Combining Rhythmic and Morphological ECG Features
for Automatic Detection of Atrial Fibrillation
Gennaro Laudato
1
, Franco Boldi
2
, Angela Rita Colavita
3
, Giovanni Rosa
1
, Simone Scalabrino
1
,
Paolo Torchitti
2
, Aldo Lazich
4
and Rocco Oliveto
1
1
STAKE Lab, University of Molise, Pesche (IS), Italy
2
XEOS, Roncadelle (BS), Italy
3
ASREM, Campobasso (CB), Italy
4
Ministero della Difesa, Roma (RM), Italy
Keywords:
ECG Analysis, Atrial Fibrillation, Arrhythmia, Decision Support System, Machine Learning.
Abstract:
Atrial Fibrillation (AF) is a common cardiac disease which can be diagnosed by analyzing a full electrocar-
diogram (ECG) layout. The main features that cardiologists observe in the process of AF diagnosis are (i)
the morphology of heart beats and (ii) a simultaneous arrhythmia. In the last decades, a lot of effort has been
devoted for the definition of approaches aiming to automatic detect such a pathology. The majority of AF
detection approaches focus on R-R Intervals (RRI) analysis, neglecting the other side of the coin, i.e., the
morphology of heart beats. In this paper, we aim at bridging this gap. First, we present some novel features
that can be extracted from an ECG. Then, we combine such features with other classical rhythmic and mor-
phological features in a machine learning based approach to improve the detection accuracy of AF events.
The proposed approach, namely MORPHYTHM, has been validated on the Physionet MIT-BIH AF Database.
The results of our experiment show that MORPHYTHM improves the classification accuracy of AF events by
correctly classifying about 4,400 additional instances compared to the best state of the art approach.
1 INTRODUCTION
Atrial Fibrillation (AF) is a quite common yet dan-
gerous cardiac pathological condition. The numbers
say that in the UK, almost 534k people have con-
tracted this disease, in 1995 (Stewart et al., 2004).
In 2010, the estimated numbers of men and women
who were affected by AF world-wide were respec-
tively 20.9 and 12.6 million. Moreover, the incidence
was higher in developed countries, such as Europe
and US. Indeed, it is expected that - by 2030 - the
number of AF patients will be between 14 and 17 mil-
lion only in Europe (Kirchhof et al., 2016). Besides,
such a condition is very expensive: the direct cost
of healthcare for patients affected by AF was about
∼655M in 2000, equivalent to 0.97% of the total UK
National Health System (NHS) expenditure (Stewart
et al., 2004). While in US, it has been estimated that
the medical cost caused by AF is $26 billion annu-
ally (January et al., 2014). Also, the prevalence of the
disease is expected to more than double in the next 50
years as the population grows older (Miyasaka, 2006).
Most of the cost of healthcare for patients affected
by AF is due to hospitalizations and home nursing. In
this context, telemedicine would be very helpful. In-
deed, telemedicine would allow to remotely and con-
tinuously monitoring thousands of patients. However,
telemedicine alone is not enough: remote monitoring
could help reducing the global cost, but physicians
and nurses would be still required to perform such a
task.
The best way for reducing the cost of AF for NHS
through telemedicine would be by employing auto-
mated approaches for AF detection: a software sys-
tem constantly acquires data from the patient and,
when an anomalous condition is detected, physicians
are warned (Balestrieri et al., 2019). This would allow
to reduce the number of specialized personnel that is
required to monitor the patients.
Many automated approaches for AF detection
were proposed in the literature (Asgari et al., 2015;
Lee et al., 2013; Petr
˙
enas et al., 2015; Zhou et al.,
2014, 2015). Such approaches acquire and transform
the electrocardiogram (ECG) signal to detect, for each
156
Laudato, G., Boldi, F., Colavita, A., Rosa, G., Scalabrino, S., Torchitti, P., Lazich, A. and Oliveto, R.
Combining Rhythmic and Morphological ECG Features for Automatic Detection of Atrial Fibrillation.
DOI: 10.5220/0008982301560165
In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 5: HEALTHINF, pages 156-165
ISBN: 978-989-758-398-8; ISSN: 2184-4305
Copyright
c
 2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

heart beat, if it is fibrillating or non-fibrillating. One
of the best approaches available, i.e., the one intro-
duced by Zhou et al. (2015), still classifies about 20k
fibrillant heart beat signals as non-fibrillant. This
means that if we assign half a second to each mis-
classified beat—which implies 120 BPM, i.e., a quite
high heart rate—three hours of fibrillating recordings
were completely ignored by the approach. This shows
that, even if the accuracy of AF detectors is very high,
there is still room for improvement. Indeed, in this
context, even a small advance would be important and
it would possibly help saving human lives.
Since arrhythmia is one of the most prominent
symptoms of AF, most of the state of the art ap-
proaches are based on rhythmic features, i.e., mea-
sures that try to capture the regularity of the heart
beat. On the other hand, another main indicator of AF
is the absence of the P-wave, which is visible through
the ECG. However, only a minority of studies consid-
ered morphological features, which aim at capturing
the shape of a single heart beat.
In this paper, we present MORPHYTHM, an ap-
proach based on machine learning techniques that
combines rhythmic and morphological features to de-
tect AF events. MORPHYTHM uses the most promis-
ing state of the art rhythmic and morphological fea-
tures and some novel features.
We compare MORPHYTHM with the approach in-
troduced by Zhou et al. (2015), the best state of the
art approach. The results show that MORPHYTHM
can improve the classification accuracy, reducing the
number of false negatives (i.e., instances classified
as non-fibrillating that are, actually, fibrillating) by
about 4.4k instances, which corresponds to about 35
minutes of ECG (at 120 BPM).
