Detection of Abnormalities in Electrocardiogram (ECG) using Deep

Learning

ao Pestana

, David Belo

and Hugo Gamboa

LIBPHYS-UNL / FCT, New University of Lisbon, Portugal

Keywords:

Electrocardiogram, Signal Processing, Deep Learning, Artiﬁcial Intelligence, Arrhythmia Detection, Noise

Detection.

Abstract:

The Electrocardiogram (ECG) cyclic behaviour gives insights on a subject’s emotional, behavioral and car-

diovascular state, but often presents abnormal events. The noise made during the acquisition, and presence of

symptomatic patterns are examples of anomalies. The proposed Deep Learning framework learns the normal

ECG cycles and detects its deviation when the morphology changes. This technology is tested in two different

settings having an autoencoder as base for learning features: detection of three different types of noise, and

detection of six arrhythmia events. Two Convolutional Neural Network (CNN) algorithms were developed

for noise detection achieving accuracies of 98.18% for a binary-class model and 70.74% for a multi-class

model. The development of the arrhythmia detection algorithm also included a Gated Recurrent Unit (GRU)

for grasping time-dependencies reaching an accuracy of 56.85% and an average sensitivity of 61.13%. The

process of learning the abstraction of a ECG signal, currently sacriﬁces the accuracy for higher generalization,

better discriminating the presence of abnormal events in ECG than detecting different types of events. Further

improvement could represent a major contribution in symptomatic screening, active learning of unseen events

and the study of pathologies to support physicians in the future.

1 INTRODUCTION

In the context of medicine and healthcare in gen-

eral, physiological signals offer information about the

health state. For the ECG, in particular, the morpho-

logical and spectral components of each cycle pro-

vides hints of the emotional, behavioral and the health

state of the individual (Silipo and Marchesi, 1998;

Brown, 1999).

With a deeper understanding of pathologies and

the development of diagnostic methods and therapies

allied to the evolution of technology in the medical

ﬁeld (e.g. wearables), greater healthcare expectations

emerged in terms of efﬁciency. The merged ﬁelds of

machine learning and the medical ﬁeld provide tech-

nologies that are useful in assisting medical practi-

tioners not only by decision making processes, diag-

nosis and treatment, but also by continuous monitor-

ing (Johnson et al., 2018; Coiera, 2003; Faust et al.,

2018). On this account Deep Neural Networks (DNN)

https://orcid.org/0000-0002-1760-5255

https://orcid.org/0000-0002-5337-0430

https://orcid.org/0000-0002-4022-7424

present themselves as a tool that learns higher abstrac-

tions of data, while dealing with a large amount of

data and the decrease the need for feature engineer-

ing (Johnson et al., 2018; Faust et al., 2018).

We propose a framework that learns the abstrac-

tion of the default morphology of a normal ECG and

detects the divergence when it is modiﬁed, in two dif-

ferent scenarios: contamination due to noise and arti-

facts, and; the presence of symptomatic occurrences.

2 ELECTROCARDIOGRAM

The ECG measures the electrical activity generated

by the heart activity in relation to time by inserting

electrodes on the skin. This signal is used to diag-

nose the cardiac health state and can be accessed by

understanding the fundamentals of its cycle, i.e. the

morphological sequence its characteristic waves: P,

QRS complex, T and U (Acharya et al., 2007).

The P wave is due to the depolarization of the

atrial myocardium, the following QRS-complex, a

fast spiking wave that stimulates the ventricular con-

traction. The end of the cycle is made by the T wave

236

Pestana, J., Belo, D. and Gamboa, H.

Detection of Abnormalities in Electrocardiogram (ECG) using Deep Learning.

DOI: 10.5220/0008967302360243

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 4: BIOSIGNALS, pages 236-243

ISBN: 978-989-758-398-8; ISSN: 2184-4305

that indicates preparation to a new cycle with the re-

polarization of the ventricular myocardium. The U

may indicate the late depolarisation of the ventricular

myocardium (Acharya et al., 2007). The Signal-to-

Noise Ratio (SNR) comprises a heart rate between 60

and 100 bpm at rest while maintaining this sequence.

2.1 ECG Noise

Noise can be deﬁned as any signal that is not related

to the heart’s electrical activity that obstructs the ECG

signal through interference (Rodrigues et al., 2017).

Usually, the signal is obtained in a controlled envi-

ronment, although the usage of ambulatory ECG is

more associated with certain variables that inﬂuence

the noise (Xiong et al., 2019).

