Morphological Classiﬁcation of Heartbeats in Compressed ECG

Gennaro Laudato

, Francesco Picariello

, Simone Scalabrino

Ioan Tudosa

, Luca De Vito

and Rocco Oliveto

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

Department of Engineering, University of Sannio, Benevento (BN), Italy

Keywords:

ECG Analysis, Arrhythmia, Decision Support System, Compressed Sensing, Machine Learning.

Abstract:

The number of connected medical devices that are able to acquire, analyze, or transmit health data is continu-

ously increasing. This has allowed the rise of Internet of Medical Things (IoMT). IoMT-systems often need to

process a massive amount of data. On the one hand, the colossal amount of data available allows the adoption

of machine learning techniques to provide automatic diagnosis. On the other hand, it represents a problem

in terms of data storage, data transmission, computational cost, and power consumption. To mitigate such

problems, modern IoMT systems are adopting machine learning techniques with compressed sensing meth-

ods. Following this line of research, we propose a novel heartbeat morphology classiﬁer, called RENEE, that

works on compressed ECG signals. The ECG signal compression is realized by means of 1-bit quantization.

We used several machine learning techniques to classify the heartbeats from compressed ECG signals. The

obtained results demonstrate that RENEE exhibits comparable results with respect to state-of-the-art methods

that achieve the same goal on uncompressed ECG signals.

1 INTRODUCTION

Nowadays, the use of Internet of Medical Things

(IoMT) systems for remote health monitoring is play-

ing a pivotal role in improving both the effectiveness

of medical devices and the accessibility to medical

services (Hassanien et al., 2018). Remote health mon-

itoring refers to a process where the patient’s health is

continuously checked, thus allowing the identiﬁcation

and the prevention of diseases. To this aim, the use of

wearable devices for continuous monitoring is receiv-

ing increasing interest from both the health services

and the manufactures. For example, in the case of

electrocardiogram (ECG) monitoring, several IoMT

systems based on wearable devices have been pro-

posed (Balestrieri et al., 2019; Wang et al., 2019).

With the spread use of IoMT systems, the com-

plex and time-consuming steps of pre-diagnosis and

diagnosis — usually manually performed by special-

ized medical staff — can be supported or undertaken

by such systems. For this reason, in the recent years,

several methods for the automatic detection of cardiac

diseases from an ECG trace and, more speciﬁcally,

automatic classiﬁcation of heartbeats have been pro-

posed (Mond

ejar-Guerra et al., 2019; Kandala et al.,

2019; Rajesh, 2018; Chen et al., 2017; David et al.,

2011b; David et al., 2011a; Garcia et al., 2017; Xu

et al., 2018; Mar et al., 2011). All these methods

— when applied in contexts of long-term continu-

ous monitoring — require that the physical devices

of the IoMT system continuously send the monitored

data to a gateway or directly to a server. Since wear-

able devices are battery-powered and since they use

a wireless connection to exchange data, it is very im-

portant to decrease the amount of transmitted data for

reducing the energy consumption of the device and, as

a consequence, increasing its battery life. (Balestri-

eri et al., 2020) proposed an ECG data acquisition

system that performs data compression according to

compressed sensing, a theoretical framework that ex-

ploits the sparsity of a signal in a speciﬁc domain

without computationally load the physical device that

performs the compression. The authors also showed

that — by using such kind of systems — it is possible

to reduce the number of transmitted data and increase

the battery life by ∼12 %, while keeping a good ECG

signal reconstruction quality.

In this paper, we propose RENEE (heaRtbEat

classiﬁcatioN in comprEssed ECG), a novel method

for the automatic classiﬁcation of heartbeats that

works on a compressed ECG signal, through the in-

volvement of a method based on a 1-bit signal quan-

tization (Picariello et al., 2021). The advantage of us-

ing RENEE with respect to others available heartbeat

classiﬁers is that it allows reducing the signal data rate

and performs the heartbeat classiﬁcation directly on

the quantized samples, therefore without reconstruct-

ing the signal waveform. Especially, RENEE was de-

386

Laudato, G., Picariello, F., Scalabrino, S., Tudosa, I., De Vito, L. and Oliveto, R.

Morphological Classiﬁcation of Heartbeats in Compressed ECG.

