=Paper=
{{Paper
|id=Vol-2667/paper32
|storemode=property
|title=Transfer Learning for tuberculosis screening by single-channel ECG
|pdfUrl=https://ceur-ws.org/Vol-2667/paper32.pdf
|volume=Vol-2667
|authors=Valeriia Guryanova
}}
==Transfer Learning for tuberculosis screening by single-channel ECG ==
https://ceur-ws.org/Vol-2667/paper32.pdf
Transfer Learning for tuberculosis screening by
single-channel ECG
Valeriia Guryanova
Faculty of Computational Mathematics and Cybernetics
Lomonosov Moscow State University
Moscow, Russia
kibitovavaleria@gmail.com
ORCID: 0000-0002-3613-8103
Abstract—Tuberculosis is one of the leading causes of death in The motivation for this study was the work that shows the
the world. The majority of the population is not able to regularly relationship between cardiological and tuberculosis diseases
conduct specific e x aminations, s u ch a s x - ray e x aminations, for [4]. Besides, often the tuberculosis bacteria themselves affect
the presence of tuberculosis. Currently, there are mobile devices
for measuring ECG, which allow taking measurements without the heart and thus affect its electrical activity. At the moment,
leaving home. This article explores the possibility of determining this is the first work that is devoted to the problem of
tuberculosis based on a single-channel mobile ECG. One of the tuberculosis detection via single-channel mobile ECG.
general top-performance neural networks is used as a classifier. Now neural networks are one of the most popular methods
This article also explored the possibility of such a classification for analyzing medical signals and images. There are plenty
based not on raw data, but the generated image. The image
interprets the prediction of the neural network and makes it of works in which medical images and ECGs are analyzed
possible for the doctor to understand the model’s decision better. using neural networks (NN), for example, [5], [6], [7], [8],
The article shows the promising performance and provides proof [9]. Usually, raw ECGs are analyzed using recurrent [10] or
of concept of such screening. Different ratios of precision and convolutional neural [8] networks with 1d convolutions. In this
recall are provided, which can be adjusted depending on the paper, it is proposed to analyze an ECG visualization and not a
situation.
Keywords—Neural Networks, ECG classification, Tubercu- raw signal. So the image but not a signal is analyzed. There are
losis, Deep Learning, Transfer Learning some works in which neural networks analyze the image of an
ECG signal [11] [12] and show good performance. However,
I. I NTRODUCTION in these articles, architectures were trained from scratch and
were specially selected for the task. There are large number of
Tuberculosis (TB) is caused by bacteria (Mycobacterium
pretrained architectures for images [13], [14], [15]. Since, for
tuberculosis) that most often affect the lungs. Tuberculosis is
many tasks, pretrained architectures show significantly better
curable and preventable. TB is easily spread from person to
performance than architectures trained from scratch [16], there
person through the air. TB is one of the ten leading causes of
is an assumption that they will show better performance on
death in the world [1]. It is crucial to determine tuberculosis
the paper’s task. There is work [17] that shows that pretrained
in time to prevent its spread and begin treatment of an ill
architectures can be used as a feature extractor for phono-
person. It takes a long time for both the physician and the
cardiogram signals (PCG), which also provides additional
patient to carry out specific tests for TB (e.g., such as X-rays or
motivation for exploring the possibility of using pretrained
Xpert MTB/RIF), and as such, they cannot be conducted with
architectures.
relatively high regularity. So, it would be useful to have some
This paper aims to investigate the possibility of determining
pre-screening, which could be done quite often and which
tuberculosis from ECG images, the possibility of using pre-
would allow identifying people with a high chance of having
trained models for such tasks, and of the interpretation of such
tuberculosis and conduct specific t e sts o n ly f o r them.
models. The main scientific novelty of this paper is the use of
ECG is a signal that displays the electronic activity of
pre-trained architecture for ECG classification. It is also the
the heart. Each ECG recording shows the potential difference
first article that explores the feasibility of determining tubercu-
between two electrodes located on the surface of the body.
losis by a single-channel ECG and explores the relationships
Each of the measured potential differences is called a lead.
of metrics that a solution can achieve.
