Int'l Conf.

Bioinformatics and Computational Biology | BIOCOMP'18 | 71

Supervised ECG Interval Segmentation Using

LSTM Neural Network
Hedayat Abrishami Matthew Campbell Richard Czosek
Chia Han Children’s Hospital of Philadelphia Heart Institute
Xuefu Zhou Philadelphia, PA, USA Children’s Hospital Medical Center
Department of EECS Email: [email protected] Cincinnati, OH, USA
University of Cincinnati Email: [email protected]
Cincinnati, OH, USA
Email: [email protected]
Email: [email protected]
Email: [email protected]

Abstract—Segmenting electrocardiogram (ECG) into its im-

portant components is crucial to the field of cardiology and
pharmaceutical studies, because analyses of ECG segments can
be used to predict heart symptoms and the effects of cardiac
medications. For each study, thousands of ECG signal points need
to be analyzed and segmented. Despite of the success of using
deep learning (DL) methods in multiple studies on classifying
the heart condition, there are still lacking DL-based methods
to characterize ECG temporal features. This paper describes a
novel ECG segmentation method based on the recurrent neural
network (RNN) with long short-term memory (LSTM) layers. In
this model, each ECG sample is classified into one of the four
categories: P-wave, QRS-wave, T-wave, and neutral (others). Our
work shows that DL sequence learning methods outperform a
traditional Markov model in terms of accuracy and using simple
local features instead of complicated features, such as wavelet
encoding. Particularly on T-wave segmentation, our approach
can achieve an accuracy of 90%, compared to that of 74.2% Fig. 1. Typical cardiac complex with annotated waves and peaks [1]
using Markov models.

I. I NTRODUCTION eventual death relies on finding the T-wave inversion, where

Analyzing critical segments of the ECG waveform in ECG the T-wave is inverted rather than in its normal rising form
signals is crucial in determining the conditions of heart for [2]. In another case, ST depression, a cause of sudden cardiac
diagnosing diseases or analyzing the effects of cardiac medica- arrest, relies on finding the end of the QRS-wave and the
tions. All these analytical tasks require finding a large quantity beginning of the T-wave [2]. Thus, a reliable automated ECG
of waveform patterns that can be identified as abnormalities signal processing approach should be capable of segmenting
in ECGs. Many of the pattern features are subtle and would ECG cardiac complexes accurately.
require an expert clinician to recognize and spot them in ECG ECG segmentation has two main challenges. The first
wave intervals. Utilizing clinicians is not economically feasible challenge is the variety of wave formations and various ECG
on a large scale. Thus, an automated approach to segment ECG abnormality patterns [3]. The second challenge is the noise
signals is very promising. In addition, an accurate automated generated by ECG monitoring devices that sample the elec-
ECG segmentation method would provide more cost-effective trical activities of the heart muscle [4]. A robust and reliable
and reliable results for patient diagnostics, especially for mass automated approach should overcome these two barriers. Over
screening applications and in drug development. the years, the research community has developed two general
A cardiac complex is composed of several wave compo- approaches for ECG segmentation. The first type relies on
nents, and three of them are of high significance. These three identification of important peaks in the ECG waveforms and
waves are the P-wave, the QRS-wave, and the T-wave, as determination of other points relative to those peaks. The other
shown in Figure 1. The main goal of ECG segmentation is to type is to classify every ECG data point into one of the cardiac
detect and localize the QRS-wave, meanwhile, other segments waves [5]. Our work focuses on segmenting ECG signals by
such as T-wave and P-wave are also of high importance. For labeling each data point into one of the four ECG cardiac
instance, one of the symptoms of sudden cardiac arrest and waves, P-wave, QRS-wave, T-wave, and neutral.

