You are on page 1of 10

Biomedical Signal Processing and Control 23 (2016) 42–51

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control


journal homepage: www.elsevier.com/locate/bspc

Detection of PVC in ECG signals using fractional linear prediction


Mohamed Lamine Talbi a , Philippe Ravier b,∗
a
Université de Bordj Bou Arréridj, 34030 Bordj Bou Arréridj, Algérie
b
Laboratoire PRISME, Université d’Orléans, 12 rue de Blois, BP 6744, 45067 Orléans, France

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we propose a modeling technique for the QRS complex based on the fractional linear
Received 24 April 2015 prediction (FLP). As a result of FLP modeling, each QRS complex is represented by a vector of three
Received in revised form 11 June 2015 coefficients. The FLP modeling evaluation is achieved in two steps. In the first step, the ability of the FLP
Accepted 15 July 2015
coefficients to efficiently model QRS complex waves is assessed by comparison with the Linear Prediction
(LP) coefficients through the signal-to-error (SER) values evaluated between the original waves and pre-
Keywords:
dicted ones. In the second step, the performance of several classifiers is used to evaluate the effectiveness
Fractional linear prediction (FLP)
and robustness of FLP modeling over LP modeling. Classifiers are fed by the three estimated coefficients in
QRS complex modeling
PVC detection
order to discriminate premature ventricular contraction (PVC) arrhythmia from normal beats. The study
ECG classification has successfully demonstrated that FLP modeling can be an alternative to the LP modeling in the field of
QRS complex modeling.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction signal modeling. An adaptive Hermite model estimation system


was presented in [7] for on-line beat-to-beat estimation of the fea-
The electrocardiogram (ECG) is an electrical signal generated by tures that describe the QRS complex with the Hermite model. In [8]
the heart’s muscular activity recorded by electrodes placed on the authors introduced a model-based dynamic algorithm for tracking
body surface. Modeling is a very useful tool for depicting this elec- the ECG characteristic waveforms using an extended Kalman
trical activity. Basically, modeling is a mathematical representation filter. Sayadi et al. [9] presented a technique based on a modified
of a process, action, or an experimental outcome [1]. extended Kalman filter structure for ECG signals denoising and
Signal modeling is one of the most widely used techniques in compression. In [10] authors presented a method for morpholog-
the automatic ECG signal analysis. Depending on the application, ical modeling of cardiac signals. A set of desired properties for a
there are many ways where ECG modeling is utilized, for example generic decomposition-based model was presented; the method
in noise removal, in signal compression or in features extraction was typically applied to the problem of ECG signal compression. In
for arrhythmia classification. A considerable amount of literature [11] authors used discrete Hermite functions for the compression
has been published on ECG signal modeling. In the next paragraphs of the QRS complexes. In [12] authors proposed an approach
we briefly present some published methods focusing on the ECG based on Generalized Orthogonal Forward Regression for ECG
modeling techniques [2–20]. modeling. In [13] an ECG modeling technique using the Sequential
In [2] authors proposed an approach based on the use of a sum Markov Chain Monte Carlo filter that could perform simultaneous
of N Gaussian shapes to model the normal and the PVC beats. model selection have been proposed. In [14] authors proposed
An ECG modeling technique based on the sequential Bayesian a method based on a dynamical model made of three coupled
methods was presented in [3], the aim of the method was to effec- ordinary differential equations which was capable of generating
tively model and adaptively select parameters of ECG signals. The realistic synthetic ECG signals. In [15] a method for dynamically
ECG signal was modeled using a parametric modeling technique modeling, estimating and classifying ECG signals was proposed.
based on signal dependent orthogonal transform in [4]. Numerous The key of the method was the use of the interacting multiple
studies [5,6] have used Hidden Markov Model approaches for ECG model framework. In [16] authors proposed an approach based on
the Hilbert Transform and a low order polynomial approximation
of the QRS complex waves. In [17] an integrated method for
∗ Corresponding author. Tel.: +33 238494863. clustering of QRS complexes was presented, each QRS complex
E-mail addresses: mltalbi@yahoo.fr (M.L. Talbi), philippe.ravier@univ-orleans.fr was decomposed into Hermite basis functions and the resulting
(P. Ravier). coefficients and width parameter were used to represent the QRS

http://dx.doi.org/10.1016/j.bspc.2015.07.005
1746-8094/© 2015 Elsevier Ltd. All rights reserved.
M.L. Talbi, P. Ravier / Biomedical Signal Processing and Control 23 (2016) 42–51 43

