Int'l Conf.

Bioinformatics and Computational Biology | BIOCOMP'19 | 3

Semantic ECG Interval Segmentation Using Autoencoders

Hedayat Abrishami1 , Matthew Campbell2 , Chia Han1 , Richard Czosek3 , and Xuefu Zhou1
1 Department of EECS, University of Cincinnati, Cincinnati, OH, USA
2 Division of Cardiology, Children’s Hospital of Philadelphia, Philadelphia, PA, USA
3 Heart Institute, Cincinnati Children’s Hospital Medical Center, Cincinnati, OH, USA

Abstract— Key cardiac waves in electrocardiogram (ECG)

signals including P-wave, QRS-complex, and T-wave provide
important information for various heart conditions and
effects of cardiac medications. Segmenting ECG into its
important components is a well-known yet challenging prob-
lem. Recently, deep learning (DL) has attracted significant
attention in a variety of fields. This paper presents a hybrid
DL-method using deep convolutional autoencoders to extract
the hierarchical structure of cardiac waves and long short-
term memory (LSTM) cells to characterize the ECG temporal
attributes. Results show that the proposed method performs
well and can achieve a segmentation accuracy of 93.99%
defined as the ratio of correct positive and correct negative
classification for ECG data points over the total positive and Fig. 1: Typical cardiac complex with annotated waves and
negative ground truth. peaks.

Keywords: ECG segmentation, cardiac waves, deep learning,

convolutional networks, LSTM layers. demonstrates that this approach has the potential to capture
long-term and short-term dependencies in ECG signals,
1. Introduction extract the complex hierarchical structure of ECG signals,
A reliable automated ECG diagnosis system is of high and show its effectiveness in ECG segmentation.
importance for the cardiology community, especially, in The rest of the paper is organized as follows. Section 2
high-throughput population-based screening applications [1]. briefly reviews related works. Section 3 introduces our novel
Effective ECG signal processing, analysis, and localization hybrid method using a convolutional autoencoder and LSTM
of the cardiac waves rely on the accurate segmentation of the layers to segment ECG cardiac waves. Section 4 presents
major wave patterns in the cardiac complex. Every cardiac experimental results and discussions. Finally, Section 5 con-
wave cycle can be divided into distinct intervals, each of cludes the results and the study.
which has various shapes and formations. The various key
types of cardiac wave formations, such as P-wave, QRS- 2. Related Works
complex, T-wave as shown in Fig. 1, and their combinations Generally speaking, there are two research efforts in
with other cardiac waves, in a long sequence of wave cycles, ECG signal analysis. The first one is to develop suitable
can lead to a myriad of complex structures of signals. The methodologies for extracting features that best represent the
structure of these waves, when properly extracted, can lead structure of cardiac waves. The second one is to classify
to cardiac symptom diagnoses. The goal of the segmentation features based on the known diagnoses of diseases. An
is to label each ECG data point into a specified category. extensive literature review on ECG signal analysis can be
Four major categories are defined in this study: 1) Neutral, found in [5].
2) P-wave, 3) QRS-complex, and 4) T-wave. The latter three The common approaches to extract ECG features consist
categories are critical in providing important information for of derivative-based [6], wavelet-filters [7] and amplitude-
diagnoses. For instance, one of the symptoms of sudden based methods [8]. Pan and Tompkins developed one of
cardiac arrest and eventual death relies on finding the T- the most well-known derivative-based methods to find QRS-
wave inversion, i.e., detecting a dip shape instead of a rising complexes [9]. Derivative-based methods identify higher
form [1]. variations in the slope of the ECG signal. Martinez et al. used
This paper introduces a real-time cardiac wave segmenta- a discrete wavelet-based method to extract temporal features
tion method using both deep convolutional autoencoders [2], [7]. Other transformation methods such as Fourier Transform
[3] and Long Short-Term Memory (LSTM) cells [4], and and Discrete Wavelet Transform (DWT) have been used in