ISBN: 1-60132-471-5, CSREA Press ©

72 Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'18 |

ECG segmentation can be performed in two steps. The III. M ETHODOLOGY

first step is to obtain relevant features that discriminates In this work, a LSTM RNN architecture is proposed to
the wave locations from the rest of the signal. The second segment ECG intervals. In Subsection III-A, the order and
step is to classify these points with the information revealed attributes of ECG intervals are discussed. Subsection III-B
from the features. These features can be obtained by various introduces the data sets including training, validation and
filters or transformation methods [6], [7], [8]. Usually filters test data used in this work [23]. Subsection III-C describes
are smoothing filters [9], first-order derivative and second- the features extracted from ECG signal to feed into our
order derivative [10], and etc. Transformations include Fourier neural network model. Section III-D reviews the Bidirectional
transform [7], wavelet transform [11], and etc. The second Long Short-Term Memory Neural Network (BLSTM-RNN).
step in ECG segmentation has used rule-based methods [12], Section E introduces the novel architecture of ECG-SegNet
Hidden Markov Model (HMM) [5], neural networks [13], and its post-processing step. Unlike the traditional Recurrent
evolutionary algorithms [14], and Bayesian techniques [15]. Neural Networks (RNN), this type of network is capable of
This paper is organized as follows: Section II briefly reviews learning long temporal dependencies, which makes it suitable
the related works. Section III introduces ECG-SegNet, our for ECG segmentation [24]. Finally, Subsection III-E describes
novel long short-term memory neural network model for ECG experiments and the convergence of ECG-SegNet. The result
segmentation. Section IV presents experimental results and demonstrates the strength of ECG-SegNet compared to the
discussion, followed by conclusion in Section V. other sequence learners such as HMM for the same task of
II. R ELATED W ORKS ECG interval segmentation.

Scientists have developed multiple automated approaches A. ECG Intervals

to detect various waveforms in ECG signals. An extensive The wave complexes contain essential information, such
literature review on ECG segmentation can be found in as shape formation, interval duration, and amplitudes. Any
[16]. Generally, there are three main approaches, namely, the abnormality in the waves, such as ST depression, T-inversion,
derivative-based [4], [17], wavelet-filters [18], and amplitude- long QT-interval, and so on, can provide meaningful informa-
based methods [3]. Martinez et al. [18] used a discrete wavelet- tion for cardiac diseases. With characterization of key values
based method for extracting temporal features. Pan-Tompkins of wave parameters and the information that they carry, it
[10], developed one of the most famous derivative-based is possible for physicians to make decisions that lead to
methods to find QRS-waves. Furthermore, for the purpose of differential diagnoses. There are many different shapes that
classification, machine learning methods have been used in can be found in ECG data. Normal QRS-waves can have nine
this field of study, which include Neural Network (NN) [13], different shapes [3]. P-waves and T-waves can also appear in
Random Forest [19], Support Vector Machine (SVM) [19], different forms and amplitudes. In addition, all these waves
Naive Bayes [15], HMM [5], rule-based methods [12], linear can have abnormal shapes and formations. Detecting wave
discriminants [20], and logistic regression [21]. complexes that can relate to serious illnesses, such as sudden
With recent emergence of DL methods in various aspects cardiac death (SCD), is critical.
of signal processing and data analysis, DL methods have been A cardiologist identifies cardiac waves in relation to each
applied to ECG pattern recognition. In [22], DL methods were other. For example, following a QRS-wave location is probably
used to classify ECG signals into normal and abnormal ECG. an S-wave. Thus, knowledge of the waves prior location
In these works, the focus was to extract hierarchical features is essential to predicting the consequent wave. Likewise,
of ECG signals in order to classify heart disease symptoms. each individual waves formation also affects the formation of
However, there is no specific DL study for finding and other waves. Therefore, a method that is capable of keeping
locating the major wave components or segmenting the cardiac persistent memory, i.e., a recurrent neural network, can be a
complexes. This task is of high importance in cardiology viable solution to this type of time-dependent problem. The
communities since every cardiologist refers to them in their recurrent neural network creates a loop to pass the information
diagnoses. Furthermore, the rule-based classification methods from one timestamp to another, which allows it to learn a time
have to come with different criteria with the changes in the series [25].
formation of cardiac waves. HMM is a capable time series
learner and ECG signal is a time series, thus, this method B. Data Sets
can be utilized to learn cardiac wave formations. However, The data used for this study is the QT database (QTDB). It is
HMM has its limitation of learning a combination of various designed to detect and segment ECG waveforms. The QTDB
patterns as typically shown in ECG signals. On the other hand, was produced by PhysioNet [23] and has a large collection
LSTM RNN is a time series learner capable of learning time of recorded physiological signals sampled at 250Hz. This
series patterns and their combinations. This paper introduces database includes over 105 two-channeled ECG recordings,
a newly-proposed LSTM RNN architecture to segment ECG each 15-minutes in duration and it is chosen to include a broad
intervals. Our findings indicated that this new architecture variety of P, QRS, ST, and T morphologies [23]. This dataset
outperforms the traditional ECG segmentation analysis using thus allows researchers to perform research on ECG signal
HMMs. delineation.