DOI: 10.5220/0010236003860393

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 5: HEALTHINF, pages 386-393

ISBN: 978-989-758-490-9

signed to be used in an IoMT system based on wear-

able devices, where it is needed to reduce the data

transmitted by the physical device and to automat-

ically classify the heartbeats, without analyzing the

signal waveform. In such a context, using a com-

pressed signal is beneﬁcial both in terms of data rate

and memory occupied.

We evaluated RENEE on a public data set, the

MIT-BIH Arrhythmia Database, a reference point in

the literature. The evaluation has respected the rec-

ommended practice provided by a well-known stan-

dard, the ANSI/AAMI EC57:1998

. The achieved re-

sults provide evidence that RENEE allows to keep a

comparable overall accuracy — in the classiﬁcation of

heartbeats — with respect to state-of-the-art methods

that works with the full (i.e., uncompressed) original

ECG signal. Indeed, one of the best approaches from

the literature shows a global accuracy of 0.947 while

RENEE achieves 0.94 in the compressed domain.

The remainder of the paper is structured as fol-

lows. Section 2 provides details about the recent state

of the art on (i) the morphological automatic classi-

ﬁcation of heartbeats and (ii) the compression algo-

rithms applied to ECG signals. Section 3 presents

RENEE, our novel approach for the morphological

classiﬁcation of heartbeats in the compressed domain.

Section 4 and Section 5 reports on the design and the

results of the empirical study that we conducted to

evaluate RENEE. Finally, Section 6 concludes the pa-

per and provides suggestions for possible future re-

search directions.

2 BACKGROUND AND RELATED

WORK

This section provides details on (i) the automatic

classiﬁcation of ECG through machine learning tech-

niques and (ii) approaches proposed for ECG com-

pression.

2.1 Heartbeat Classiﬁcation

In the last years, several methods have been pro-

posed for the automatic classiﬁcation of ECG heart-

beats (Mond

ejar-Guerra et al., 2019; Kandala et al.,

2019; Rajesh, 2018; Chen et al., 2017; Garcia et al.,

2017; Xu et al., 2018; Mar et al., 2011). Most of

them have shown good results. They all dealt with the

full ECG, i.e., without any type of compression. In

addition, most of them involve complex algorithms,

American National Standard prepared by the Association

for the Advancement of Medical Instrumentation.

i.e., that need high-computational cost for the creation

of the features. In details, these approaches classify

each heartbeat in ﬁve output categories: normal beat

(N), ventricular ectopic beat (V), supraventricular ec-

topic beat (S), fusion of a normal and a ventricular ec-

topic beat (F) and unknown beat type (Q). Also, they

have been validated on the public MIT-BIH arrhyth-

mia database

(Mond

ejar-Guerra et al., 2019) proposed a method

for the automatic classiﬁcation of ECG based on the

combination of multiple Support Vector Machines.

The method relies on the time intervals between con-

sequent beats and their morphology for the ECG char-

acterization. Several features based on wavelets, local

binary patterns (LBP), higher order statistics (HOS)

were employed. The designed methodology approach

was tested classifying four kinds of abnormal and nor-

mal beats. The authors achieved an overall accuracy

of approximately 0.945 and, in some cases, they have

obtained better results than the related state of the art

approaches.

(Kandala et al., 2019) presented an inter-patient

heartbeat classiﬁcation algorithm. The foundation of

this work is based on the consideration that the ECG

is a non-stationary, non-Gaussian signal derived from

nonlinear systems (Rajesh, 2018). Therefore, the au-

thors employed a decomposition method, namely im-

proved complete ensemble empirical mode decompo-

sition (ICEEMD), to obtain features from the ECG

beats. Then, nonlinear measures such as entropies and

higher-order statistics (HOS) were determined from

the modes obtained after ICEEMD. These were used

as features for the discrimination of the heartbeats. To

handle with class unbalance, the authors employed

a type of ensemble classiﬁcation based on a major-

ity voting scheme. Finally, the authors demonstrated

good improvement — with respect to the state of the

art methods — especially for the minority classes.

(Garcia et al., 2017) proposed a heartbeat rep-

resentation, called the temporal vector-cardiogram

(TVCG), and an optimized feature extraction pro-

cess with complex networks and particle swarm op-

timization (PSO). The authors show that their method

presents an overall accuracy in the classiﬁcation equal

to 0.924.

