Content uploaded by Dante Mújica-Vargas

Author content

All content in this area was uploaded by Dante Mújica-Vargas on Sep 19, 2022

Content may be subject to copyright.

Implementation of Time-Frequency

Moments for the Classiﬁcation of Atrial

Fibrillation Sequences Through

a Bidirectional Long-Short Term Memory

Network

Christian Garc´ıa-Aquino1, Dante M´ujica-Vargas1(B

),

Manuel Matuz-Cruz2, Nimrod Gonzalez-Franco1,

and Gabriel Gonz´alez-Serna1

1Tecnol´ogico Nacional de M´exico, CENIDET, Cuernavaca, Mexico

{m21ce012,dante.mv}@cenidet.tecnm.mx

2Tecnol´ogico Nacional de M´exico, Campus Tapachula, Tapachula, Mexico

Abstract. This article proposes a method to classify atrial ﬁbrillation

signals using time-frequency characteristics through a BiLSTM network.

The experiment was performed with the ECG signals, which are part of

the PhysioNet CinC 2017 database. In addition to the BiLSTM network,

machine learning algorithms such as k Nearest Neighbors, Linear SVM,

RBF SVM, Decision Tree, Random Forest, Neural Net, AdaBoost, Naive

Bayes and QDA were used for the classiﬁcation experiments. To measure

the eﬃciency and quality of the proposed method, the Accuracy, Preci-

sion, Recall, F1 Score metrics were used, as well as the Cohen Kappa

score and the Mathews correlation coeﬃcient. The results obtained show

a better classiﬁcation performance in the BiLSTM Network with 93.57%,

92.86%, 94.20%, 93.53%, 1.0 and 1.0 of the mentioned metrics.

Keywords: Atrial ﬁbrillation ·Feature extraction ·BiLSTM ·

Time-frequency ·ECG ·PhysioNet

1 Introduction

Abnormalities of the circulatory system are the most common cardiac disorders

addressed with Electrocardiography (ECG). The importance of these studies is

motivated by the prevalence, which is relatively high considering that around

3% of the world population suﬀers from some anomaly. Such is the case that

studies have been carried out focused on the classiﬁcation of Atrial Fibrillation

(AF), this being one of the most common types of arrhythmias [1].

In this sense, the use of deep learning has become popular for its classiﬁca-

tion, abstraction and, above all, learning capabilities that make it attractive in a

wide range of approaches that can hardly be solved with conventional comput-

ing equipment [2,3]. Consequently, works are emerging that emphasize proposing

c

The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. F. Mata-Rivera et al. (Eds.): WITCOM 2022, CCIS 1659, pp. 1–14, 2022.

https://doi.org/10.1007/978-3-031-18082-8_13

2C.Garc´ıa-Aquino et al.

increasingly complex neural network models, so that the computational cost to

train these models ends up increasing and ends up resulting in low performance.

These limitations can be resolved by focusing on the ﬁeld of digital signal pro-

cessing, which would help improve the performance of deep models and would

lead to the use of models with a relatively low computational cost for classiﬁca-

tion tasks.

In the literature, there are approaches to AF classiﬁcation using LSTM con-

volutional neural networks (CNN-LSTM) [4], deep residual-jump convolutional

neural networks [5], one-dimensional convolutional neural networks (1D-CNN)

[6], Hierarchical Attention Networks (HAN) [7], Multiscale Fusion Neural Net-

works (DMSFNet) [8] and Deep Neural Networks with attention mechanisms [9].

As for classical learning algorithms, there are approaches using AdaBoost [10]

and Support Vector Machines (SVM) [11].

Regarding the context of signal processing, there are works that focus on

extracting RR peaks from ECGs [12,13], the characterization of heart rate short-

age (HRV) [14], the use of the random process correlation function [15], ECG

language processing [16], the decomposition of the signal in multiple scales [17],

among other existing works in the literature just to mention those that are con-

sidered most relevant to the investigation.