In medical institutions, 12-lead ECGs are commonly used.
The paper is organized as follows. Initially, the dataset and
Currently, on the market, there are mobile devices that are
data preprocessing methods are described. Then, the process
capable of reading an electrocardiogram (ECG) of a person,
of generation of the ECG image and the architecture of the
for example, AliveCor [2], CardioQvark [3]. Such
neural network are given. Then the results of experiments and
devices read only one of the 12 leads. Such devices can be used
interpretation of the neural network are reported. Finally, the
as often as required. Anybody can have just a few devices per
limitations of the article are analyzed, and the conclusion is
institution or organization to conduct such pre-screenings
given.
every day. People who have a high chance of having a disease
will need to see a TB specialist as soon as possible.
Copyright © 2020 for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0)
Data Science
II. DATASET Since the signals were taken from a device that can have
noise, and the signal could disappear for various reasons, it is
All data used in this study was collected using the Cardio-
necessary to remove signals that contain much noise or do not
Qvark device [3]. The dataset consists of 1232 ECGs from
contain any useful information. R-peaks that represent heart-
136 people with TB and 3609 ECGs from people without
beat were calculated using the Pan–Tompkins algorithm [19].
TB. Dataset sex distribution among people with tuberculosis
Then the distances between the two peaks were calculated.
and people without tuberculosis is shown in Fig.1. Dataset
It was measured heuristically that the signal section almost
age distribution among people with tuberculosis and people
certainly does not contain useful information if the distance
without tuberculosis is shown in Fig.2.
between the two R-peaks is more than 2.6s or less than 0.3s.
The signal was removed from the dataset if there were more
than 55 percent of such distances in it.
After dividing the long signals into sub-signals and remov-
ing the signals that contain mainly noise, 13776 signals were
obtained. Ten thousand ninety-seven of them belong to 367
TB-positive people, and 3679 of them belong to 136 TB-
negative people. More than 90 percent of the resulting signals
is 100 seconds long. The number of signals per TB-positive
patients varies from 1 to 1540. The number of signals per TB
negative patients varies from 3 to 198.
The image was generated using the Matplotlib plot function
[20]. The whole ECG signal was divided into seven parts
Fig. 1. Dataset sex distribution. of 2857 points each, with dots per inch of 70. The size of
the figure was set to 7.4 by 7.4 inches. Since the matplotlib
generates an image in the RGBA format, and most neural
networks use RGB format, only the 4th channel was used. This
channel was multiplied to create three channels. The example
of the generated image is given in Fig. 3.
Fig. 2. Dataset age distribution.
ECG recordings were sampled as 1000 Hz, and they have
been filtered b y t h e C a rdioQvark d e vice. T h e l e ngth o f ECG
recordings varies from 30.5 seconds to 900 seconds. For trend
subtraction, the median filter w i th a k e rnel s i ze o f 1 8 7 was
used.
III. I MAGE G ENERATION P ROCESS AND N EURAL
N ETWORK A RCHITECTURE
Signals were resampled to a frequency of 200 Hz via the
polyphase resample method from the SciPy Python library [18]
to reduce the image size. Signals longer than 100 seconds were Fig. 3. Generated ECG image.
divided into sub-signals. The signal is divided sequentially into
sections of 100 seconds to receive the sub-signals, if the length ResNext 101 [15] was used as the main model for the
of the last section is less than 30 seconds, then it was not added experiments. The main idea of this model is the introduction
to the sample set. of the new dimension, which authors called ”cardinality” -
the size of the set of transformations. The main function of
VI International Conference on "Information Technology and Nanotechnology" (ITNT-2020) 150
Data Science
this dimension is the control over complex transformations. The test set contains original 1130 ECGs from 144 patients.