Noise interference has many possible sources,

which can be classiﬁed as biological, like motion

artifact and muscle contraction, and environmental

or non-physiological noise, related to nearby electri-

cal devices and variables associated with the equip-

ment (Rodrigues et al., 2017).

baseline-wander (BW) artifacts are caused by

movement such as the contraction and relaxation of

the thoracic cage muscles, which is evident if the

event is periodic. Noise affected signals can also orig-

inate from the misplacement of electrodes and loss

of contact between the sensor and the skin, called

electrode motion (EM). Additionally, the activation of

the muscles cause the EMG signal interference in the

ECG signal, revealing motion artifact (MA).

2.2 Arrhythmias

Arrhythmias can be deﬁned as rhythms that are

considered abnormal, considering the Normal Sinus

Rhythm (NSR), that are related to malfunctions in the

heart’s electrical activity and changes in the heart tis-

sue (arr, ). They can be detected in ECG by analysing

the heart rate, the presence or absence of certain

waves, the duration of intervals, wave overlap, ab-

normal timing of cardiac events, the shape, size and

direction of the waves, between others.

Common physiological arrythmias can be de-

tected from variations of the intervals between R

peaks (RR intervals). These changes are correlated

with the respiration rhythm and the sympathetic ner-

vous system. When RR interval is lower than 60 bpm

the arrhythmia is named tachycardia and higher than

100bpm bradycardia. (arr, ).

The pathophysiological arrhythmias may be

caused due to blocks in the electrical impulse

conduction, enlargement of the myocardium, peri-

carditis, electrolyte imbalance, respiratory diseases,

drugs, hypothermia and accessory conduction path-

ways(Acharya et al., 2007).

3 STATE OF THE ART

Over the years, various methods have been studied

for ECG noise and arrhythmia detection. The follow-

ing sections show an overview of the general devel-

opments in these areas.

3.1 Noise Detection in ECG

A decision rule-based algorithm is proposed by

(Satija et al., 2018) that calculates the maximum ab-

solute amplitude, number of zero crossings and local

maximum peak amplitude of the autocorrelation func-

tion are calculated, after using a modiﬁed ensemble

empirical mode decomposition. This work was able

to discriminate between six signal groups with an ac-

curacy of 98.93%. (Ansari et al., 2018) proposed a

16 layer Convolution Neural Networks (CNN) which

predicts once per second on a 10 second inputs. An

AUC of 0.977 was reached for this binary classiﬁ-

cation model with a 88,7% sensitivity (John et al.,

2018). Ansari et al. (2018) also developed a CNN

to detect usable and unasable ECG segments, in terms

of calculating the heart rate variability (HRV). The ﬁl-

tering using the CNN resulted in an area under curve

(AUC) of 0.96 for the classiﬁcation of noise affected

segments, compared to an AUC of 0.87 for a Support

Vector Machine model.

3.2 Pathological Event Detection in

ECG

More recently, (Acharya et al., 2017) used a CNN

composed of 11 layers that analyse with 2 or 5 sec-

onds for detection of atrial ﬁbrillation, atrial ﬂutter,

and ventricular ﬁbrillation. This study achieved ac-

curacy, sensitivity and speciﬁcity values of 92.50%,

98,09% and 93,13%, respectively, were achieved,

and for the two seconds ECG segments and 94,90%,

99.13% and 81.44% in the same order.

Also, a combination of CNN with Recurrent Neu-

ral Networks (RNN) was developed by Andersen et

al. (2018) in order to distinguish between atrial ﬁbril-

lation and normal segments for ECG recordings of 24

hours. The extracted features using the CNN module

were processed by the RNN achieving a sensitivity of

98.98% and a speciﬁcity of 96.95%, making predic-

tions for 24 hours of ECG signal in under one second.

This algorithm was sensible to noise, reducing signif-

Detection of Abnormalities in Electrocardiogram (ECG) using Deep Learning

237

icantly its results when present (S. Andersen et al.,

2018).

Recently, Hannun et al. (2019) developed a 34

layer DNN that could detect between 10 different

arrhythmias and NSR using raw signal, while also

detecting noise corrupted segments, constituting 12

prediction classes in total. The results were higher

than average cardiologists in terms of sensitivity and

matching speciﬁcity values. The average AUC was

of 0.97 and the confusion matrices were similar, ac-

centuating the same problematic rhythm classes for

both (Hannun et al., 2019).

