Automatic Real-time Beat-to-beat Detection of Arrhythmia Conditions

Giovanni Rosa

1 a

, Gennaro Laudato

1 b

, Angela Rita Colavita

Simone Scalabrino

1 c

and Rocco Oliveto

1 d

STAKE Lab, University of Molise, Pesche (IS), Italy

ASREM, Regione Molise, Italy

angelaritacolavita@asrem.molise.it

Keywords:

Arrhythmia, ECG, Machine Learning, Decision Support Systems.

Abstract:

With the spread of Internet of Medical Things (IoMT) systems, the scientiﬁc community has dedicated a lot of

effort in the deﬁnition of approaches for supporting specialized staff in the early diagnosis of pathological con-

ditions and diseases. Several approaches have been deﬁned for the identiﬁcation of arrhythmia, a pathological

condition that can be detected from an electrocardiogram (ECG) trace. There exist many types of arrhythmia

and some of them present a great impact on the patients in terms of worsening of physical conditions or even

mortality. In this work we present NEAPOLIS, a novel approach for the accurate detection of arrhythmia con-

ditions. NEAPOLIS takes as input a heartbeat signal, extracted from an ECG trace, and provides as output a

5-class classiﬁcation of the beat, namely normal sinus rhythm and four main types of arrhythmia conditions.

NEAPOLIS is based on ECG characteristics that do not need a long-term observation of an ECG for the classi-

ﬁcation of the beat. This choice makes NEAPOLIS a (near) real-time detector of arrhythmia because it allows

the detection within few seconds of ECG observation. The accuracy of NEAPOLIS has been compared to one

of the best and most recent work from the literature. The achieved results show that NEAPOLIS provides a

more accurate detection of arrhythmia conditions.

1 INTRODUCTION

The Internet of Things (IoT) is a neologism referring

to the extension of the Internet to the world of objects

allowing them to collect and exchange data. In the

healthcare sector, IoT plays an important role and rep-

resents a fertile ground. Indeed, healthcare is evolv-

ing, moving from a traditional model in which care

was only provided in hospital centers, to a new model,

where care is accessible from anywhere. This tran-

sition is supported by sensor technology. Nowadays

sensors are able to track almost every parameter of

the human body, such as blood oxygen level, insulin

level, blood pressure, temperature or even chemical

balance, and they can be easily used by patients since

they do not require special training for use (Dimitrov,

2016).

The main advantages of using IoMT (Internet of

Medical Things) are (i) preventive care, because the

https://orcid.org/0000-0002-5241-1608

https://orcid.org/0000-0002-3776-2848

https://orcid.org/0000-0003-1764-9685

https://orcid.org/0000-0002-7995-8582

data collected from patients can help to identify the

ﬁrst symptoms and possible health risks, allowing to

act promptly, and (ii) long-term care and chronic dis-

eases, because the fact of being able to collect patient

data and make them available to health professionals

makes treatment procedures much easier, faster and

more comfortable. In cases of chronic diseases, being

connected is of great help because the devices allow

patients to constantly monitor health status indicators,

follow therapy independently with higher security and

collect biometric data in real-time during therapy.

ATTICUS is an example of IoMT system—

recently proposed by Balestrieri et al. (2019)—that

constantly monitors electrocardiogram (ECG), respi-

ration, temperature, skin response and dynamics of a

patient. In ATTICUS vital signals are acquired by a

smart wearable and automatically analyzed by an Ar-

tiﬁcial Intelligence (AI) component to detect anoma-

lies and critical health conditions. Such alarms are

forwarded to a specialist doctor or can even alert a

prompt intervention of hospital staff. Thus, it is of

vital importance in ATTICUS to have accurate and

real-time analysis of the acquired data.

212

Rosa, G., Laudato, G., Colavita, A., Scalabrino, S. and Oliveto, R.

Automatic Real-time Beat-to-beat Detection of Arrhythmia Conditions.

DOI: 10.5220/0010267902120222

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 5: HEALTHINF, pages 212-222

ISBN: 978-989-758-490-9

In the context of ATTICUS we devised NEAPO-

LIS, a NovEl APproach for the autOmatic reaL-time

beat-to-beat detectIon of arrhythmia conditionS, such

as Bundle Branch Block (BBB), Premature Ventric-

ular Contractions (PVC) and Atrial Premature Beats

(APB). Arrhythmia can describe a disorder that af-

fects the regularity of the heart rhythm, by observing

too fast or too slow rhythm. Arrhythmia can be cat-

egorized into two types: atrial and ventricular. Es-

pecially this latter kind of arrhythmia may be very

dangerous. Therefore, without a continuous monitor-

ing and the right attention, ventricular arrhythmia can

lead to sudden cardiac arrest (Elhaj et al., 2016).

NEAPOLIS performs the classiﬁcation of heart

beat by extracting a set of features from an ECG trace

and providing them to a machine learning component.

The common characteristic among all the features that

NEAPOLIS extracts from the ECG is that they are

real-time, i.e., they do not need any long-term obser-

vation of the ECG.

A lot of effort has been dedicated by the sci-

entiﬁc community to the deﬁnition of methods for

the automatic detection of arrhythmia conditions (Bai

et al., 2019; Jung and Kim, 2017; Pandey and Janghel,

2020; Smisek et al., 2018; Talbi and Ravier, 2016).