˛
complex. In [18] an Extended Kalman Filter was proposed for the a Dt f (t)
denotes fractional-order Differintegrals operator applied
filtering of noisy ECG signals. The method was based on a modified on the continuous function f (t).
nonlinear dynamic model. A number of studies [19,20] also used When applied to the discrete function f (n), a simplified formula
linear prediction (LP) for ECG modeling and feature extraction. of the fractional-order Differintegrals operator using the first-order
In recent years, there has been an increasing interest in frac- backwards approximation and the MacLaurin series truncation can
tional calculus and it has been more and more introduced into be written as [32]:
various engineering science. We can quote for example control [21],  
signal and image processing [22], biology [23], robotics [24], and 
n
˛
D˛ f (n) = (−1)j f (n − j) (3)
others. j
Recent studies in physiological signal modeling have introduced j=0

the use of the fractional linear prediction (FLP) to model speech


This approximation is rough and several other discretization
[25] and EEG signal [26]. Motivated by the effectiveness of the
methods can be used for digital implementation of the fractional-
FLP method for speech and EEG signals modeling, we propose in
order operator, some of those can be found in [33].
this study to investigate this technique as a new tool for efficiently
In the present work, we have used the first-order backwards
modeling ECG signals.
finite difference method. This approximation yields the follow-
The main objective of this study is to demonstrate the effective-
ing link between the Laplace variable s (used for continuous time
ness of the FLP modeling in the ECG signal modeling over linear
functions) and the variable z (used for T-sampled discrete time
prediction (LP) modeling, with attempt to show the utility of the
functions):
extracted FLP coefficients on the discrimination and the classifica-
 
tion of the PVC arrhythmia. To that end, a series of experiments k=0 ˛
will be done to discriminate the PVC arrhythmia from other beats.  ˛ +∞ (−1)
k
zk
˛ 1 − z −1 k
In addition, different types of classifiers for PVC discrimination will s = = (4)
be implemented and evaluated. T T˛

Finally, the discretization formula for the fractional deriva-


2. Methods
tion/integration operator is obtained using the Mac Laurin power
expansion as reported in Eq. (4). However, at each time n, the bino-
2.1. Fractional calculus
mial coefficient values make the old values of f (n) negligible in the
expansion and only recent values of f (n) are kept by truncation of
Fractional calculus is a mathematical discipline which deals
the sum expression to order v:
with derivatives and integrals of real or complex orders [27]. It
 
can be considered as a generalization of the integer-order calcu- k=0 ˛
lus. Several methods can be used to calculate the fractional-order  ˛  (−1)k zk
1 − z −1 k
derivation/integration. For continuous functions, the most usual ≈ (5)
definition is given by Riemann–Liouville. It permits to calculate the T T˛
fractional-order derivative of function f (t) as follow:
For each fractional order ˛, a linear filter which coefficients are

1
 d m t those of the power series expansion can be obtained. The fractional
˛
a Dt f (t) = (t − )m−˛−1 f () d (1) operator can therefore be characterized by its corresponding trans-
 (m − ˛) dt fer function. In order to simplify the regeneration of the results of
a
our work by other researchers, we have used the Ninteger toolbox
where  is the Gamma function, a Dt−˛ is the left-sided fractional [34] to calculate the fractional-order derivation/integration. In this
derivative of order ˛ of the function f (t) with m − 1 < ˛ < m. The work, the fractional-order operator has been approximated to the
fractional-order derivative is estimated in the interval [a; t], where order 6 of the Mac Laurin series expansion as shown later in Section
a is a real positive number. For more details about fractional cal- 2.4.2 [35].
culus, especially on the historical survey we refer the reader to
[28,29].
2.2. ECG FAR modeling
Note that the several definitions of the fractional operators can
lead to different results in engineering applications. In this paper,
In this section, we introduce the Fractional Autoregressive
we adopt the definitions of the Grünwald–Letnikov backward
model as an extension of the famous Autoregressive model. Both
derivatives because those definitions lead to a correct generaliza-
models are thus described in parallel.
tion of the current linear systems theory [30].
The fractional Differintegrals of continuous function f (t) accord-
ing to the Grünwald–Letnikov definition is given by [31]: 2.2.1. Autoregressive (AR) versus Fractional Autoregressive (FAR)
modeling
[(t−a)/h
]   Autoregressive (AR) modeling is a very useful mean for simply
˛ 1 j ˛
a Dt f (t) = lim ˛ (−1) f (t − jh) (2) approximating a stochastic process with a linear model [36]. The
h→0 h j
j=0 AR model of any process x (n) can be expressed by the following
formula:
where ˛ is a real value with 0 < ˛ < 1, [.] represents the nearest inte-
 the value between the brackets, h is the sampling period and
ger of 
P