As can be seen throughout this section, important approaches have been

generated that contribute to the ﬁeld of AF. However, the use of hybrid neural

networks may give the impression of oﬀering good classiﬁcation performance, but

by not employing signiﬁcant or discriminating features, the models are unable

to learn the dependencies of the ECG signals. In this sense, it would also entail

an enormous computational cost generated by the model. On the other hand, a

correct feature extraction method is not speciﬁed since the nature of ECG data

is time dependent and such property cannot be exploited in depth.

There is a previous work where the classiﬁcation of arrhythmias was car-

ried out by extracting artisanal features using the wavelet packet transform and

classifying them with classical machine learning algorithms [18]. Therefore, the

purpose of this article is to propose a new method to classify AF this time using

time-frequency characteristics. The main idea of this approach is to carry out a

robust and eﬃcient treatment of ECG signals by generating Spectrograms, which

will allow us to obtain time-frequency moments to improve the performance of

a Deep Hybrid Neural Network and reduce the computational cost.

The rest of this document is organized as follows. In Sect. 2, a brief introduc-

tion to the short-term fourier transform, as well as the classiﬁcation algorithms

used during the experimentation. The method of implementation for the classi-

ﬁcation of Atrial Fibrillation is indicated in Sect. 3. The results obtained from

the experiments and a comparative analysis are presented in Sect.4. Conclusions

are mentioned in the ﬁnal section and future work is described.

Classiﬁcation of Atrial Fibrilation 3

2 Background

2.1 Recurrent Neural Networks

Recurrent Neural Networks (RNN) are used to analyze data that changes over

time. Commonly used for speech recognition, image subtitles or character pre-

diction for source code generation [19]. To allow the network to store and access

input histories, recursive connections are introduced to predict the current time

step and transferring that prediction to the next step as an input. According

to Fig. 2, an RNN model has the same structure as the classic ANN models,

with an input layer, nhidden layers and an output layer, without forgetting the

parameter tcorresponding to time, being xt−1,xtand xt+1 are the inputs of the

RNN model at diﬀerent instants of time. (Fig. 1).

Fig. 1. Basic structure of an RNN [19].

In the most basic form of an RNN, its learning function is found in the Eq. 1

for the hidden layers and in the Eq.2for the output layer.

h[t]=f(Wh

x(x[t]+bh)+Wh

h(h[t−1] + bh)) (1)

y[t]=g(Wo

h(h[t]+bo)) (2)

2.2 Short Time Fourier Transform

The Short Time Fourier Transform (STFT) is responsible for analyzing non-

stationary signals through the Fourier Transform (TF). Where the STFT consists

of dividing the signal into small time segments in such a way that it can be

assumed that for each segment the signal is stationary, and thus calculate the

TF in each portion of the signal, which is taken as a window that slides along

the time axis, resulting in a two-dimensional representation of the signal [20].

Mathematically, it is written as:

STFT x(t)=X(τ,ω)=∞

−∞

x(t)w(t−τ)e−jωtdt (3)

where w(t) is the Hann or Gaussian hill window function initialized at 0 and x(t)

is the input signal to transform, X(τ,ω) is essentially the TF of x(t), w(t−τ)is

4C.Garc´ıa-Aquino et al.

a complex function that represents the phase and magnitude of the signal over

time and frequency. Concurrently the instantaneous phase is used in conjunction

with the time axis τand the frequency axis ωto suppress any resulting phase

jump discontinuities in the STFT. The time index τis normally considered a

“slow” time and is usually not expressed with as high a resolution as the time t.

3 Implementation

In order to classify arrhythmias, a methodology for classifying electrocardio-

graphic signals is analyzed in this section, starting with their acquisition and the

segmentation of the QRS complexes present. Likewise, the process of extraction

of statistical characteristics is analyzed, which are subjected to a dimension-

ality reduction. Finally, the classiﬁcation is performed with machine learning

algorithms. A description of the aforementioned can be seen in Fig 2.

