Utilizing Autoencoder Based Frameworks for Detecting

Anomalous ECG Signal

Nisha Kumari Singh1, Soumya Bhattacharyya2 , Shambhu Nath Saha3, Sourav
Saha4, Rahul Das Gupta5
2,3 Department of Information Technology, Narula Institute of Technology, Kolkata

1,4Department of Computer Science & Engineering, Narula Institute of Technology, Kolkata

5Department of Computer Science & Engineering, SNU, Kolkata

soumya.bhattacharyya@nit.ac.in

Abstract The electrocardiogram signal (ECG) is the basis of cardiac diagnosis

due to its ability to detect the electrical pulse of the heart. Detecting
abnormalities in ECG is important for diagnosing heart disease. Traditional
ECG analysis relies on interpretation by trained medical professionals which is
time consuming and monotonous at times. Detecting anomalies poses a
challenging task to the researchers while they attempt software-based
techniques to develop an automated framework, regardless of general
understanding of the anomalous patterns. Medical anomaly identification in
time series signals, such as electrocardiograms (ECG), faces similar challenges
as other sectors of application. A significant issue arises from the inconsistent
behaviour of different anomalous ECG signals, as well as the lack of a
universally accepted criteria for formulation of abnormality. The emergence of
deep learning has led to the potential use of software-based diagnostic tools for
ECG abnormality detection. In particular, a popular deep learning architecture
known as the autoencoder model has recently become a useful tool to identify
anomalous patterns in time series data. Such autoencoder model-based
frameworks, while developed using ECG time series data, are also expected to
detect anomalies in ECG. This article explores the application of autoencoder
models by developing an ECG-anomaly detection framework to predict the
possibility of heart disease. The promising outcome of our proposed framework
in terms of detecting anomaly in ECG with around 95% accuracy rate while
experimented with real ECG dataset highlights the potential use of autoencoder
models for automating clinical diagnosis for cardiac care.
Keywords: Electrocardiogram, Anomaly Detection, Abnormal Heart Rate, Autoencoder

1 Introduction

An electrocardiogram signal (ECG) based technique is a frequently used

clinical diagnostic method that analyzes the recorded electrical impulses of
physiological activities of the heart over time. The cells within the cardiac
conduction system have the ability to produce electrical impulses and
subsequently transmit the signal across our heart. Anomalous electrical
impulses recorded as ECG signals can indicate many abnormal heart
conditions, including arrhythmias, myocardial infarction, and heart attack.
Early detection of these abnormalities on analyzing ECG signal is critical for
prompt intervention and effective treatment.
 A typical ECG signal consists of five distinct morphological waveforms,
namely the P, Q, R, S, and T waves [1]. The PQRST waveform is generated
by the depolarization and repolarization activities that take place in each
cardiac cycle, corresponding to the working and resting phases, respectively.
Figure 1 illustrates the waveform. The P wave is the result of the atrial
depolarization wave, which is initiated by the sinoatrial node (SA), the
heart's intrinsic pacemaker responsible for generating electrical impulses.
The QRS complex is formed by the sequential occurrence of the Q wave, R
wave, and S wave. The atrioventricular node (AV) is responsible for
generating these depolarization waves. This complex signifies the
depolarization of the ventricles. The T wave in the cardiac cycle represents
ventricular repolarization. A cardiac cycle consists of the depolarization and
repolarization that occur as an electrical signal travel from the atrium to the
ventricle. The morphology of ECG reports can vary based on the heart's
vector, typically observed across distinct leads, such as the 12 leads
commonly used. A typical ECG report displays a regular cardiac rhythm, but
an atypical ECG lacks a continuous rhythmic pattern. The abnormality in the
electrocardiogram (ECG) indicates myocardial injury or disturbances in the
heart's rhythm. Figure 1 displays the ECG wave and a single interval of a
normal Heart Cycle.