4 METHODS

Before the signals are fed to the deep learning algo-

rithms, preprocessing steps are implemented in order

to transform the raw data. These are imperative for

the models to learn the underlying patterns, to pro-

vide better generalization and overall quality. There-

fore, the signal is submitted to the decimation method

in order to reduce the number of samples, while miti-

gating the loss of information. This technique applies

a low pass anti-aliasing ﬁlter (order 8 Chebyshev type

I) that suppresses all frequencies that may cause alias-

ing, before the subsampling process.

In order to reduce the baseline noise that is con-

tained in the normal ECG, the clean signal was con-

volved with a Hanning window. All records were nor-

malized and the average was removed, in order to be

interpretable by the network, with the following rule:

x − x

max(x) − min(x)

(1)

where the normalized signal is x

and x denotes the

raw signal. Finally the signals are segmented in win-

dows of 64 samples, since the mean duration of a cy-

cle is of 1 second, most of the widows have one QRS

complex.

4.1 Architectures

In order to perform ECG classiﬁcation according to

the two tasks at hand, two different architectures were

chosen and optimized to classify noise affected seg-

ments and to classify different types of arrhythmia

and NSR. Since autoencoders are good in learning

general features from the input, they were selected as

the base for the implementation of both detection al-

gorithms.

In a ﬁrst stage, an autoencoder will lean the char-

acteristics of a normal ECG signal in the purest form

possible, i.e. with a good SNR and no symptomatic

events. Consequently, the latent vector will contain

the coded information of the input, and thus also

called ”feature vector”. When this structure is fed

with a modiﬁed ECG it will produce a different la-

tent space. As the state changes it can be analysed

and detected using a classiﬁer.

The training of each algorithm was performed us-

ing batches of 256 shufﬂed windows, each without an

overlap for the noise detection experiment, and 50 %

for the arrhythmia. The testing was made with the

rest of each signal. Since the normal ECG cycle oc-

curs during one second it is expected that each win-

dow contains at least one QRS complex.

4.1.1 Autoencoder

CNN autoencoders are able to encode the input data,

as well as reconstruct it at the output layer with the

goal of reproducing the input signal as similar as pos-

sible. As the convolutional layers are capable learning

feature maps, which extracts the essential features to

reconstruct the input, these algorithms are an unsuper-

vised feature learning mechanism (Mao et al., 2017).

The diagram for the encoder is depicted in Fig. 1,

which comprises three convolutional blocks, each

containing one convectional layer with Rectiﬁed Lin-

ear Unit (ReLU) activation and a max pool, producing

the ﬁnal latent vector with 256 elements.

Figure 1: The encoder is composed by three blocks com-

posed by a convolutional (”Conv”) with a ReLU activation

and a max pooling ”MP” layers.

Even though the decoder won’t be used speciﬁ-

cally for the classiﬁcation of abnormal events in ECG,

this component is vital for the training phase. The

training, monitoring and optimization of the encoder

is made by training comparing the input and output

with the mean squared error and RMSProp as the op-

timizer (initial learning rate of 0.001).

4.1.2 Noise Detection Neural Network

Two different detection networks were developed: (1)

a binary noise detection model capable of classify-

ing between Noise Affected Signals (NAS) or Nor-

mal Signal (NS); (2) multi-class detection using three

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

238

types commonly found in ECG: BW, MA and EM

noise.

Both algorithms follow the same structure, as de-

picted in Fig. 2. The difference between the al-

gorithms is in the number of layers in the fully

connected (FC) network because of the number of

classes. As this network progresses toward the end,

each layer decreases by half the number of neurons,

starting in 256, with ReLU activations in the ﬁrst lay-

ers and softmax in the last one, and since the last layer

contains the number of classes for each system, the

ﬁrst settled with 9 while the second with 10 layers.

In both setups, a dropout of 50% is implemented in

the ﬁfth layer. The training was also made with RM-

SProp, but with cross entropy loss function.

Figure 2: Model architecture for the noise detection, with

number of classes.

4.1.3 Arrhythmia Detection Neural Network

The sequential approach for detection and prevention

of the arrythmia events is portrayed in the diagram in

Fig. 3. It is composed by an an encoder, FC layers and

a RNN comprising three sequential Gated Recurrent

Units (GRU). A RNN is a particular type of neural

network that can be described as a dynamic sequen-

tial data processing model where its current internal

state is affected by the previous one (Pascanu et al.,

2013). The GRU is a particular type that manages the

learning process using gates (Ghimes et al., 2018).