ISBN: 1-60132-492-8, CSREA Press ©

4 Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'19 |

[10]. Tafreshi et al. introduced an amplitude-based method ECG signal. Finally, Subsection 3.6 describes experiments
to identify QRS-complexes specifically [8]. and the convergence of the Hybrid-ECG-SegNet. The result
Furthermore, depending on the extracted ECG features, demonstrates the strength of Hybrid-ECG-SegNet compared
various algorithms have been developed for ECG cardiac to the other sequence learners such as HMM for the same
wave classification. In particular, algorithms based on Neural task of ECG interval segmentation using raw ECG signals.
Network (NN) [11], Support Vector Machine (SVM) [12],
Naive Bayes [13], Hidden Markov Model (HMM) [14], 3.1 ECG Intervals
logistic regression [15] and ruled-based [16] methods have Normal wave complexes contain well-known patterns that
been found in research literature. can be characterized by features, such as shape formation,
Recently, Deep Learning (DL) methods have been applied interval duration, and amplitudes. However, there are many
in ECG signal processing and analysis as well [17], [18], different shapes that can be found in ECG data. Normal
[19]. In [17], DL methods were used to classify ECG signals QRS-complexes can have nine different shapes [8] and P-
into normal and abnormal ECG related to certain heart waves and T-waves can also appear in different forms and
disease symptoms directly. In [19], Abrishami, et al., used amplitudes. When the characterization of key values of
convolutional networks to localize P-wave, QRS-complex, the wave parameters and the information that they carry
and T-wave in a single cardiac complex and utilized a are understood by the clinicians, then as abnormalities are
secondary algorithm to extract cardiac complexes. Then in observed, it is possible for physicians to make decisions
[20], Abrishami, et al., used local derivative-based features that lead to differential diagnoses of cardiac diseases. Car-
and LSTM networks to capture the temporal attributes of diologists do not make diagnoses based solely on single
ECG signals. Following these works, this paper is to present features in wave complexes, such as ST depression, T-
a comprehensive hybrid deep NN, called the Hybrid-ECG- inversion, and long QT-interval, but they also identify cardiac
SegNet, which is capable of automatic feature learning and wave components in relation to each other. For example,
capturing the temporal attribute of the extracted features following a QRS-complex location is probably an S-wave.
simultaneously, to segment real-time ECG signals. Thus, knowledge of the waveâĂŹs prior location is essential
to predicting the consequent wave. Likewise, each individual
3. Methodology waveâĂŹs formation also affects the formation of other
Deep convolutional autoencoders have shown their advan- waves. Therefore, a method that is capable of keeping
tages in segmentation [2]. They have the ability to extract persistent memory, i.e., a Recurrent Neural Network (RNN),
or compress the essential information from the input and can be a viable solution to this type of time-dependent
reconstruct the desired output based on the extracted infor- problem. The recurrent neural network creates a loop to pass
mation. In addition, LSTM networks can capture temporal the information from one timestamp to another, which allows
long-term and short-term dependencies in the input. it to learn a time series [22].
The proposed method used in the Hybrid-ECG-SegNet
scheme consists of four components. The first component is 3.2 Data Preparation and Dat Set
the input layer to the normalized ECG signals. The second The data preparation step consists of preprocessing of a
component is a deep convolutional autoencoder. The convo- given data set of ECG signals to get the raw data ready for
lutional autoencoder is capable of generating feature vectors NN training. The data used for this study is the QT database
for the next layer, the LSTM layer. The third component (QTDB). QTDB was produced by PhysioNet [23] and has
is a sequence learner. The sequence learner includes two a large collection of recorded physiological signals sampled
Bidirectional LSTM (BLSTM) layers. BLSTMs have two at 250Hz. This database includes over 105 two-channeled
different hidden LSTM layers including one forward hidden ECG recordings, each 15-minutes in duration and it is chosen
layer and one backward hidden layer to receive inputs in both to include a broad variety of P, QRS, ST, and T pattern
forward and backward directions [21]. The fourth and the morphologies [23]. This dataset thus allows researchers to
last component forms the output layer, which classifies every perform research on ECG signal delineation.
data point in the time-stamp into one of the four categories, The preprocessing of raw ECG signals takes care of
namely, Neutral, P-wave, QRS-complex and T-wave. two problems: wander drift baseline caused by sampling
This Section, Section 3, contains several subsections devices and not having a defined unit to measure amplitude.
and they are organized as follows. In Subsection 3.1, the Therefore, in this step, the ECG wander drift baseline is
order and attributes of ECG intervals are discussed. The removed, and the range of ECG amplitudes range is brought
data preparation and the dataset are to be introduced in to the interval of (−1, 1). Dohare et al. [24] proposed a
Subsection 3.2. Subsections 3.3, 3.4, and 3.5 introduce the successful median filtering approach to remove the wander
novel architecture of the Hybrid-ECG-SegNet. The network drift baseline from ECG. This approach applies two median
is capable of extracting the hierarchical ECG structure, long filters, each with half the size of the sampling frequency
temporal dependencies, and short temporal dependencies of (fs ), to the ECG signal. The result of applying these two