The achieved results confirm our initial conjec-
ture: morphological features combined with rhythmic
ones should be considered for AF detection. Thus,
this paper does the first step towards the definition
of novel approaches for AF detection that simultane-
ously combine morphological and rhythmic features.
The rest of the paper is structured as follows. Sec-
tion 2 provides details on AF and its ECG diagnostic
features. Section 3 presents MORPHYTHM, our novel
approach for AF detection. Section 4 reports the de-
sign and the results of the empirical study we con-
ducted to evaluate MORPHYTHM. Finally, Section 5
concludes the paper and provides suggestions for pos-
sible future research directions.
Figure 1: ECG theoretical waveform.
2 BACKGROUND
This section provides details on AF and how it can be
identified through the manual analysis of an electro-
cardiogram. We also discuss the approaches proposed
in the literature for AF automatic detection.
2.1 Atrial Fibrillation
AF is a pathological heart rhythm which results in a
rapid and irregular beating of the atria. The conse-
quences of this cardiac disease are very adverse. In-
deed, contracting AF may lead to stroke, dementia,
and death. Thus, a precise diagnosis of this pathology
needs to become a priority (Schnabel et al., 2015).
During AF, the hearts atria are quicker than nor-
mal beating. This leads to the condition that the blood
is not ejected completely out of atria and there might
be chances of formation of blood clots in the atria.
The result is an increased risk of stroke. Electrocar-
diograms (ECGs) are useful tools for AF detection.
ECGs are recordings of heart’s electrical activity and
are widely used by physicians to diagnose pathologies
related to the heart. Patients with or at risk of cardio-
vascular diseases often present ECGs that are irreg-
ular in rate and in morphology of the signal (Chou
et al., 2016).
According to the official international guidelines
(Kirchhof et al., 2016), AF can be detected by ob-
serving three main features in the ECG, as shown in
Figure 1, i.e., (i) absence of the P-wave, (ii) presence
of fluctuating waveforms (f-waves) instead of the P-
wave, and (iii) heart rate irregularity. The first two
features can be defined as morphological, while the
third one is rhythmic.
Since AF is often asymptomatic (Camm et al.,
2012; Kearley et al., 2014), a reliable device com-
bined with an accurate, real-time, and automatic AF
detection algorithm is desirable for improving detec-
tion of AF (Camm et al., 2012; Capucci et al., 2012;
Censi et al., 2013; Kearley et al., 2014).
Combining Rhythmic and Morphological ECG Features for Automatic Detection of Atrial Fibrillation
157

Table 1: Literature Detector Performances on MIT-BIH
AFDB. In AFDB
1
records “00735” and “03665” excluded,
while in AFDB
2
records “04936” and “05091” excluded.
Method Year DB Se[%] Sp[%]
Zhou et al. (2015) 2015 AFDB 97.4 98.4
Petr
˙
enas et al. (2015) 2015 AFDB 97.1 98.3
Asgari et al. (2015) 2015 AFDB
2
97.0 97.1
Zhou et al. (2014) 2014 AFDB 96.9 98.3
Lee et al. (2013) 2013 AFDB
1
98.2 97.7
Huang et al. (2010) 2011 AFDB 96.1 98.1
2.2 Automatic Detection of AF
In the last decade, several methods have been pro-
posed for the automatic detection of AF. Most of
them have shown good results by exploiting only the
analysis of heart beat rhythm (Colloca et al., 2013;
Mohebbi and Ghassemian, 2008; Sepulveda-Suescun
et al., 2017; Xiong et al., 2017; Yuan et al., 2016).
Morphological features were used in the patent by
Kurzweil et al. (2016) and, even not specifically fo-
cused only on AF detection, in the work by Xu et al.
(2018).
For sake of space limitation, in the following we
focus the attention on the most accurate methods re-
ported in the literature, i.e., the ones summarized in
Table 1. These methods represent our baseline, due
to the common evaluation on the Physionet MIT-BIH
AF Database (Goldberger et al., 2000). This database
includes 25 long-term ECG recordings of patients
with atrial fibrillation (mostly paroxysmal
1
). Of these
25 long-term ECG recordings, 23 include the ECG
signals while for records (i.e., patients) 00735 and
03665 only information on the rhythm are available.
The individual recordings are 10 hours each in du-
ration and contain two ECG signals each sampled at
250 samples per second with 12-bit resolution over a
range of ±10 millivolts.
Huang et al. (2010) propose a method to detect
the transition between AF and sinus rhythm, based
on RRI. In the proposed method the authors first ob-
tain the delta RR interval distribution difference curve
from the density histogram of delta RRI, and then de-
tect its peaks, which represent the AF events. Once
an AF event was detected, four successive steps have
been used to classify its type.
1
AF can be classified into specific types depending on
the duration and ability to self-terminate or to be terminated
by some therapeutic technique (Kirchhof et al., 2016). AF
is named as paroxysmal when it is self-terminating (in most
cases within 48 hours). Some AF paroxysmal episodes may
continue up to 7 days. Thus, also AF episodes that are car-
dioverted within 7 days are considered paroxysmal.
Lee et al. (2013) introduce a method for auto-
matic detection of AF using time-varying coherence
functions (TVCF). The TVCF is estimated by the
multiplication of two time-varying transfer functions
(TVTFs). The first TVTF is obtained by consider-
ing two adjacent data segments (as input and output
signals); the second TVTF is computed by reversing
these signals. They found that the resultant TVCF
between two adjacent normal sinus rhythm segments
shows high coherence values (near 1) while lower
than 1 if either or both segments partially or fully con-
tain AF, throughout the entire frequency range. They
have also combined TVCF with Shannon entropy. In
this case, the approach shows even more accurate AF
detection rate: 97.9% for the MIT-BIH AF database
(considering 23 records) with 128 beat segments.