After the ECG features from each widow are ex-

tracted using the previous CNN encoder, N number

of feature vectors are stacked, resulting in matrices of

size (N × F

), where F

is the dimension of the fea-

ture vector. Since the RNN module has its own set of

hidden units (H

), the FC layer, with linear activation,

makes the connection between both modules. The pa-

rameters used for this experiment were F

= 256 and

= 128 and N = 32

After the three GRU layers the output is carried

into a FC layer, classifying into one of the seven

classes (C

= 7). The ﬁnal module starts with a batch

normalization layer but is similar to the previous FC

layers used for classiﬁcation of the noise algorithms

from H

= 128, until the size of C

= 7.

The training of this algorithm was made with RM-

SProp with the initial learning rate of 0.001, which

was automatically adjusted during its training phase.

The input data is separated into batches of 256 sig-

nal windows ensuring that it was within the computa-

tional limitations.

4.2 Datasets

All the datasets used are public and can be accessed

through Physionet, a biosignal database that was cre-

ated under the auspices of the National Institutes of

Health (Goldberger et al., 2000).

The Fantasia dataset was acquired from twenty

young (21-34 yr) and twenty elderly (68-81 yr) while

exposed to 120 min of continuous supine resting ECG

recordings (250 Hz sampling rate) while watching the

Disney’s movie Fantasia (Iyengar et al., 1996). The

signals collected by this database are mostly clean and

free from symptoms, which reﬂects a good baseline

for training the autoencoder.

For the noise detection model, both Fantasia

and MIT-BIH Noise Stress Database were consid-

ered (Moody et al., 1984). The last includes 12 half-

hour ECG recordings and 3 half-hour recordings typ-

ical noise in ambulatory ECG recordings with 250Hz.

For the last task, the MIT-BIH Arrhythmia

Database was used, displaying a sampling frequency

of 360 Hz. The chosen 6, out of 14, had enough

number of episodes to provide a good balance be-

tween classes while training the algorithm: Atrial

Fibrillation (AFIB), Atrial Flutter (AFL), Ventricular

Bigeminy (B), Paced Rhythm (P), Wolff-Parkinson-

White Syndrome, or Pre-excitation (PREX), and Si-

nus Bradycardia (SBR) (Moody and Mark, 2001).

5 RESULTS

All models were developed, trained and tested using

Keras API on top of the Tensorﬂow library. The hard-

ware included a NVIDIA GeForce GTX 960.

5.1 Autoencoder

The autoencoder was trained and tested using all 40

individuals in the Fantasia Database, being that 70%

were used for training and 30% for testing. After

approximately 2100 epochs of training the minimum

reached loss was 0.0001. For the testing phase, the

mean squared error was calculated for each signal

reaching a mean value of 0.0026 and a standard devi-

ation of 0.0012 during ±140 minutes. The results for

Detection of Abnormalities in Electrocardiogram (ECG) using Deep Learning

239

Figure 3: Model architecture for the arrhythmia detection, with C

= 7 classes.

the ﬁrst 10 individuals of the dataset, which have ages

between 21 and 34 years old, are depicted in Fig. 4.

Figure 4: Mean and standard deviation values for the ﬁrst

10 subjects within the age group of 21 to 34 years old.

The low values exhibited by the results suggest

that the model was able reconstruct the main charac-

teristics of the signal, that can be seen in the example

shown in Fig. 5.

Figure 5: Portion of the signal ECG 9, from the Fantasia

Database (a) and reconstruction of the same signal (b).

After analyzing the results, it was concluded that

the encoder could be applied for the following classi-

ﬁcation models.

5.2 Noise Detection Neural Network

The training set for both noise detection models the

data was separated in 70% for training and 30% for

testing. Furthermore, the testing set had three indi-

viduals that the training set did not contain. Both al-

gorithms took ±120min to train.

5.2.1 Binary Noise Detection Model

As stated before, the binary noise detection model

classiﬁes ECG segments according to two classes: NS

or NAS. The training proceeded until 600 epochs,

where an abrupt increase of error was observed,

reaching a minimum cross entropy of 0.51 with an

accuracy of 98,56%.

The classiﬁcation results, shown in Table 1, pro-

vides evidence that the model was able to successfully

detect the presence of noise in the ECG. Further sup-

port for this claim is presented in the confusion matrix

exhibited in Fig. 6.