ISBN: 1-60132-492-8, CSREA Press ©

Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'19 | 5

filters is the wander drift baseline signal. Thus, a wander- max-pooling, and up-sampling layers. While convolutional
baseline-free ECG signal will be obtained by subtracting and max-pooling layers are used for encoding, convolutional
ECG signal from its wander drift baseline. The mathematical and up-sampling layers are used for decoding. The decoder
representation of removing wander drift is design follows the architecture of the Super-Resolution Con-
volutional Neural Network (SRCNN) [2]. SRCNN upsam-
XW F = ER − M (M (ER )) (1)
ples the signal to a higher resolution from a convolution layer
where XW F is wander free ECG, ER is the raw ECG and has demonstrated its strengths in reconstructing a high-
from QTDB, and M (.) operation is the result of applying a resolution image from the low-resolution image without any
median filter operation to a signal, and the size of the median obvious image artifact [2].
filter equals fs /2. Given that QTDB is sampled at 250Hz;
thus, fs = 250Hz.
The second task of data preparation is to normalize the
amplitudes to the range of (−1, 1), after removing the spikes.
For this task, Eq. 2 is used. Therefore, even when there are
still some amplitudes that can be above one they do not
affect the learning process.
XW F − μ
X= (2)
σ
where μ is the XW F mean, σ is XW F standard deviation,
and XN orm is the normalized ECG signal.
After performing these signal preprocessing steps, the
ECG signals are ready for generating a dataset.
In this study, every recording is divided into 1, 000 data
points sampled at 250Hz. Within every segment, one or
more cardiac complexes can be found, which makes the
segmenting task more challenging. The 1, 000 ECG data
points are the inputs to Hybrid-ECG-SegNet model, and
correspondingly, the related annotations are the output tar-
gets for the Hybrid-ECG-SegNet model. In total, there are
46, 690 sets of 1, 000 data points of ECG that have been
extracted from the QTDB to be used as inputs/outputs.
More specifically, for our experiments, three different sets –
training, validation, and test sets – have been created using
all the extracted ECG segments. These sets are mutually
exclusive, indicating there are no identical segments from
one recording to another and subject independent. Table 1
illustrates the training, validation and test data sets.

Table 1: Dataset
Dataset Number of samples Percentage
Training set 28,014 60%
Validation set 4,669 10%
Testing set 14,007 30%

Fig. 2: Autoencoder feature extractor

3.3 Autoencoder to Extract Features
An autoencoder model is based on an encoder-decoder A brief discussion of the structure and functionalities of
paradigm where the encoder compresses data into a lower- autoencoder layers is given as follows.
dimensional representation and the decoder is tuned to either • Convolutional layer: This type of layer consists of
reconstruct the initial input from the compressed data with trainable filters and feature maps. Trainable filter is
minimum loss or a representation of the input. The desired a set neurons with weights as its input and a value
output is a feature vector for every timestamp. as its output. The feature map is the result of a
The proposed autoencoder for feature extraction includes filter applied to the previous layer using convolution
three different neural network layers including convolutional, operator.Mathematical representation of a convolutional