The accuracy of NEAPOLIS has been compared to

the approach proposed by Pandey and Janghel (2020)

since—to the best of our knowledge—this approach

is one of the most accurate in the literature and pro-

vides the same 5-class classiﬁcation of heart beat

of NEAPOLIS. However, the method proposed by

Pandey and Janghel (2020) requires a long-term ob-

servation of the ECG, by extracting from features the

ECG trace that are computed on the past 20 minutes

of the ECG. The unique characteristics of NEAPO-

LIS allows to obtain a classiﬁcation in a much shorter

time. Indeed, in NEAPOLIS eleven beats are required

to compute all the features used by our approach to

perform the classiﬁcation. Therefore, for a subject

with a heart rate value of 60 bpm the ﬁrst classiﬁ-

cation can be performed after 11 seconds + t, where

t represents the computational time of NEAPOLIS to

build and classify the features vector (that is, however,

negligible). An empirical evaluation conducted on the

Physionet MIT-BIH arrhythmia database provides ev-

idence of the beneﬁts provided by NEAPOLIS also in

terms of classiﬁcation accuracy.

The rest of the paper is structured as follows.

Section 2 provides background information on ap-

proaches for heart beat classiﬁcation. Section 3

presents NEAPOLIS, while Section 4 and Section 5

report the design and the results of the empirical study

we conducted to evaluate NEAPOLIS, respectively.

Finally, Section 6 concludes the paper.

2 BACKGROUND

This section discusses (i) the incidence of arrhythmia

conditions on the health status; and (ii) the approaches

proposed in the literature for the automatic classiﬁca-

tion of arrhythmia conditions. The chosen baseline

method used in the evaluation of NEAPOLIS is de-

scribed in more details in a dedicated subsection.

2.1 Incidence of Arrhythmia Conditions

A bundle branch block can be deﬁned as an abnor-

mality of the electrical conduction system of the heart

(Fahy et al., 1996). In case the defect is originated

in the left or right ventricles the blocks are further

classiﬁed into Right BBB (RBBB) and Left BBB

(LBBB). Scientiﬁc research studies have reported that

BBB has been observed in 8% to 18% of subjects

with acute myocardial infarction. It has also been

associated with an increased risk of complete heart

block and sudden death (Kones and Phillips, 1980;

Newby et al., 1996). Before the involvement of

thrombolytic treatment—that limits infarct size, im-

proves ventricular morphology and function, and de-

creases mortality—several studies had reported on the

incidence of RBBB in patients with acute myocar-

dial infarction (Melgarejo-Moreno et al., 1997). The

range of incidence rate was found to be between the

3% and 29% (Col and Weinberg, 1972; Julian et al.,

1964).

It was also found that RBBB is usually the mani-

festation of infarctions. These latter are often accom-

panied by heart failure, complete AV block, arrhyth-

mias, and a high mortality rate (Atkins et al., 1973;

Mullins and Atkins, 1976; Rizzon et al., 1974). With

regard to the LBBB, the incidence in the general pop-

ulation is low, approximately 0.6% of subjects devel-

oping it over 40 years (Clark et al., 2008; Imanishi

et al., 2006). The incidence rate changes if consid-

ering patients with chronic heart failure. Indeed, ap-

proximately one third of these patients have left bun-

dle branch block (LBBB) on their 12-lead ECG (Bal-

dasseroni et al., 2002; Shenkman et al., 2002).

In the absence of structural heart disease, frequent

PVCs have traditionally been considered a benign

phenomenon, only requiring medical attention when

symptomatic. This understanding has undergone a

substantive evolution over the last decade. So-called

benign PVCs are now known to have malignant po-

tential in susceptible patients and can manifest as trig-

gers for ventricular ﬁbrillation (VF) and sudden car-

diac death (Ip and Lerman, 2018).

Ranging from 20% to 25% of ischemic strokes

occur due to embolic complications caused by atrial

Automatic Real-time Beat-to-beat Detection of Arrhythmia Conditions

213

ﬁbrillation (Evans et al., 2000; Hart, 2003). In ad-

dition, for patients that have experienced ischemic

stroke or transient ischemic attacks, in presence of AF

they can be exposed to recurrent strokes (Wallmann

et al., 2007). Therefore, it is vital to detect paroxys-

mal atrial ﬁbrillation after stroke or transient ischemic

attack and involve anticoagulation treatment in such

patients (Hart et al., 2003; van Walraven et al., 2003).

This diagnose typically includes a 24 hours continu-

ously monitoring. One of the clues that can lead to a

early diagnosis of paroxysmal atrial ﬁbrillation are the

occurrence of atrial premature beats (APB). Indeed,

in 24-hour ECG recordings frequent APB are corre-

lated to an increased incidence of paroxysmal AF in

patients with ischemic stroke(Wallmann et al., 2003).

2.2 Classiﬁcation of Heartbeats

Zhao and Zhang (2005) proposed an approach for

the extraction of features that allows a reliable heart

rhythm recognition. They basically used two tech-

niques for the features generation: wavelet was used

to extract the coefﬁcients of the transform and au-

toregressive modelling (AR) to obtain the temporal

structures of ECG waveforms. Then, wavelet and AR

coefﬁcients were concatenated together to form the

feature vector for the classiﬁcation. They evaluated a

large set of outputs that include also our target con-

ditions, but they chose to experiment the method on

a subset of the available recordings from the MIT-

BIH Arrhythmia

, a freely accessible and common

database of the scientiﬁc literature with annotation at

heartbeat level. The results showed that the approach

provided good performances of classiﬁcation reach-

ing an accuracy of 99.68%.

Li and Zhou (2016) proposed a method for ECG

classiﬁcation using entropy on Wavelet packet de-

composition (WPD) and random forests. The au-

thors also experimented the devised method on the

MIT-BIH Arrhythmia database but with a differ-

ent output because they conducted another kind of

experiment, focused on a medical standard, i.e.,

the EC57:1998 standard (ANSI/AAMI-EC57, 1998).

The authors stated that although the coefﬁcients by

Discrete Wavelet Transform (DWT) or WPD can re-

veal the local characteristics of an ECG signal, the

number of such coefﬁcients is usually so huge that

it is hard to use them as features for classiﬁcation

directly. Therefore, they extracted some high-level

features from these coefﬁcients for better classiﬁca-

tion. In the proposed method, they chose the entropy

as high level features extractor from a DWT. The re-

https://archive.physionet.org/physiobank/database/mitdb/

sults reported on an obtained overall accuracy approx-

imately equal to 94.5%.