Zhou et al. (2014) devise a method for real-time,
automated detection of AF episodes in ECGs. This
method utilizes RR intervals, and it involves sev-
eral basic operations of nonlinear/linear integer filters,
symbolic dynamics and the calculation of Shannon
entropy.
Asgari et al. (2015) employ a stationary wavelet
transform and a support vector machine to detect AF
episodes. The proposed method eliminates the need
for P-peak or R-Peak detection (a pre-processing step
required by many existing algorithms), and hence its
performance (sensitivity, specificity) does not depend
on the performance of beat detection.
Petr
˙
enas et al. (2015) propose a RR-based AF de-
tector with a low complexity structure. The detec-
tor involves blocks for pre-processing, bigeminal sup-
pression, characterization of RR irregularity, signal
fusion and threshold detection.
Zhou et al. (2015) adopt heart rate sequence and
apply symbolic dynamics and Shannon entropy. Us-
ing novel recursive algorithms, a low-computational
complexity can be obtained. With this approach, the
authors were able to slightly improves their previous
work. The approach proposed by Zhou et al. (2015)
is the most accurate method of AF detection on the
MIT-BIH AF Database proposed so far.
The method proposed by Zhou et al. (2015) will be
deeply explained in the next subsection for two main
reasons: (i) the method represents our baseline in the
evaluation of MORPHYTHM; (ii) the entropy measure
used in Zhou et al. (2015) has been exploited as fea-
ture in MORPHYTHM.
2.3 Baseline Method for AF Detection
This section provides details on the method proposed
by Zhou et al. (2015), i.e., our baseline in the evalua-
tion of MORPHYTHM. Such an approach consists in
HEALTHINF 2020 - 13th International Conference on Health Informatics
158

the following steps (see Figure 2):
• The HR sequence is converted to a symbolic se-
quence in a fixed interval;
• A probability distribution is constructed from the
word sequence which is transformed from the
symbolic sequence;
• A coarser version of Shannon entropy is em-
ployed to quantify the information size of HR se-
quence using the probability distribution of word
sequence;
• Discrimination of the heart beat type (AF or no-
AF) using a threshold.
Step 1: Converting the HR Sequence. Consid-
ering a preliminary stage of RRI analysis and thus
known the HR sequence, the first step expected in the
method is to evaluate a symbolic dynamic. This quan-
tity encodes the information of hr
n
to a series with
fewer symbols, with each symbol aims at representing
an instantaneous state of heart beating. The mapping
function is the following:
sy
n
=
(
63, if n hr ≥ 315
bhr
n
c, other cases
where [·] represents a floor operator.
Step 2: Building the Symbolic Sequence. The
authors apply a 3-symbols template in order to ex-
plore the entropic properties of the symbolic series
sy
n
. Thus, to examine the chaotic behavior, the word
value can then be calculated by the operator as defined
below:
wv
n
= (sy
n−2
× 2
1
2) + (sy
n−1
× 2
6
) + sy
n
Step 3: Computing the Entropy. The authors
define a coarser version of Shannon entropy H
00
(A) to
quantitatively calculate the information size of wv
n
.
In this study, the dynamic A comprises of 127 con-
secutive word elements from wv
n−126
to wv
n
, as pro-
posed in the function below:
H
00
(A) = −
k
Nlog
2
N
k
∑
i=1
p
i
log
2
p
i
where N and k are total number of the elements and
characteristic elements in space A, respectively.
Step 4: Classification. Based on the obtained
entropy value, a final beat-to-beat classification (fib-
rillant or non-fibrillant) is presented by applying a
threshold discrimination. The optimal threshold was
empirically identified at 0.639.
2.4 Usage Scenarios of AF Detectors
AF detection methods might be useful in two different
scenarios: offline and online. In an offline scenario,
Figure 2: Graphical representation of the main steps in the
method by Zhou et al. (2015).
the ECG of a patient is recorded and, later, an AF
detection method is used to find possible AF events
occurred in the recorded period. This can help physi-
cians discovering AF events in possibly long EGCs.
AF detection methods could be also particularly
valuable in an online scenario: while the ECG is ac-
quired, it is immediately passed to the AF detector,
which promptly detects AF events. Online AF detec-
tion can be useful in tele-medical applications, where
patients are constantly monitored.
The application context we take into account in
this paper is the online monitoring. In other words,
we assume that we have chunks of ECG incremen-
tally available. Therefore, we specifically focus on
real-time (or near real-time) approaches. Online mon-
itoring is useful for tele-medical applications.
Several tele-medical projects were proposed in the
literature. Zhu et al. (2015) introduced the SPHERE
system, which combines several sensors which ac-
quire data through wearable, environment, and video
devices. Villar et al. (2015) introduced Hexoskin, a
line of cutting-edge smart clothing that include body
sensors into garments for health monitoring. Balestri-
eri et al. Balestrieri et al. (2019) recently introduced
ATTICUS, an innovative Internet of Medical Things
(IoMT) system for implementing personalized health
services.
3 MORPHYTHM OVERVIEW
In this section we present MORPHYTHM, a novel ap-
proach for the detection of AF events. The proposed
Combining Rhythmic and Morphological ECG Features for Automatic Detection of Atrial Fibrillation
159

approach combines rhythmic and morphological fea-
tures through machine learning techniques.
3.1 Preprocessing of ECG Data
Before extracting features, the ECG data (from the
AFDB) have to be pre-processed according to Pan and
Tompkins (1985) and Clifford et al. (2006). The main
steps involved in this phase are the followings:
• Detrend of ECG Signal: the offset has been re-
moved from the raw ECG signal by removing the
mean from the signal.
• Filtering Stage: first, a low and high pass filters
have been applied to get rid of baseline wander
and discard high frequency noise, respectively.