(Xu et al., 2018) proposed an end-to-end method

with a deep neural network (DNN) for both feature

extraction and classiﬁcation based on aligned heart-

beats. This method avoids any further elaboration

for the features creation and produces optimized ECG

representation for heartbeat classiﬁcation. The overall

working principle of this approach can be resumed as

follows: the system buffers raw single lead ECG sig-

https://physionet.org/content/mitdb/1.0.0/

Morphological Classiﬁcation of Heartbeats in Compressed ECG

387

nals — at one end — and produces heartbeat classiﬁ-

cation, at the other end. The pre-processing concerns

with the selection of heartbeats from continuous ECG

signals. In the proposed approach a heartbeat signal is

represented by a segment of the ECG that comprises

the consecutive sample points of a complete heartbeat

cycle, which includes not only the QRS complex but

also the P and T waves. The Neural Network (NN)

is used for both feature extraction and classiﬁcation,

which are achieved by the lower part and the upper

part of the network, respectively. Two steps must be

performed in order to extract ﬁxed-length feature vec-

tors from raw ECG signals: (i) heartbeat segmentation

and (ii) heartbeat alignment. To the best of our knowl-

edge, the approach proposed by (Xu et al., 2018) rep-

resents one of the best approaches presented in the

state of the art with an overall accuracy of 0.947 in

the 5-class classiﬁcation. For this reason, we used

such an approach as baseline in the evaluation of RE-

NEE.

2.2 ECG Compression

In the literature, several compression methods of ECG

signals have been proposed with the aim of reduc-

ing the data rate of the IoMT physical device and its

energy consumption. The proposed approaches can

be classiﬁed in hardware-based methods and digital-

based methods (Picariello et al., 2021).

The hardware-based compression methods exploit

the sparsity of the ECG signal in the time domain to

design speciﬁc Analog-to-Digital Converter (ADC).

According to (Picariello et al., 2021), the digital-

based ECG compression methods can be classiﬁed in:

(i) direct methods, (ii) parameter extraction methods,

and (iii) transform domain methods.

The transform domain methods have gained sig-

niﬁcant attention due to their good capability of rep-

resenting the ECG signal even at high compression

ratios (Picariello et al., 2021). However, most of them

require a high computational load to be implemented

in real-time on data acquisition systems having low

resources (Picariello et al., 2021).

Alternatively, Compressed Sensing (CS) has been

proposed in the literature for ECG data compression

(Picariello et al., 2021). The advantage of CS —

as compared to other methods — relies in its capa-

bility of achieving performance comparable with the

transform-domain methods, while moving the com-

putational load from the data acquisition system to the

node that receives the compressed samples. Thus, this

solution has been widely used for implementing data

compression on devices with constrained resources,

such as wearable devices.

In some cases, the data rate reduction can be

obtained by optimizing the resolution of the data,

thus introducing a controlled quantization (Jha and

Kolekar, 2018; Bera et al., 2019).

The aim of this paper is to apply a low-complexity

compression algorithm based on a 1-bit quantization

of the ECG signal and to assess the capability of some

machine learning algorithms to successfully classify

the heartbeat from the quantized samples. It is worth

noting that the proposed classiﬁcation approach op-

erates directly on the compressed data and does not

require a reconstruction of the ECG waveform before

the classiﬁcation. As compression algorithm, we have

chosen a simple 1-bit quantization, as proposed by

(Picariello et al., 2021). The algorithm is applied to

heartbeat signals. During this phase, (i) the data is

normalized, (ii) the dither is applied, and (iii) the 1-

bit quantization is performed. With normalization we

simply refer to the application of the formula

hbs

− min(hbs)

max(hbs) − min(hbs)

for each sample i of the heartbeat signal hbs; in other

words, we normalize the data between the minimum

and the maximum value for every heartbeat signal. As

a result, all the values will be in the interval [0, 1].

The application of dither consists in applying a

Gaussian dithering noise to the heartbeat signals with

power σ. Thus, let be (i) hbs a heartbeat signal,

(ii) ds the noising signal obtained by imposing ds

σ ×random(0, 1) so that |ds| = |hbs|. The noised ver-

sion of the original signal is given by nhbs = hbs +ds.