Another very important set of features is the one

proposed by Leonarduzzi et al. (2010), i.e., a set of

features derived from the multifractal analysis. The

authors stated that this analysis highly suits the anal-

ysis of the Heart Rate Variability (HRV) ﬂuctuations,

since it gives a description of the singular behavior of

a signal. Therefore, the main features of this work are

based on the multifractal wavelet leader estimates of

the second cumulant of the scaling exponents and the

range of Holder exponents, or singularity spectrum.

The results demonstrated how these features can be

involved in a tool for a precise detection of myocar-

dial ischemia.

Many works from the scientiﬁc literature have

involved the Fast Fourier Transform (FFT) in their

methods for the classiﬁcation of ECG segments. For

instance, Haque et al. (2009) proposed a combination

of FFT-based and wavelet features. The main ﬁndings

achieved by the authors was that the wavelet can pro-

vide better indicators—rather than the FFT—of small

abnormalities in ECG signals.

2.3 The Selected Baseline

We chose as baseline for the evaluation of NEAPOLIS

the approach proposed by Pandey and Janghel (2020).

The choice is not random: the selected approach pro-

vides a complete automatic detection of heartbeats in

ﬁve heartbeat types, including the LBBB, RBBB and

PVC, i.e., the same of NEAPOLIS. The selected ap-

proach is based on a single Long Short-Term Memory

(LSTM) Neural Network as model. The inputs to the

model were based on higher-order statistics, wavelets,

morphological descriptors, and R–R intervals. Thus,

45 features were in charge of describing the electro-

cardiogram signals. In details, to extract the features,

the authors designed a temporal window of 180 sam-

ples sized (half of a second on the MIT-BIH Arrhyth-

mia). The window was centered on each R peak, pre-

viously obtained thanks to the annotations of each R

wave position available from this database. The fea-

tures have been evaluated only inside this interval.

A 2-fold cross validation was used to evaluate the

accuracy of the classiﬁcation: The entire MIT-BIH ar-

rhythmia database was divided in two folds, i.e., two

sub-dataset. Their LSTM model was trained on 40 %

(80 % of 50 %) sub-dataset, and 10 % (20 % of 50

%) sub-dataset was dedicated to a preliminary valida-

tion phase. The remaining 50 % of the data set was

used for testing. After the performance evaluation, the

model obtained an overall accuracy equal to 99.37%.

HEALTHINF 2021 - 14th International Conference on Health Informatics

214

3 NEAPOLIS IN A NUTSHELL

In this section, we present NEAPOLIS, an online de-

tector of important arrhythmia conditions, such as

BBB and PVC, based on the analysis of heartbeat sig-

nals. The high-level workﬂow of NEAPOLIS is de-

picted in Figure 1.

Input ECG signal

Count of R-peaks in

input signal> 11 ?

Buffering

Feature extraction

Beat segmentation

QRS detector for

R-peaks

Yes

2-step median ﬁlter

Patient

Beat classiﬁcation

Figure 1: The workﬂow of NEAPOLIS for online beat clas-

siﬁcation.

Once buffered a small segment—i.e., at least

11 heartbeats—of a single lead digital ECG signal,

NEAPOLIS operates to compute a beat-to-beat seg-

mentation. Then, a 2-step median ﬁlter is applied

to get rid of baseline drifts. Finally, NEAPOLIS—

through speciﬁc algorithms—evaluates the features

on the signal, scale them and creates the ﬁnal feature

vector to be submitted to the machine learning model

as input. Last task of NEAPOLIS is to provide a label

for the most probable classiﬁcation among N (Normal

Sinus Rhythm), RBBB (Right Bundle Branch Block),

LBBB (Left Bundle Branch Block), PVC (Premature

Ventricular Contraction), and APB (Atrial Premature

Beat). Next subsections describe the main compo-

nents of NEAPOLIS in detail.

3.1 ECG Digital Processing

The digital signal processing embedded in NEAPO-

LIS can be conceptually divided in preprocessing and

main processing. Both these procedures are triggered

only when a long enough portion of a digital single

lead ECG is buffered. Once these two steps are com-

pleted, the features can be extracted from the obtained

signal.

3.1.1 Preprocessing

The preprocessing step of NEAPOLIS is the same pro-

posed by Pandey and Janghel (2020). Therefore, only

the baseline removal has been performed. Speciﬁ-

cally, it concerns with the application of two median

ﬁlters: a median ﬁlter of 200 ms is applied on the raw

signal, a second median ﬁlter of 600ms is applied on

the resulting signal from the previous step.

3.1.2 Beat-to-beat Segmentation

This procedure is the same proposed by Pandey and

Janghel (2020). Especially, NEAPOLIS needs to em-

bed a QRS detector, such as the consolidated algo-

rithm proposed by Pan and Tompkins (1985). Once

evaluated each R peak position in the buffered ECG,

the segmentation process can start. The procedure is

based on the evaluation of a window of 180 samples to

be centered on an R peak. After, the selection of the

samples included in the window is performed. This

leads to the deﬁnition of a heartbeat signal, i.e., a sam-

ple vector of length 180 centered on an R peak.

3.2 Heartbeat Features

Due to their promising performance in prior simi-

lar works, we combined a set of morphological fea-

tures already used in literature for ECG classiﬁcation.

NEAPOLIS differs from the state of the art approaches

because of the constraint on the real-time detection.

Indeed, only a very limited buffering of an ECG sig-

nal is needed so that the detection of arrhythmia is

promptly offered. Next subsections describe in detail

the features extracted by NEAPOLIS.