Table 1: Binary noise detection model: classiﬁcation per-

formance (%).

Accuracy Sensitivity Speciﬁcity

98,18 98,21 98,15

Figure 6: Normalized confusion matrix of the dataset used

for the training phase, where NS is the positive label and

NAS is the negative label.

The encoded data for both classes were submitted

to t-Distributed Stochastic Neighbour Embedding (t-

SNE) in order to visualize the proximity between the

feature vectors. Therefore, each point in the result-

ing graph represents the tensor created by the encoder

originated by a single window input. Both classes are

clearly separated in two clusters when each time win-

dow is (equally distributed by label) submitted to t-

SNE(Fig. 7). This conﬁrms that the encoder extracts

different values for each feature represented by the

two classes.

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

240

Figure 7: Binary noise detection algorithm - t-SNE repre-

sentation of normal signal (NS) (red) and Noise affected

signal (NAS) (green) encoded data windows.

5.2.2 Multi-class Noise Detection Model

During the training phase, a minimum cross entropy

error of 0.85 was reached after 600 epochs, result-

ing in a training accuracy of 86,17% after ±180min.

However, testing accuracy was only able to reach

70,74%.

The evaluation of this model for each class are

conveyed in Table 2.

Table 2: Multi-class noise detection model: classiﬁcation

performance for each class (%)

Class Sensitivity Speciﬁcity

NS 89,77 96,63

BW 72,70 84,26

EM 65,19 91,35

MA 59,11 89,47

By inspecting the normalized confusion matrix in

Fig. 8, it is can be stated that even though the algo-

rithm is able to correctly identify normal signal most

of the times, it is harder to correctly separate the iden-

tity of each type of noise. However, the model was

able to discern between EM and MA characteristics.

Figure 8: Multi-class Noise detection algorithm - Confu-

sion matrix for the classes: Normal Signal (NS), baseline-

wander (BW), electrode motion (EM) and motion artifact

(MA).

Comparing these results with the previous exper-

iments, we come to the conclusion that even though

it is possible to detect noise corrupted ECG with high

accuracy, but differentiating between the several types

of noise reveals to be a harder task. This may be due

the fact that most of the times signals are affected by

different types of noise sources at once and because

the encoder was not trained to deal with these differ-

ences.

5.3 Arrhythmia Detection Model

Cross-validation was performed with 90% for training

the data and the rest for testing. The minimum cross

entropy error of 0.65 was reached after 1750 epochs,

ending this training phase with a 90,74% accuracy.

However, for the test data, the accuracy was only of

56,85%.

The confusion matrix, displayed in Fig. 9, shows

that the model was able to correlate between certain

characteristics with the correspondent arrhythmia,

but underperforming between distinguishing some of

them. This DNN architecture had a high classiﬁcation

rate for the P in contrast with the detection of AFIB,

often confusing its morphology with AFL, since both

of them are quite similar.

Figure 9: Arrithmia detection algorithm - Confusion matrix

for the classes of Atrial Fibrillation (AFIB), Atrial Flutter

(AFL), Ventricular Bigeminy (B), Paced Rhythm (P), Pre-

excitation (PREX), and Sinus Bradycardia (SBR).

These results are promising and reveal enough

facts to embrace the development of this algorithm by

ﬁne-tuning its parameters and have some considera-

tions in the training phase of the autoencoder. Unfor-

tunately at this current stage it cannot be used in a real

life context, since the model often considers patho-

logical ECG segments as NSR, giving too may false

positives and negatives. By analysing the classiﬁca-

tion performance of each class presented in Table 3, it

is possible to conclude that the algorithm often strug-

gles to correctly identify ECG segments with B, AFL

and AFIB, reaching an average sensitivity and speci-

Detection of Abnormalities in Electrocardiogram (ECG) using Deep Learning

241

ﬁcity for this model of 61.13% and 93.10%, respec-

tively. But we are conﬁdent that these results can be

improved with more powerful tools. Options for these

improvement are the increase of the window size, that

would help the algorithm understand the differences

of AFL and AFIB that, between other aspects, change

the distance between peaks and extending the period

of time, including more cycles, would make these eas-

ier to perceive. In a future work, this algorithm will

be tested with a more powerful hardware setting to be

able to encompass this parameter change.

Table 3: Arrhythmia detection algorithm: classiﬁcation per-

formance for each class (%).

