Simulating the Doctor’s Behaviour: A Preliminary Study on the

Identiﬁcation of Atrial Fibrillation through Combined Analysis of Heart

Rate and Beat Morphology

Gennaro Laudato

1 a

, Giovanni Rosa

1 b

, Giovanni Capobianco

, Angela Rita Colavita

Arianna Dal Forno

, Fabio Divino

1 d

, Claudio Lupi

1 e

, Remo Pareschi

1 f

, Stefano Ricciardi

Luca Romagnoli

, Simone Scalabrino

1 g

, Cecilia Tomassini

1 h

and Rocco Oliveto

1 i

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

ASREM – Regione Molise, Italy

Keywords:

Recommender System, Deep Learning, ECG Analysis, Atrial Fibrillation, Arrhythmia.

Abstract:

Atrial ﬁbrillation (AF) is a medical disorder that affects the atria of the heart. AF has emerged as a world-

wide cardiovascular epidemic affecting more than 33 million people around the world. Several automated

approaches based on the analysis of the ECG have been proposed to facilitate the manual identiﬁcation of AF

episodes. Especially, such approaches analyze the heartbeat morphology (absence of P-wave) or the heart rate

(presence of arrhythmia). In this article, we present AMELIA (AutoMatic dEtection of atriaL fIbrillation for

heAlthcare), an approach that simulates the doctor’s behavior by considering both the sources of information

in a combined way. AMELIA is basically composed of two components; one integrating a LSTM (Long

Short-Term Memory) Recurrent Neural Network (RNN) and the second integrating a rhythm analyzer. When

the RNN reveals a heartbeat with abnormal morphology, the rhythm analyzer is activated to verify whether

or not there is a simultaneous arrhythmia. AMELIA has been experimented by using well-known datasets,

namely Physionet-AF and NSR-DB. The achieved results provide evidence of the potential beneﬁts of the ap-

proach, especially regarding sensitivity. AMELIA has an incredibly high potential to be used in applications

of continuous monitoring, where the detection of AF episodes is a fundamental and crucial activity.

1 INTRODUCTION

In modern healthcare systems the vital signals of pa-

tients are acquired, collected, and analyzed within the

system itself.

This is the case of ATTICUS (Laudato et al.,

2021), an innovative tele-service and remote monitor-

ing system for ambient-assisted living based on the

analysis of vital and behavioral parameters. The data

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

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

https://orcid.org/0000-0003-0500-3852

https://orcid.org/0000-0003-4107-3727

https://orcid.org/0000-0001-5166-1130

https://orcid.org/0000-0002-4912-582x

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

https://orcid.org/0000-0002-2819-7779

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

are acquired through a smart t-shirt (Balestrieri et al.,

2019; De Vito et al., 2021) and then transmitted to an

Ambient Intelligence device located nearby, which,

in turn, predicts potentially anomalous situations and

it communicates them to a Decision Support System

(DSS). Such a system can perform deeper and more

accurate analysis and, if it conﬁrms the anomaly, it

can alert a monitoring station in which human experts

(e.g., doctors) manually analyze the data and plan an

intervention.

In this paper, we present an approach that we aim

at integrating into the DSS of ATTICUS. The ap-

proach is in charge of analyzing the ECG to iden-

tify atrial ﬁbrillation (AF) events, an abnormal heart

rhythm characterized by rapid and irregular beating

of the atria. The process of AF episodes diagno-

sis involves two ECG sources of information: (i)

morphology-based, because during an AF episode,

446

Laudato, G., Rosa, G., Capobianco, G., Colavita, A., Forno, A., Divino, F., Lupi, C., Pareschi, R., Ricciardi, S., Romagnoli, L., Scalabrino, S., Tomassini, C. and Oliveto, R.

Simulating the Doctor’s Behaviour: A Preliminary Study on the Identiﬁcation of Atrial Fibrillation through Combined Analysis of Heart Rate and Beat Morphology.