ISBN: 1-60132-492-8, CSREA Press ©

6 Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'19 |

layer is  long-term dependencies [25], i.e., the derivative of the error

hkl+1 = σ( xjl ⊗ wlk +b )
k
(3) with respect to weights gets close to zero or infinity after a
j∈J short period of the time. This problem makes the networks
hard to train for long term dependencies. Hochreiter and
where hkl+1 is the latent representation of kth feature
Schmidhuber [26] were able to overcome this problem in
map in layer l + 1, σ is a non-linear activation function
their well-known work on LSTM RNN. LSTM RNN uses
(in this research it is a ReLu function), xjl is the jth
trainable memory cells called LSTM cells instead of simple
feature map of the output in layer l, wlk is the kth filter
neurons. These memory cells have three trainable gates
weight for the layer l and bk is the bias parameter.
including input, output, and forget gates. These gates have
In this equation, ⊗ represents the 2D convolution
the ability to add or remove information, thus, avoid long
operation.
term dependencies. A large number of applications have
• Max-Pooling layer: The max-pooling layer takes n × m
performed better than their competitors using such LSTM
region (in this research 2×1) and output a single value,
networks [21]. Fig. 3 shows an LSTM cell, the gates,
which is the maximum in that region. For instance, if
and their output computations are given in the following
their input layer is an N × M layer, they will then
equations.
output a N/n × M/m layer, as each m × n block is
reduced to just a single value via the max function. it = σ(Wxi xt + Whi ht−1 + Wci ct−1 + bi ) (6)
• UpSampling: Our upsampling function repeats the rows
and columns of the data by 2 and 1 respectively. ft = σ(Wxf xt + Whf ht−1 + Wcf ct−1 + bf ) (7)
In summary, the only input to the proposed neural network ct = ft ct−1 + it tanh(Wxc xt + Whc ht−1 + bc ) (8)
is the normalized ECG data points. Furthermore, convo-
lutional autoencoder model has been used to extract the ot = σ(Wxo xt + Who ht−1 + Wco ct−1 + bo ) (9)
hierarchical structure of ECG cardiac waves. The encoder is
ht = ot tanh(ct ) (10)
two convolutional layers with alternating max-pooling layers
that convert the input dimensions from 1000×1 to 250×32. where, σ is the logistic sigmoid function, and i, f , o, a
An SRCNN has been adopted for the decoder to extract the and c are the input gate, forget gate, output gate, cell input
hierarchical structure of ECG signal which converts 250×32 activation, and cell state vectors, respectively, and all of them
shrunk data to 1000 × 16 feature vector. Fig. 2 shows the are the same size as the hidden vector h. Wci , Wcf , Wco
details of the designed deep convolutional autoencoder. are weight matrices for peephole connections [22].
3.4 Bidirectional Long Short-Term Memory
Recurrent Neural Network
A shortcoming of traditional neural networks is that it
fails to classify an event based on prior observations [22];
however, recurrent neural networks are intrinsically fit for
segmenting ECG signals. In a conventional recurrent neu-
ral network, the input x = (x1 , x2 , ..., xT ) feeds to the
network, and RNN computes the hidden vector sequence,
h = (h1 , h2 , ..., hT ), and the output vector sequence, y =
(y1 , y2 , ..., yT ), from t = 1, ..., T while T is the number of
timestamps. The hidden vector h and the output vector y
are as follows Eq. 4 and Eq. 5 respectively.
ht = H(Wxh xt + Whh ht−1 + bh ) (4)
yt = Why ht + by (5)
where, W denotes weight matrices, b denotes bias vectors, Fig. 3: LSTM cell [22]
and H denotes hidden layer function.
The novelty of this work is to use recurrent neural Another aspect of RNN is that only the prior data is used.
networks to classify each data timestamp of an ECG signal However, in many cases, future data is available and can be
into one of the four categories, namely, the P-wave, the QRS- used as an informational source. Schuster et al. introduced
complex, the T-wave, and the Neutral. Thus, as a result of a Bidirectional RNN [27], which uses both directions of the
ECG segmentation, classification of every data timestamp data, prior and future data points, in two separate hidden
can be achieved. Because of a problem called vanishing- layers. Graves et al. used LSTM in a Bidirectional RNN and
exploding gradients, conventional RNNs fails in dealing with introduced BLSTM [28]. Therefore, BLSTM is an RNN that