Even while maintaining complexity, increasing ”cardinality” Among them, 37 patients with 288 ECGs have TB. After the
can improve classification accuracy. In this paper, a model preprocessing stage, the validation set has 3205 ECGs and
with a ”cardinality” value of 32 was used. This model is one among them 857 from TB-positive persons. The number of
of the top best models for classification on ImageNet [21] and patients after the preprocessing stage stays the same in all
is easy to use and retrain on new data. datasets.
This model has four blocks of convolutional layers. The For reproducibility, all neural network configurations were
following neural network configurations were used to explore trained with a random state of 10. This random state was fixed
the possibilities of transfer learning: the completely untrained for NumPy, PyTorch CPU, and GPU variables and also for
network (UN), pretrained on ImageNet network with no frozen standard Python random library. Also, the CUDA state was
layers (PN), pretrained on ImageNet network with frozen layer set to deterministic. For all configurations, Adam was used as
1 (or 2, 3, 4) and all layers before it (L1(2, 3, 4)). In the case an optimizer, and batch size was set to 16. For the untrained
of freezing up to layer four, only one fully connected classifier network, the learning rate was set to 0.0005 and is dropped
layer is trained. by a factor of 0.0001 after the 8th epoch. For the pretrained
Augmentation is a method in which artificially generated network, the learning rate was set to 0.0005 and is dropped
data is added to the original data. This method is one of the by a factor of 0.001 after the 8th. For the pretrained network
methods that is widely used to improve the quality of neural with unfrozen and frozen layers, the learning rate was set to
network models [22]. Online augmentation is a method that 0.0001 and is dropped by a factor of 0.001 after the first and
replaces the part of the data to the synthetic one during the fourth epoch. All parameters that are not described were set
training stage. A study was conducted that investigated online to default values of PyTorch library [24]. The metric that was
augmentation for the classification quality improvement of used for early stopping was the area under the precision-recall
single-channel ECGs by using a neural network [23]. The data (PR AUC) curve since classes are not balanced in samples. The
that was used in that article contains the dataset discussed in precision-recall curve illustrates the ratio between precision
this study. However, that study did not use the representation of and recall of different thresholds. A high value of precision
the signal as an image, nor did it study the possibilities of such indicates that a model has a low false-positive rate, and a high
training. Research in an article on online augmentation has value of recall shows that model has a low false-negative rate.
shown that it helps in improving the quality of classification A high value of the PR AUC corresponds to both high recall
of single-channel signals. and high precision.
One of the most promising methods was the ResampleParts Another important metric that is useful to take into con-
method. This method resamples parts of the signal with a sideration during model evaluation is an area under the ROC
specific resampling coefficient using the polyphase resample curve (ROC AUC). ROC curve shows the ratio between the
method from the SciPy Python library [18]. The specific size true positive rate and the false positive rate. This metric could
of the signal part and resampling coefficient is set at random be useful as it shows the probability that the model ranks
each time from a certain interval. Due to the reason mentioned a random positive example higher than a random negative
above, this augmentation method was chosen to increase the example. However, this metric can be sensitive to unbalanced
quality of tuberculosis detection. classes in data, and so it should not be the only metric
for performance evaluation. Table I shows the results from
IV. E XPERIMENTS AND R ESULTS
different network configurations v ia t he ROC AUC s core and
All experiments were conducted using Python language. PR AUC score.
PyTorch [24] was used as the main neural network framework.
Signals were read and transformed using NumPy [25] and
TABLE I. EXPERIMENT RESULTS
SciPy [18] libraries.