Fig. 2. Proposed classiﬁcation method

3.1 Data Acquisition/Segmentation

For the experimentation of this research, the ECG data of the PhysioNet 2017

Challenge [21] were used, which consist of a set of ECG signals sampled 300Hz

and divided by a group of experts, which are classiﬁed as Normal (N), AFib (A),

Noisy signal (∼), and Other signal (O). The reason for using this data set is

because it mainly contains the two classes that interest us for the classiﬁcation

tasks, these being: Normal (N), AFib (A), added to the fact that the character-

istic of the ECG signals is that they are of a single derivation and the duration

in seconds of an individual signal is mostly around 30 s. A graphic representation

of the signals present in the ECG’s can be seen in Fig. 3.

Classiﬁcation of Atrial Fibrilation 5

(a) Normal Signal (b) Signal with AF

Fig. 3. Signals from the selected database.

3.2 Feature Extraction

The next step after acquiring the signals and segmenting them is the generation

of Spectrograms using the STFT, which will serve to extract speciﬁc character-

istics such as the Time-Frequency moments, which for the work in question was

considered the ﬁrst moment as the Frequency Instantaneous and the second as

Spectral Entropy.

Spectrograms. The spectrogram is a visual representation to identify varia-

tions in the frequency and intensity of a signal over a period of time. In this

sense, to generate the spectrogram of an ECG signal, the STFT (Eq. 3)isused

for continuous signals due to the nature of ECGs. To represent the series of spec-

tra located in time by the STFT, the spectrogram is expressed as the squared

series of the STFT of said spectra as |X(k, l)|2.

Therefore, for the construction of the spectrogram, the power of each spec-

trum is shown segment by segment, representing the magnitudes side by side as

an image with a magnitude-dependent color map. (Fig. 4).

Fig. 4. Representation of a spectrogram of a normal signal and with FA.

6C.Garc´ıa-Aquino et al.

Instantaneous Frequency. The ﬁrst moment that is extracted from a spec-

trogram is the Instantaneous Frequency (IF), which is considered as a variable

parameter in time that is related to the average of the frequencies present in the

signal as it evolves. From a probabilistic perspective [22] and [23] show that the

instantaneous frequency is the ﬁrst moment or mean of the spectral density of

the signal at an instant of time. Now, considering that a complex signal can be

expressed as ˆs(t)=s(t)+s∗(t), where s(t) is the magnitude of the signal and

s∗(t) its phase, a better representation can be made for this investigation, so the

complex signal representation is rewritten as:

ˆs(t)=A(t) + expθ(t)(4)

Inferring that if the real terms of a complex signal are known or can be

calculated, then the magnitude (time) A(t) and instantaneous phase θ(t)ofthe

signal can be found as:

A(t)=s2(t)+s∗2(t)=|S(t)|and (5)

θ(t) = arctan(s∗(t)

s(t)) (6)

The phase variation with time is related to the angular frequency of the signal

and is called the instantaneous angular frequency [24], Expressed another way,

the instantaneous angular frequency is the derivative with respect to the phase

time, expressed as: d

dtθ(t)=ω(t) (7)

In the case of this investigation, it is necessary to know the frequency, instead

of the angular frequency, so the relationship between frequency and angular

frequency f=ω/2πis used, with which the instantaneous frequency of a spec-

trogram using the expression [25]:

f(t)= 1

2π

d

dtθ(t) (8)

Spectral Entropy. The second moment that is extracted from a spectrogram

is the spectral entropy (SE), which is a measure of its spectral power distri-

bution. Therefore, Spectral Entropy treats the normalized power distribution

of the signal in the frequency domain as a probability distribution and calcu-

lates the uncertainty or entropy of the spectrum [26]. Therefore, knowing the

time-frequency power spectrogram, the Spectral Entropy in time is obtained by:

SE(m)=−

w

k=1

pm

klog2pm

k(9)

where again mis the window index, wis the duration of the windows and pm

k

is the Probability Mass Function (PMF) that corresponds to the m-th window,

Classiﬁcation of Atrial Fibrilation 7

obtained from the power spectral density or also known as the spectrogram

(Sm(ωk))

pm

k=Sm(ωk)

w

j=1 Sm(ωj)(10)

In this way, the function pm

kis in charge of verifying the necessary condition of

the FMPs: kpm

k= 1. With the normalization carried out in (4.15), a measure

is obtained that is independent of the total power of the signal, and therefore

the SE is independent of the amplitude.