Figure. 1. Heart Cycle

in a typical ECG signal

Anomaly detection involves the identification of instances that fall outside the normal
distribution. In essence, it aims to identify instances that deviate from the overall
trend observed in the dataset (Fig. 2). The detection of anomaly is of utmost
importance as aberrant observations are closely linked to various issues or flaws, such
as structural deficiencies, system or malware invasions, manufacturing failures,
financial fraud, or health complications. Regardless of the specification of anomalous
patterns, detecting anomalies poses a difficult challenge while designing an automated
framework using any software-based strategy. An important problem stems from the
variable conduct exhibited by various anomalies, as well as the absence of a standard
definition for what qualifies as an abnormality. Medical anomaly detection in time
series signals such as ECG has the same obstacles as other application sectors.
The advent of deep learning has given rise to the possibility of utilising software-
based diagnostic tools for the detection of abnormalities in electrocardiogram (ECG)
readings [2]. The autoencoder model, which is a widely adopted deep learning
architecture, has emerged as a valuable tool for detecting anomalous patterns in time
series data. The utilisation of autoencoder model-based frameworks, which have been
specifically designed and developed using electrocardiogram (ECG) time series data,
holds great promise in the field of anomaly detection in ECG signals [12]. This study
investigates the utilisation of autoencoder models for the purpose of predicting the
likelihood of heart disease through the analysis of electrocardiogram (ECG) signals.
The potential use of autoencoders for anomaly detection in ECG has shown promising
outcomes, which suggests that these models have the ability to automate clinical
diagnosis and enhance accessibility to cardiac care.

Figure 2: Diagram of normal and abnormal ECG

The rest of the article begins with Section 2 which highlights the related research
works carried out so far for developing automated tools in respect of ECG signal
analysis. Section 3 describes the proposed methodology with a detailed illustration of
the employed autoencoder model. Section 4 reports the results of our experimentation
with an analysis of the proposed framework’s performance. Finally, some conclusive
remarks are made in Section 5 indicating the probable future work as an extension of
the proposed work.

2 Related Works

Research on ECG signal analysis generally focuses on the identification of

different heart disorders, including atrial fibrillation, tachycardia,
bradycardia, arrhythmia, and other related concerns [1, 2]. Additional
research endeavours to predict mortality rates or demographic characteristics
[3, 4]. The customary input for these models often considers the raw
electrocardiogram (ECG) data of a single lead or multiple leads. However, in
other instances, only electrocardiogram (ECG) images were utilized for
analysis [5]. The ECG data could be converted into useful features such as
time-domain features, nonlinear-domain features, distance-based features,
and time-series features [6]. Alternatively, a deep learning model could
process the data automatically to create a high-dimensional representation.
Presently, widely used artificial intelligence (AI) techniques for diagnosing
ECG abnormalities involve the utilisation of machine learning algorithms
(namely feature extraction-based classifiers) and deep neural networks.
These methods primarily rely on classification models to identify aberrant
ECG patterns. Machine learning techniques such as self-organizing map
(SOM) and C-means clustering have also proven to be effective to some
extent in identifying aberrant ECG signals as outliers [7]. The features
extracted from the ECG signal are usually consisting of wavelet coefficients
and autoregressive coefficients. The Euclidean distance-based ECG signal
analysis approach was proposed by Alodia Yusuf and Hidayat [8] for ECG
classification. The suggested approach employs the MFCC (Mel Frequency
Cepstrum Coefficient) and Discrete Wavelet transformation, along with the
KNN algorithm, to extract and classify features while Huang et al. [9]
utilised Short-Time Fourier Transform (STFT) to classify ECG Arrhythmia
in their investigation.
Recently, the literature on anomaly detection in sequences (e.g., ECG) has
been utilising the successes achieved with deep learning models, primarily in
the context of supervised as well as semi-supervised learning [12]. In this
context, ANN models serve as a foundation for subsequent research efforts
for ECG signal classification, including anomaly detection.
The supervised learning-based ECG signal classification algorithms outlined
above have demonstrated excellent performance in earlier research.
However, these classification frameworks necessitate the inclusion of a
comprehensive dataset containing all categories of heart disease data,
accompanied by precise manual annotations by expert physicians. The
clinical electrocardiogram (ECG) data usually exhibit an imbalance, with a
scarcity of aberrant ECG samples. This poses a challenge in developing a
robust classification model to detect abnormal ECG patterns. Furthermore, it
is challenging to compile an extensive dataset encompassing all categories of
abnormal electrocardiograms (ECGs) for practical clinical applications.
Although DNN classifiers have mainly been developed using supervised
methods [6, 7, 8], there is currently ongoing research on one-class
classification ECG anomaly detection [9, 10, 11]. This includes training a
classifier using data from only one class [12]. It is commonly employed in
situations where there is a significant disparity in class distribution. This is a
common phenomenon observed in ECG datasets, where there is a large
number of normal ECG examples and a significantly smaller number of
various types of problematic ECG examples. A first endeavour to identify
abnormalities in ECG utilised One-Class Support Vector Machines to
analyse the features of a typical ECG. However, this approach did not
demonstrate satisfactory accuracy.
The autoencoder model, a widely recognised deep learning architecture, has
emerged as a valuable tool for detecting abnormal patterns in time series data
[10]. A framework utilising an autoencoder model can be constructed only
 with normal ECG time series data that can accurately reconstruct the input
data under a specified error threshold. Essentially, a well-trained autoencoder
will generate a significant amount of error when trying to reconstruct an
abnormal ECG signal, which may help us in detecting the presence of any
unusual patterns in the signal. This article examines the efficacy of
autoencoder models in detecting the anomalous electrocardiogram (ECG)
signals.