To correctly evaluate the results, the entire dataset was
divided into training, test, and validation set. The training set Quality Network Configuration
Criteria UN PN L1 L2 L3 L4
was used to train the model. The validation set was used to ROC AUC 0.9079 0.93 0.934 0.9307 0.9138 0.8089
select the hyperparameters and for early stopping. A test set PR AUC 0.7956 0.83 0.84 0.8399 0.8081 0.615
was used to check the final performance. The full dataset was
partitioned into three parts in a specific way, such that each
of the three groups contained different patients. The train set As can be seen from table I, the best performance is
contains 2648 original ECGs from 247 patients. Among them, obtained with configurations L1 and L2. This table shows that
69 patients with 689 ECGs have TB. After the preprocessing transfer learning is useful for classifying ECG signals. These
stage the train set has 7495 ECGs and among them 2058 from results also correspond to the logic that the first layers of a
TB-positive persons. The validation set contains original 1062 neural network learn simple dependencies such as lines and
ECGs from 112 patients. Among them, 30 patients with 255 simple shapes. When freezing a neural network up to the last
ECG have TB. After the preprocessing stage, the validation set layers, a worse performance can be gotten than on a network
has 3076 ECGs and among them 765 from TB-positive person. without pretraining (UN), since the features on the last layers
VI International Conference on "Information Technology and Nanotechnology" (ITNT-2020) 151
Data Science
are not relevant for this specific task. Thus, based on metrics, Table II, table III, and table IV show the performance of
L1 can be considered the best model. models trained with augmentation and without it. As shown in
For the improvement of this best model earlier described, these tables, the results are quite promising and customizable,
online augmentation was used. The interval for the resampling depending on the situation. These tables show that augmen-
part was set to (40, 200). The two types of resampling were tation increases the values of Fβ and F Cβ when β equals to
used: frequency decreasing (FD) and frequency increasing 0.25 and 1, but does not improve the values when β equals to
(FI). The resulting sample rate is calculated as the upsample 2. It can be explained by the fact that augmentation parameters
factor divided on the downsampling factor. For the frequency, were selected to maximize PR AUC value. This value is most
increasing the upsample factor was getting from 2 to 3, and correlated with Fβ when β equals to 1. As can be seen from
the downsampling factor was gotten from 3 to 5. For the the ratio between precision and recall, the main increase in the
frequency decreasing upsample, factor was gotten from 3 to PR AUC metric was associated with an increase in precision.
6, and the downsampling factor was gotten from 1 to 3. It explains the increase in performance when β equal to 0.25.
The probabilities of augmentations was chosen from values Perhaps it is necessary to focus on other metrics during the
[0.1, 0.2, 0.3, 0.4, 0.5]. The best model was determined based process of selection of augmentation parameters to increase
on the PR AUC metric on the validation set. The best model the value of F2.
was the model with the probability of augmentation of 0.5.
The best model has 0.9475 ROC AUC and 0.875 PR AUC TABLE II. EXPERIMENT RESULTS
on the test set. Based on the results of the experiments, it
is reasonable to conclude that augmentation improves model β P R Fβ
performance. M1 M2 M1 M2 M1 M2
0.25 0.8918 0.9147 0.471 0.55075 0.617 0.687
In most real-life scenarios, the probability of certain events 1 0.756 0.785 0.7911 0.782 0.773 0.784
can not be used to make a decision. A specific label is 2 0.665 0.6274 0.878 0.9393 0.757 0.7523
assigned to each patient using probability thresholding. To
explore the balance between precision and recall in a situation
of unbalanced classes, Fβ score metric is used:
TABLE III. EXPERIMENT RESULTS
precision × recall
Fβ = (1 + β 2 ) .
β 2 precision + recall β PC RC F Cβ
M1 M2 M1 M2 M1 M2
0.25 0.8688 0.894 0.5559 0.6223 0.678 0.734
The β is in range (0, +∞). The greater the β is, the larger the 1 0.726 0.76 0.863 0.856 0.789 0.8059
weight the recall has. β can be selected differently based on the 2 0.621 0.5711 0.937 0.9825 0.7475 0.7223
situation. If we want to find all TB-positive people and the cost
of additional procedures is not that important, then we give the
recall the greater weight. If the cost of additional procedures
TABLE IV. EXPERIMENT RESULTS
is crucial, then the precision is much more important even at
the cost of skipping some TB-positive people.