Dither can be considered as a kind of noise, but it

is typically and intentionally applied to randomize

quantization error and thus to improve the next quan-

tization step (Pohlmann, 1995).

Finally, 1-bit quantization step performs a com-

parison with a pre-deﬁned threshold γ: if a given sam-

ple value nhbs

exceeds the threshold γ, 1 is assigned

to the output vector, while 0 is assigned otherwise.

Formally, for each nhbs

qhbs

(

1, if nhbs

≥ γ

0, otherwise

3 THE PROPOSED APPROACH

In this section, we present RENEE, a novel approach

for the classiﬁcation of heartbeats in the compressed

domain. The proposed method is contextualized in

an IoMT system and in an healthcare scenario, where

a wearable device (transmitter) acquires an ECG sig-

nals and needs to send the acquired signal to a deci-

HEALTHINF 2021 - 14th International Conference on Health Informatics

388

sion support system (receiver) for its automatic anal-

ysis.

The pre-processing stage is composed by an R-

peak detection algorithm and a consecutive selection

of a complete heartbeat signal. Once executed these

steps, the further processing consists in the compres-

sion. Finally, the compressed data is provided as in-

put to the machine learning classiﬁer. This latter is in

charge of providing the ﬁnal multi-class classiﬁcation

on the heartbeat types. We provide more details about

each step of our approach below.

3.1 Pre-processing of ECG Data

The pre-processing steps expected from our proposed

approach are composed of an R-Peak detector and a

heartbeat selection technique. Such steps may be per-

formed on the transmitter device.

The R-peak detector is in charge of accurately

evaluating the R-peak positioning in a single-lead

ECG signal. For our purposes, we used the R-peak

annotations available from the database but — for an

online scenario — a R-peak detection algorithm has

to be involved in RENEE, such as the Pan-Tompkins

algorithm (Pan and Tompkins, 1985; Sedghamiz,

2014).

The heartbeat selection technique needs the R-

peak positioning information provided by the previ-

ous algorithm in order to properly select the heartbeat.

According to the chosen baseline, a heartbeat signal is

deﬁned as the samples included between two middle

points of three successive R-peaks. In other words, a

heartbeat is not computed as an ECG signal included

between two R-peaks, but as a signal composed of (i)

an individual QRS complex, and (ii) the previous and

successive dynamics.

After all these steps, we compressed the data

through the method proposed by (Picariello et al.,

2021), based on a 1-bit signal quantization.

3.2 Features Creation & Classiﬁcation

In order to create informative features for the classi-

ﬁcation stage, we ﬁrst deﬁned a windowed accumu-

lation of samples by imposing a ﬁxed window length

winLen: given the qhbs signal, we deﬁne a new sig-

nal whbs so that whbs

∑

j=i−winLen

qhbs

. In this

way, RENEE is able to represent the dynamics of the

original heartbeat signal in the compressed domain.

The classiﬁcation component of RENEE—

through the use of machine learning techniques —

is in charge of providing the ﬁnal classiﬁcation of

the heartbeat in ﬁve types, according to the AAMI

standard (ANSI/AAMI-EC57, 1998): normal beat

(N), ventricular ectopic beat (V), supraventricular

ectopic beat (S), fusion of a normal and a ventricular

ectopic beat (F) and unknown beat type (Q).

The features we use for the automatic classiﬁca-

tion of the heartbeats are the samples of the ﬁnal sig-

nal we obtained, i.e., whbs. It is worth noting that the

number of samples may vary among different heart-

beats. However, the machine learning model, ap-

pointed for the classiﬁcation of each heartbeat sig-

nal, needs the data to be aligned in terms of features

across all the instances. Therefore, we needed to se-

lect a maximum number D of whbs samples to use as

features. To do this, we used the same method used

by (Xu et al., 2018), i.e., we apply zero-padding and

truncation in case the heartbeat signal contains less

or more samples as compared to a ﬁxed threshold D,

respectively.

Figure 1 depicts an example of how the signal

changes after each of the main stages of RENEE.