DOI: 10.5220/0010823900003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 5: HEALTHINF, pages 446-453

ISBN: 978-989-758-552-4; ISSN: 2184-4305

ﬂuctuating waveforms instead of P waves can be ob-

served and (ii) rhythm-based, because of the heart rate

irregularity, which may appear.

A lot of effort was devoted by the research com-

munity to the deﬁnition of approaches for the auto-

matic detection of AF by using Machine Learning

(ML) techniques based on one of the above sources

of information. Indeed, a widespread approach is

the Support Vector Machines (SVM) (Sepulveda-

Suescun et al., 2017; Islam et al., 2017; Padma-

vathi and Ramakrishna, 2015), while other authors

chose Neural Network (NN) to classify ECG seg-

ments (Yuan et al., 2016; Xiong et al., 2017) or novel

recursive algorithms (Zhou et al., 2015). The most

common features for the ML tools used in these

methods are based on RR Intervals (RRI) analysis.

There are also approaches, such as the one by Padma-

vathia and Ramakrishnab (Padmavathi and Ramakr-

ishna, 2015), where it is proposed the use of autore-

gressive coefﬁcients as derived morphological ECG

features.

Despite the high accuracy of the above methods,

we conjecture that there is still room for improve-

ment. We believe that an approach – based on the

combination of morphological and rhythmic analysis

– can reproduce the exact procedure used by cardiol-

ogists when manually checking an ECG for AF diag-

nosis. Our conjecture is supported by the results of

previous works proposed in Laudato et al. (2020b,a),

where a machine learning approach, named MOR-

PHYTHM was deﬁned to combine morphological and

rhythmic information to support the identiﬁcation of

AF events. Despite the promising results achieved,

the main limitation of such an approach was repre-

sented by the difﬁculties to explain a speciﬁc predic-

tion. Indeed, all the features extracted from the ECG

were put together in one single learning algorithm.

This made it difﬁcult to identify the event that trig-

gered the prediction.

Based on the willingness to have an accurate

and explainable method for detecting AF events, we

present AMELIA (AutoMatic dEtection of atriaL

fIbrillation for heAlthcare). AMELIA is an auto-

mated AF detector based on (i) an LSTM Recurrent

Neural Network (Hochreiter and Schmidhuber, 1997)

for the ECG Morphology classiﬁcation and on (ii)

statistical heuristics to identify arrhythmia. The pro-

posed approach aims at simulating as much as pos-

sible the doctor’s behaviour during the detection of

AF episodes. Especially, AMELIA ﬁrst analyzes the

morphology of the heartbeat (as it is and not in terms

of derived features as shown in the literature) in order

to identify the absence of p-wave and then conﬁrm the

anomaly by checking the presence of arrhythmia.

The proposed approach was experimented on

two publicly accessible sets of clinical data (MIT-

BIH Atrial Fibrillation Database

and MIT-BIH Nor-

mal Sinus Rhythm Database

). The accuracy of

AMELIA was compared with the work by Zhou

et al. (2015), one of the most accurate methods of

AF detection based on RRI analysis (therefore based

on only rhythmic features) and with the work re-

cently proposed by Laudato et al. (2020b) which, as

AMELIA, also takes in consideration morphological

and rhythmic features. The results reported in Sec-

tion 4 show that it is possible to achieve beneﬁts with

respect to the chosen baselines.

We believe that AMELIA can be better em-

ployed in telemedicine applications, where e-AI

(explainable-Artiﬁcial Intelligence) is often a strong

requirement. Indeed, AMELIA– thanks to the con-

ceptual and de facto separation of the data sources

between rhythmic and morphological – can provide

highly accurate information in the process of diag-

nosis. For example, AMELIA– beyond the genera-

tion of a warning indicating a potential AF episode –

can provide the additional information of which is the

heartbeats not showing a P wave with a high accuracy.

The premise that AMELIA can reproduce the ex-

act procedure used by cardiologists is based on the

consideration that the approach takes as input a se-

quence of heartbeats that are submitted only to noise-

removal and downsampling processing stages, there-

fore preserving the ECG shape. AMELIA aims

at simulate the manual diagnosis by observing the

rhythm and the shape of a pattern of heartbeats.