This article investigated the following beta threshold op- β A A1 PS
M1 M2 M1 M2 M1 M2
tions: 0.25, 1, 2. The validation set was used for the threshold 0.25 0.843 0.866 0.8655 0.885 0.8389 0.8498
calculation. The following metrics were investigated: accuracy 1 0.876 0.884 0.8824 0.895 0.819 0.831
(A), which is the number of right guesses, as well as precision 2 0.849 0.834 0.8388 0.8076 0.7611 0.7246
(P) and recall (R). Since each cardiogram was divided into
several images, the following rule was used to determine the
label for the entire cardiogram. If at least one cardiogram is V. I NTERPRETATION
labeled as TB-positive, then the entire cardiogram is labeled Doctors often want to understand why a neural network
TB-positive. The metrics that are determined for the entire made a confident decision. Therefore, the interpretation of the
ECG are called: AC, PC, and RC for accuracy for ECG, neural network is quite useful. Besides, interpretation can be
precision for ECG, and recall for ECG, respectively. Since used to identify previously unknown signs of the disease to
one patient can have multiple ECG, the patient score for think through treatment in the future.
cardiogram labels was calculated. The patient score (PS) Grad-CAM [26] shows quite good results in interpreting
equals to mean accuracy for person ECGs averaged through neural networks. Grad-CAM uses the gradient information
all people. flowing into the last convolutional layer of CNN to understand
Models without augmentation (M1) and with it (M2) were the impact on each neuron for a decision of interest. Fig. 4
compared using the metrics described above. Fβ metric was and 5 show the interpretation results on the pretrained model
used to determine which model has a better ratio between without augmentation and with it. The lighter the points on
recall and precision. Fβ , which shows the ratio between ECG the image, the more important these points are. Both images
precision and ECG recall is shown as F Cβ . represent the same ECG of TB-positive person. As can be seen
VI International Conference on "Information Technology and Nanotechnology" (ITNT-2020) 152
Data Science
from the images, the neural network draws attention to the
PQST complexes, the distance between the peaks, and also the
height of some peaks. Features selected by the network with
and without augmentation are almost identical. However, the
network with augmentation does not take into consideration
some features as one at the end of the 4th row. Perhaps the
elimination of some noisy features might help to improve the
classification quality.
Fig. 5. Interpretation of NN with augmentation on TB-positive person.
VII. CONCLUSION
The new method for classifying the visualization of ECGs
using pre-trained architecture was proposed. This method
was used on a single-channel ECG to explore the possibility
of determination of tuberculosis from such images. It was
Fig. 4. Interpretation of NN without augmentation on TB-positive person. determined that the neural network could detect tuberculosis
by a cardiogram with moderately high performance. The
VI. DISCUSSION new method help to improve the quality of the ECG image
The reader of this article should know that this article classification. ROC AUC s core w as i ncreased f rom 0 .9079 to
has the following limitations: the sample size of TB-positive 0.934, and PR AUC score was increased from 0.7956 to 0.84.
people is quite small, so for further investigation of perfor- The significance o f t he r esults s hows t hat t his m ethod should
mance, the bigger sample should be used. It should also be be used to classify the ECG for other tasks. For example, to
noted that the sample contains patients who received treatment detect coronary heart disease and arrhythmia.
for tuberculosis. Due to the existing database structure, it is Another paper novelty was the confirmation o f t he possi-
impossible to say whether the patient took medicine, and if bility of increasing the classification q uality b y u sing online
he did, how recent the use is. For more robust experiments, a augmentation with such architectures and data. It has been
new database must be compiled. However, even in the current shown that online augmentation can improve the performance
circumstances, this study is useful for the following reasons. of the classification. I n t his a rticle, t he p erformance was
Let us imagine an organization that employs a TB-positive improved from 0.934 ROC AUC to 0.9475 ROC AUC, and
person, who has already started taking medicine and hides PR AUC was increased from 0.84 to 0.875. The increase
his condition, thus increasing the chance of infecting people shows that online augmentation should be used when using
around him. In such a situation, it is still beneficial t o screen the proposed method.