The plot in the upper row shows the original signal

waveform of a sample heartbeat. The second sub-

plot shows the signal after the application of dither-

ing noise. Then, the samples obtained by the 1-bit

quantization procedure are depicted in the third sub-

plot. Finally, the fourth line of plot shows the sig-

nal once applied the windowed accumulation of sam-

ples. Such a signal contains the features used by the

machine learning method for the classiﬁcation of the

heartbeat.

4 EMPIRICAL STUDY DESIGN

The goal of this study is to evaluate the accuracy of

RENEE in classifying heartbeat types from a highly

compressed version of an ECG. The literature shows

that the uncompressed trace of an ECG allows to ob-

tain a very accurate classiﬁcation (Xu et al., 2018).

Thus, the study is steered by the following research

question:

Can RENEE provide a heartbeat

classiﬁcation comparable to state-of-the-art

methods based on uncompressed ECG?

The perspective of the study is both (i) of a researcher

who wants to understand if machine learning tech-

niques are able to classify heartbeat also in the com-

pressed domain, and (ii) of a practitioner who wants

to use a method in a telemedicine application that is

able to balance accuracy and data storage and trans-

mission.

Morphological Classiﬁcation of Heartbeats in Compressed ECG

389

Figure 1: The four main steps performed by RENEE after the pre-processing stage: (1) the original heartbeat signal, (2) the

signal after the pre-processing and noising, (3) the compressed signal through 1-bit quantization and (4) the ﬁnal elaboration

— applied to the compressed signal — that consists of a windowed accumulation of binary samples.

4.1 Context of the Study

The context of this study is represented by the MIT-

BIH Arrhythmia Database (Moody and Mark, 2001;

Goldberger et al., 2000), a commonly used bench-

mark which contains 48 half-hour two-channel am-

bulatory ECG recordings, obtained from 47 subjects.

These ECG were digitized at 360 Hz with 11-bit res-

olution over a 10 mV range. Approximately 110,000

annotations are included in the database.

Each heartbeat is classiﬁed by using 15 different

classes. These 15 types of heartbeat in the MIT-

BIH arrhythmia database have been categorized in

ﬁve classes, reported in (ANSI/AAMI-EC57, 1998).

In most of the recordings in the MIT-BIH

database, the ﬁrst channel is a modiﬁed limb lead II

(MLII), and the second one is a modiﬁed lead V1. In

our experiments, only the signal from the ﬁrst channel

was used for ECG classiﬁcation because, typically,

QRS complexes are usually prominent (Xu et al.,

2018).

Finally, according to the AAMI recommendation

(ANSI/AAMI-EC57, 1998), we removed from the

dataset four recordings containing paced beats. The

ﬁnal dataset was composed of a total of 44 records.

4.2 Experimental Procedure

We experimented a large set of machine learning

techniques to train the model embedded in RE-

NEE. In order to execute a complete experimenta-

tion, we chose at least one classiﬁer from each cat-

egory of classiﬁers available from the Weka ma-

chine learning toolkit (Hall et al., 2009), i.e., J48

(Quinlan, 2014), Replication Tree (Devasena, 2014),

Random Forest (Barandiaran, 1998), Logistic regres-

sion (Cramer, 2002), AdaBoost M1 (Freund and

Schapire, 1997), BayesNet

, J48 (Cohen, 1995) and a

3-layer long short-term memory (LSTM) Neural Net-

work (NN) (Lang et al., 2019). We chose the Long

Short-Term Memory (LSTM) NN (instead of DNN,

for example) because ECG signals are time series data

and LSTM is capable of learning long-term depen-

dencies (Xu et al., 2018). We have also included

in our study several classiﬁers implemented in the

Matlab Classiﬁcation Learner app

. Basically, they

are speciﬁc implementation of different categories

of classiﬁers. Examples are the k-nearest neighbors

(KNN)(Dasarathy, 1991) and the Support Vector Ma-

chine (SVM)(Noble, 2006).

The whole dataset, composed by 44 records, was

split into two datasets, i.e., DS1 and DS2: this al-

lows to perform a patient-independent classiﬁcation

in which each patient appears either in the training or

in the test set, but never in both of them. Each dataset

contains approximately 50,000 beats from 22 record-

ings. We used DS1 as the training set and DS2 as the

https://bit.ly/3bYCFcR

https://bit.ly/35oMcZ9

HEALTHINF 2021 - 14th International Conference on Health Informatics

390

test set. This is a consolidated procedure from the lit-

erature. Indeed, it was used in many previous works

(De Chazal et al., 2004; Ye et al., 2012; Raj and Ray,

2018; Xu et al., 2018).