˛ x (n) = − ai x (n − i) + w (n) (6)


is the combinatorial:
j i=1

  
1 for j=0 where w (n) is a zero mean white noise with variance  2 , the ai
˛
= ˛ (˛ − 1) . . . (˛ − j + 1) values are the AR coefficients of the model and P is the model order.
j for j = 1, 2, 3. . . Using the fractional-order Differintegrals operator applied on
j!
discrete time functions D˛ x (n) which are as defined in the
44 M.L. Talbi, P. Ravier / Biomedical Signal Processing and Control 23 (2016) 42–51

previous section, a similar expression can be derived for the Frac- However, in order to simplify the resolution of Eq. (11), let us
tional Autoregressive (FAR) model of the signal x (n): set the following matrix notations:


P x = [x (0) , x (1) , . . ., x (N − 1)]T
pi
x (n) = − ai D x (n) + w (n) (7)
i=1 w = [w (0) , w (1) , . . ., w (N − 1)]T
where w (n) is a zero mean white noise with variance  2 , the ai
values are the FAR coefficients of the model, the pi values are the

T
␪ = ˇ1 , ˇ2 , . . ., ˇQ
derivative orders and P is the model order.

pi = Dpi x
2.2.2. Linear prediction (LP) versus fractional-order linear
prediction (FLP)

The linear prediction (LP) of any stationary AR model signal H = p1 , p2 , . . ., pQ
x (n) can be interpreted as the prediction x̂ (n) of x (n) knowing the
previous samples x (n − i) , i = 1, . . ., P as: where the superscript ‘T’ denotes the transpose operation, Q is the
FAR model order (the number of fractional integral operations),  is

P
the Q-dimension vector which components are the FLP coefficients
x̂ (n) = − ai x (n − i) . (8) to estimate. The parameters p1 , . . ., pQ are the fractional integral
i=1 values for each operator Dpi in the model. For each pi , the signals
resulting from the application of the fractional operator Dpi on the
The difference between the predicted sample and the true one
data x (n) are collected in the pi vectors.
is a prediction error. This error also corresponds to the white noise
Then we need to solve the following equation to get the required
w (n) of Eq. (6).
coefficients:
The fractional-order linear prediction (FLP) is a generalization of
the notion of linear prediction applied to FAR stationary processes. x = H + w
The FLP has proven its effectiveness to model speech signals [25] or
EEG signals [26] and we propose to use this method for ECG signals The least-squares estimation of the FLP coefficients vector gives
modeling. ˆ as:
Similarly to the LP modeling, the FLP modeling can be viewed as −1
ˆ = H T H HT x (12)
the prediction x̂ (n) of x (n) when expressed as a linear combination
of its fractional derivatives or integrals [26]. The FLP (or LP) coefficients which estimation permits to char-
In derivative terms, the estimation expression writes: acterize the FAR (or AR) model of the observed data x (n) will be
referred to FLP modeling (or AR modeling) in the following. Note

Q
here that the use of FLP modeling not only requires defining the
x̂ (n) = ˛i Dpi x (n) (9)
model order as for the LP modeling, but also requires the defini-
i=1
tion of the Q fractional integral parameter values. The FLP modeling
where pi is a strictly positive real number, Dpi denotes the frac- therefore enjoys this new high flexibility compared to LP modeling
tional derivative operator with fractional order value pi and the but at the expense of properly defining these parameters.
x (n) are the samples of the processed signal. The ˛i values are the
FLP coefficients which number is Q referred to the FLP model order. 2.2.4. Evaluation of the ECG modeling
Note that in the Grünwald–Letnikov definition of the fractional For any stationary sequence of N samples x (n), the LP and FLP
operator, the cases where pi < 0 correspond to the fractional integral coefficients can be estimated. Predicted versions x̂ (n) of x (n) can
operator for Dpi . So, the formula of Eq. (9) can also be expressed in therefore be simulated from the estimated coefficients. So the cri-
integral terms as: terion used for evaluating the LP and FLP modeling quality is based
on the comparison of the true data with the simulated one thanks

Q
to the signal-to-error ratio (SER). The SER is given by:
x̂ (n) = ˇi Dpi x (n) (10)
N 2
i=1
i=1 [x (i)]
SER = 10 ∗ log N
2 (13)
where the ˇi values are the FLP coefficients. x (i) − x̂ (i)
i=1