3 Methodology

3.1 Proposed Works

The proposed framework consists of three modules: 1) an autoencoder

module, 2) a discriminator module, and 3) an anomaly detector module, as
shown in Figure 3. The autoencoder is exclusively trained with normal ECG
signals with the objective of reconstructing every input ECG signal as its
output with minimal error. The discriminator module uses the original
dataset and the corresponding reconstructed dataset generated by the trained
autoencoder model to examine the distribution of reconstruction errors and
based on the error distribution; it determines an error threshold limit for the
reconstructed ECG signals.
During the deployment phase, the ECG signal is fed to the autoencoder
module, which tries to reconstruct the signal as its output. The error between
the input signal and the reconstructed signal is computed as an anomaly
score, which is subsequently fed to the anomaly detector module. Finally, the
anomaly detector module compares the anomaly score with the threshold
value determined by the discriminator module, and if it finds the score larger
than the threshold value, then the input ECG signal is detected as anomalous.

Figure 3. Proposed Autoencoder Based ECG Anomaly Detection Framework

Autoencoder Module

The autoencoder model consists of three parts: encoder block, hidden layer, and
decoder block as illustrated in Fig. 4. It is trained exclusively using normal ECG data.
Initially, the input data X is compressed and encoded into hidden layer, which is then
used to generate the reconstructed ECG signal X′. During the training phase, the loss
function is formulated based on the reconstruction error. This error quantifies the
discrepancy between the input data X and the corresponding output data X'. The
activation functions employed in the autoencoder module of neural networks are

presented below:
Here, the symbols δ and δ' are used to represent non-linear activation functions while
the symbols W, b, W', and b' denote the weights and biases that are utilized in linear
transformations. The minimization of the loss function, as represented by equation 3
is aimed by adjusting the model parameters iteratively in both the encoder and
decoder of the autoencoder model.

Figure 4. Architecture of the employed Autoencoder Model

Discriminator Module

The discriminator (D) probes the original dataset and the corresponding reconstructed
dataset generated by the trained autoencoder model to analyze the distribution of
reconstruction errors and based on the error distribution; it determines an error
threshold limit for the reconstructed ECG signals. Once training of the autoencoder
model is accomplished, the distribution frequency of reconstruction errors can be
analyzed by the discriminator module for error threshold determination. The threshold
value of reconstruction error “T” is determined appropriately so that the reconstructed
signal with error values below T are classified as normal ECG signals and above
which are classified as abnormal ECG signals, as shown in Fig. 5. The reconstruction
error represents the difference between the original test signal and their reconstructed
counterpart generated by the autoencoder module.