ISBN: 1-60132-471-5, CSREA Press ©

Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'18 | 73

For our current study, every recording is divided into 500 yt = Why ht + by (2)
data points sampled at 250Hz. Within every segment, one
where, W denotes weight matrices, b denotes bias vectors,
or more cardiac complexes can be found, which makes the
and H denotes hidden layer function.
segmenting task more challenging. The 500 ECG data points The novelty of this work is to use recurrent neural networks
and their extracted features are the inputs to ECG-SegNet to classify each data point of an ECG signal into one of the
model, and correspondingly, the related annotations are the four categories, namely, the P-wave, the QRS-wave, the T-
output targets for the ECG-SegNet model. In total, there are wave, and the neutral. Thus, as a result of ECG segmentation,
64, 040 sets of 500 data points of ECG that have been extracted classification of every data point can be achieved. One of
from the QTDB to be used as inputs. the main shortcomings of the conventional recurrent neural
In our experiments, three different sets - training, validation,
networks in dealing with long term dependencies is that these
and test sets - have been created using all the extracted ECG
networks encounter a problem called vanishing-exploding gra-
parts. These sets are mutually exclusive, indicating there are
dients [26], i.e., the derivative of error with respect to weights
no identical segments from one recording to another. Table I
gets close to zero or infinity after a short period of the time.
illustrates the training, validation and test data we used in this
This problem makes the networks hard to train for long term
paper.
dependencies. Hochreiter and Schmidhuber [27] were able to
overcome this problem in their well-known work on Long
TABLE I
DATASET Short-Term Memory recurrent neural networks (LSTM RNN).
LSTM RNN uses trainable memory cells called LSTM cells
Dataset Number of samples Percentage instead of simple neurons. These memory cells have three
Training set 51,419 55%
Validation set 9,350 10% trainable gates including input, output, and forget gates. These
Testing set 32,721 35% gates have the ability to add or remove information, thus,
avoid long term dependencies. A large number of applications
have performed better than their competitors using such LSTM
C. Extracted Features
networks [28]. Figure 2 shows a LSTM cell, the gates and their
In this work, in addition to raw ECG data points, three output computations are given in the following equations:
other features are extracted using different filtering kernels,
such as the local average of a data point and the first and it = σ(Wxi xt + Whi ht−1 + Wci ct−1 + bi ) (3)
second derivatives of a data point. Hence, for every data point, ft = σ(Wxf xt + Whf ht − 1 + Wcf ct−1 + bf ) (4)
a feature vector of size four is created to feed to the network, as
explained in Subsection III-E. As a result of applying feature ct = ft ct−1 + it tanh(Wxc xt + Whc ht−1 + bc ) (5)
extraction to 500 ECG data points, the complete input to the ot = σ(Wxo xt + Who ht−1 + Wco ct−1 + bo ) (6)
ECG-SegNet becomes a matrix of 500 × 4 dimension. Table II
ht = ot tanh(ct ) (7)
illustrates the kernels that are used to convolve with the raw
ECG signal. where, σ is the logistic sigmoid function, and i, f , o, a and c
are the input gate, forget gate, output gate, cell input activation,
TABLE II and cell state vectors, respectively, and all of them are the
K ERNELS TO EXTRACT FEATURES same size as the hidden vector h. Wci , Wcf , Wco are weight
No. Type Kernel Length matrices for peephole connections [25].
1 Raw ECG value None Not applicable Another aspect of RNN is that only the prior data is used.
2 Smoothing [1,1,1,1,1,1,1,1,1,1] 10 However, in many cases, the future data is available and can
3 First Derivative [1, 1, 1, 0, -1, -1, -1] 7
4 Second Derivative [1, 1, 1, -6, 1, 1, 1] 7
be used as an informational source. Schuster et al. introduced
a Bidirectional RNN (BRNN) [29], which uses both directions
of the data, prior and future data points, in two separate hidden
D. Bidirectional Long Short-Term Memory Recurrent Neural layers. Graves et al. use LSTM in a BRNN and introduce
Network Review Bidirectional LSTM (BLSTM) [24]. Therefore, BLSTM is a
As mentioned earlier, a shortcoming of traditional neural recurrent neural network that uses LSTM cells and computes
networks is that it fails to classify an event based on prior both forward and backward hidden sequences. By stacking up
observations [25]; however, recurrent neural networks are these types of layers, a new deep network, called the Deep
intrinsically fit for segmenting ECG signals. In a conventional Bidirectional LSTM (DBLSTM), is obtained.
recurrent neural network, the input x = (x1 , x2 , ..., xT ) feeds Utilizing DBLSTM in ECG segmentation helps to classify
to the network, and RNN computes the hidden vector se- a signal sample based on prior and future sample points, i.e.,
quence, h = (h1 , h2 , ..., hT ), and the output vector sequence, finding QRS-wave using prior sample points such as P-wave
y = (y1 , y2 , ..., yT ), from t = 1, ..., T while T is the number samples and future data points such as S-wave samples. Thus,
of timestamps as, DBLSTM becomes a very viable approach to be explored
for ECG segmentation task. We will use this to define new
ht = H(Wxh xt + Whh ht−1 + bh ) (1) network architecture in the next section.