3.3 Training/Classiﬁcation

The network with which we worked in this research is given by a Deep Learning

Neural Network model, speciﬁcally, a Network with Bidirectional Short and Long

Term Memory (BiLSTM). The architecture is initially formed with a sequence

input layer, a bidirectional LSTM layer (BiLSTM) proposed by [27] formed by

two LSTM hidden layers [28] to learn the bidirectional long-term dependencies

between time steps, a third fully connected layer to obtain the probabilities of

belonging to the diﬀerent classes, a Softmax layer to represent the probabilities

through a categorical distribution, and ﬁnally a classiﬁcation layer. The archi-

tecture diagram can be seen in Fig. 5.

Fig. 5. Bidirectional LSTM hybrid neural network.

The model has two inputs x=(x1,x

2, ..., xt) according to the two extracted

features, where tis the length of the input signal to through time. These serve

8C.Garc´ıa-Aquino et al.

as input for the BiLSTM layer, which consists of 2 hidden LSTM layers with 100

neurons each. Since a single LSTM layer can only memorize past tenses which

makes it impossible to memorize future tenses, to overcome this characteristic

deﬁciency, in they proposed Bidirectional Recurrent Neural Networks (BiRNN)

to be able to combine two hidden LSTM layers separated in directions opposite

but always pointing to the same exit. The internal structure of a BiLSTM layer

can be seen in Fig. 6.

Fig. 6. Internal structure of a BiLSTM layer.

With this structure, the BiLSTM layer computes the sequence of inputs

x=(x1,x

2, ..., xn) from the opposite direction to a hidden forward sequence

−→

h=(

−→

h1,−→

h2, ..., −→

hn) and a backward hidden sequence ←−

h=(

←−

h1,←−

h2, ..., ←−

hn). The

encoded vector of outputs htis generated from the concatenation of the ﬁnal

forward and backward outputs, ht=[

−→

ht,←−

ht]. Expressing the aforementioned by

means of the following expressions.

−→

ht=σ(Wx−→

hxt+W−→

h−→

h

−→

ht−1+b−→

h) (11)

←−

ht=σ(Wx←−

hxt+W←−

h←−

h

←−

ht+1 +b←−

h) (12)

ht=Wy−→

h

−→

ht+Wy←−

h

←−

ht+bh(13)

where h=(h1,h

2, ..., ht, ..., hn) is the output sequence of the BiLSTM layer. The

output of the BiLSTM layer serves as input data for a fully connected layer with

100 perceptron-like neurons with a sigmoidal activation function σ(·) to gener-

ate outputs bounded at [0,1], this activation function is the one that is regularly

used, however, other types of functions can be used depending on the criteria

of each researcher. The output of the fully connected layer is passed through a

softmax function to transform (normalize) these outputs to a probability distri-

bution representation such that the sum of all probabilities of the outputs is 1

Classiﬁcation of Atrial Fibrilation 9

[31]. Deﬁned in the following expression.

f(yn)= exp(yn)

K

k=1 exp(yn)(14)

Finally, there is a classiﬁcation layer to determine if the output obtained

belongs to the class of signals with Atrial Fibrillation or signals with normal

beats.