3.2.1 Energy of Maximal Overlap Discrete

Wavelet Transform

The wavelet transform (WT) is a mathematical oper-

ator that can be used for the decomposition of time

series signals into distinct subsignals. One of the

two forms of WT is the DWT. The maximum overlap

discrete wavelet transform (MODWT) is a modiﬁed

DWT. In the MODWT, there is no process of subsam-

pling, therefore leading to a higher level of informa-

tion in the resulting wavelet and scaling coefﬁcients,

when compared to the DWT (Ghaemi et al., 2019).

For our purposes, we evaluated the MODWT and then

extracted the energy features according to the follow-

ing steps: (i) selection of a mother wavelet function W

Automatic Real-time Beat-to-beat Detection of Arrhythmia Conditions

215

and the decomposition level L; (ii) decomposition of

the original heartbeat signals according to the spec-

iﬁed W and L; and (iii) calculation of the energy of

each coefﬁcient in each node in the last level L. This

procedure has also been partially considered in the

feature extractor proposed by Li and Zhou (2016). In

our case, we used db2 as Daubechies wavelet function

and three levels of decomposition.

3.2.2 Autoregressive Model (AR)

As suggested in the method proposed by Zhao and

Zhang (2005), we involved the calculation of the Au-

toregressive model (AR) coefﬁcients of order 4. As

outcomes, we evaluated the AR coefﬁcients and the

reﬂection coefﬁcients, using the Yule-Walker estima-

tor (Friedlander and Porat, 1984).

3.2.3 Multifractal Wavelet Leader

The goal of multifractal analysis is to study signals

that present a point-wise Holder regularity variable,

i.e., that may largely vary from point to point. When

dealing with a signal, performing the multifractal

analysis refers to the estimation of its spectrum of sin-

gularities. Therefore, the determination of the spec-

trum of singularities of a signal is important to an-

alyze its singularities (Leonarduzzi et al., 2010). In

case of a real-life signal, it cannot be numerically

evaluated due to constraint like ﬁnite resolution and

the sampling of signals (Lashermes et al., 2005). To

overtake this limitation, a multifractal formalism was

introduced: the wavelet leaders (Jaffard et al., 2006).

In NEAPOLIS, we involved the multifractal wavelet

leader estimates of the log-cumulants of the scaling

exponents.

3.2.4 Fast Fourier Transform

Our approach embeds the evaluation of the Fast

Fourier Transform on the heartbeat signal. Indeed,

FFT represents a method for extracting helpful infor-

mation out of statistical features of ECG signal.

3.2.5 R-R Interval Descriptors

This set of features is basically composed of three fea-

tures:

• pre-RR interval: the distance between the actual

and previous heartbeat;

• post-RR interval: the distance between the actual

and next heartbeat;

• local-RR interval: the average of 10 previous pre-

RR values.

These features have been proposed by Pandey and

Janghel (2020), where they belonged to a larger set

of R-R statistical descriptors. We opted to embed in

NEAPOLIS only the features with an acceptable ECG

buffering. Indeed, we avoid to integrate in NEAPO-

LIS the global-RR interval presented by Pandey and

Janghel (2020) because it represented the average of

all the pre-RR values present in the last 20 min. This

would have compromised the constraint of NEAPO-

LIS to be a real-time detector.

3.3 Beat Classiﬁcation

Once extracted, the features described in Section 3.2

are normalized, in order to transform the features in

a predeﬁned range of values. We also apply a tech-

nique of sampling of the instances to deal with data

unbalance.

After these further elaborations, the features are

provided to a machine learning classiﬁer for the ﬁ-

nal classiﬁcation of the heartbeat in N (Normal Si-

nus Rhythm), RBBB (Right Bundle Branch Block),

LBBB (Left Bundle Branch Block), PVC (Prema-

ture Ventricular Contraction), and APB (Atrial Pre-

mature Beat). NEAPOLIS has not been designed for

a speciﬁc machine learning technique. The only con-

straint is represented by the use of a supervised tech-

nique. During the evaluation of NEAPOLIS we exper-

imented several machine learning techniques.

4 STUDY DESIGN

The goals of this study are (i) understanding which

are the most important descriptors of a heartbeat sig-

nal in applications of automatic detection of arrhyth-

mia conditions, such as the LBBB, RBBB, PVC and

APB and (ii) comparing NEAPOLIS with the selected

baseline. Thus, our study is steered by the following

research questions:

: What are the most important features

for the beat-to-beat classiﬁcation of arrhyth-

mia conditions?

: Which is the accuracy of NEAPOLIS?

With these research questions, we can distinguish

two objectives. With RQ

, we want to understand if

some of the features we deﬁne can be discarded to ob-

tain a higher classiﬁcation accuracy while with RQ

we want to see if NEAPOLIS can reach a classiﬁca-

tion accuracy comparable to similar state of the art

methods, especially to those that can be classiﬁed as

off-line approaches, i.e., that embed features requir-

ing a long-term observation of an ECG.

HEALTHINF 2021 - 14th International Conference on Health Informatics

216

4.1 Context of the Study

The context of our study is represented by the Phy-

sionet MIT-BIH arrhythmia database (Goldberger

et al., 2000; Moody and Mark, 2001), a state-of-art

database widely used in literature as reference data

set for arrhythmia detection (Moody and Mark, 2001).

It is composed of 48 ambulatory ECG recordings.

The acquisition was performed with a sampling fre-

quency of 360 Hz. Each recording has two channels

available: one is the modiﬁed lead II (MLII) and the

other can vary between V1, V2, V4 or V5. Heart-

beat annotations were provided by cardiologists. The

total number of labelled heartbeats is approximately

110,000 divided into 15 different beat types.