Fig. 5. Reconstruction Error threshold (T=0.065) determination

Anomaly Detector Module

During the deployment phase, the ECG signal is fed into the autoencoder module,
which attempts to reconstruct the signal as its output. An anomaly score is calculated
using equation (1) as the discrepancy between the input signal and the reconstructed
signal (Fig. 6). This score is then passed to the anomaly detection module. Ultimately,
the anomaly detector module assesses the anomaly score against the threshold value
(T) established by the discriminator module. If the score exceeds the threshold value
by a predefined margin, the input ECG signal is identified as anomalous.

n
1
Anomaly Score =¿ T −
n ∑ (Xinput data [i]−Reconstructed data [i])2∨(5)
i=1

Where:

 n is the number of samples in the test signal

 Xinput_data[i] represents the i-th sample in the test signal
 Reconstructed_data[i] represents the reconstructed output of the i-th
sample of the output signal from the autoencoder model
 Fig. 6. Discrepancy between the input signal and the reconstructed signal

4. Experimental Results and Discussion

The effectiveness of our framework is validated using data from the MIT-BIH and
CMUH datasets. We use one standard deviation above the mean as the anomaly score
in the training phase to determine the threshold. The T values for the MIT-BIH and
CMUH datasets are calculated to be 0.025 and 0.01, respectively. The training set
considers normal ECG signals only while the abnormal scores are kept within the
range of the threshold.

MIT-BIH Dataset
In this research, we analyzed and used 46 Lead II signal data from the MIT-BIH
Dataset. Of note, files 112, 114, and 124 were removed because they did not contain
lead II data or Heart related data for our study. To improve the quality of the data, we
use wavelet transform to reduce noise and base shift following the procedure
described by [5]. Each beat has a total of 260 sample points, including 100 points and
150 points before the R peak. R points after the peak.

CMUH dataset

The CMUH dataset provided ECG signals from 44,173 individuals for our proposed
study. The data was resampled at 360 Hz to match the MITBIH dataset. Beat
segmentation was performed using the same method as previously described. For data
set partitioning, 20,000 normal ECG sample points were selected as the training set.
The test set consisted of 10,000 normal ECG sample points and 10,000 abnormal
ECG sample points, as detailed in the following statistics.

Dataset Type Type of Number Numb Sampl Numb Numb

heartbeats of er of e size er er
heartbea cases of of test
ts traini set
MIT- Normal N 74,962 40 15,000 ng
10,000 5,000
BIH Abnormal A 2,545 — 5,000 0 5,000
L 8,068 —
R 7,254 —
V 7,034 —
Total 99,863 47 20,000 10,000 10,000
CMUH Normal N 20,000 20,000 15,000 10,000 5,000
Abnormal A 6,811 6,811 5,000 0 5,000
L 1,247 1,247
R 8,268 8,268
 V 7,847 7,847
Total 44,173 44,173 20,000 10,000 10,000
TABLE 1

Number of heartbeats involved in each dataset and the division of datasets.

Evaluation Metrics

The evaluation metrics such as accuracy (Acc), precision (Pre), recall (Rec), F1 score
(F1), and AUC (area under the ROC curve were used to evaluate the performance of
our model compared to other models. In the confusion matrix, abnormal ECGs are
designated as positive and normal ECGs as negative.

In the clinical setting, accuracy refers to the correct identification of patients with true
ECG abnormalities, while recall indicates the correct diagnosis of these patients. High
detection accuracy reduces misdiagnosis, and a model with high recall decreases
missed diagnoses. In this study, the F1 score (the weighted ratio of recall and
precision) and the AUC value were used as the primary indicators to evaluate the
effectiveness of the detection method.