4 Experimentation and Results

Computational performance and cost for FA classiﬁcation are quantiﬁed and

compared using the CinC 2017 database obtained from PhysioNet. The exper-

imentation was carried out in a computer equipment that consists of an Intel

(R) Core (TM) i7-10870H CPU @ 2.20 GHz with 8 cores and 16 GB of RAM;

as well as an NVIDIA RTX 3060 GPU, with 3840 CUDA cores and 6 GB of

dedicated VRAM. The implementation was developed in MATLAB and other

experimentations with classic machine learning algorithms were carried out in

Python using the Scikit-learn library. To know the behavior of the proposed

method in the proposed Hybrid Architecture compared to other machine learn-

ing approaches, algorithms such as Nearest Neighbors, Linear SVM, RBF SVM,

Decision Tree, Random Forest, Neural Net, AdaBoost, Na¨ıve Bayes and QDA.

4.1 Metrics

In order to carry out the objective evaluation of the proposed method for the

classiﬁcation of cardiac arrhythmias, 3 aspects were considered, which are men-

tioned below:

The confusion matrix was considered to evaluate the performance of a classi-

ﬁcation model, the weighting of correct and incorrect predictions are summarized

with the count values and separated by class. This set of predictions are inter-

preted through metrics derived from the confusion matrix such as: Accuracy,

Precision, Recall and F1-Score that are detailed in [32]. The Cohen’s Kappa

Score (KCS) is used to compare the observed agreement in a data set with

respect to the expected agreement as mentioned in [33]. The Mathews Cor-

relation Coeﬃcient (MCC) is a contingency matrix method used to calculate

Pearson’s product-moment correlation coeﬃcient between actual and predicted

values, as discussed in [34].

4.2 Results

Taking into account the signals of the data set, the division was made in a data

set for training, designating 90% and the rest in a test set. Both the training set

and the test set were augmented in order to normalize the amount of data in both

classes, and as can be seen in Table4, 4438 training instances were obtained, as

well as 490 test instances. (Table 1).

10 C. Garc´ıa-Aquino et al.

Table 1. Split signals for training and testing.

Label Data train Data test

Normal 4438 490

Atrial ﬁbrillation 4438 490

A ﬁrst experimentation was carried out using the raw time series of the

database in order to observe the classiﬁcation behavior of the algorithms when

using unprocessed signals.

Table 2. Performance results with raw data.

Algorithm/Arquitecture Accuracy Precision Recall F1-Score CKS MCC

Nearest Neighbors 0.5051 0.5149 0.5051 0.4079 0.0 0.0

Linear SVM 0.4949 0.4934 0.4949 0.4646 0.0 0.0

RBF SVM 0.5 0.25 0.5 0.3333 0.0 0.0

Decision Tree 0.5582 0.5734 0.5582 0.534 0.0 0.0

Random Forest 0.5214 0.5228 0.5214 0.5143 0.0 0.0

Neural Net 0.5204 0.5529 0.5204 0.4334 0.0 0.0

AdaBoost 0.5265 0.53 0.5265 0.5125 0.0 0.0

Naive Bayes 0.5102 0.5112 0.5102 0.4994 0.0 0.0

QDA 0.5 0.25 0.5 0.3333 0.0 0.0

Proposed 0.5786 0.7000 0.5632 0.6242 0.0 0.0

The average quantitative summary for each method and metric considered

in the experimentation is presented in Table 2. From the results obtained, it is

noteworthy to observe that the hybrid model used had the best classiﬁcation

performance despite being raw data, in contrast to classical machine learning

algorithms where it can be seen that in some metrics I have a performance

below 0.30, which shows that the algorithms did not reach a convergence in

training and according to the literature, the value for an implementation to be

acceptable it must be at least 0.80.

Now, for the second experimentation, the time-frequency characteristics pro-

posed in this research work were used to observe the behavior and demonstrate

an increase in the classiﬁcation performance of the algorithms.

Classiﬁcation of Atrial Fibrilation 11

Table 3. Performance results of the proposed classiﬁcation method.