According to a consolidated procedure on this

database (Xu et al., 2018), the records with paced

beats, namely 102, 104, 107 and 217 have been ex-

cluded from the study. The experiment was conducted

on the remaining 44 records and considering 5 types

of beats annotations: N, LBBB, RBBB, APB and

PVC. Figure 2 shows the distribution of such types

of beats in the dataset.

Figure 2: Count of selected heartbeat types from the MIT-

BIH arrhythmia database (Moody and Mark, 2001).

4.2 Experimental Procedure

This section details the experimental procedure we

follow to answer our research questions.

4.2.1 RQ

: Feature Analysis

Using wfdb

toolkit we extracted raw signals and an-

notations from the arrhythmia database. Since the

annotations contain both R-peak positions and beat

types, we used the former information to split the sig-

nals in beat segments and the latter to ﬁlter beats by

https://archive.physionet.org/physiotools/wfdb.shtml

the selected types for this study. After this, we pre-

processed the signals following the procedure detailed

in Section 3.1.1. Finally, we subtracted the ﬁltered

signal from the raw one, obtaining a signal with cor-

rected baseline, as depicted in Figure 3.

Figure 3: An example of a raw beat (on top) and the same

beat with the 2-step median ﬁlter applied.

For each ECG segment obtained from the above

elaboration steps, we computed the features gener-

ation through the algorithms described in section 3.

The features vector was therefore composed of the

record id (a code used by Physionet to indicate a pa-

tient), the computed features and the label indicating

the heartbeat class.

To answer RQ

, we conducted a features analysis

on this data set. The ﬁrst step has been focused on an

analysis based on the Pearson correlation coefﬁcient

r. Indeed, we removed the features having r greater

than 0.9. Afterwards, we did another step of features

selection based on importance weights using a tree-

based classiﬁer as estimator. The features importance

is computed as the contribution of a feature to max-

imize the split criterion used by the algorithm, also

deﬁned as the minimization of the impurity of child

nodes, i.e., Gini impurity (Breiman et al., 1984).

In this way, starting from an initial set of 160 fea-

tures, we selected only 39 and ﬁltered the data set ac-

cordingly.

Automatic Real-time Beat-to-beat Detection of Arrhythmia Conditions

217

4.2.2 RQ

: NEAPOLIS Accuracy

With the purpose at answering RQ

, we ﬁrst evalu-

ated the accuracy of NEAPOLIS by using different

Machine Learning algorithms such as Random Forest

(Ho, 1998), Support Vector Machine (Noble, 2006),

k-nearest neighbors (Cunningham and Delany, 2020)

and Multi-layer Perceptron (Hinton, 1990). In ad-

dition, we distinctly involved in the experimentation

two consolidated state of the art approaches for han-

dling with the problem of data unbalance. Specif-

ically, we used (i) SMOTE (Chawla et al., 2002),

which makes an over-sampling of the minority class

by creating synthetic minority class examples and

(ii) Tomek’s links, an undersampling techniques pre-

sented by Tomek (1976). We also tested standard-

ization and scaling techniques based on the type of

classiﬁer used. For example, we used standardization

with Support Vector Machine and min-max scaling

with Random Forest.

Once identiﬁed the best conﬁguration for NEAPO-

LIS, we compare its accuracy with our baseline

(Pandey and Janghel, 2020). The two approaches

have been compared by using the following class-

level metrics:

• Sensitivity, i.e., the number of correctly classiﬁed

positive instances divided by the sum between the

number of instances correctly classiﬁed as pos-

itive and the instances misclassiﬁed as negative,

computed as

T P

T P+FN

• Speciﬁcity, i.e., the number of correctly classiﬁed

negative instances divided by the sum between the

number of instances correctly classiﬁed as neg-

ative and the instances misclassiﬁed as positive,

computed as

T N

T N+FP

• Precision, i.e., the number of correctly classi-

ﬁed positive instances divided by the total num-

ber of instances classiﬁed as positive, computed

T P

T P+FP

• F1, i.e., the harmonic mean of precision and re-

call, computed as

2×TP

(2×TP)+FN+FP

As for the validation, we followed the same pro-

tocol as the one proposed in our baseline (Pandey and

Janghel, 2020), i.e., a stratiﬁed split of the data set in

two sub data sets, namely DS1 and DS2. The result of

the stratiﬁed split procedure is that both DS1 and DS2

contains a proportional number of instances, based on

classes (i.e., the beat types). Such a decomposition of

the data set is depicted in Table 1.

In this way, we obtained two sub data sets where

DS1 was used for training and DS2 for testing only.

According to the validation protocol exhibited by

Pandey and Janghel (2020), the training set in turn

Table 1: Stratiﬁed split of the data set used for the classiﬁ-

cation experiment.

Beat type DS1 DS2

APB 1,269 1,269

LBBB 4,023 4,023

N 37,109 37,109

RBBB 3,606 3,607

PVC 3,440 3,440

Total 49,447 49,448

was further split in 80% and 20% for a preliminary

validation. In this way, for each model, in the train-

ing phase it is performed a preliminary validation on

DS1. Then, the ﬁnal testing was performed on DS2.

To avoid any convenient split of the original data

set into DS1 and DS2, we have repeated the splitting

process several times, in order to have results less af-

fected by the randomness. Especially, we selected

1,000 random seeds and then for each seed we re-

peated (i) the stratiﬁed split in DS1 and DS2 and (ii)

the individual split of DS1. This means that we chose

to repeat the complete validation protocol for 1,000

times and average the results accordingly.

5 ANALYSIS OF THE RESULTS

This section describes the results achieved aiming at

answering our research questions.

5.1 RQ

: Feature Analysis

The main results of the experiment conducted to an-

swer RQ

are depicted in Figure 4. We used a Ran-

dom Forest classiﬁer with a threshold of 1.25 * me-

dian of the features importance. Speciﬁcally, in the

ﬁgure we exhibit the ﬁve features with the highest

weight.