The rest of the paper is structured as follows: Sec-

tion 2 presents the proposed approach for AF detec-

tion, Section 3 describes the experimental choices for

the design of the study, while Section 4 presents the

results of the evaluation of the proposed approach on

the Physionet data set. Finally, Section 6 concludes

the paper by discussing the results and by reporting

all the potential future works which can be undertaken

with AMELIA in the context of ATTICUS (Balestri-

eri et al., 2019; Laudato et al., 2021).

2 AMELIA

An AF episode is diagnosed by a doctor when the

morphology of the heartbeat is abnormal (no P-wave),

RR intervals are irregularly irregular, and f-waves ap-

pear. AMELIA aims at simulating as much as possi-

ble such a behavior. The workﬂow of AMELIA is

https://physionet.org/physiobank/database/afdb/

https://physionet.org/physiobank/database/nsrdb/

Simulating the Doctor’s Behaviour: A Preliminary Study on the Identiﬁcation of Atrial Fibrillation through Combined Analysis of Heart

Rate and Beat Morphology

447

depicted in Fig. 1. In the preprocessing stage,

AMELIA extracts all the heartbeat signals and all the

R peak positions from a raw single lead digital ECG

acquired with a given sampling frequency. These sig-

nals are submitted to a Morphology analyzer.

Figure 1: AMELIA workﬂow.

If the morphology of the heartbeat is abnormal,

the Morphology analyzer triggers the Rhythm ana-

lyzer. The Rhythm analyzer takes as an input the

extracted R peak positions and tries to consolidate

the initial warning identiﬁed by the Morphology an-

alyzer. If the Rhythm analyzer identiﬁes through

the analysis of ten consecutive R-R intervals an ar-

rhythmia, then an AF episode is identiﬁed. Otherwise,

the initial warning of the Morphology analyzer is re-

jected. In this case, the abnormal morphology of the

heartbeat could be due to the wrong classiﬁcation of

the NN or just to some noise in the ECG.

2.1 Deﬁnition of a Heartbeat

It is necessary to clarify the concept of heart beat sig-

nal. In AMELIA, a heartbeat signal is a raw ECG

segment included between two successive R peaks

(see Fig. 2). The choice to deﬁne a heartbeat signal

in this way is due to the consideration that the mor-

phological features – observable during AF episodes

– are (i) the absence of P wave and (ii) the poten-

tial ﬁbrillation waves in its place. The concept of the

heartbeat signal is also faced in the work by Xu et

al. (Xu et al., 2018) with the difference that the au-

thors deﬁne it as the signal between the two middle

points of three consecutive R peaks. We decided to

work with heart dynamics included between two suc-

cessive R peaks because the ﬁbrillating phenomena

are inscribed between those waves. We used the Pan-

Figure 2: Deﬁnition of heartbeat signal in AMELIA.

Figure 3: Representation of a complete heartbeat in

AMELIA.

Tompkins method (Pan and Tompkins, 1985) to ob-

tain all the expected heartbeat signals of a given full

ECG signal. We opted for the validation of AMELIA

in the online scenario (worst case), therefore without

using the beat annotations available from Physionet.

A complete heartbeat is represented by a vector

deﬁned as:

hbs = [u

, u

, ..., u

] (1)

where u

and u

are the raw amplitudes of the

samples corresponding to the position of the left and

right R peak, respectively (see Fig. 3).

In order to provide ﬁxed-length instances to the

Morphology analyzer component – all the heartbeat

signals were submitted to a process of down-sampling

(in section 4.3, more details are provided). N is the

ﬁxed length of each complete heartbeat signal.

2.2 Morphology Analyser

The Morphology Analyser is in charge of analyzing

the morphology of a heartbeat. The input of this com-

ponent is represented by a heartbeat. Ideally, the out-

HEALTHINF 2022 - 15th International Conference on Health Informatics

448

put is no-AF if the morphology of the heartbeat is

normal and AF if the morphology of the heartbeat

does not have the P wave and shows ﬂuctuating wave-

forms (f-waves), i.e., the morphological characteris-

tics of a heart beat in the presence of an AF event. The

morphology classiﬁcation of the heartbeat is based on

a Recurrent Neural Network (RNN) (Hochreiter and

Schmidhuber, 1997) with multiple LSTM cells. The

choice is justiﬁed by the consideration that LSTM

cells better adapt to time-series classiﬁcation (Karim

et al., 2017) (as in the case of ECG).