Algorithm/Arquitecture Accuracy Precision Recall F1-Score CKS MCC

Nearest Neighbors 0.8199 0.8199 0.8199 0.8199 1.0 1.0

Linear SVM 0.574 0.5794 0.574 0.5665 0.0 0.0

RBF SVM 0.5 0.25 0.5 0.3333 0.0 0.0

Decision Tree 0.6357 0.636 0.6357 0.6355 0.0 0.0

Random Forest 0.6168 0.6236 0.6168 0.6115 0.0 0.0

Neural Net 0.6724 0.757 0.6724 0.6431 0.0 0.0

AdaBoost 0.6393 0.6425 0.6393 0.6372 0.0 0.0

Naive Bayes 0.6061 0.6093 0.6061 0.6033 0.0 0.0

QDA 0.5378 0.7067 0.5378 0.419 0.0 0.0

Proposed 0.9357 0.9286 0.9420 0.9353 1.0 1.0

The average quantitative summary for each method and metric considered

in the experimentation is presented in Table 3. From the results obtained, it

is noteworthy to observe that the hybrid model used once again had the best

classiﬁcation performance, surpassed even in all the metrics used the 0.90 in clas-

siﬁcation performance, in contrast to the classical machine learning algorithms

where it can be seen that most of them had a performance below 0.80, due to

the fact that the extracted features are in the time and frequency domain, are

time dependent for which these classiﬁcation algorithms were not designed. In

addition to the above, it can also be veriﬁed that classical algorithms have a

particular problem known as the performance plateau, which consists in that

the greater the data load to train and evaluate said algorithms, the performance

is truncated.

As a ﬁnal part of the experimentations and another of the objectives of this

research, the computational cost of each algorithm was calculated at the time of

classifying the set of tests.

Table 4. Computational cost of classiﬁcation.

Algorithm/Arquitecture With raw data (in

seconds)

With the proposed

method (in seconds)

Nearest Neighbors 9.96 1.95

Linear SVM 13.99 5.27

RBF SVM 64.28 8.14

Decision Tree 0.08 0.008

Random Forest 0.08 0.009

Neural Net 0.39 0.02

AdaBoost 1.17 0.07

Naive Bayes 0.18 0.01

QDA 9.73 0.06

Proposed 6.41 0.35

12 C. Garc´ıa-Aquino et al.

As can be seen in Table4, the computational cost of the algorithms and

architectures considered was measured. The experimentation was performed by

classifying the raw test dataset and the test dataset with the applied method.

According to the results obtained, it was observed that the proposed architec-

ture had a lower computational cost compared to classical machine learning

algorithms. Although it is true that there were algorithms that seemed to have a

lower computational cost, such as the case of the Decision Tree, however, accord-

ing to the poor classiﬁcation performance that can be seen in the previous table,

the malfunction of said algorithm when classifying AF is evident.

5 Conclusions

In this research work, a method for the classiﬁcation of Atrial Fibrillation and Nor-

mal Beats was proposed. The results obtained showed that the correct treatment

of the signals, speciﬁcally, the use of time-frequency characteristics, improves the

training process of the classiﬁcation algorithms used during the investigation, also

suggesting that the proposed Hybrid Neural Architecture obtained the best per-

formance. For the classiﬁcation of arrhythmias, since although algorithms such

as k -Nearest Neighbors, the Decision Trees, despite being characterized as mul-

ticlass classiﬁcation algorithms, do not manage to overcome the performance of

the Hybrid Neural Architecture proposed when analyzing ECG signals, due to the

aforementioned performance plateau that reduces their performance, which, apart

from using other types of features, another of the objectives with respect to the

previous work was to use deep learning algorithms. As future work, the implemen-

tation of the proposed method in low consumption embedded cards will be carried

out in order to conceive a portable and remote system, in addition to improving

processing times, in such a way that the system can work as close as possible to

real time.

Acknowledgments. This work was supported by the Tecnol´ogico Nacional de

M´exico/CENIDET trough the project entitled “Clasicador para detectar brilaci´on

auricular en se˜nales electrocardiogr´acas utilizando una red recurrente profunda entre-

nada con momentos de tiempo-frecuencia”, as well as by CONACYT.