In details, we obtained that the feature with the

highest weight is the ﬁrst reﬂection coefﬁcient from

the AR model. Almost with the same weights, we can

ﬁnd the fourth descriptor from the MODWT model

and the pre-RR interval. Finally, the ﬁrst and third

coefﬁcients, from the FFT, are also included in the

top-5 ranking.

5.2 RQ

: NEAPOLIS Accuracy

As designed, we experimented several machine learn-

ing technique to identify the best conﬁguration for

NEAPOLIS. The best conﬁguration found is the one

composed of SMOTE, min-max scaler and Random

Forest, this latter set with 100 estimator trees. The

classiﬁcation accuracy achieved by NEAPOLIS using

HEALTHINF 2021 - 14th International Conference on Health Informatics

218

Figure 4: Top ﬁve selected features using importance

weight.

such a conﬁguration is reported in Table 2. It is worth

noting that such a conﬁguration of NEAPOLIS is used

for the comparison with our baseline.

Table 2: NEAPOLIS’s classiﬁcation metrics computed on

the validation set DS2. Those values are averaged among

the 1,000 runs of our validation protocol.

Beat type Sensitivity Speciﬁcity Precision F1

APB 90.48 99.81 92.49 91.47

LBBB 98.53 99.96 99.50 99.01

N 99.34 98.29 99.43 99.39

RBBB 99.18 99.97 99.68 99.43

PVC 98.28 99.61 95.02 96.62

avg 97.16 99.53 97.22 97.18

Table 3 reports the comparison—in terms of over-

all accuracy—between NEAPOLIS and the selected

baseline. Considering the average of the overall met-

rics, NEAPOLIS outperforms the state of the art base-

line method in terms of sensitivity, speciﬁcity, preci-

sion and F1 score. In particular—with regards to the

sensitivity and F1 score—the improvement is greater

than 2% and 1% respectively.

Table 3: Comparison of NEAPOLIS with the chosen base-

line (Pandey and Janghel, 2020) in terms of Sensitivity,

Speciﬁcity, Precision and F1 score.

Avg metrics NEAPOLIS (Pandey et al., 2020) Delta

Sensitivity 97.16 94.89 + 2.27

Speciﬁcity 99.53 99.14 + 0.39

Precision 97.22 96.73 + 0.49

F1 score 97.18 95.77 + 1.41

Performing a class level analysis (see Table 4),

what emerges from the classiﬁcation results is that

Table 4: Comparison of NEAPOLIS with the chosen base-

line (Pandey and Janghel, 2020) at class level in terms of

Sensitivity, Speciﬁcity, Precision and F1 score.

Class N

Metrics NEAPOLIS (Pandey et al., 2020) Delta

Sensitivity 99.34 99.31 + 0.03

Speciﬁcity 98.29 96.45 + 1.84

Precision 99.43 98.84 + 0.59

F1 score 99.39 99.07 + 0.32

Class LBBB

Metrics NEAPOLIS (Pandey et al., 2020) Delta

Sensitivity 98.53 97.52 + 1.01

Speciﬁcity 99.96 99.92 + 0.04

Precision 99.50 99.05 + 0.45

F1 score 99.01 98.28 + 0.73

Class RBBB

Metrics NEAPOLIS (Pandey et al., 2020) Delta

Sensitivity 99.18 98.97 + 0.21

Speciﬁcity 99.97 99.93 + 0.04

Precision 99.68 99.05 + 0.63

F1 score 99.43 99.01 + 0.42

Class PVC

Metrics NEAPOLIS (Pandey et al., 2020) Delta

Sensitivity 98.28 95.18 + 3.10

Speciﬁcity 99.61 99.63 -0.02

Precision 95.02 95.07 -0.05

F1 score 96.62 95.13 + 1.49

Class APB

Metrics NEAPOLIS (Pandey et al., 2020) Delta

Sensitivity 90.48 83.48 + 7.00

Speciﬁcity 99.81 99.79 + 0.02

Precision 92.49 91.64 + 0.85

F1 score 91.47 87.37 + 4.10

NEAPOLIS, with regard to the LBBB class, shows an

improvement greater than 1% and 0.5% only for Sen-

sitivity and F1-score respectively while for the other

metrics the results are almost the same. As far as

RBBB class, NEAPOLIS shows a slight improvement

for all the classiﬁcation metrics except for the Preci-

sion that has a delta greater than 0.5%. PVC Class is

the only one that has registered a decrease—that how-

ever does not exceed 0.05%—in terms of Speciﬁcity

and Precision. On the contrary, NEAPOLIS shows a

signiﬁcant impact in terms of Sensitivity and F1 score

for the same class, i.e., greater than 3% and 1% re-

spectively. With respect to the APB class, the im-

provement of NEAPOLIS is not signiﬁcant in terms

of Speciﬁcity and Precision but very high in terms of

Sensitivity and F1 score, i.e., equal to 7% and greater

than 4%, respectively. Finally, for what concerns

the N class, i.e., the normal heart beats, NEAPOLIS

outperforms—even slightly—the baseline method in

terms of all the classiﬁcation metrics.

Automatic Real-time Beat-to-beat Detection of Arrhythmia Conditions

219

6 CONCLUSION

We have presented NEAPOLIS, an automatic real-

time detector of arrhythmia conditions that works at

heartbeat level. Thanks to the combination of tech-

niques of (i) signal processing, (ii) features analysis

and (iii) machine learning, NEAPOLIS has shown bet-

ter results than one of the most accurate state of the

art method. Speciﬁcally, in terms of average classiﬁ-

cation metrics, NEAPOLIS outperforms the baseline

work presented by Pandey and Janghel (2020).

The main advantage of NEAPOLIS—with respect

to state of the art tool—is that it can be easily involved

in online scenarios of modern IoMT systems. Indeed,

the proposed approach embeds only features that al-

low to obtain a prompt early diagnosis of arrhythmia

conditions. In few words, NEAPOLIS does not embed

features that need a long-term buffering and elabora-

tion of the ECG.