Class Sensitivity Speciﬁcity

NSR 63,35 85,49

P 97,12 99,73

B 36,31 93,67

SBR 62,60 96,33

PREX 81,33 94,45

AFL 38,14 91,88

AFIB 49,06 90,12

AFIB+AFL 84,80 90,91

Hannun et al (2019) were able to reach an average

sensitivity of 75.22% for a 12 class model, compared

to the average sensitivity of 61.13% for the proposed

model. However, is it important to note that in the

mentioned study, AFIB and AFL classes were com-

bined into one unique class (Hannun et al., 2019).

Figure 10 represents the merged results for a direct

comparative study Figure 9, and when the sensitivity

and speciﬁcity values are re-calculated (Table 3) the

ﬁrst mean increases to 84,80%, but the second aver-

age decreases 70,91%, a value that only differs from

the study mentioned by 4,31%.

6 CONCLUSIONS

As for the last example of application of DNN archi-

tectures, the objective was to develop a framework

that detects abnormalities in the normal pattern of the

ECG signal, by creating a model that learns the key

characteristics of a normal ECG cycle. The resultant

models were capable of detecting two types of abnor-

malities: noise and pathological events.

A machine-learned extractor was developed in or-

der to be included in the architectures for noise and

arrhythmia detection, by using the encoder module

of a trained autoencoder. From the shown results it

is suggested that by successfully reproduce the input

signals, the autoencoder learns the key characteristics

of a normal ECG cycle.

Figure 10: Arrhythmia detection algorithm -Normalized

confusion matrix of the dataset used for the testing phase,

with classes: Atrial Fibrillation and Atrial Flutter (AFIB +

AFL), Ventricular Bigeminy (B), Paced Rhythm (P), Pre-

excitation (PREX), and Sinus Bradycardia (SBR).

For the noise detection models, reaching state-of-

the-art performance in the case of binary classiﬁca-

tion, despite achieving lower results in discriminating

different types of noise. In the case of the arrhythmia

decision system, the model did not perform as well as

the compared state-of-the-art accuracy, but the sensi-

tivity and speciﬁcity showed to be promising.

The improvement of the system could rely on the

need to increase the power for the encoder by sup-

plying it with more ECGs. As the autoencoder learns

and replicates an increasing quantity of different seg-

ments of asymptomatic signals, the more this algo-

rithm will embrace the abstract notion of what is an

ECG. The presence of more noisy and symptomatic

episodes would also contribute to the classiﬁcation

blocks, since the dataset was limited in the number

of provided segments.

In sum, the performance of these algorithms

proves that it is possible to detect generic abnor-

mal events in ECG using the proposed architectures.

Therefore, the implementation of an encoder for clas-

siﬁcation algorithms proves to be effective in the de-

tection of abnormal events by learning the signals key

characteristics.

As for the detecting each type of abnormal events

the model did not reach state-of-the-art results as ex-

pected since the algorithms developed sacriﬁce accu-

racy to reach better generalizations for all ECG sig-

nals to be applicable in real life. Applications that

understand the deviation from the normal signal into

unseen pathologies and the possibility of creating new

classes that did not exist before, could be extremely

beneﬁcial for future not only for ECG, but also other

physiological signals.

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

242

One example is the application of this framework

to active learning mechanisms as most of the available

algorithms are speciﬁc to speciﬁc conditions, while

not detecting others. The ability to detect similar ab-

normalities in the ECG signal, the classiﬁer could ac-

tively create new labels for patterns that diverge from

the normal ECG signal, while increasing new diag-

nostic possibilities. That said, this technique could

not only aid in the diagnostic of arrhythmias, but also

contribute to the study of these pathological anoma-

lies, by ﬁnding correlations between the different ex-

pressions of the ECG morphology, and also by ex-

ploring these manifestations with different variables

in mind such as the patient’s gender, age, medications,

among other factors. Another example is by analysing

billions of data points in wearable data, the medical

doctor could focus only in the parts which contained

different aspects of the signal, so that he could diag-

nose without wasting hours or days of analysis.

This method helps to increase the possibilities, not

only for detection, but also for studying what exactly

is a normal cycle and which are the deviation patterns.

By having early detection procedures, one could early

seek for medical help without developing a symp-

tomatic episode. This preventive point of view could

represent a major change in perspective in how medi-

cal care should be delivered in the future.

REFERENCES

Arrhythmia — National Heart, Lung, and Blood Institute

(NHLBI).