2.3 Rhythm Analyser

The Rhythm Analyser aims at identifying normal

rhythm or arrhythmia by evaluating a buffer of ten

successive heartbeats. It is worth noting that– based

on a consolidated opinion from cardiologists – ten

consecutive heartbeats can be can be deemed enough

to diagnose atrial ﬁbrillation. This number is also

conﬁrmed by the works in (Kurzweil et al., 2009;

Zurro et al., 1995) where a minimum of 3 and 6 suc-

cessive heartbeats was evaluated.

In details, the Rhythm Analyser classiﬁes each

beat as short, long, or normal. Considering that the

normal heart rate during rest for teenagers is around

70-120 beat per minute (bpm) and adults is around

60-90 bpm (D. Limmer, 2005), each beat is classi-

ﬁed as follows: short if bpm > 120, long if bpm <

50 and normal otherwise. Once the Rhythm Analyser

has buffered and labeled ten consecutive heartbeats,

computes the entropy of the buffer B.

2.4 Putting All Together

Algorithm 1 shows how the Morphology Analyser

and the Rhythm Analyser are combined in order to de-

tect AF events.

For each heartbeat signal, a ﬁxed-length buffer

hbs

is instantiated, containing the amplitudes of the

signal. The buffer hbs

then is submitted to the Mor-

phology Analyser, which provides its classiﬁcation.

When the morphology is classiﬁed as AF, a new

buffer of heartbeats is created. Once the buffer of

heartbeats has reached the max size (set as 10, in our

case), it is submitted to the Rhythm Analyser. Based

on the entropy information evaluated on the buffer, a

classiﬁcation in terms of rhythm is provided. If also

the rhythm is identiﬁed as IRREGULAR, a warning is

generated.

Algorithm 1: Detection of Arrhythmia.

Require: ECG  Raw ECG

HBS = ExtractHeartBeatSignals(ECG)

RRI = ExtractRRInterval(ECG)

for each hbs

∈ HBS do

Morphology = MorphologyAnalyser(hbs

)

if Morphology == AF then

buffer

←

0  new buffer for the i

heart

beat

BUFFERS ← BUFFERS ∪ buffer

end if

for each buffer

∈ BUFFERS do

buffer

← buffer

∪ RRI

if size(buffer

) == MAX SIZE then

Rhythm = RhythmAnalyser(buffer

)

if Rhythm == ABNORMAL then

GenerateWarning()

end if

BUFFERS ← BUFFERS \ buffer

end if

end for

3 STUDY DESIGN

We compared AMELIA to the method proposed in

(Zhou et al., 2015), where AF episodes are identiﬁed

by using only an RRI analysis. Thus, in the context of

the study, we formulated the following research ques-

tion:

Does AMELIA outperform state-of-the-art

AF detection approaches?

We chose as baseline the approach by Zhou (2015)

et al. (Zhou et al., 2015) because in the state of the art,

it is one of the most accurate approaches based on RRI

analysis. We also keep a recent tool MORPHYTHM

(Laudato et al., 2020b) as a reference, because it is

based on a combination of rhythmic and morphologi-

cal analysis, too.

3.1 Context of the Study

The proposed approach was experimented on the

MIT-BIH AFDB (Goldberger et al., 2000). For

this DB, Physionet offers 25 2-lead ECG recordings.

These were acquired with a sampling frequency of

250 Hz, 12-bit resolution over a range of ± 10 mil-

livolts. In this preliminary study, we performed a de-

tection based on a single-lead ECG, thus we took into

account only the ﬁrst lead. Furthermore, for this data

set, Physionet does not provide distinction among

beat types (but only in terms of rhythm); indeed, all

Simulating the Doctor’s Behaviour: A Preliminary Study on the Identiﬁcation of Atrial Fibrillation through Combined Analysis of Heart

Rate and Beat Morphology

449

Figure 4: The records chosen for this study.