As part of our future agenda, we aim at improving

the validation technique by involving a scheme that

avoids the random split, i.e., that separates the data

between train and test belonging to the same subject.

In addition, we will try to improve the accuracy of

NEAPOLIS by performing a ﬁne tuning of the param-

eters of the machine learning models. We also plan

to experiment Artiﬁcial Neural Networks as machine

learning technique.

ACKNOWLEDGMENT

Angela Rita Colavita, Rocco Oliveto, Giovanni Rosa

and Simone Scalabrino have been supported by the

project PON 2014-2020—ARS01 00860 “ATTICUS:

Ambient-intelligent Tele-monitoring and Telemetry

for Incepting and Catering over hUman Sustainabil-

ity” funded by the Ministry of Education, University

and Research—RNA/COR 576347.

REFERENCES

ANSI/AAMI-EC57 (1998). Testing and reporting perfor-

mance results of cardiac rhythm and ST segment mea-

surement algorithms. Standard, Association for the

Advancement of Medical Instrumentation, Arlington,

VA.

Atkins, J. M., Leshin, S. J., Blomqvist, G., and Mullins,

C. B. (1973). Ventricular conduction blocks and

sudden death in acute myocardial infarction: poten-

tial indications for pacing. New England Journal of

Medicine, 288(6):281–284.

Bai, J., Mao, L., Chen, H., Sun, Y., Li, Q., and Zhang, R.

(2019). A new automatic detection method for bun-

dle branch block using ecgs. In International Confer-

ence on Health Information Science, pages 168–180.

Springer.

Baldasseroni, S., Opasich, C., Gorini, M., Lucci, D., Mar-

chionni, N., Marini, M., Campana, C., Perini, G., De-

orsola, A., Masotti, G., et al. (2002). Left bundle-

branch block is associated with increased 1-year sud-

den and total mortality rate in 5517 outpatients with

congestive heart failure: a report from the italian net-

work on congestive heart failure. American heart jour-

nal, 143(3):398–405.

Balestrieri, E., Boldi, F., Colavita, A. R., De Vito, L.,

Laudato, G., Oliveto, R., Picariello, F., Rivaldi, S.,

Scalabrino, S., Torchitti, P., and Tudosa, I. (2019).

The architecture of an innovative smart t-shirt based

on the internet of medical things paradigm. In 2019

IEEE International Symposium on Medical Measure-

ments and Applications (MeMeA), pages 1–6.

Breiman, L., Friedman, J., Stone, C. J., and Olshen, R. A.

(1984). Classiﬁcation and regression trees. CRC

press.

Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer,

W. P. (2002). Smote: synthetic minority over-

sampling technique. Journal of artiﬁcial intelligence

research, 16:321–357.

Clark, A. L., Goode, K., and Cleland, J. G. (2008). The

prevalence and incidence of left bundle branch block

in ambulant patients with chronic heart failure. Euro-

pean journal of heart failure, 10(7):696–702.

Col, J. J. and Weinberg, S. L. (1972). The incidence

and mortality of intraventricular conduction defects in

acute myocardial infarction. The American journal of

cardiology, 29(3):344–350.

Cunningham, P. and Delany, S. J. (2020). k-nearest neigh-

bour classiﬁers–. arXiv preprint arXiv:2004.04523.

Dimitrov, D. V. (2016). Medical internet of things and big

data in healthcare. Healthcare informatics research,

22(3):156–163.

Elhaj, F. A., Salim, N., Harris, A. R., Swee, T. T., and

Ahmed, T. (2016). Arrhythmia recognition and classi-

ﬁcation using combined linear and nonlinear features

of ecg signals. Computer methods and programs in

biomedicine, 127:52–63.

Evans, A., Perez, I., Yu, G., and Kalra, L. (2000). Sec-

ondary stroke prevention in atrial ﬁbrillation: lessons

from clinical practice. Stroke, 31(9):2106–2111.

Fahy, G. J., Pinski, S. L., Miller, D. P., McCabe, N., Pye,

C., Walsh, M. J., and Robinson, K. (1996). Natural

history of isolated bundle branch block. The American

journal of cardiology, 77(14):1185–1190.

Friedlander, B. and Porat, B. (1984). The modiﬁed yule-

walker method of arma spectral estimation. IEEE

Transactions on Aerospace and Electronic Systems,

(2):158–173.

Ghaemi, A., Rezaie-Balf, M., Adamowski, J., Kisi, O.,

and Quilty, J. (2019). On the applicability of max-

imum overlap discrete wavelet transform integrated

with mars and m5 model tree for monthly pan evap-

oration prediction. Agricultural and Forest Meteorol-

ogy, 278:107647.

HEALTHINF 2021 - 14th International Conference on Health Informatics

220

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., Ali, M. H., Kiber, M. A., Hasan, M. T., et al.

(2009). Detection of small variations of ecg features

using wavelet. ARPN Journal of Engineering and Ap-

plied Sciences, 4(6):27–30.

Hart, R. G. (2003). Atrial ﬁbrillation and stroke prevention.

New England Journal of Medicine, 349(11):1015–

1016.

Hart, R. G., Halperin, J. L., Pearce, L. A., Anderson, D. C.,

Kronmal, R. A., McBride, R., Nasco, E., Sherman,

D. G., Talbert, R. L., and Marler, J. R. (2003). Lessons

from the stroke prevention in atrial ﬁbrillation trials.

Hinton, G. E. (1990). Connectionist learning procedures. In

Machine learning, pages 555–610. Elsevier.

Ho, T. K. (1998). The random subspace method for con-

structing decision forests. IEEE transactions on pat-

tern analysis and machine intelligence, 20(8):832–

844.