ISBN: 1-60132-471-5, CSREA Press ©

74 Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'18 |

Fig. 2. LSTM cell [25]

E. ECG-SegNet Architecture Model and Post-Processing Step

Based on the aforementioned rationale for a suitable net-
work using DBLSTM to classify ECG waveform data points,
a new architecture, called ECG-SegNet, is proposed, as shown
in Figure 3.
This architecture contains the following layers.
From bottom up, the first layer is the input layer,
x = (x1 , ..., xt , xt+1 , ..., x500 ), which takes the raw ECG
signal of size 500 and three additional extracted features per
data point, explained in Section III.C. Therefore, the input is
500 × 4 time series. The hidden layer is a BLSTM layer. On
every timestamp , there are two different hidden LSTM layers
including forward hidden layer and one backward hidden
layer. On every set, this layer has 250 LSTM cells, which Fig. 3. Proposed deep bidirectional RNN architecture
suggests that it has a total of 500 hidden LSTM cells. This
is followed by another BLSTM of size 250, and each hidden
LSTM layer has 125 LSTM cells. The next layer is the output where ȳ t is the vector of output activations before they have
layer that classifies every data point in time series into four been normalized with the softmax function. These derivatives
categories. This layer is called time distributed output layer are then fed back through the network using backpropagation
and is applied on every timestamp . Each ECG signal has 500 through time to determine the weight gradient [25], [28].
samples, so the output dimension is 500 × 4. For ECG signal, As explained earlier, every data point in the ECG signal is
x of length T , the network produces a length of T output classified into one of the four different categories. Segmenta-
sequence y, where each yt defines a probability distribution tion implies that if the input has the temporal dimension of
over the |K| possible states where K = {1, 2, 3, 4} and kK: size T , the output has the same dimension of size T . In this
that is, ytk (the k th element of yt ) is the network’s estimate work, the ECG input has 500 data points, which means output
for the probability of observing state k at time t, given x. The dimension is equal to 500 × |K|, where |K| is the number of
network is trained to minimize the negative log-probability of categories. Eq. 9 computes the probability of input x at time t,
the target sequence using a softmax output layer, Eq. 8 and belonging to class k, with k being one of the four categories.
Eq. 9. With length T and target sequence z, the network is The final output for each data point is the class with the highest
trained to minimize the error function shown in Eq. 9. probability. In summary, Table III illustrates the architecture of
T
the ECG segmentation network, the ECG-SegNet. It is three
X layers deep, consisting of the input layer, two bidirectional
−log P r(z|x) = − log ytzt (8)
t=1
LSTM layers, and an output layer. More specifically, the ECG-
SegNet input is a 500 × 4 dimension, which includes 500 data
The error derivatives at the output can be obtained as points each of raw ECG signals, smoothed ECG signals, the
∂log P r(z|x) first derivative of the ECG signal, and the second derivative
− = ytk − δk,zt (9) of the ECG signal. The output of the signal comprises of a
∂ȳtk