As for the parameters of RENEE, we used the con-

ﬁgurations reported in Table 1 and determined with a

trial & error approach—on a different data set—than

the one used for the classiﬁcation experiment.

We compare RENEE with the chosen baseline

work, i.e., a state-of-the-art method designed to auto-

matically classify uncompressed ECG heartbeats with

high accuracy(Xu et al., 2018).

To compare the approaches, we use several met-

rics. First, we use the accuracy, i.e., the number

of correctly classiﬁed instances divided by the total

number of instances. Then, we use also some class-

level metric—designed for a given class—among the

ones we consider for our study i.e., N, S, V, F. We

list below the class-level metrics we compute for each

class:

• 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

• Recall

, i.e., the number of correctly classiﬁed

positive instances divided by the total number of

instances actually positive, computed as

T P

T P+FN

• F-Measure

, i.e., the harmonic mean of precision

and recall, computed as

2×Precision

×Recall

Precision

+Recall

• AUC

, i.e., the deﬁnite integral used as a measure

of the two-dimensional area underneath the ROC

curve.

• MCC

(Matthews Correlation Coefﬁcient), a very

reliable statistical rate which returns a higher

value the better are the four confusion matrix in-

dicators: true and false positives, true and false

negatives.

We report the comparison for the classes N, S, V, F.

Similarly to other studies (De Chazal et al., 2004; Ye

et al., 2012; Raj and Ray, 2018; Xu et al., 2018), we

exclude the class Q, because such a class contains

paced beats (that were excluded) and unclassiﬁable

beats (only 15).

Table 1: Conﬁguration parameters of RENEE used in the

experimentation.

Parameter Description Value

σ Power of Gaussian dithering 0.1

γ Threshold for the quantization 0.2

winLen Window size for whbs 20

D Number of features 417

4.3 Threats to Validity

A limitation of this study may be represented by the

validation. Even if our kind of validation takes care

to appropriately separate the data of distinct subjects

for the training and testing phases, a more appropri-

ate validation would have been a typical L1SO-CV

(Leave 1 Subject Out Cross Validation). This impli-

cates that the data related to an individual patient will

be included once in the test data set and n-1 times in

the training data set. However, we decided to adopt

the validation method used in previous study to fa-

cilitate the comparison of the results achieved. The

replication of the study on larger data sets and with

different validation methods is part of the agenda of

our future works.

5 ANALYSIS OF THE RESULTS

We report in Table 2 the accuracy of RENEE ob-

tained by using the 10 top performing classiﬁers ex-

perimented in our study. The achieved results show

that RENEE allows to keep a comparable overall ac-

curacy in the classiﬁcation of heartbeats when com-

pared to the baseline, which, as previously mentioned,

uses the uncompressed original ECG signal. Espe-

cially, the Random Forest, the Bagged Trees and the

Medium Gaussian classiﬁers achieve the highest ac-

curacy (between 0.93 and 0.94).

Table 2: Overall accuracy of RENEE by using the 10 top

performing classiﬁers experimented in our study. At the

bottom we also report the accuracy achieved by the ap-

proach proposed by (Xu et al., 2018).

Classiﬁer Accuracy

Random Forest 0.940

Bagged Trees 0.937

Medium Gaussian SVM 0.932

Boosted Trees 0.928

LSTM NN 0.928

Fine Gaussian SVM 0.926

Fine Tree 0.925

Quadratic SVM 0.922

JRip 0.919

Cubic SVM 0.912

Baseline (Xu et al., 2018) 0.947

Table 3 reports the detailed results achieved by

RENEE when using the best performing classiﬁer,

i.e., Random Forest. The results reported in Table

3 highlight a clear outcome: RENEE is able to cor-

rectly classify with a high accuracy the classes la-

Morphological Classiﬁcation of Heartbeats in Compressed ECG

391

beled as N and V. In details, RENEE has correctly

classiﬁed respectively 43,978 out of 44,259 and 2,713

out of 3,221 total instances. For what concerns the

class S, RENEE still needs to improve in terms of

classiﬁcation accuracy. The machine learning model

with the highest classiﬁcation accuracy speciﬁc for

class S is the Quadratic SVM; such model was able

to correctly classiﬁes 128 instances out of approxi-

mately 1,800 total instances. Finally, for the F class,

the results obtained in terms of Precision, Recall and

F-Measure are not satisfying. It is worth noting,

however, that the classiﬁcation performance on these

classes has a low impact on the overall performance.