Figure 5: The records ignored for the study.

beats are labeled as normal. Considering the diver-

sity of the ECG shapes – in order to guarantee con-

sistency of information – we manually selected only

those recordings with a common shape. This choice is

due to speciﬁcity of AMELIA’s morphology compo-

nent. Thus, in this preliminary study, we considered

records #04015, #04048, #04126, #05091, #05261,

#06995, #08219, #08455. Indeed, this group of pa-

tients presents a high similarity between the ECG

waveform shapes (see Fig. 4). For the moment, the

other records were ignored. Examples of ignored

records are #06453 and #08455, where in the avail-

able signal 1 from the database, the shapes differ from

the ones of the above group of recordings (Fig. 5).

Each heartbeat signal was manually observed and

analyzed, with the help of a medical equipe. Only sig-

nals presenting a clear AF effect were selected. The

operation was carried out for all the chosen records.

A total of 1637 heartbeat signals from the 8 different

recordings were manually extracted. The minimum

length for each of these signals varies from 33 to 111

samples. By doing that, we obtained two types of sig-

nal: AF and Normal (no-AF) heart-beat signals.

To conclude the experiment, we tested the ﬁ-

nal version of AMELIA also on the MIT-BIH Nor-

mal Sinus Rhythm Databases because it contains pa-

tients with a healthy ECG except for some no signif-

icant arrhythmia episodes. Therefore, at the end of

the study, we will experiment AMELIA under sev-

Figure 6: The dendrogram for the manually selected records

from AF Database, based on an AF heartbeat.

eral circumstances: the detection of AF and NO-AF

episodes. These latter include normal sinus rhythm

and pathological rhythm. Indeed, both the chosen

database include pathological rhythm different from

AF episodes. Speciﬁcally, the AFDB contains atrial

ﬂutter episodes and the NSRDB arrhythmia episodes

(Goldberger et al., 2000).

3.2 Patient-centered Data Clustering

The 8 records were manually observed and selected.

Therefore, before reporting the classiﬁcation results,

we aimed at validating the manual selections per-

formed in the previous steps. To do so, the data was

submitted to a clustering algorithm to assess effec-

tively if the 8 patients group together in an individ-

ual cluster. To this aim, we selected one AF-labelled

heartbeat from each of the chosen records, and we de-

scribed each of these heartbeats using several descrip-

tors, such as: entropy measures (e.g. the one proposed

in (Zhou et al., 2015)), Statistical Features (e.g. the

mean, variance, and norm of the amplitude samples),

Fast Fourier Transform and AR model coefﬁcients.

After creating this data set, we applied a technique of

Hierarchical clustering with Euclidean distance as the

similarity function and the average as the agglomera-

tion method. We obtained the dendrogram depicted in

Fig. 6. Thus, even if they seemed to have a common

ECG waveform shape – when they are observed from

an AF heartbeat perspective – the clustering process

assigns them to distinct groups. This separation could

be due to physiological aspects that are embedded in

an ECG.

From this step of clustering, we obtained a reﬁned

view of the groups to which our records belong. This

will be used to represent the results for each patient,

according to the reference cluster.

3.3 Training of the LSTM RNN

The training of the LSTM RNN was performed on

a balanced data set composed of 1637 instances for

the AF Class and 1653 for the no-AF Class. To the

HEALTHINF 2022 - 15th International Conference on Health Informatics

450

Figure 7: Examples of the manually selected instances from

the Physionet AFDB.

aim of guaranteeing an alignment, all the instances

were downsampled to 33 data points. An example

of selected and downsampled instances is depicted in

Fig. 7. The LSTM parameters were experimentally

deﬁned through a trial & error approach. In order

to validate the network, a classical Leave-one-person-

out (LOPO-CV) cross-validation was applied to the

data set. LOPO-CV means that one person at a time

is left out from the training set, so that the training

set contains no data speciﬁc to the individual who is

being tested (the classiﬁer was not tuned with the test

data of that person).