ISBN: 1-60132-471-5, CSREA Press ©

Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'18 | 75

500 × 4 neuron-producing probability over a 500-unit timeline

and is classified into four categories: neutral, P-wave, QRS-
wave, and T-wave.

TABLE III
D EEP BIDIRECTIONAL LSTM FOR ECG SEGMENTATION

Layer Description of the layer LSTM Cells

Input ECG raw signal and its features 500 ×4
Layer 1 Bidirectional LSTM with 250 cells each 2 ×250
Layer 2 Bidirectional LSTM with 125 cells each 2 ×125
Output Time distributed dense layer 500 ×4

After generating the output from the network, a post-

processing step is applied. A filter of size 17 is applied to
it. If the beginning and the end of the output under this filter
belong to the same cardiac wave class, any output under this
window is then assigned to the class of the start and end of
this filter.
F. Training Experiment
Fig. 5. Loss curve
The training set includes 51, 419 ECG segments of size
500 × 4. The target data is a 500 × 4 matrix, which is the
annotated class obtained from QTDB. If a data point belongs Hughes et al. [5] used HMM to solve ECG segmentation with
to the first class, neutral, the output data at that timestamp is two approaches. The first approach used raw ECG signal and
[1, 0, 0, 0] vector. It gave the probability of 1 to the first class the second approach used wavelet encoded ECG. Comparison
and the rest were 0. The ECG-SegNet is trained with Adam of ECG-SegNet and both HMM approaches is provided in
Optimizer [30] through 68 epochs using mini-batch procedure Table V. It was shown that ECG-SegNet performs better
of batch size 250. The training stopped after 68 epochs because in term of accuracy. Using HMM with wavelet encoded
the gap between training set error was getting smaller and in segmenting QRS-wave and T-wave gave better accuracy.
validation set error was getting larger and this divergence is However, in all other cases and also in overall results, ECG-
a sign of overfitting. After training, the results showed 94.6% SegNet outperforms HMM approaches.
accuracy for training set, 93.8% accuracy for validation set, Figure 6 shows two samples from the test set and their
and 93.7% accuracy for test set. Figure 4 shows the accuracy related results. It shows the accuracy of ECG-SegNet. P-wave,
rates and Figure 5 shows the error rates through 68 epochs for QRS-wave, and T-wave areas are represented by red, blue,
validation and training sets. and green regions. In the example positioned at top, the effect
of post-processing step can be seen. In the bottom example,
the original result of ECG-SegNet matches the ground truth
without any practical use of post-processing.

TABLE IV
ECG SEGMENTATION RESULTS

Precision Recall F1-Score

Neutral 0.95 0.95 0.95
P-wave 0.92 0.90 0.91
QRS-wave 0.94 0.95 0.94
T-wave 0.90 0.92 0.91
avg/total 0.94 0.94 0.94

TABLE V
S EGMENTATION ACCURACY COMPARISON

Method P (%) QRS (%) T (%) Overall (%)

ECG-SegNet 92.0 94.0 90.0 92.0
Fig. 4. Accuracy curve HMM on raw ECG [5] 5.5 79.0 83.6 56.03
HMM on wavelet en-
74.2 94.4 96.1 88.23
coded ECG [5]
IV. R ESULTS
The detailed result of this segmentation is reported in Table The majority of the other researches focus on finding the
IV. For comparison purposes, two other approaches are used. cardiac complex fiducial points and not segmenting every