Acharya, R., Krishnan, S. M., Spaan, J. A., and Suri,

J. S. (2007). Advances in cardiac signal processing.

Springer.

Acharya, U. R., Fujita, H., Lih, O. S., Hagiwara, Y., Tan,

J. H., and Adam, M. (2017). Automated detection

of arrhythmias using different intervals of tachycar-

dia ECG segments with convolutional neural network.

Information Sciences, 405:81–90.

Ansari, S., Gryak, J., and Najarian, K. (2018). Noise de-

tection in electrocardiography signal for robust heart

rate variability analysis: A deep learning approach.

In 2018 40th Annual International Conference of the

IEEE Engineering in Medicine and Biology Society

(EMBC), pages 5632–5635. IEEE.

Brown, B. H. (cop. 1999). Medical physics and biomedi-

cal engineering. Medical Science Series. Institute of

Physics Publishing, Bristol.

Coiera, E. (2003). Guide to Health Informatics. Oxford

University Press, second edition.

Faust, O., Hagiwara, Y., Hong, T. J., Lih, O. S., and

Acharya, U. R. (2018). Deep learning for healthcare

applications based on physiological signals: A review.

Computer Methods and Programs in Biomedicine,

161:1 – 13.

Ghimes, A.-M., Avram, A.-M., and Vladuta, V.-A. (2018).

A character prediction approach in a security context

using a recurrent neural network. In 2018 Interna-

tional Symposium on Electronics and Telecommuni-

cations (ISETC), pages 1–4. 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.

Hannun, A. Y., Rajpurkar, P., Haghpanahi, M., Tison, G. H.,

Bourn, C., Turakhia, M. P., and Ng, A. Y. (2019).

Cardiologist-level arrhythmia detection and classiﬁ-

cation in ambulatory electrocardiograms using a deep

neural network. Nature Medicine, 25(1):65–69.

Iyengar, N., Peng, C., Morin, R., Goldberger, A. L., and

Lipsitz, L. A. (1996). Age-related alterations in the

fractal scaling of cardiac interbeat interval dynam-

ics. American Journal of Physiology-Regulatory, Inte-

grative and Comparative Physiology, 271(4):R1078–

R1084.

John, J. N., Galloway, C., and Valys, A. (2018). Deep con-

volutional neural networks for noise detection in ecgs.

arXiv preprint arXiv:1810.04122.

Johnson, K. W., Soto, J. T., Glicksberg, B. S., Shameer,

K., Miotto, R., Ali, M., Ashley, E., and Dudley, J. T.

(2018). Artiﬁcial intelligence in cardiology. Journal

of the American College of Cardiology, 71(23):2668 –

2679.

Mao, Z., Yao, W. X., and Huang, Y. (2017). Eeg-based

biometric identiﬁcation with deep learning. In 2017

8th International IEEE/EMBS Conference on Neural

Engineering (NER), pages 609–612.

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.

Moody, G. B., Muldrow, W., and Mark, R. G. (1984). A

noise stress test for arrhythmia detectors. Computers

in cardiology, 11(3):381–384.

Pascanu, R., Gulcehre, C., Cho, K., and Bengio, Y. (2013).

How to construct deep recurrent neural networks.

Rodrigues, J., Belo, D., and Gamboa, H. (2017). Noise de-

tection on ecg based on agglomerative clustering of

morphological features. Computers in Biology and

Medicine, 87:322 – 334.

S. Andersen, R., Peimankar, R., and Puthusserypady, S.

(2018). A deep learning approach for real-time de-

tection of atrial ﬁbrillation. Expert Systems with Ap-

plications, 115.

Satija, U., Ramkumar, B., and Manikandan, M. S.

(2018). Automated ecg noise detection and classiﬁ-

cation system for unsupervised healthcare monitoring.

IEEE Journal of Biomedical and Health Informatics,

22(3):722–732.

Silipo, R. and Marchesi, C. (1998). Artiﬁcial neural net-

works for automatic ecg analysis. IEEE Transactions

on Signal Processing, 46(5):1417–1425.

Xiong, F., Chen, D., Chen, Z., and Dai, S. (2019). Cancel-

lation of motion artifacts in ambulatory ECG signals

using TD-LMS adaptive ﬁltering techniques. Journal

of Visual Communication and Image Representation,

58:606–618.

Detection of Abnormalities in Electrocardiogram (ECG) using Deep Learning

243