Indeed, as suggested by the standard ANSI/AAMI

EC57 (ANSI/AAMI-EC57, 1998), it is recommended

to focus the attention on the two majority arrhythmia

classes, i.e., classes S and V.

Table 3: Detailed classiﬁcation evaluation of RENEE when

using the best performing classiﬁer, i.e., Random Forest.

Class Precision Recall F-Measure AUC

N 0.945 0.994 0.969 0.966

S 0.375 0.016 0.031 0.565

V 0.907 0.842 0.873 0.987

F 0.074 0.018 0.029 0.812

Table 4 shows the MCC achieved by the two com-

pared approaches for each class. The achieved results

indicate that RENEE achieves similar results (even if

it performs slightly better) compared to the baseline

for the classes N and V. Indeed, for these classes,

the difference in terms of MCC is below 0.05. The

greatest delta has been obtained for the class S, which

touches the amount of 0.6. Thus, further improve-

ments are required for the classiﬁcation in the com-

pressed domain of this particular class.

Table 4: Comparison between RENEE (Random Forest)

and the approach proposed by (Xu et al., 2018).

Class (Xu et al., 2018) RENEE Delta

N 0.69 0.67 -0.02

S 0.67 0.07 -0.60

V 0.91 0.87 -0.04

F 0.22 0.03 -0.19

6 CONCLUSION

We have presented RENEE, an automatic approach

for (i) compressing an ECG signal through 1-bit quan-

tization and (ii) classifying the heartbeats in com-

pressed ECGs using machine learning techniques. An

empirical evaluation conducted on the MIT-BIH Ar-

rhythmia Database indicates that the overall classiﬁ-

cation accuracy of RENEE is comparable to the accu-

racy of the best state-of-the-art method for the clas-

siﬁcation of heartbeats based on uncompressed ECG

signal (0.940 vs. 0.947).

Future work will be devoted to replicate the evalu-

ation of RENEE by using more robust validation, such

as the L1SO-CV to corroborate our ﬁndings.

ACKNOWLEDGMENT

This work has been supported by the project PON

2014-2020—ARS01 00860 “ATTICUS: Ambient-

intelligent Tele-monitoring and Telemetry for

Incepting and Catering over hUman Sustainability”

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.

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.

Balestrieri, E., Daponte, P., De Vito, L., Picariello, F., Ra-

puano, S., and Tudosa, I. (2020). A wi-ﬁ iot proto-

type for ecg monitoring exploiting a novel compressed

sensing method. ACTA IMEKO, 9:38.

Barandiaran, I. (1998). The random subspace method for

constructing decision forests. IEEE Trans. Pattern

Anal. Mach. Intell, 20(8):1–22.

Bera, P., Gupta, R., and Saha, J. (2019). Preserving Abnor-

mal Beat Morphology in Long-term ECG Recording:

An Efﬁcient Hybrid Compression Approach. IEEE

Transactions on Instrumentation and Measurement,

69(5).

Chen, S., Hua, W., Li, Z., Li, J., and Gao, X. (2017). Heart-

beat classiﬁcation using projected and dynamic fea-

tures of ecg signal. Biomedical Signal Processing and

Control, 31:165–173.

Cohen, W. W. (1995). Fast effective rule induction. In

Twelfth International Conference on Machine Learn-

ing, pages 115–123. Morgan Kaufmann.

Cramer, J. S. (2002). The origins of logistic regression

(technical report). In Tinbergen Institute.

HEALTHINF 2021 - 14th International Conference on Health Informatics

392

Dasarathy, B. V. (1991). Nearest neighbor (nn) norms: Nn

pattern classiﬁcation techniques. IEEE Computer So-

ciety Tutorial.

David, V., Adochiei, N., Adochiei, F., and Tudosa, I.