4 ANALYSIS OF THE RESULTS

We experimented with the proposed approach on

two freely accessible data set, the Physionet MIT-

BIH AFBD and the Normal Sinus Rhythm Database

(NSRDB) (Goldberger et al., 2000). We considered

the following classiﬁcation metrics: TP (beat labeled

as AF and classiﬁed as AF), FP (beat labeled as no-

AF and classiﬁed as AF), TN (beat labeled as no-AF

and classiﬁed as no-AF), and FN (beat labeled as AF

and classiﬁed as no-AF).

Unfortunately – for the chosen baseline (Zhou

et al., 2015) – the authors did not report classiﬁca-

tions at patient-level. Therefore, we replicated the

method and obtained the desired level of classiﬁca-

tion. The classiﬁcation results are shown in Tables

1, 2, 3 according to the clusters previously obtained.

The difference in the number of heartbeats is due to

the nature of AMELIA. Our online tool embeds the

Pan-Tompkins algorithm as a peak detector (while the

methods of the state of the art use the peak annota-

tions provided from Physionet). Therefore, even if

highly accurate, the performances of AMELIA inte-

grate an additive error due to potential wrong classi-

ﬁcations of this algorithm. Therefore, the only way

to compare AMELIA—with respect to the chosen

baselines—is to evaluate the overall statistics as: Sen-

Table 1: AMELIA classiﬁcation performance compared to

the chosen baseline on MIT-BIH AF-db cluster 1.

Cluster 1 Method TP TN FP FN

#04015

AMELIA 500 42088 2886 25

MORPHYTHM 491 40650 2836 27

(Zhou et al., 2015) 478 40707 2779 40

#04048

AMELIA 792 38967 165 90

MORPHYTHM 443 38982 145 363

(Zhou et al., 2015) 419 38990 137 387

#04126

AMELIA 3345 39581 960 51

MORPHYTHM 3154 38149 1424 132

(Zhou et al., 2015) 3082 38743 830 204

Method Sens Spec Acc

AMELIA 0,965 0,968 0,968

MORPHYTHM 0,887 0,964 0,961

(Zhou et al., 2015) 0,863 0,969 0,965

Table 2: AMELIA classiﬁcation performance compared to

the chosen baseline on MIT-BIH AF-db cluster 2.

Cluster 2 Method TP TN FP FN

#05091

AMELIA 44 35470 986 98

MORPHYTHM 0 36640 4 133

(Zhou et al., 2015) 0 36644 0 133

#05261

AMELIA 881 43739 1953 51

MORPHYTHM 766 43015 1595 157

(Zhou et al., 2015) 655 44215 395 268

Method Sens Spec Acc

AMELIA 0,861 0, 964 0, 963

MORPHYTHM 0, 725 0, 980 0, 977

(Zhou et al., 2015) 0, 620 0,995 0,990

sitivity =

T P

T P+F N

, Speciﬁcity =

T N

T N+FP

and Accuracy =

T P+T N

T P+T N+FP+FN

From the achieved results, it is possible to observe

that for records belonging to: Cluster 1: AMELIA

outperforms both the baseline and MORPHYTHM

in terms of all the metrics of validation; Cluster

2: AMELIA provides signiﬁcantly higher sensitiv-

ity with slightly lower speciﬁcity and accuracy; Clus-

ter 3: AMELIA presents a signiﬁcant loss, mostly in

terms of sensitivity and accuracy.

From the LOPO-CV cross-validation, we chose

the best network in terms of accuracy on the test data

set. With this, we experimented with the proposed

approach also on the MIT-BIH Normal Sinus Rhythm

Table 3: AMELIA classiﬁcation performance compared to

the chosen baseline on MIT-BIH AF-db cluster 3.