Results

Figure 7 presents examples of normal and abnormal ECG signal along with
reconstructed signals produced by the proposed autoencoder model. For normal ECG
signal, the reconstructed output closely resembles the input waveform, with an
average error of 0.0063. In contrast, for abnormal ECG signal, there is a noticeable
difference between the input waveform and the reconstructed waveform as the error
increases to 0.0289.
 (A) (B)

Figure 7. Reconstruction of ECG Signal: (A) Normal and (B) Anomalous

Figure 8 displays the confusion matrix illustrating the detection results. In the MIT-
BIH dataset, our model correctly identified 4,930 abnormal ECGs but misclassified
257 normal ECGs as abnormal. Similarly, in the CMUH dataset, 4,917 abnormal
ECG recordings were correctly detected, while 559 normal ECGs were erroneously
classified as abnormal.

Figure 8: Confusion matrix of ECG dataset (A) MIT-BIH (B) CMUH (1 = Normal 0 =
Anomalous)
Regarding the MIT-BIH dataset, it is worth mentioning that the proposed model
achieved notable performance metrics. Specifically, the accuracy, precision, recall, F1
score, and AUC values were recorded as 96%, 98%, 94%, and 96%, respectively, as
indicated in Table 3. Regarding the CMUH dataset, it is worth noting that our model
has achieved remarkable scores in terms of accuracy, F1-score, and AUC.
Specifically, our model has achieved leading scores of 93% in accuracy, 93% in F1-
score, and 93% in AUC, as indicated in Table 4. Although the GMM, iForest, and
LOF models demonstrate exceptional precision scores of 1.000, it is worth noting that
their recall rates are relatively lower in comparison. On the other hand, the AE model
demonstrates an impressive recall rate of 99%. However, it is worth noting that its F1-
score and AUC value are comparatively lower.

TABLE 3
Average Classification performance for different methods on the MIT-BIH
dataset.
Methods Acc ± SD Pre ± SD Rec ± SD F1-score AUC ±SD
±SD
OURS 0.9673 ± 0.9854± 0.948 ± 0.9666 ± 0.9672 ±

0.0005 0.0003 0.0001 0.0014 0.0015

AnoGAN (Schlegl et 0.9257 ± 0.8829± 0.9876 ± 0.9323 ± 0.9283 ±
al.)
0.0101 0.0167 0.0027 0.0085 0.0101
 AE (K.Wanget al.) 0.9282 ± 0.8733± 0.9902 ± 0.9281 ± 0.9233 ±

0.0180 0.2042 0.02333 0.1490 0.0049

VAE (X.Wang et al.) 0.8048 ± 0.7196± 0.9874 ± 0.8325 ± 0.8013 ±

0.0028 0.0029 0.0002 0.0157 0.0028

Stack LSTM 0.8875 ± 0.8013± 0.9740 ± 0.8970 ± 0.8882 ±
0.0021
(Chauhan et al.) 0.0017 0.0007 0.0019 0.0052
GRU (Cowton et al.) 0.8764 ± 0.8128± 0.9746 ± 0.8864 ± 0.8751 ±

0.0040 0.0064 0.0017 0.0031 0.0040

RNN(Latif et al.) 0.8568 ± 0.7826± 0.9798 ± 0.8702 ± 0.8538 ±

0.0031 0.0040 0.0003 0.0024 0.0031

DEEP-SVDD (Ruff 0.8039 ± 0.7221± 0.8342 ± 0.8342 ± 0.8037 ±
et al.)
0.0035 0.0037 0.0002 0.0025 0.0033

AE + OCSVM (Mo et 0.8624 ± 0.7965± 0.9788 ± 0.8783 ± 0.8644 ±

al.)
0.0036 0.0046 0.0003 0.0029 0.0050
DAGMM (Song et 0.7646 ± 0.9992± 0.5304 ± 0.6930 ± 0.7650 ±
al.)
0.0007 0.0008 0.0004 0.0019 0.0019
GMM (Dai et al) 0.6462 ± 0.9986± 0.2924 ± 0.0004 ± 0.6460 ±

