor in atrial fibrillation and sinus
rhythm,” BioMed Research International, vol. 2015, no. 840356, pp.
1–10, Feb 2015.
[36] M. A. Garcia-Gonzalez, A. Argelags-Palau, M. Fernndez-Chimeno,
and J. Ramos-Castro, “A comparison of heartbeat detectors for the
seismocardiogram,” in Computing in Cardiology 2013, Sept 2013, pp.
461–464.
[37] A. D. Wiens, M. Etemadi, S. Roy, L. Klein, and O. T. Inan, “Toward
continuous, noninvasive assessment of ventricular function and hemo-
dynamics: Wearable ballistocardiography,” IEEE journal of biomedical
and health informatics, vol. 19, no. 4, pp. 1435–1442, 2015.
[38] M. J. Tadi, T. Koivisto, M. Pankaala, and A. Paasio, “Accelerometer-
based method for extracting respiratory and cardiac gating information
for dual gating during nuclear medicine imaging,” International Journal
of Biomedical Imaging, vol. 2014, no. 690124, pp. 1–11, 2014.
[39] R. Peters, E. James, and M. Russell, “Amplitude Modulation
Effects in Cardiac Signals,” 2010. [Online]. Available:
http://arxiv.org/abs/1011.1880
[40] Electrophysiology, Task Force of the European Society of Cardiology the
North American Society of Pacing, “Heart rate variability: Standards of
measurement, physiological interpretation, and clinical use,” Circulation,
vol. 93, no. 5, pp. 1043–1065, 1996.
[41] M. Tadi, E. Lehtonen, T. Koivisto, M. Pankaala, A. Paasio, and M. Teras,
“Seismocardiography: Toward heart rate variability (hrv) estimation,” in
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS , APRIL 2016 10
Medical Measurements and Applications (MeMeA), 2015 IEEE Interna-
tional Symposium on, May 2015, pp. 261–266.
[42] J. Jezewski, D. Roj, J. Wrobel, and K. Horoba, “A novel technique for
fetal heart rate estimation from doppler ultrasound signal,” BioMedical
Engineering OnLine, vol. 10, no. 1, p. 92, 2011.
[43] C. Bru¨ser, S. Winter, and S. Leonhardt, “Robust inter-beat interval
estimation in cardiac vibration signals,” Physiological Measurement,
vol. 34, no. 2, p. 123, 2013.
[44] ECG Holter,Guide to Electrocardiographic Interpretation. Springer US,
2008.