people with the suggested algorithm. Besides, it is known that The method for interpretation of such models is described,
people taking medicine may experience an increase in QT- which can be used further for drug and disease effect explo-
interval [27]. However, in this dataset, the average value of ration purposes.
the QT-interval showed a low predictive quality for TB of Various ratios of quality metrics of the obtained solution for
about 0.51 ROC AUC. Based on this, we can conclude that tuberculosis detection have been received. They show that it is
the neural network finds some other defining signs. Also, with reasonable to conduct TB pre-screening using mobile ECG in
the help of the neural network interpretation in the future, the the future to identify potential patients and recommend further
effect of TB-drugs on the heart can be explored. specific t ests. T his s tudy i s a n i mportant p roof o f concept.
VI International Conference on "Information Technology and Nanotechnology" (ITNT-2020) 153
Data Science
Some disadvantages of the dataset were indicated in the [18] P. Virtanen, "SciPy 1.0: fundamental algorithms for scientific computing
in Python," Nature Methods, vol. 17, no. 3, pp. 261-272, 2020.
article. In the future, to build a production solution, it is
[19] J. Pan and W. J. Tompkins, "A Real-Time QRS Detection Algorithm,"
reasonable to increase and carefully verify the dataset. IEEE Transactions on Biomedical Engineering, vol. BME-32, no. 3, pp.
230-236, 1985.
ACKNOWLEDGMENT [20] J. Hunter, "Matplotlib: A 2D Graphics Environment", Computing in
This article contains the results of a project carried out Science & Engineering, vol. 9, no. 3, pp. 90-95, 2007.
[21] D. Mahajan, R. Girshick, V. Ramanathan, K. He, M. Paluri, Y. Li,
within the imple-mentation of the Program of the Center for A. Bharambe and L. van der Maaten, ”Exploring the limits of
Competence of the National Tech-nology Initiative “Center for weakly supervised pretraining,” Proceedings of the European
Storage and Analysis of Big Data”, supported by the Ministry Conference on Computer Vision (ECCV), pp. 181-196, 2018.
of Science and Higher Education of the Russian Federation [22] A. Mikołajczyk and M. Grochowski, ”Data augmentation for improving
deep learning in image classification problem,” International
under the Lomonosov Moscow State University Project Interdisciplinary PhD Workshop (IIPhDW), Swinoujście, pp. 117-122,
Support Fund 13/1251/2018 from 11.12.2018. 2018.
[23] V. Guryanova, ”Online Augmentation for Quality Improvement of Neu-
ral Networks for Classification of Single-Channel Electrocardiograms,”
REFERENCES Communications in Computer and Information Science, pp. 37-49, 2020.
[1] ”Tuberculosis (TB)”, Who.int, 2020 [Online]. URL: [24] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan and
https://www.who.int/news-room/fact-sheets/detail/tuberculosis. T. Killeen, ”PyTorch: An imperative style, high-performance deep
[2] ”AliveCor”, Alivecor.com, 2020 [Online]. URL: learning library,” Advances in Neural Information Processing Systems,
https://www.alivecor.com. pp. 8024-8035, 2019.
[3] ”Cardiomonitor (in Russian) CardioQVARK”, Cardioqvark.ru, 2020 [25] S. van der Walt, S. Colbert and G. Varoquaux, ”The NumPy Array: A
[Online]. URL: http://www.cardioqvark.ru. Structure for Efficient Numerical Computation,” Computing in Science
[4] M. Huaman, D. Henson, E. Ticona, T. Sterling and B. Garvy, ”Tu- & Engineering, vol. 13, no. 2, pp. 22-30, 2011.
berculosis and cardiovascular disease: linking the epidemics,” Tropical [26] R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh and D.