0.0463 0.1603 0.0068 0.2924 0.0042

OCSVM (ScholkopF 0.8376 ± 0.9982± 0.6760 ± 0.8061 ± 0.8374 ±
et al.)
0.0009 0.0006 0.0005 0.0018 0.0019
iForest (Liu at al.) 0.6521 ± 0.9987± 0.3046 ± 0.4668 ± 0.6521 ±

0.0106 0.2119 0.3468 0.1334 0.0106

LOF (Bin Yao et al.) 0.5050 ± 0.5027± 0.9170 ± 0.6494 ± 0.5050 ±

0.0006 0.0007 0.0025 0.0018 0.0020

TABLE 4

Average classification performance for different methods on the CMUH dataset.

 Methods Acc ± SD Pre ± SD Rec ± SD F1-score AUC ±
± SD SD

OURS 0.9358 ± 0.9816 ± 0.8882 ± 0.9325 ± 0.9358 ±

0.0004 0.0002 0.0010 0.0008 0.0010

AnoGAN (Schlegl et 0.8985 ± 0.8396 ± 0.9852 ± 0.9066 ± 0.8985 ±

al.)
0.0092 0.0128 0.0018 0.0078 0.0092

AE (K.Wang et al.) 0.9103 ± 0.8504 ± 0.9946 ± 0.9169 ± 0.9098 ±

0.0181 0.0253 0.0012 0.0148 0.0181

VAE (X.Wang et 0.7744 ± 0.6885 ± 0.9910 ± 0.8125 ± 0.7713 ±

Al.) 0.0040 0.0039 0.0015 0.0027 0.0041

Stack LSTM 0.8875 ± 0.8097 ± 0.9772 ± 0.8856 ± 0.8738 ±

(Chauhan et al.) 0.0033 0.0051 0.0019 0.0025 0.0033

GRU (Cowton etal.) 0.8779 ± 0.8156 ± 0.9748 ± 0.8881 ± 0.8772 ±

0.0038 0.0052 0.0019 0.0030 0.0038

RNN(Latif et al.) 0.8221 ± 0.7414 ± 0.9860 ± 0.8464 ± 0.8210 ±

0.0037 0.0041 0.0021 0.0026 0.0037

DEEP-SVDD (Ruff et 0.7649 ± 0.6794 ± 0.9908 ± 0.8061 ± 0.7616 ±

al.)
0.0050 0.0047 0.0017 0.0032 0.0050
AE + OCSVM (Mo et 0.8245 ± 0.7436 ± 0.9864 ± 0.8479 ± 0.8231 ±
al.)
0.0030 0.0034 0.0022 0.0021 0.0030

DAGMM (Song et al.) 0.7260 ± 0.9991 ± 0.4520 ± 0.6224 ± 0.7258 ±

0.0012 0.0004 0.0024 0.0023 0.0012

GMM (Dai et al) 0.6057 ± 1.0000 ± 0.2148 ± 0.3536 ± 0.6074 ±

0.0037 0.0005 0.0074 0.0101 0.0037

OCSVM (ScholkopF 0.7600 ± 0.9985 ± 0.5208 ± 0.6845 ± 0.7600 ±

et al.)
0.0024 0.0010 0.0048 0.0041 0.0024
 iForest (Liu at al.) 0.6303 ± 1.0000 ± 0.2606 ± 0.4135 ± 0.6303 ±