Subsequently, a derivative filtering has been oper-
ated on the signal aiming at emphasizing the high
frequency components of the ECG.
• Sample Amplitude Normalization: each record-
ing has been normalized in terms of sample ampli-
tude around the maximum.
3.2 Rhythmic Features
Rhythmic features are based on one or more heart
beats and they aim at capturing aspects that mostly
regard the regularity of the heart beat signal. Zhou
et al. (2015) state that the detection methods based on
RRI are more useful to produce a precise and accu-
rate identification of AF because the R-wave peak of
the QRS complex is the most prominent characteris-
tic feature of an ECG recording. Such a characteristic
is less subject to noise (Huang et al., 2010; Lake and
Moorman, 2010; Lee et al., 2012; Lian et al., 2011).
In MORPHYTHM we use two features based on the
observation of a single heart beat signal, i.e., HBL
and HBDL, and two additional rhythmic features that
consider the information of a sequence of consecutive
heart beats, i.e., HBR and Entropy:
• Heart Beat Length (HBL). This feature repre-
sents how long a single heart beat signal lasts. We
measure HBL as the number of samples from a
peak R to the next peak R;
• Heart Beat Discrete Length (HBDL). Such a
feature is a classification of the heart beat signal
in three classes, based on its length. A beat is (i)
short if it takes less than 0.5 seconds, (ii) long if
it takes more than 1.2 seconds, and (iii) regular
otherwise;
• Heart Beat Regularity (HBR). This feature is
based on HBDL. It considers a rhythmic pattern of
10 consecutive discrete heart beats lengths. Once
obtained the pattern, we compute HBR simply
counting the number of regular heart beats. It is
worth noting that there are approaches in the lit-
erature which consider a very short windowed se-
quence of heart beats (Petr
˙
enas et al., 2015);
• Entropy, as defined by Zhou et al. (2015) and de-
scribed in Section 2.3.
While HBL and Entropy have been previously used in
the literature (Zhou et al., 2015), HBDL and HBR are
two new rhythmic features defined in this paper.
3.3 Morphological Features
Even if the acquisition of rhythmic features can be
very reliable, such features can only help detecting
arrhythmia, which is just one of the possible signs
of AF. On the other hand, morphological features are
necessary to detect anomalies in the shape of a single
heart beat signal.
In MORPHYTHM we propose three different mea-
sures that—given a sequence of samples provided for
a heart beat signal
2
—return a single numeric value:
• Mean Signal Intensity (MSI). Such a feature is
measured as the mean of all the samples acquired
in a heart beat signal. The mean signal intensity,
alone, provides a very rough indication of regular-
ity of the heart beat signal. If there is any anomaly
in any part of the heart beat signal, such a feature
may help identifying it. For example, if the P-
wave is missing, the MSI may be slightly affected;
• Signal Intensity Variance (SIV). This feature is
measured as the variance of all the samples ac-
quired in a heart beat signal. The SIV helps char-
acterizing the heart beat signal: again, a low SIV
might indicate the absence of the P-wave.
• Signal Intensity Entropy (SIE). This feature is
computed as the entropy (Moddemeijer, 1989) of
the distribution of the sample values in a heart beat
signal. This feature is similar to SIV, i.e., it is
aimed at representing the variations in the signal
of a heart beat.
It is worth noting that extracting features by con-
sidering a whole heart beat might compress too much
the information in the ECG data. To extract richer
information, we also propose a novel descriptor of a
heart beat signal by (i) dividing the whole heart beat
in n segments; and (ii) computing the above defined
features on each segment:
2
We consider as a heart beat a digital signal which goes
from a R-peak to the successive. Such an interpretation
is very suitable for AF detection, because it highlights the
atrial activity.
HEALTHINF 2020 - 13th International Conference on Health Informatics
160

• Segmented Mean Signal Intensity (S-MSI
i
):
given the i-th segment of the heart beat signal, S-
MSI
i
is computed as the mean of the sample val-
ues of such a segment.
• Segmented Signal Intensity Variance (S-SIV
i
):
given the i-th segment of the heart beat signal, S-
SIV
i
is computed as the variance of the sample
values of such a segment.
• Segmented Signal Intensity Variance (S-SIE
i
):
given the i-th segment of the heart beat signal, S-
SIE
i
is computed as the entropy (Moddemeijer,
1989) of the sample values of such a segment.
All such features allow to roughly represent the shape
of the signal of the heart beat. We reduce the resolu-
tion of the heart beat signal to just 30 values (n=10 for
each feature) to reduce the noise of the samples.
Besides the aforementioned features, we also inte-
grate in MORPHYTHM other state of the art morpho-
logical features:
• Fast Fourier Transform (FFT
i
): we include the
features introduced by Haque et al. (2009) by cal-
culating the Fast Fourier Transform of the heart
beat signal on 32 points.
• Auto-Regressive Model (ARM
i
): we include the
features introduced by Zhao and Zhang (2005) by
estimating the coefficients of the Auto-Regressive
model of order 16.
3.4 Putting All Together
MORPHYTHM combines all the features we previ-
ously described using supervised machine learning
techniques. After the training phase, MORPHYTHM
is able—given a heart beat signal—to classify it as
fibrillating or not fibrillating. In the MORPHYTHM
evaluation, we experimented several classifiers.
4 EMPIRICAL EVALUATION
The goal of this study is to evaluate the accuracy of
MORPHYTHM is classifying AF events in a patient.
The perspective is both (i) of a researcher who wants
to understand if combining rhythmic and morpholog-
ical features is useful for detecting AF events, and (ii)
of a practitioner who wants to use the most accurate
and precise approach in a telemedicine application.