ISBN: 1-60132-471-5, CSREA Press ©

76 Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'18 |

single data point of ECG independently. Even though the V. C ONCLUSION

ECG-SegNet task is different than finding ECG cardiac waves To the best of our knowledge, there is not a DL-based
location, it provides competitive accuracy in finding cardiac method for ECG signal segmentation yet. Our work demon-
wave locations. Table VI shows the accuracy of finding waves strated that ECG-SegNet is a powerful network capable of
regardless of segmentation using ECG-SegNet. understanding the temporal ECG using only a few local
features to yield very competitive results.
The ability to delineate ECG cardiac waves augments the
TABLE VI
WAVE IDENTIFICATION RESULTS possibility of contributing future research in cardiology. By
combining the vital information of waveforms with other
Accuracy
P-wave 0.95
methods in recognizing symptoms, more accurate heart related
QRS-wave 0.98 diseases can be diagnosed, and high-throughput, automated
T-wave 0.97 ECG diagnostic systems can be developed to serve the need
of large population screening for disease prevention [2].

ACKNOWLEDGMENT
We thank Ohio Supercomputer Center [31] for providing
high performance computation power for this research.

R EFERENCES
[1] Ecg wave. http://lifeinthefastlane.com/ecglibrary/ basics/t-wave/. Ac-
cessed: 2017-09-30.
[2] M. J. Campbell, X. Zhou, C. Han, H. Abrishami, G. Webster, C. Y. M.,
C. T. Sower, J. B. Anderson, T. K. Knilans, and R. J. Czosek. Pilot study
analyzing automated ecg screening of hypertrophic cardiomyopathy.
Heart Rhythm, 14(6):848–852, 6 2017.
[3] R. Tafreshi, A. Jaleel, J. Lim, and L. Tafreshi. Automated analysis of
ECG waveforms with atypical QRS complex morphologies. Biomedical
Signal Processing and Control, 10:41–49, 2014.
[4] A. Karimipour and M. R. Homaeinezhad. Real-time electrocardiogram
P-QRS-T detection-delineation algorithm based on quality-supported
analysis of characteristic templates. Computers in Biology and Medicine,
52:153–165, 2014.
[5] N. P. Hughes, L. Tarassenko, and S. J. Roberts. Markov models for
automated ecg interval analysis. In Proceedings of the 16th International
Conference on Neural Information Processing Systems, NIPS’03, pages
611–618, Cambridge, MA, USA, 2003. MIT Press.
[6] B. Frénay, G. de Lannoy, and M. Verleysen. Emission Modelling for
Supervised ECG Segmentation using Finite Differences, pages 1212–
1216. Springer Berlin Heidelberg, Berlin, Heidelberg, 2009.
[7] I. S. N. Murthy and U. C. Niranjan. Component wave delineation of ecg
by filtering in the fourier domain. Medical and Biological Engineering
and Computing, 30(2):169–176, Mar 1992.
[8] S. K. Mukhopadhyay, M. Mitra, and S. Mitra. Time plane ecg feature
extraction using hilbert transform, variable threshold and slope reversal
approach. pages 1–4, December 2011.
[9] P. Laguna, N. V. Thakor, P. Caminal, R. Jané, Hyung-Ro Yoon,
A. Bayés de Luna, V. Marti, and Josep Guindo. New algorithm for
qt interval analysis in 24-hour holter ecg: performance and applications.
Medical and Biological Engineering and Computing, 28(1):67–73, Jan
1990.
[10] J. Pan and W. J. Tompkins. A real-time qrs detection algorithm. IEEE
Transactions on Biomedical Engineering, BME-32(3):230–236, March
1985.
[11] C. Li, C. Zheng, and C. Tai. Detection of ecg characteristic points using
wavelet transforms. IEEE Transactions on Biomedical Engineering,
42(1):21–28, Jan 1995.
[12] W. Kaiser, T. S. Faber, and M. Findeis. Automatic learning of rules. A
practical example of using artificial intelligence to improve computer-
based detection of myocardial infarction and left ventricular hypertrophy
in the 12-lead ECG. Journal of electrocardiology, 29 Suppl:17–20, 1996.
[13] 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.
[14] R. V. Andreao, B. Dorizzi, and J. Boudy. Ecg signal analysis through
hidden markov models. IEEE Transactions on Biomedical Engineering,
Fig. 6. Two sample results 53(8):1541–1549, Aug 2006.
[15] O. Sayadi and M. B. Shamsollahi. A model-based bayesian framework
for ecg beat segmentation. Physiological Measurement, 30(3):335, 2009.