0.0043 0.0009 0.0086 0.0107 0.0043

LOF (Bin Yao et al.) 0.5767 ± 0.10000± 0.1700 ± 0.2906 ± 0.5850 ±

0.0059 0.0010 0.0117 0.0174 0.0059

5 Conclusion

This study investigates the application of autoencoder models in the

development of a deep learning framework for the detection of anomalies in
electrocardiogram (ECG) data. The framework being proposed in this study
comprises three distinct modules, namely: 1) an autoencoder, which attempts
to reconstruct the input data, 2) a discriminator, which is designed for
determining a threshold value to distinguish between normal and abnormal
data, and 3) an anomaly detector, which aims to identify and flag anomalous
patterns. The autoencoder is trained using normal electrocardiogram (ECG)
signals in order to reconstruct each input ECG signal as its output, while
minimizing the error between the original and reconstructed signals. The
discriminator module utilizes the training dataset and the corresponding
reconstructed dataset by the autoencoder model to analyze the distribution of
reconstruction errors. By studying the error distribution, the discriminator
module establishes a reconstruction error threshold limit for the normal
electrocardiogram (ECG) signals. The anomaly detector module performs a
comparison between the anomaly score of a reconstructed ECG signal and a
predetermined threshold value, which is determined by the discriminator
module. If the anomaly score exceeds the threshold value, the input ECG
signal is identified as anomalous. The proposed framework achieved a
promising anomaly detection accuracy of approximately 95% when tested
with real ECG data. The study highlights the promising potential of utilizing
autoencoder models in the field of automating clinical diagnosis for cardiac
care through the detection of anomaly in recorded ECG signal.

References

[1] H. Li, P. Boulanger, A survey of heart anomaly detection using ambulatory

electrocardiogram (ecg), Sensors 20 (5) (2020) 1461.
[2] R. J. Martis, U. R. Acharya, and L. C. Min, ‘‘ECG beat classification using
PCA, LDA, ICA and discrete wavelet transform,’’ Biomed. Signal Process.
Control, vol. 8, no. 5, pp. 437–448, Sep. 2013.

[3] V. Gupta and M. Mittal, ‘‘KNN and PCA classifier with autoregressive
modelling during different ECG signal interpretation,’’ Procedia Comput. Sci.,
vol. 125, pp. 18–24, Jan. 2018.

[4] S. M. Woo, H. J. Lee, B. J. Kang, and S. W. Ban, ‘‘ECG signal monitoring

using one-class support vector machine,’’ in Proc. WSEAS, Penang, Malaysia,
Mar. 2010, pp. 23–25.

[5] K. Li, N. Du, and A. Zhang, ‘‘Detecting ECG abnormalities via

transductive transfer learning,’’ in Proc. ACM Conf. Bioinf., Comput. Biol.
Biomed. BCB, 2012, pp. 210–217.

[6] J. An and S. Cho, ‘‘Variational autoencoder based anomaly detection using

reconstruction probability,’’ Special Lect. IE, vol. 2, no. 1, pp. 1–18, 2015.

[7] A. Gogna, A. Majumdar, and R. Ward, ‘‘Semi-supervised stacked label

consistent autoencoder for reconstruction and analysis of biomedical signals,’’
IEEE Trans. Biomed. Eng., vol. 64, no. 9, pp. 2196–2205, Sep. 2017

[8] S. A. Alodia Yusuf and R. Hidayat, ‘‘MFCC feature extraction and KNN
classification in ECG signals,’’ in Proc. 6th Int. Conf. Inf. Technol., Comput.
Electr. Eng. (ICITACEE), Sep. 2019, pp. 1–5.

[9] G. B. Moody and R. G. Mark, ‘‘The impact of the MIT-BIH arrhythmia

database,’’ IEEE Eng. Med. Biol. Mag., vol. 20, no. 3, pp. 45–50, Jun. 2001

[10] S. S. Xu, M.-W. Mak, C.-C. Cheung, Towards end-to-end ecg

classification with raw signal extraction and deep neural networks, IEEE
journal of biomedical and health informatics 23 (4) (2018) 1574{1584

[11] M. Braei, S. Wagner, Anomaly detection in univariate time-series: A

survey on the state-of-the-art, arXiv preprint arXiv:2004.00433 (2020).

[12] Farady, Isack, et al. "ECG Anomaly Detection with LSTM-Autoencoder

for Heartbeat Analysis." 2024 IEEE International Conference on Consumer
Electronics (ICCE). IEEE, 2024.