ISBN: 1-60132-492-8, CSREA Press ©

Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'19 | 7

uses LSTM cells and computes both forward and backward Table 2: Hybrid-ECG-SegNet Layers
hidden sequences. Layer Category Description of the layer Size
As mentioned earlier, it is useful to find QRS-complex Input Input ECG raw signal 1000 x 1
Convolutional layer
using prior sample points such as P-wave samples and future Layer 1
with 16 filters
1000 x 16
data points such as S-wave samples. Thus, BLSTM becomes Layer 1 Max Pooling of Size 2 x 1 500 x 16
a very viable approach to be explored for ECG segmentation Convolutional layer
Layer 2 500 x 32
with 32 filters
task. Given the rationale of the components, we define a new Layer 2 Max Pooling of Size 2 x 1 250 x 32
neural network architecture in the next section. Convolutional layer
Layer 3 Atuoencoder 250 x 32
with 32 filters
3.5 New Hybrid-ECG-SegNet Architecture Layer 3 Upsampling of Size 2 x 1 500 x 32
Layer 4 Conv. layer with 16 filters 500 x 32
Model Layer 4 Upsampling of Size 2 x 1 1000 x 16
Time Distributed
Layer 4 1000 x 16
Dense Layer
Layer 5 Sequence BDLSTM Layer 150
Layer 6 Learner BDLSTM Layer 75
Time Distributed
Layer 7 Output 1000 x 4
Dense Layer

consisting of an encoder and a decoder is applied. The

encoder compresses the output of the first layer data to
250×32 dimension sizes through its bottleneck architecture.
The encoder consists of two convolutional layers with alter-
nating max-pooling layers. A decoder is positioned after the
encoder. The decoder decompresses the encoder data back
to 1, 000 × 16 dimension sizes. The process of encoding and
decoding allows extracting the hierarchical structure of the
ECG signal that represents the essential information of the
ECG signal. As a result, The autoencoder generates a feature
vector of size 16 for every timestamp and the feature vectors
are the input to the sequence learner.
Considering the sequence learner component, the first
layer is the feature vector series layer which takes the
autoencoder feature vector series of size 1, 000 × 16. The
next hidden layer is a BLSTM layer of size 150 cells.
This is followed by another BLSTM of size 75 cells. The
next layer is the output layer that classifies every data
point in the time series into four categories. This layer is
called time distributed output layer and is applied to every
timestamp. Each ECG signal has 1, 000 samples, so the
output dimension is 1, 000 × 4. The details and overall
structure of the autoencoder and sequence learner are shown
in Fig. 4.
In summary, the Hybrid-ECG-SegNet is seven-layer deep
and consists of a deep convolutional autoencoder and a
sequence learner. Specifically, the Hybrid-ECG-SegNet input
is a 1, 000 × 1 dimension, which includes 1, 000 sample
points of normalized ECG signals. The output of the signal
comprises of a 1, 000 × 4 matrix producing probability over
Fig. 4: Hybrid-ECG-SegNet Architecture four categories for a 1, 000 data samples and Table 2 explains
the details of every layer in the hybrid model.
As discussed in Subsection 3.3, the input to the network
is the normalized 1, 000 × 1 ECG data points. The first 3.6 Training Experiment
convolutional layer has 16 feature maps which transform the As mentioned earlier, the training set includes 28, 014
input to 1, 000 × 16 dimension sizes. Then an autoencoder ECG segments of size 1, 000 × 1. The target data is a