Cluster 3 Method TP TN FP FN

#06995

AMELIA 11215 27160 490 17784

MORPHYTHM 27240 25901 1767 280

(Zhou et al., 2015) 27072 25648 2020 448

#08219

AMELIA 7946 42595 4286 6207

MORPHYTHM 13420 40934 4203 735

(Zhou et al., 2015) 12627 42637 2500 1528

#08455

AMELIA 32705 15265 22 12470

MORPHYTHM 44111 15244 45 151

(Zhou et al., 2015) 44103 15250 39 159

Method Sens Spec Acc

AMELIA 0, 587 0,947 0, 768

MORPHYTHM 0,986 0, 932 0, 959

(Zhou et al., 2015) 0, 975 0,948 0,962

Simulating the Doctor’s Behaviour: A Preliminary Study on the Identiﬁcation of Atrial Fibrillation through Combined Analysis of Heart

Rate and Beat Morphology

451

Table 4: AMELIA’s accuracy on MIT-BIH NSR-db.

Record Rhythm Analyser AMELIA

ID # FP # FP

16265 2 0

16272 0 0

16273 1 0

16420 1 0

16483 0 0

16539 13 0

16773 0 0

16786 0 0

16795 4 0

17052 29 0

17453 0 0

18177 5 0

18184 0 0

19088 10 1

19090 0 0

19093 0 0

19140 0 0

19830 16 14

Databases. In this DB, all the recordings present a

shape with high similarity with respect to the ones

used in this study. The individuals included in the

NSRDB were found to have no signiﬁcant arrhyth-

mia. Tables 4 show the results achieved on this DB.

This experimentation represents a boundary valida-

tion for our proposed methods because the goal is to

avoid all the phenomena with arrhythmia (different

from AF) by using the useful information provided

by the morphological module of AMELIA.

In this validation of AMELIA, we expected that

the proposed tool do not get confused with (not signif-

icant) arrhythmia episodes affecting the patient from

this data set. As we can see from the tables - by using

our rhythm analyzer - an arrhythmia can be detected

as an Atrial Fibrillation episode. With the introduc-

tion of the morphological Analyzer in AMELIA we

reduced the chance of misclassifying several heart-

beats.

5 THREATS TO VALIDITY

One of the limitations of the present study is that the

evaluation is performed on a reduced number of se-

lected recordings. This choice was due to the consid-

eration that the Neural Network—used in this context

as a morphology analyzer —is strictly dependent on

several features related to an ECG recording, such as

the lead, physiological aspects (smoker, BMI, etc.),

and the instrumentation used to acquire the ECG.

Therefore, we opted for manually selecting the

available recordings from the AF database to be in-

volved in this preliminary study. Of course, in the

real-world application of arrhythmia detection, the

detectors are subject to a broad range of beat mor-

phologies, including patients with ectopic beats, dif-

ferent types of arrhythmia, recordings corrupted by

noise, and so on. The evaluation results of this study

are to be intended only for a limited number of ECGs,

such as the ones with common features as the ones

selected for the validation of AMELIA. We also de-

cided to validate the manual selection of ECG record-

ings in order to assess the similarity between them.

From a reﬁned perspective offered by the dendro-

gram, we observed that the 8 recordings could be fur-

ther grouped into three sub-clusters. Thanks to this re-

sult, we could report the evaluation in terms of ’clus-

ters’ in order to highlight the speciﬁcity of AMELIA

in the detection of AF episodes

The comparison with state of the art should be

considered only illustrative because the validation

was performed with different procedures: (i) Zhou

et al. (2015) used the Physionet Long-Term AF

Database

to tune their entropy threshold and the

MIT-BIH Atrial Fibrillation Database to validate it,

(ii) in the validation of MORPHYTHM (Laudato et al.,

2020b) a LOPO-CV among all the patients from the

MIT-BIH Atrial Fibrillation Database was performed

while (iii) in AMELIA a LOPO-CV between only the

selected patients from the MIT-BIH Atrial Fibrillation

Database was executed.

6 CONCLUSION

We have presented AMELIA, an approach for au-

tomated detection of atrial ﬁbrillation in the context

of real-time monitoring of vital parameters. The ap-

proach is based on the combined use of two different

sources of information that have paramount impor-

tance in the detection of AF events: (i) morpholog-

ical analysis of the heartbeats; and (ii) RRI analysis.