Imanishi, R., Seto, S., Ichimaru, S., Nakashima, E., Yano,

K., and Akahoshi, M. (2006). Prognostic signiﬁ-

cance of incident complete left bundle branch block

observed over a 40-year period. The American jour-

nal of cardiology, 98(5):644–648.

Ip, J. E. and Lerman, B. B. (2018). Idiopathic malignant

premature ventricular contractions. Trends in Cardio-

vascular Medicine, 28(4):295–302.

Jaffard, S., Lashermes, B., and Abry, P. (2006). Wavelet

leaders in multifractal analysis. In Wavelet analysis

and applications, pages 201–246. Springer.

Julian, D. G., Valentine, P. A., and Miller, G. G. (1964).

Disturbances of rate, rhythm and conduction in acute

myocardial infarction: a prospective study of 100 con-

secutive unselected patients with the aid of electro-

cardiographic monitoring. The American journal of

medicine, 37(6):915–927.

Jung, Y. and Kim, H. (2017). Detection of pvc by using

a wavelet-based statistical ecg monitoring procedure.

Biomedical Signal Processing and Control, 36:176–

182.

Kones, R. and Phillips, J. (1980). Bundle branch block in

acute myocardial infarction. current concepts and in-

dications. Acta cardiologica, 35(6):469–478.

Lashermes, B., Jaffard, S., and Abry, P. (2005). Wavelet

leader based multifractal analysis. In Proceed-

ings.(ICASSP’05). IEEE International Conference on

Acoustics, Speech, and Signal Processing, 2005., vol-

ume 4, pages iv–161. IEEE.

Leonarduzzi, R. F., Schlotthauer, G., and Torres, M. E.

(2010). Wavelet leader based multifractal analysis

of heart rate variability during myocardial ischaemia.

In 2010 Annual International Conference of the IEEE

Engineering in Medicine and Biology, pages 110–113.

IEEE.

Li, T. and Zhou, M. (2016). Ecg classiﬁcation using

wavelet packet entropy and random forests. Entropy,

18(8):285.

Melgarejo-Moreno, A., Galcer

a-Tom

as, J., Garc

ıa-

Alberola, A., Vald

es-Chavarri, M., Castillo-Soria,

F. J., Mira-S

anchez, E., Gil-S

anchez, J., and Allegue-

Gallego, J. (1997). Incidence, clinical characteristics,

and prognostic signiﬁcance of right bundle-branch

block in acute myocardial infarction: a study in the

thrombolytic era. Circulation, 96(4):1139–1144.

Moody, G. B. and Mark, R. G. (2001). The impact of the

mit-bih arrhythmia database. IEEE Engineering in

Medicine and Biology Magazine, 20(3):45–50.

Mullins, C. B. and Atkins, J. M. (1976). Prognoses and

management of venticular conduction blocks in acute

myocardial infarction. Modern Concepts of Cardio-

vascular Disease, 45(10):129–133.

Newby, K. H., Pisano, E., Krucoff, M. W., Green, C.,

and Natale, A. (1996). Incidence and clinical rele-

vance of the occurrence of bundle-branch block in pa-

tients treated with thrombolytic therapy. Circulation,

94(10):2424–2428.

Noble, W. S. (2006). What is a support vector machine?

Nature biotechnology, 24(12):1565–1567.

Pan, J. and Tompkins, W. J. (1985). A real-time qrs de-

tection algorithm. IEEE transactions on biomedical

engineering, (3):230–236.

Pandey, S. K. and Janghel, R. R. (2020). Automatic arrhyth-

mia recognition from electrocardiogram signals using

different feature methods with long short-term mem-

ory network model. Signal, Image and Video Process-

ing, pages 1–9.

Rizzon, P., Di Biase, M., and Baissus, C. (1974). Intraven-

tricular conduction defects in acute myocardial infarc-

tion. British Heart Journal, 36(7):660.

Shenkman, H. J., Pampati, V., Khandelwal, A. K., McK-

innon, J., Nori, D., Kaatz, S., Sandberg, K. R., and

McCullough, P. A. (2002). Congestive heart failure

and qrs duration: establishing prognosis study. Chest,

122(2):528–534.

Smisek, R., Viscor, I., Jurak, P., Halamek, J., and Plesinger,

F. (2018). Fully automatic detection of strict left

bundle branch block. Journal of Electrocardiology,

51(6):S31–S34.

Talbi, M. L. and Ravier, P. (2016). Detection of pvc in ecg

signals using fractional linear prediction. Biomedical

Signal Processing and Control, 23:42–51.

Tomek, I. (1976). Two modiﬁcations of cnn.

van Walraven, C., Hart, R. G., Singer, D. E., Koudstaal,

P. J., and Connolly, S. (2003). Oral anticoagulants

vs. aspirin for stroke prevention in patients with non-

valvular atrial ﬁbrillation: the verdict is in. Cardiac

electrophysiology review, 7(4):374–378.

Wallmann, D., T

uller, D., Kucher, N., Fuhrer, J., Arnold,

M., and Delacretaz, E. (2003). Frequent atrial prema-

ture contractions as a surrogate marker for paroxys-

mal atrial ﬁbrillation in patients with acute ischaemic

stroke. Heart, 89(10):1247–1248.

Wallmann, D., T

uller, D., Wustmann, K., Meier, P., Iseneg-

ger, J., Arnold, M., Mattle, H. P., and Delacr

etaz, E.

Automatic Real-time Beat-to-beat Detection of Arrhythmia Conditions

221

(2007). Frequent atrial premature beats predict parox-

ysmal atrial ﬁbrillation in stroke patients: an opportu-

nity for a new diagnostic strategy. Stroke, 38(8):2292–

2294.

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, 23(4):1574–1584.

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.

HEALTHINF 2021 - 14th International Conference on Health Informatics

222