References

1. Miyasaka, Y., Barnes, M.E., Gersh, B.J., Cha, S.S., Bailey, K.R., Abhayaratna,

W.P., et al.: Secular trends in incidence of atrial ﬁbrillation in Olmsted County,

Minnesota, 1980 to 2000, and implications on the projections for future prevalence.

Circulation 114, 119–25 (2006)

2. Bartlett, P.L.: The sample complexity of pattern classiﬁcation with neural net-

works: the size of the weights is more important than the size of the network.

IEEE Trans. Inf. Theory 44(2), 525–536 (1998)

Classiﬁcation of Atrial Fibrilation 13

3. Akesson, B.M., Toivonen, H.T.: A neural network model predictive controller. J.

Process Control 16(9), 937–946 (2006)

4. Wu, X., Sui, Z., Chu, C.-H., Huang, G.: Detection of atrial ﬁbrillation from short

ECG signals using a hybrid deep learning model. In: Chen, H., Zeng, D., Yan, X.,

Xing, C. (eds.) ICSH 2019. LNCS, vol. 11924, pp. 269–282. Springer, Cham (2019).

https://doi.org/10.1007/978-3-030- 34482-5 24

5. Sanjana, K., Sowmya, V., Gopalakrishnan, E.A., Soman, K.P.: Performance

improvement of deep residual skip convolution neural network for atrial ﬁbrilla-

tion classiﬁcation. In: Bhateja, V., Peng, S.L., Satapathy, S.C., Zhang, Y.D. (eds.)

Evolution in Computational Intelligence. AISC, vol. 1176, pp. 755–763. Springer,

Singapore (2021). https://doi.org/10.1007/978-981-15-5788-0 71

6. Tutuko, B., Nurmaini, S., Tondas, A.E., et al.: AFibNet: an implementation of

atrial ﬁbrillation detection with convolutional neural network. BMC Med. Inform.

Decis. Mak. 21, 216 (2021). https://doi.org/10.1186/s12911-021-01571-1

7. Mousavi, S., Afghah, F., Acharya, U.R.: Han-ECG: an interpretable atrial ﬁbril-

lation detection model using hierarchical attention networks. Comput. Biol. Med.

127, 104057 (2020)

8. Wang, R., Fan, J., Li, Y.: Deep multi-scale fusion neural network for multi-class

arrhythmia detection. IEEE J. Biomed. Health Inform. 24(9), 2461–2472 (2020)

9. Kuvaev, A., Khudorozhkov, R.: An attention-based CNN for ECG classiﬁcation.

In: Arai, K., Kapoor, S. (eds.) CVC 2019. AISC, vol. 943, pp. 671–677. Springer,

Cham (2020). https://doi.org/10.1007/978-3-030- 17795-9 49

10. Wu, C., Hwang, M., Huang, T.H., et al.: Application of artiﬁcial intelligence ensem-

ble learning model in early prediction of atrial ﬁbrillation. BMC Bioinform. 22,93

(2021). https://doi.org/10.1186/s12859-021-04000- 2

11. Kostka, P., Tkacz, E.: Support vector machine classiﬁer with feature extraction

stage as an eﬃcient tool for atrial ﬁbrillation detection improvement. In: Kurzynski,

M., Puchala, E., Wozniak, M., Zolnierek, A. (eds.) Computer Recognition Systems

2. ASC, vol. 45, pp. 356–363. Springer, Heidelberg (2007). https://doi.org/10.1007/

978-3-540-75175-5 45

12. Kisohara, M., Masuda, Y., Yuda, E., et al.: Optimal length of R-R interval seg-

ment window for Lorenz plot detection of paroxysmal atrial ﬁbrillation by machine

learning. Biomed. Eng. Online 19, 49 (2020). https://doi.org/10.1186/s12938-020-

00795-y

13. Hickey, B., Heneghan, C., De Chazal, P.: Non-episode-dependent assessment of

paroxysmal atrial ﬁbrillation through measurement of RR interval dynamics and

atrial premature contractions. Ann. Biomed. Eng. 32, 677–687 (2004). https://doi.