ISBN: 1-60132-492-8, CSREA Press ©

8 Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'19 |

1, 000×4 matrix, which is the annotated class obtained from Table 4: Segmentation Accuracy Comparison
QTDB. If a data point belongs to the first class, neutral, the Method P (%) QRS (%) T (%) Overall (%)
output data at that timestamp is [1, 0, 0, 0] vector. It gave Hybrid-ECG-
91.0 94.0 93.0 93.99
SegNet
the probability of 1 to the first class and the rest were 0. HMM on raw ECG
The Hybrid-ECG-SegNet is trained with Root Mean Square 5.5 79.0 83.6 56.03
[14]
Propagation (RMSProp) Optimizer [29] through 47 epochs HMM on wavelet
74.2 94.4 96.1 88.23
using mini-batch procedure of batch size 75. The training encoded ECG [14]
stopped after 47 epochs because the validation set error was
not improving after 10 epochs and this was the description of
the early stopping policy. RMSProp maintains per parameter The majority of research focuses on finding the cardiac
learning rate based on the average of the recent magnitude complex fiducial points and not segmenting ECG data points.
of the gradients for the parameter (weight). This approach Even though the Hybrid-ECG-SegNet task is different than
is recommended for non-stationary signals [29]. After 68 finding ECG cardiac waves location, it provides competitive
epochs training, the results showed 94.73% accuracy for accuracy in finding cardiac wave locations. The accuracies of
training set, 93.58% accuracy for the validation set, and finding waves regardless of segmentation for P-wave, QRS-
93.99% accuracy for the test set. complex, and T-wave are 96%, 99%, and 98%, respectively.

4. Results and Discussion

Accuracy, precision, recall and F1-score metrics are used
to compare our results with other research [30]. Accu-
racy is the ratio of correct positive and correct negative
classifications over the total positive and negative ground
truth. Precision is the ratio of the correct positive classified
ECG data points over all the correct and incorrect positive
classifications. Accuracy measures the performance of cor-
rect positive and correct negative classification. Precision
measures the performance of incorrect classifications. The
recall metric is the ratio of correct positive classifications
over all the positive ground truth. Thus, recall measures
the performance of correct classification. Finally, F1-score
metric is the weighted average of precision and recall, thus
it provides a benchmark metric to compare the performance
of different algorithms.
The results are presented in Table 3. For comparison
purposes, two other approaches are used. Hughes et al. [14]
used HMM to solve ECG segmentation with two approaches.
The first approach used the raw ECG signal and the sec-
ond approach used wavelet encoded ECG. Comparison of
Hybrid-ECG-SegNet and both HMM approaches is provided Fig. 5: Two sample results
in Table 4. As illustrated in Table 4, HMM with wavelet
encoded achieved slightly better accuracy in segmenting The majority of the other research focuses on finding
QRS-complex and T-wave, however, in all other cases and the cardiac complex fiducial points and not segmenting
overall, Hybrid-ECG-SegNet achieved better results. Fig. 5 every single data point of ECG independently. Even though
shows two samples of ground truth from the test set and the Hybrid-ECG-SegNet task is different than finding ECG
their results. P-wave, QRS-complex, and T-wave areas are cardiac waves location, it provides competitive accuracy in
represented by red, blue, and green areas, respectively. finding cardiac wave locations. Table 5 shows the accuracy
of finding cardiac waves.
Table 3: ECG segmentation results
Precision (%) Recall (%) F1-Score (%) Table 5: Wave identification results
Neutral 95 93 95 Cardiac wave Accuracy
P-wave 90 92 91 P-wave 0.96
QRS-complex 90 95 94 QRS-complex 0.99
T-wave 92 93 93 T-wave 0.98