Thus, the study is steered by the following research
question:
Can the combination of rhythmic and morphological
features improve the classification accuracy
of Atrial Fibrillation events?
4.1 Context Selection
The context of this study is represented by MIT-BIH
AF Database (Goldberger et al., 2000), a commonly
used benchmark which contains recordings of 25 pa-
tients. Due to the embedding of morphology descrip-
tors, our overall study has been performed on the
AFDB
1
, i.e., the AFDB without records 00735 and
03665 because, for such records, only information
on the rhythm is available (Goldberger et al., 2000).
Each recording in the dataset lasts 10 hours and con-
tains two ECG signals sampled at 250 samples per
second (12-bit resolution).
In the context of our study, we experimented a
large set of machine learning technique to train MOR-
PHYTHM. Especially, we experimented tree-based
classifiers, i.e., J48 (Quinlan, 2014), Replication Tree
(Devasena, 2014), and Random Forest (Barandiaran,
1998). Such approaches, indeed, can build models
that are also easy to understand by a human. We also
experimented Logistic regression (Cramer, 2002) and
AdaBoost M1 (Freund and Schapire, 1997).
4.2 Experimental Procedure
To evaluate the accuracy of MORPHYTHM, we
used a classical Leave-1-Person Out (L1PO) cross-
validation: we divided all the data in n folds, one for
each patient, and we use one at a time each of such
folds as test set and the union of the remaining folds
as training set. This means that the data related to a
single patient were embedded once in the test dataset
and n-1 times in the training dataset. This technique
allows to build a classifier which is not trained and
tested on the data belonging to the same patient. We
did this to evaluate the technique in the most chal-
lenging scenario: the ECG of different patients can
be very different.
We compared MORPHYTHM to the approach pro-
posed by Zhou et al. (2015), previously presented in
Section 2.3. The instances to be classified were all
the single heart beat signals provided in the dataset,
labeled as fibrillating or non-fibrillating. The work
by Zhou et al. (2015) just reported the performance of
the approach globally, i.e., for all the patients. Instead,
we provide the performance of the approaches with a
finer grain, i.e., on patient-by-patient base. Since we
do not have the patient-by-patient results for the base-
line, it was necessary to re-implement the approach
and to re-compute the results.
To answer our research question we compared two
critical aspects: True Positives (TP), i.e., the number
of instances classified as fibrillating by the approach
and that were actually fibrillating, and the False Neg-
Combining Rhythmic and Morphological ECG Features for Automatic Detection of Atrial Fibrillation
161

Table 2: Comparison of MORPHYTHM and the approach proposed by Zhou et al. (2015). In boldface the results achieved by
MORPHYTHM that are better than the baseline.
Approach TP TN FP FN ∆TP ∆FN
Zhou et al. (2015) 489,834 603,216 17,188 19,911
MORPHYTHM — Random Forest 490,810 584,692 35,612 18,935 +976 -976
MORPHYTHM — J48 479,411 560,567 57,049 33,122 -10,423 +13,211
MORPHYTHM — Logistic 494,255 595,664 24,789 15,445 +4,421 -4,466
MORPHYTHM — AdaBoost M1 494,384 601,974 18,430 22,362 +4,550 +2,451
MORPHYTHM — RepTree 481,397 571,262 49,142 32,348 -8,437 +12,437
atives (FN), i.e., the number of instances classified
as non-fibrillating which were, actually, fibrillating.
A high number of TP is desirable, because it indi-
cates the number of AF episodes correctly detected.
Also, ideally, a perfect approach does not lose any AF
episode: thus, keeping the number of FN low is very
important.
We use a Wilcoxon signed-rank test to verify if
MORPHYTHM achieves statistically significant better
results than the approach proposed by Zhou et al.
(2015). To do this, we use the results achieved pa-
tient by patient in terms of TP and FN. Formally, our
null hypotheses are:
• H0
1
: MORPHYTHM does not identify a higher
number of TP as compared to the approach pro-
posed by Zhou et al. (2015);
• H0
2
: MORPHYTHM does not identify a lower
number of FN as compared to the approach pro-
posed by Zhou et al. (2015);
We reject a null hypothesis if the p-value is lower than
α = 0.05.
Even if we evaluate the possible improvement
only on TP and FN, we also report the global results
in terms of True Negatives (TN — i.e., instances cor-
rectly classified as non-fibrillating) and False Posi-
tives (FP — i.e., instances classified as fibrillating that
are, actually, non-fibrillating).
4.3 Analysis of the Results
We show the global performance of the compared
approaches in Table 2. For MORPHYTHM, we also
specifically report the difference in terms of TP and
FN with the baseline, i.e., ∆TP (the higher, the better)
and ∆FN (the lower, the better), and we put in bold-
face the cases in which MORPHYTHM achieves better
results.
The first consideration that can be derived from
the analysis of Table 2 is that three (Random Forest,
Logistic and AdaBoost M1) of the five chosen ma-
chine learning are able to achieve better results than
the baseline. Furthermore, the Logistic method per-
forms definitely better than its competitors by show-
ing an improvement of around 4,400 heart beats com-
pared to the method by Zhou et al. (2015). If we
would assign an inter-beat interval of 0.5 seconds, an
improvement of 4,400 indicates more than 35 minutes
of AF rhythm improved in the classification with re-
spect to the baseline. Even if this could appear as a
negligible result, it should be noticed that the accu-
racy level achieved by such approaches is very high
and, therefore, even achieving a small improvement
is very difficult.