org/10.1023/B:ABME.0000030233.39769.a4

14. Mandal, S., Sinha, N.: Prediction of atrial ﬁbrillation based on nonlinear modeling

of heart rate variability signal and SVM classiﬁer. Res. Biomed. Eng. 37, 725–736

(2021). https://doi.org/10.1007/s42600-021-00175- y

15. Wang, J., Wang, P., Wang, S.: Automated detection of atrial ﬁbrillation in ECG

signals based on wavelet packet transform and correlation function of random pro-

cess. Biomed. Sig. Process. Control 55, 101662 (2020)

16. Mousavi, S., Afghah, F., Khadem, F., Acharya, U.R.: ECG language processing

(ELP): a new technique to analyze ECG signals. Comput. Methods Programs

Biomed. 202, 105959 (2021)

17. Cao, X.-C., Yao, B., Chen, B.-Q.: Atrial ﬁbrillation detection using an improved

multiscale decomposition enhanced residual convolutional neural network. IEEE

Access 7, 89152–89161 (2019)

14 C. Garc´ıa-Aquino et al.

18. Garc´ıa-Aquino, C., M´ujica-Vargas, D., Matuz-Cruz, M.: Classiﬁcation of cardiac

arrhythmias using machine learning algorithms. In: Mata-Rivera, M.F., Zagal-

Flores, R. (eds.) WITCOM 2021. CCIS, vol. 1430, pp. 174–185. Springer, Cham

(2021). https://doi.org/10.1007/978-3-030- 89586-0 14

19. Vogt, N.: CNNs, LSTMs, and attention networks for pathology detection in medical

data (2019). arXiv preprint arXiv:1912.00852

20. Smith, J. O.: Mathematics of the Discrete Fourier Transform (DFT). W3K Pub-

lishing (2007). www.w3k.org/books/

21. Goldberger, A.L., et al.: PhysioBank, PhysioToolkit, and PhysioNet. Circulation

101(23) (2000)

22. Flandrin, P.: Time-frequency/time-scale analysis (1998)

23. Barnes, A. E.: Instantaneous spectral bandwidth and dominant frequency with

applications to seismic reﬂection data (1993)

24. Taner, M.T., Koehler, F., Sheriﬀ, R.: Complex seismic trace analysis. Geophysics

44(6), 1041–1063 (1979)

25. Boashash, B.: Estimating and interpreting the instantaneous frequency of a signal.

i. fundamentals. Proc. IEEE 80(4), 520–538 (1992)

26. Pan, Y.N., Chen, J., Li, X.L.: Spectral entropy: a complementary index for rolling

element bearing performance degradation assessment. Proc. Inst. Mech. Eng. C J.

Mech. Eng. Sci. 223(5), 1223–1231 (2008)

27. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Trans.

Sig. Process. 45(11), 2673–2681 (1997)

28. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8),

1735–1780 (1997)

29. Elman, J.: Language as a dynamical system (2001)

30. Pattanayak, S.: Pro Deep Learning with TensorFlow. Apress, New York (2017)

31. Svens´en, M., Bishop, C. M.: Pattern recognition and machine learning (2007)

32. Chen, H., et al.: GasHis-transformer: a multi-scale Visual Transformer approach

for gastric histopathology image classiﬁcation (2021). arXiv:2104.14528

33. Samiuc. www.samiuc.es/estadisticas-variables-binarias/medidas-de-concordancia/

kappa-de-cohen/. Accessed 05 July 2021

34. Chicco, D., Jurman, G.: The advantages of the Matthews correlation coeﬃcient

(MCC) over F1 score and accuracy in binary classiﬁcation evaluation. BMC

Genom. 21(1), 6 (2020)