ISBN: 1-60132-492-8, CSREA Press ©

Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'19 | 9

5. Conclusion [10] I. S. N. Murthy and U. C. Niranjan. Component wave delineation

of ecg by filtering in the fourier domain. Medical and Biological
Our work demonstrated that Hybrid-ECG-SegNet is a Engineering and Computing, 30(2):169–176, Mar 1992.
powerful network to yield very good results, as shown in [11] N. Ouyang, K. Yamauchi, and M. Ikeda. Training a NN with ECG to
diagnose the hypertrophic portions of HCM. 1:306–309 vol.1, 1998.
Table 3 and Table 4. Therefore, deep convolutional autoen- [12] Q. A. Rahman, L. G. Tereshchenko, M. Kongkatong, T. Abraham,
coder method can be used to extract features necessary for M. R. Abraham, and H. Shatkay. Identifying hypertrophic cardiomy-
delineating ECG cardiac waves and a sequence learner such opathy patients by classifying individual heartbeats from 12-lead ecg
signals. In 2014 IEEE International Conference on Bioinformatics
as LSTM model can be utilized to learn the temporal relation and Biomedicine (BIBM), pages 224–229, Nov 2014.
of the obtained features. The ability of detecting, delineating, [13] O. Sayadi and M. B. Shamsollahi. A model-based bayesian framework
and segmenting cardiac waves augments the possibility of for ecg beat segmentation. Physiological Measurement, 30:335, 2009.
[14] N. P. Hughes, L. Tarassenko, and S. J. Roberts. Markov models
making an impact on future research in cardiology. Fur- for automated ecg interval analysis. In Proceedings of the 16th
thermore, our approach has achieved satisfactory results, as International Conference on Neural Information Processing Systems,
shown in Table 3, Table 4, and Table 5. Therefore, deep NIPS’03, pages 611–618, Cambridge, MA, USA, 2003. MIT Press.
[15] R. Warner, Y. Ariel, M. Gasperina, and P. Okin. Improved electro-
convolutional autoencoder method can be used to extract cardiographic detection of left ventricular hypertrophy. Journal of
features necessary for delineating ECG cardiac waves and electrocardiology, 35 Suppl:111–5, 2002.
a sequence learner such as LSTM model can be utilized to [16] W. Kaiser, T. S. Faber, and M. Findeis. Automatic learning of
rules. A practical example of using artificial intelligence to improve
learn the temporal relation of the obtained features. The abil- computer-based detection of myocardial infarction and left ventricular
ity of detecting, delineating, and segmenting cardiac waves hypertrophy in the 12-lead ECG. Journal of electrocardiology, 29
augments the possibility of making an impact on future Suppl:17–20, 1996.
[17] S. Kiranyaz, T. Ince, R. Hamila, and M. Gabbouj. Convolutional
research in cardiology. By combining the vital information neural networks for patient-specific ecg classification. In 2015 37th
of waveforms with other methods in recognizing symptoms, Annual International Conference of the IEEE Engineering in Medicine
more accurate heart-related diseases can be diagnosed and and Biology Society (EMBC), pages 2608–2611, Aug 2015.
[18] P. Rajpurkar, A. Y. Hannun, M. Haghpanahi, C. Bourn, and A. Y.
high-throughput, automated ECG diagnostic systems can be Ng. Cardiologist-level arrhythmia detection with convolutional neural
developed to serve the need of large population screening networks. CoRR, abs/1707.01836, 2017.
for disease prevention [1]. [19] H. Abrishami, M. Campbell, C. Han, R. Czosek, and X. Zhou. P-
qrs-t localization in ecg using deep learning. In 2018 IEEE EMBS
International Conference on Biomedical Health Informatics (BHI),
Acknowledgment pages 210–213, March 2018.
[20] H. Abrishami, M. Campbell, C. Han, R. Czosek, and X. Zhou.