It can be noticed that the global accuracy of the
approach by Zhou et al. (2015) slightly differs from
the global accuracy reported in the original paper. Es-
pecially, in the original paper the authors reported the
following values for sensitivity, specificity, and accu-
racy—they just report aggregated measures: 97.31%,
98.28%, and 97.89%. With our replication of the ap-
proach by Zhou et al. (2015) we achieve the following
results: 96.09% of sensitivity, 97.22% of specificity,
and 96.71% of accuracy. We are confident that the
different results are not due to implementation errors,
but to different choices in the evaluation design. The
different results could be due to the following reasons:
• the transient: to avoid any error due to interpreta-
tion, the first 128 (126 coming from the entropy
compression + 2 from the word sequence evalua-
tion) beats have not been considered in the repli-
cation of the work by Zhou et al. (2015). Unfortu-
nately, in the paper by Zhou et al. (2015) there is
no clear indication on how the authors deal with
the initial 128 beats;
• the timestamps: Physionet offers two different
timestamps: one for each beat classification an-
other one for each AF events (rhythm annotation).
There are cases where there is a mismatch be-
tween the two timestamps, i.e.,, the AF event does
not start (or does not end) with the beginning of a
(or the end) of beat. In other words beats and AF
events are not always synchronized. This causes
an ambiguity regarding the interpretation of “hy-
HEALTHINF 2020 - 13th International Conference on Health Informatics
162

Figure 3: A graphical example of hybrid heart beats.
Table 3: Patient-level comparison between MORPHYTHM
and Zhou et al. (2015). In boldface the best results for each
patient.
Record
Zhou et al. (2015) MORPHYTHM
TP FN TP FN
04015 478 40 491 27
04043 8,690 5,862 9,608 4,944
04048 419 387 443 363
04126 3,082 204 3,154 132
04746 30,731 137 30,624 244
04908 5,443 359 5,557 245
04936 32,833 6,812 33,725 5,920
05091 0 133 0 133
05121 32,575 1,164 33,563 176
05261 655 268 766 157
06426 52,104 1,006 52,633 477
06453 126 313 130 309
06995 27,072 448 27,240 280
07162 39,297 0 39,297 0
07859 61,891 0 61,891 0
07879 39,944 89 39,939 49
07910 6,499 266 6,440 325
08215 32,958 170 32,912 216
08219 12,627 1,528 13,420 735
08378 10,995 478 10,969 504
08405 45,005 88 45,041 52
08434 2,307 0 2,301 6
08455 44,103 159 44,111 151
brid beats”, i.e., beats that are not aligned with an
AF event (see Figure 3).
It is worth noting that the different results achieved
does not represent a threat for the final message of
this paper. Indeed, improving the accuracy of the ap-
proach by Zhou et al. (2015) by performing a different
evaluation design likely results in an improvement of
MORPHYTHM as well, since the approach by Zhou
et al. (2015) is one of the features exploited by MOR-
PHYTHM.
Table 3 shows a patient-by-patient comparison be-
tween the approach by Zhou et al. (2015) and MOR-
PHYTHM (with the classifier that achieves the best re-
sults globally, i.e., Logistic Regression). Even if the
method proposed by Zhou et al. (2015) is incredibly
accurate, as it can be observed from Table 3, MOR-
PHYTHM achieves much better results for some pa-
tients and comparable results on some other patients.
Specifically, MORPHYTHM identifies a higher num-
ber of TPs for 15 out of 23 patients and it identifies a
lower number of FNs for 15 out of 23 patients. The
results of the Wilcoxon signed-rank test show that we
Table 4: Comparison between the proposed classifier on the
record 05091.
Approach TP TN FP FN
Zhou et al. (2015) 0 36,644 0 133
MORPHYTHM — Random Forest 25 36,633 11 108
MORPHYTHM — J48 19 36,592 52 114
MORPHYTHM — Logistic 0 36,640 4 133
MORPHYTHM — AdaBoost M1 0 36,644 0 133
MORPHYTHM — RepTree 14 36,620 24 119
Table 5: Features ranking using Information Gain.
InfoGain Attribute Type
0.86 Entropy from Zhou et al. (2015) Rhythmic
0.20 Entropy from the rhythmic pattern Rhythmic
0.18 Heart beat absolute length Rhythmic
0.14 coeff. no. 10 from AR model Morphological
0.13 coeff. no. 11 from AR model Morphological
0.12 coeff. no. 7 from AR model Morphological
0.12 coeff. no. 9 from AR model Morphological
0.11 coeff. no. 8 from AR model Morphological
can reject both our null hypotheses: MORPHYTHM
identifies a significantly higher number of TPs (p =
0.021) and a significantly lower number of FNs (p =
0.014).
Table 4 shows the comparison between MOR-
PHYTHM and the approach by Zhou et al. (2015) for
one of the patients, i.e., 05091. For such a patient,
the baseline has never detected any fibrillating event.
This means that this record has been classified as not
affected by AF, overall, even if it was. On the other
hand, most of the classifiers we consider are able
to detect some fibrillating events. However, unfor-
tunately, this is not true for the best classifier (i.e.,
Logistic regression), which, similarly to the baseline,
does not identify any heart beat of the specific patient
as fibrillating.
Finally, Table 5 shows the best attributes and their
worth computed with Information Gain. From the
achieved ranking, we can observe that, as expected,
the classifiers use the entropy values as their main
source of information. After that, the length of the
signal and the central coefficient obtained from the
AR model of order 4 seem to be the features which
help the classifiers to slightly improve its accuracy as
compared to the baseline. Very interesting is the result
on the AR model because it seems that the central co-
efficient, which should describe the atrial activity, are
mostly taken in consideration.
Combining Rhythmic and Morphological ECG Features for Automatic Detection of Atrial Fibrillation
163

5 CONCLUSIONS
We have presented MORPHYTHM, an approach based
on machine learning that combines rhythmic and mor-
phological features extracted from ECG data to de-
tect AF events. We compared MORPHYTHM with the
method introduced by Zhou et al. (2015), the most
accurate on the AFDB in the literature. The results
show that (i) MORPHYTHM globally achieves better
results compared to the baseline, since it is able to
correctly classify about 4,400 more heart beats, and
(ii) that some of the patients for which all the fibril-
lating heart beat were mis-classified by the baseline
were correctly classified by MORPHYTHM.