An empirical study conducted on two different well-

known public data sets has shown the potential of the

proposed approach compared with the state-of-the-art

methods based on (i) just RRI analysis and (ii) on both

sources of information.

Future works will be devoted to extending the ex-

perimentation of AMELIA on other datasets.

Also, we plan to further improve the accuracy of

AMELIA, by replacing the manual selection of ECG

recordings with a fully automated process.

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

HEALTHINF 2022 - 15th International Conference on Health Informatics

452

Finally, we plan to improve the classiﬁcation per-

formances by reﬁning some clue parameters in our

proposed method. We speciﬁcally refer to the down-

sampling resolution, the Neural Network structure,

and the length of the rhythm pattern.

ACKNOWLEDGMENT

The authors have 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

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.

D. Limmer, M. O’Keefe, H. G. B. M. D. B. (2005). Emer-

gency Care Workbook. Pearson.

De Vito, L., Picariello, E., Picariello, F., Tudosa, I., Lo-

previte, L., Avicolli, D., Laudato, G., and Oliveto, R.

(2021). An undershirt for monitoring of multi-lead

ecg and respiration wave signals. In 2021 IEEE Inter-

national Workshop on Metrology for Industry 4.0 &

IoT (MetroInd4. 0&IoT), pages 550–555. IEEE.

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.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural computation, 9(8):1735–1780.

Islam, S., Ammour, N., and Alajlan, N. (2017). Atrial ﬁb-

rillation detection with multiparametric rr interval fea-

ture and machine learning technique. In 2017 Interna-

tional Conference on Informatics, Health & Technol-

ogy (ICIHT), pages 1–5. IEEE.

Karim, F., Majumdar, S., Darabi, H., and Chen, S. (2017).

Lstm fully convolutional networks for time series clas-

siﬁcation. IEEE access, 6:1662–1669.

Kurzweil, R. C., Gibson, L., Albrecht, P., and Grimshaw,

P. (2009). Atrial ﬁbrillation detection. US Patent

7,596,405.

Laudato, G., Boldi, F., Colavita, A. R., Rosa, G., Scal-

abrino, S., Lazich, A., and Oliveto, R. (2020a). Com-

bining rhythmic and morphological ecg features for

automatic detection of atrial ﬁbrillation: Local and

global prediction models. In International Joint Con-

ference on Biomedical Engineering Systems and Tech-

nologies, pages 425–441. Springer.

Laudato, G., Boldi, F., Colavita, A. R., Rosa, G., Scal-

abrino, S., Torchitti, P., Lazich, A., and Oliveto, R.

(2020b). Combining rhythmic and morphological ecg

features for automatic detection of atrial ﬁbrillation.

In HEALTHINF, pages 156–165.

Laudato, G., Scalabrino, S., Colavita, A., Chiacchiari, Q.,

D’Orazio, R., Donadelli, R., De Vito, L., Picariello, F.,

Tudosa, I., Malatesta, R., Gallo, L., and R, O. (2021).

ATTICUS: Ambient-intelligent Tele-monitoring and

Telemetry for Incepting and Catering over hUman

Sustainability. Frontiers in Human Dynamics - Dig-

ital Impacts (Frontiers).

Padmavathi, K. and Ramakrishna, K. S. (2015). Classiﬁ-

cation of ecg signal during atrial ﬁbrillation using au-

toregressive modeling. Procedia Computer Science,

46:53–59.

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

tection algorithm. IEEE transactions on biomedical

engineering, (3):230–236.

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.

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, 23(4):1574–1584.

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.

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).

Zurro, V., Stelle, A., and Nadal, J. (1995). Detection of

atrial persistent rhythm based on p-wave recognition

and rr interval variability. In Computers in Cardiology

1995, pages 185–188. IEEE.

Simulating the Doctor’s Behaviour: A Preliminary Study on the Identiﬁcation of Atrial Fibrillation through Combined Analysis of Heart

Rate and Beat Morphology

453