We thank Ohio Supercomputer Center [31] for provid- Supervised ecg interval segmentation using lstm neural network. In
ing high-performance computation power for this research. BIOCOMP’18, pages 71–77, March 2018.
[21] A. Graves, N. Jaitly, and A. r. Mohamed. Hybrid speech recognition
This research was supported by Children’s Cardiomyopathy with deep bidirectional lstm. In 2013 IEEE Workshop on Automatic
Foundation Grant. Speech Recognition and Understanding, pages 273–278, Dec 2013.
[22] X. Li and X. Wu. Long short-term memory based convolutional
recurrent neural networks for large vocabulary speech recognition.
References CoRR, abs/1610.03165, June 2016.
[23] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. Ch.
[1] M. J. Campbell, X. Zhou, C. Han, H. Abrishami, G. Webster, C. Y. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E.
M., C. T. Sower, J. B. Anderson, T. K. Knilans, and R. J. Czosek. Stanley. PhysioBank, PhysioToolkit, and PhysioNet: Components of
Pilot study analyzing automated ecg screening of hypertrophic car- a new research resource for complex physiologic signals. Circulation,
diomyopathy. Heart Rhythm, 14(6):848–852, 6 2017. 101(23):e215–e220, June 2000.
[2] C. Dong, C. C. Loy, K. He, and X. Tang. Image super-resolution using [24] A. K. Dohare, V. Kumar, and R. Kumar. An efficient new method for
deep convolutional networks. IEEE Transactions on Pattern Analysis the detection of QRS in electrocardiogram. Computers and Electrical
and Machine Intelligence, 38(2):295–307, Feb 2016. Engineering, 40(5):1717–1730, 2014.
[3] Q. Meng, D. Catchpoole, D. Skillicom, and P. J. Kennedy. Relational [25] Y. Bengio, P. Simard, and P. Frasconi. Learning long-term dependen-
autoencoder for feature extraction. In 2017 International Joint cies with gradient descent is difficult. IEEE Transactions on Neural
Conference on Neural Networks (IJCNN), pages 364–371, May 2017. Networks, 5(2):157–166, Mar 1994.
[4] A. Graves. Supervised Sequence Labelling with Recurrent Neural [26] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural
Networks. Studies in Computational Intelligence. Springer, Berlin, Comput., 9(8):1735–1780, November 1997.
2012. [27] M. Schuster and K. K. Paliwal. Bidirectional recurrent neural
[5] I. Beraza and I. Romero. Comparative study of algorithms for ecg networks. IEEE Transactions on Signal Processing, 45(11):2673–
segmentation. Biomed. Signal Proc. and Control, 34:166–173, 2017. 2681, Nov 1997.
[6] A. Karimipour and M. R. Homaeinezhad. Real-time electrocardiogram [28] A. Graves and J. Schmidhuber. Framewise phoneme classification
P-QRS-T detection-delineation algorithm based on quality-supported with bidirectional lstm networks. In Proceedings. 2005 IEEE Interna-
analysis of characteristic templates. Computers in Biology and tional Joint Conference on Neural Networks, 2005., volume 4, pages
Medicine, 52:153–165, 2014. 2047–2052 vol. 4, July 2005.
[7] J. P. Martinez, R. Almeida, S. Olmos, A. P. Rocha, and P. Laguna. [29] S. Ruder. An overview of gradient descent optimization algorithms.
A Wavelet-Based ECG Delineator Evaluation on Standard Databases. CoRR, abs/1609.04747, 2016.
IEEE Transactions on Biomedical Engineering, 51(4):570–581, 2004. [30] Alaa Tharwat. Classification assessment methods. Applied Computing
[8] R. Tafreshi, A. Jaleel, J. Lim, and L. Tafreshi. Automated analysis of and Informatics, 2018.
ECG waveforms with atypical QRS complex morphologies. Biomed- [31] Ohio Supercomputer Center. Ohio supercomputer center.
ical Signal Processing and Control, 10:41–49, 2014. http://osc.edu/ark:/19495/f5s1ph73, 1987.
[9] J. Pan and W. J. Tompkins. A real-time qrs detection algorithm. IEEE
Transactions on Biomedical Engineering, BME-32(3):230–236, 1985.

ISBN: 1-60132-492-8, CSREA Press ©