The improvement achieved is promising; how-
ever, there is still much room for improving the ac-
curacy. In order to increase the generalizability of
our results, we aim at apply in the future at least one
classifier for each family. In this work, for exam-
ple, we did not consider Bayesian networks, Rules-
based classifiers, and Neural Networks. Also, to max-
imize the accuracy of MORPHYTHM, it would be de-
sirable to use feature selection, to remove useless fea-
tures that could decrease the classification accuracy.
Specifically, since morphological features can be very
patient-dependent, it could be useful performing fea-
ture selection for each single patient rather than glob-
ally. Finally, we plan to perform a cost-benefit anal-
ysis. Indeed, in some online applications, it could be
necessary to have some constraints, such as the total
reduction of FN, even if the FP rate increases. Thus,
we would like to study this specific scenario and ob-
serve if the application of a cost-benefit analysis can
suite some specific constraints.
REFERENCES
Asgari, S., Mehrnia, A., and Moussavi, M. (2015). Au-
tomatic detection of atrial fibrillation using stationary
wavelet transform and support vector machine. Com-
puters in biology and medicine, 60:132–142.
Balestrieri, E., Boldi, F., Colavita, A. R., De Vito, L.,
Laudato, G., Oliveto, R., Picariello, F., Rivaldi, S.,
Scalabrino, S., Torchitti, P., et al. (2019). The ar-
chitecture of an innovative smart t-shirt based on the
internet of medical things paradigm. In 2019 IEEE
International Symposium on Medical Measurements
and Applications (MeMeA), pages 1–6. IEEE.
Barandiaran, I. (1998). The random subspace method for
constructing decision forests. IEEE Trans. Pattern
Anal. Mach. Intell, 20(8):1–22.
Camm, A. J., Corbucci, G., and Padeletti, L. (2012). Use-
fulness of continuous electrocardiographic monitoring
for atrial fibrillation. The American journal of cardi-
ology, 110(2):270–276.
Capucci, A., Calcagnini, G., Mattei, E., Triventi, M., Bar-
tolini, P., Biancalana, G., Gargaro, A., Puglisi, A., and
Censi, F. (2012). Daily distribution of atrial arrhyth-
mic episodes in sick sinus syndrome patients: impli-
cations for atrial arrhythmia monitoring. Europace,
14(8):1117–1124.
Censi, F., Calcagnini, G., Mattei, E., Gargaro, A., Bian-
calana, G., and Capucci, A. (2013). Simulation of
monitoring strategies for atrial arrhythmia detection.
Annali dell’Istituto superiore di sanita, 49:176–182.
Chou, T., Tamura, Y., and Wong, I. (2016). Detection of
atrial fibrillation in ecgs. Computer Methods and Pro-
grams in Biomedicine, 136:143–50.
Clifford, G. D., Azuaje, F., McSharry, P., et al. (2006).
Advanced methods and tools for ECG data analysis.
Artech house Boston.
Colloca, R., Johnson, A. E., Mainardi, L., and Clifford,
G. D. (2013). A support vector machine approach for
reliable detection of atrial fibrillation events. In Com-
puting in Cardiology 2013, pages 1047–1050. IEEE.
Cramer, J. S. (2002). The origins of logistic regression
(technical report). In Tinbergen Institute.
Devasena, C. L. (2014). Comparative analysis of random
forest rep tree and j48 classifiers for credit risk predic-
tion. In International Conference on Communication,
Computing and Information Technology (ICCCMIT-
2014).
Freund, Y. and Schapire, R. E. (1997). A decision-theoretic
generalization of on-line learning and an application
to boosting. Journal of computer and system sciences,
55(1):119–139.
Goldberger, A. L., Amaral, L. A., Glass, L., Hausdorff,
J. M., Ivanov, P. C., Mark, R. G., Mietus, J. E., Moody,
G. B., Peng, C.-K., and Stanley, H. E. (2000). Phys-
iobank, physiotoolkit, and physionet: components of
a new research resource for complex physiologic sig-
nals. Circulation, 101(23):e215–e220.
Haque, A. F., Ali, M. H., Kiber, M. A., and Hasan+, M. T.
(2009). Automatic feature extraction of ecg signal us-
ing fast fourier transform.
Huang, C., Ye, S., Chen, H., Li, D., He, F., and
Tu, Y. (2010). A novel method for detection of
the transition between atrial fibrillation and sinus
rhythm. IEEE Transactions on Biomedical Engineer-
ing, 58(4):1113–1119.
January, C. T., Wann, L. S., Alpert, J. S., Calkins, H.,
Cigarroa, J. E., Cleveland, J. C., Conti, J. B., Ellinor,
P. T., Ezekowitz, M. D., Field, M. E., et al. (2014).
2014 aha/acc/hrs guideline for the management of pa-
tients with atrial fibrillation: a report of the american
college of cardiology/american heart association task
force on practice guidelines and the heart rhythm so-
ciety. Journal of the American College of Cardiology,
64(21):e1–e76.
Kearley, K., Selwood, M., Van den Bruel, A., Thompson,
M., Mant, D., Hobbs, F. R., Fitzmaurice, D., and
Heneghan, C. (2014). Triage tests for identifying
atrial fibrillation in primary care: a diagnostic accu-
racy study comparing single-lead ecg and modified bp
monitors. BMJ open, 4(5):e004565.