Diseases, Travel Medicine and Vaccines, vol. 1, no. 1, 2015. Batra, ”Grad-cam: Visual explanations from deep networks via
[5] T. Pashina, A. Gaidel, P. Zelter, A. Kapishnikov and A. Nikonorov, ”Au- gradient-based localization,” Proceedings of the IEEE international
tomatic highlighting of the region of interest in computed tomography conference on computer vision, pp. 618-626, 2017.
images of the lungs,” Computer Optics, vol. 44, no. 1, pp. 74-81, 2020. [27] A. A. Bykova, M. I. Chashkina, M. V. Serova, A. V. Simonov
DOI: 10.18287/2412-6179-CO-659. and F. Yu. Kopylov, "The prevalence of prolongation of the QT in-
[6] J. Agafonova, A. Gaidel, P. Zelter and A. Kapishnikov, ”Efficiency terval in patients receiving antituberculous chemotherapy," Cardiology
of machine learning algorithms and convolutional neural network for and Cardiovascular Surgery, vol. 12, no. 2, pp. 146-151, 2019.
detection of pathological changes in MR images of the brain,” Computer
Optics, vol. 44, no. 2, pp. 266-273, 2020. DOI: 10.18287/2412-6179-
CO-671.
[7] S. Kiranyaz, T. Ince and M. Gabbouj, ”Real-Time Patient-Specific
ECG Classification by 1-D Convolutional Neural Networks,” IEEE
Transactions on Biomedical Engineering, vol. 63, no. 3, pp. 664-675,
2016.
[8] M. Zihlmann, D. Perekrestenko and M. Tschannen, ”Convolutional
Recurrent Neural Networks for Electrocardiogram Classification,”
Computing in Cardiology Conference (CinC), pp. 1-4, 2017.
[9] Z. Yao, Z. Zhu and Y. Chen, ”Atrial fibrillation detection by multi-scale
convolutional neural networks,” 20th International Conference on
Information Fusion (Fusion), Xi’an, pp. 1-6, 2017.
[10] R. Banerjee, A. Ghose and S. Khandelwal, ”A Novel Recurrent Neural
Network Architecture for Classification of Atrial Fibrillation Using
Single-lead ECG,” 27th European Signal Processing Conference
(EUSIPCO), A Coruna, Spain, pp. 1-5, 2019.
[11] P. Sean and J. Huizinga, ”Robust and fast heart rate variability analysis
of long and noisy electrocardiograms using neural networks and
images,” arXiv: 1902.06151, 2019.
[12] M. Kim, H. Ko and S. Pan, ”A study on user recognition using 2D ECG
based on ensemble of deep convolutional neural networks,” Journal of
Ambient Intelligence and Humanized Computing, vol. 11, no. 5, pp.
1859-1867, 2019.
[13] H. Kaiming, X. Zhang, Sh. Ren and J. Sun, ”Deep residual
learning for image recognition,” Proceedings of the IEEE conference
on computer vision and pattern recognition, pp. 770-778, 2016.
[14] F. Chollet, ”Xception: Deep learning with depthwise
convolutions,” Proceedings of the IEEE conference on computer
vision and pattern recognition, pp. 1251-1258, 2017.
[15] X. Saining, R. Girshick, P. Dollár, Z. Tu and K. He, ”Aggregated
residual transformations for deep neural networks,” Proceedings of the
IEEE conference on computer vision and pattern recognition, pp.
1492-1500, 2017.
[16] K. Simonyan, and A. Zisserman, ”Very deep convolutional networks
for large-scale image recognition,” arXiv: 1409.1556, 2014.
[17] H. Alaskar, N. Alzhrani, A. Hussain and F. Almarshed, ”The Imple-
mentation of Pretrained AlexNet on PCG Classification,” Intelligent
Computing Methodologies, pp. 784-794, 2019.
VI International Conference on "Information Technology and Nanotechnology" (ITNT-2020) 154