ISBN: 1-60132-471-5, CSREA Press ©

Int'l Conf. Bioinformatics and Computational Biology | BIOCOMP'18 | 77

[16] I. Beraza and I. Romero. Comparative study of algorithms for ecg

segmentation. Biomed. Signal Proc. and Control, 34:166–173, 2017.
[17] M. R. Homaeinezhad, S. Atyabi, E. Daneshvar, A. Ghaffari, and M. Tah-
masebi. Discrete wavelet-aided delineation of PCG signal events via
analysis of an area curve length-based decision statistic. Cardiovascular
engineering (Dordrecht, Netherlands), 10(4):218–234, 2010.
[18] J. P. Martinez, R. Almeida, S. Olmos, A. P. Rocha, and P. Laguna.
A Wavelet-Based ECG Delineator Evaluation on Standard Databases.
IEEE Transactions on Biomedical Engineering, 51(4):570–581, 2004.
[19] Q. A. Rahman, L. G. Tereshchenko, M. Kongkatong, T. Abraham, M. R.
Abraham, and H. Shatkay. Identifying hypertrophic cardiomyopathy
patients by classifying individual heartbeats from 12-lead ecg signals. In
2014 IEEE International Conference on Bioinformatics and Biomedicine
(BIBM), pages 224–229, Nov 2014.
[20] P. De Chazal and R. B. Reilly. A patient-adapting heartbeat classifier us-
ing ECG morphology and heartbeat interval features. IEEE Transactions
on Biomedical Engineering, 53(12):2535–2543, 2006.
[21] R. Warner, Y. Ariel, M. Gasperina, and P. Okin. Improved electro-
cardiographic detection of left ventricular hypertrophy. Journal of
electrocardiology, 35 Suppl:111–5, 2002.
[22] S. Kiranyaz, T. Ince, R. Hamila, and M. Gabbouj. Convolutional neural
networks for patient-specific ecg classification. In 2015 37th Annual
International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC), pages 2608–2611, Aug 2015.
[23] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. Ch.
Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E.
Stanley. PhysioBank, PhysioToolkit, and PhysioNet: Components of a
new research resource for complex physiologic signals. Circulation,
101(23):e215–e220, June 2000.
[24] A. Graves and J. Schmidhuber. Framewise phoneme classification with
bidirectional lstm networks. In Proceedings. 2005 IEEE International
Joint Conference on Neural Networks, 2005., volume 4, pages 2047–
2052 vol. 4, July 2005.
[25] X. Li and X. Wu. Long short-term memory based convolutional recur-
rent neural networks for large vocabulary speech recognition. CoRR,
abs/1610.03165, June 2016.
[26] Y. Bengio, P. Simard, and P. Frasconi. Learning long-term dependen-
cies with gradient descent is difficult. IEEE Transactions on Neural
Networks, 5(2):157–166, Mar 1994.
[27] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural
Comput., 9(8):1735–1780, November 1997.
[28] A. Graves, N. Jaitly, and A. r. Mohamed. Hybrid speech recognition
with deep bidirectional lstm. In 2013 IEEE Workshop on Automatic
Speech Recognition and Understanding, pages 273–278, Dec 2013.
[29] M. Schuster and K. K. Paliwal. Bidirectional recurrent neural networks.
IEEE Transactions on Signal Processing, 45(11):2673–2681, Nov 1997.
[30] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization.
CoRR, abs/1412.6980, 2014.
[31] Ohio Supercomputer Center. Ohio supercomputer center.
http://osc.edu/ark:/19495/f5s1ph73, 1987.

ISBN: 1-60132-471-5, CSREA Press ©