HEALTHINF 2020 - 13th International Conference on Health Informatics
164

Kirchhof, P., Benussi, S., Kotecha, D., Ahlsson, A., Atar,
D., Casadei, B., Castella, M., Diener, H.-C., Heid-
buchel, H., Hendriks, J., et al. (2016). 2016 esc guide-
lines for the management of atrial fibrillation devel-
oped in collaboration with eacts. European journal of
cardio-thoracic surgery, 50(5):e1–e88.
Kurzweil, R. C., Gibson, L., Albrecht, P., and Grimshaw,
P. (2016). Atrial fibrillation detection and associated
methods. US Patent 9,456,762.
Lake, D. E. and Moorman, J. R. (2010). Accurate esti-
mation of entropy in very short physiological time
series: the problem of atrial fibrillation detection
in implanted ventricular devices. American Jour-
nal of Physiology-Heart and Circulatory Physiology,
300(1):H319–H325.
Lee, J., Nam, Y., McManus, D. D., and Chon, K. H. (2013).
Time-varying coherence function for atrial fibrillation
detection. IEEE Transactions on Biomedical Engi-
neering, 60(10):2783–2793.
Lee, J., Reyes, B. A., McManus, D. D., Maitas, O., and
Chon, K. H. (2012). Atrial fibrillation detection us-
ing an iphone 4s. IEEE Transactions on Biomedical
Engineering, 60(1):203–206.
Lian, J., Wang, L., and Muessig, D. (2011). A simple
method to detect atrial fibrillation using rr intervals.
The American journal of cardiology, 107(10):1494–
1497.
Miyasaka (2006). Secular trends in incidence of atrial
fibrillation in olmsted county, minnesota, 1980 to
2000, and implications on the projections for future
prevalence (vol 114, pg 119, 2006). Circulation,
114(11):E498–E498.
Moddemeijer, R. (1989). On estimation of entropy and mu-
tual information of continuous distributions. Signal
processing, 16(3):233–248.
Mohebbi, M. and Ghassemian, H. (2008). Detection of
atrial fibrillation episodes using svm. In 2008 30th An-
nual International Conference of the IEEE Engineer-
ing in Medicine and Biology Society, pages 177–180.
IEEE.
Pan, J. and Tompkins, W. J. (1985). A real-time qrs detec-
tion algorithm. IEEE Trans. Biomed. Eng, 32(3):230–
236.
Petr
˙
enas, A., Marozas, V., and S
¨
ornmo, L. (2015). Low-
complexity detection of atrial fibrillation in continu-
ous long-term monitoring. Computers in biology and
medicine, 65:184–191.
Quinlan, J. R. (2014). C4. 5: programs for machine learn-
ing. Elsevier.
Schnabel, R. B., Yin, X., Gona, P., Larson, M. G., Beiser,
A. S., McManus, D. D., Newton-Cheh, C., Lubitz,
S. A., Magnani, J. W., Ellinor, P. T., et al. (2015). 50
year trends in atrial fibrillation prevalence, incidence,
risk factors, and mortality in the framingham heart
study: a cohort study. The Lancet, 386(9989):154–
162.
Sepulveda-Suescun, J., Murillo-Escobar, J., Urda-Benitez,
R., Orrego-Metaute, D., and Orozco-Duque, A.
(2017). Atrial fibrillation detection through heart
rate variability using a machine learning approach
and poincare plot features. In VII Latin American
Congress on Biomedical Engineering CLAIB 2016,
Bucaramanga, Santander, Colombia, October 26th-
28th, 2016, pages 565–568. Springer.
Stewart, S., Murphy, N., Walker, A., McGuire, A., and Mc-
Murray, J. (2004). Cost of an emerging epidemic:
an economic analysis of atrial fibrillation in the uk.
Heart, 90(3):286–292.
Villar, R., Beltrame, T., and Hughson, R. L. (2015). Vali-
dation of the hexoskin wearable vest during lying, sit-
ting, standing, and walking activities. Applied Physi-
ology, Nutrition, and Metabolism, 40(10):1019–1024.
Xiong, Z., Stiles, M. K., and Zhao, J. (2017). Robust ecg
signal classification for detection of atrial fibrillation
using a novel neural network. In 2017 Computing in
Cardiology (CinC), pages 1–4. IEEE.
Xu, S. S., Mak, M.-W., and Cheung, C.-C. (2018). To-
wards end-to-end ecg classification with raw signal
extraction and deep neural networks. IEEE journal
of biomedical and health informatics.
Yuan, C., Yan, Y., Zhou, L., Bai, J., and Wang, L. (2016).
Automated atrial fibrillation detection based on deep
learning network. In 2016 IEEE International Con-
ference on Information and Automation (ICIA), pages
1159–1164. IEEE.
Zhao, Q. and Zhang, L. (2005). Ecg feature extraction
and classification using wavelet transform and support
vector machines. In 2005 International Conference on
Neural Networks and Brain, volume 2, pages 1089–
1092. IEEE.
Zhou, X., Ding, H., Ung, B., Pickwell-MacPherson, E., and
Zhang, Y. (2014). Automatic online detection of atrial
fibrillation based on symbolic dynamics and shannon
entropy. Biomedical engineering online, 13(1):18.
Zhou, X., Ding, H., Wu, W., and Zhang, Y. (2015). A
real-time atrial fibrillation detection algorithm based
on the instantaneous state of heart rate. PloS one,
10(9):e0136544.
Zhu, N., Diethe, T., Camplani, M., Tao, L., Burrows, A.,
Twomey, N., Kaleshi, D., Mirmehdi, M., Flach, P.,
and Craddock, I. (2015). Bridging e-health and the in-
ternet of things: The sphere project. IEEE Intelligent
Systems, 30(4):39–46.
Combining Rhythmic and Morphological ECG Features for Automatic Detection of Atrial Fibrillation
165