(2011a). ECG waves and features extraction using

Wavelet Multi-Resolution Analysis. In 2011 E-Health

and Bioengineering Conference (EHB), pages 1–4.

David, V., Adochiei, N., and Tudosa, I. (2011b). Methods

of electromagnetic interference reduction in electro-

cardiographic signal acquisition. Environmental En-

gineering and Management Journal, 10(4):553–559.

De Chazal, P., O’Dwyer, M., and Reilly, R. B. (2004). Auto-

matic classiﬁcation of heartbeats using ecg morphol-

ogy and heartbeat interval features. IEEE transactions

on biomedical engineering, 51(7):1196–1206.

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.

Garcia, G., Moreira, G., Menotti, D., and Luz, E. (2017).

Inter-patient ecg heartbeat classiﬁcation with temporal

vcg optimized by pso. Scientiﬁc Reports, 7(1):1–11.

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.

Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann,

P., and Witten, I. H. (2009). The weka data min-

ing software: an update. ACM SIGKDD explorations

newsletter, 11(1):10–18.

Hassanien, A. E., Dey, N., and Borra, S. (2018). Medical

Big Data and Internet of Medical Things: Advances,

Challenges and Applications.

Jha, C. K. and Kolekar, M. H. (2018). Electrocardiogram

data compression using DCT based discrete orthogo-

nal Stockwell transform. Biomedical Signal Process-

ing and Control, 46:174–181.

Kandala, R. N., Dhuli, R., Pławiak, P., Naik, G. R., Moein-

zadeh, H., Gargiulo, G. D., and Gunnam, S. (2019).

Towards real-time heartbeat classiﬁcation: evalua-

tion of nonlinear morphological features and voting

method. Sensors, 19(23):5079.

Lang, S., Bravo-Marquez, F., Beckham, C., Hall, M.,

and Frank, E. (2019). Wekadeeplearning4j: A deep

learning package for weka based on deeplearning4j.

Knowledge-Based Systems, 178:48 – 50.

Mar, T., Zaunseder, S., Mart

ınez, J. P., Llamedo, M., and

Poll, R. (2011). Optimization of ecg classiﬁcation

by means of feature selection. IEEE transactions on

Biomedical Engineering, 58(8):2168–2177.

Mond

ejar-Guerra, V., Novo, J., Rouco, J., Penedo, M. G.,

and Ortega, M. (2019). Heartbeat classiﬁcation fusing

temporal and morphological information of ecgs via

ensemble of classiﬁers. Biomedical Signal Processing

and Control, 47:41–48.

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.

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.

Picariello, F., Iadarola, G., Balestrieri, E., Tudosa, I., and

De Vito, L. (2021). A novel compressive sampling

method for ecg wearable measurement systems. Mea-

surement, 167:108259.

Pohlmann, K. C. (1995). Principles of digital audio.

McGraw-Hill, Inc.

Quinlan, J. R. (2014). C4. 5: programs for machine learn-

ing. Elsevier.

Raj, S. and Ray, K. C. (2018). Sparse representation of ecg

signals for automated recognition of cardiac arrhyth-

mias. Expert systems with applications, 105:49–64.

Rajesh, K.N.and Dhuli, R. (2018). Classiﬁcation of imbal-

anced ecg beats using re-sampling techniques and ad-

aboost ensemble classiﬁer. Biomed. Signal Process.

Control, 41:242–254.

Sedghamiz, H. (2014). Matlab implementation of pan

tompkins ecg qrs detector. Code Available at the File

Exchange Site of MathWorks.

Wang, C., Qin, Y., Jin, H., Kim, I., Granados Vergara, J. D.,

Dong, C., Jiang, Y., Zhou, Q., Li, J., He, Z., Zou,

Z., Zheng, L., Wu, X., and Wang, Y. (2019). A low

power cardiovascular healthcare system with cross-

layer optimization from sensing patch to cloud plat-

form. IEEE Transactions on Biomedical Circuits and

Systems, 13(2):314–329.

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.

Ye, C., Kumar, B. V., and Coimbra, M. T. (2012). Heartbeat

classiﬁcation using morphological and dynamic fea-

tures of ecg signals. IEEE Transactions on Biomedical

Engineering, 59(10):2930–2941.

Morphological Classiﬁcation of Heartbeats in Compressed ECG

393