2.2.3. Parameters model estimation where x and x̂ are the original and the simulated signals respec-
The way to estimate the parameters ai in the expression (6) is tively.
well established through the Yule–Walker equations solved by the
Durbin–Levinson algorithm [37]. 2.3. ECG FLP classification
We just focus here on the determination of the parameters in
expression (10) in the fractional case (the resolution for the expres- 2.3.1. The classification problem
sion (9) is similar). Classification is a problem of pattern recognition. In the famous
The FLP coefficients can be determined by minimizing the paper of Jain et al. [38] the authors announce four approaches
energy of the prediction error w (n) = x (n) − x̂ (n). The least squares for dealing with a pattern recognition problem, namely: template
method can be used to estimate the FLP coefficients, where the matching, syntactic or structural matching, neural networks and
objective function is the total energy given by: statistical classification which can be geometric or density based.
However, the choice of the appropriate pattern recognition

N−1
approach itself is a complex task because each approach is faced
E= w(n)2 (11)
to its own difficulties. The methods based on matching are faced to
n=0
the choice of the metric which is used as a similarity measure useful
with N representing the number of samples in the signal. for achieving the matching. The density based methods are faced to
M.L. Talbi, P. Ravier / Biomedical Signal Processing and Control 23 (2016) 42–51 45

the definitions of the correct probability density functions. The geo- [40] while the waveform limit locator is based on the algorithm
metric based methods are faced to the definition of the searching described in [41,42].
criterion for defining the frontiers between the classes. The neural Consequently, variable lengths of the QRS complexes have been
networks based methods are faced to the model selection problem used and the QRS complexes were extracted from the original ECG
(definition of the activation functions, definition of the number of signal without filtering. Each extracted QRS segment value was
neurons as well as the number of hidden layers, definition of the scaled to the range between 0 and 1.
weights estimation algorithm. . .). Even if some empirical rules exist
as a help for making the correct choices, these problems are not 2.4.2. The fractional operator definition
solved and remain research issues. Each fractional integral operators is discritized by back forward
As a consequence of these difficulties, we selected a panel of formula. In order to be able to compare LP modeling with FLP mod-
ten methods based on two approaches within the four existing eling, a third-order LP modeling is performed on each QRS complex
ones, namely the statistical and the neural network approaches. and the same order for FLP modeling is considered. The three values
The idea herein is to test the robustness of both LP modeling and for fractional integration are −0.15, −0.3, and −0.45, respectively.
FLP modeling with respect to many classifiers. The 6th order MacLaurin truncations in the series expansion of for-
In this paper, we will focus on the discrimination of premature mula (4) give the three operators used for achieving an approximate
ventricular contraction ECG beats (PVC) from all other ECG beats. fractional integration which are expressed as the following transfer
PVC is a premature contraction which takes place in the ventricles functions are:
because of abnormal electrical activity. PVCs are characterized by D−0.15 ≈ (0.4136 − 0.3515 z−1 − 0.02637 z−2 − 0.01011
the premature occurrence of large bizarre-shaped QRS complexes; −3
z − 0.005432 z − 0.003422 z − 0.002367 z )/(1 − z−1 )
−4 −5 −6
these complexes are not preceded by a P wave [2]. This subject D−0.30 ≈ (0.171 − 0.1197 z−1 − 0.01796 z−2 − 0.007783
is of interest because PVC beats are the first signs of disturbance z−3 − 0.004475 z−4 − 0.002953 z−5 − 0.002117 z−6 )/(1 − z−1 )
in the cardiac depolarization process and precede in some cases D−0.45 ≈ (0.07074 − 0.03891 z−1 − 0.008754 z−2 − 0.004231
malignant cardiac arrhythmias. z−3 − 0.002592 z−4 − 0.001788 z−5 − 0.001326 z−6 )/(1 − z−1 )

2.3.2. Evaluation of the ECG classifier 2.4.3. The proposed evaluation algorithm
To compare the obtained classification results based on the FLP The QRS complex FLP and LP modeling and evaluation process
coefficients with those based on the LP coefficients, the heart- is carried out using the 5 steps as detailed in the table below, with
beat classification performance was evaluated in terms of measures parallel processing for the comparison between FLP and LP model-
widely used in the literature: ing. Steps 1 and 2 are repeated for all the ECG beats in order to have
Numberoftotaltestedbeats − FN − FP statistical evaluation of the SER. Step 4 is repeated for all the tested
Accuracy % = ∗ 100 (14) classifier types.
Numberoftotaltestedbeats
As illustration of FLP modeling, we show on Fig. 1 the effect of
Numberofnormaltestedbeats − FP the three integration operators on a QRS complex example in the
Specificity % = ∗ 100 (15)
Numberofnormaltestedbeats case of a normal and a PVC beat. The curves show a raw QRS wave-
NumberofPVCtestedbeats − FN form and the resulting ones obtained by applying the operators Dpi
Sensitivity % = ∗ 100 (16)
NumberofPVCtestedbeats defined in Section 2.4.2. Each curve has been normalized in mag-
nitude between 0 and 1 for readability. The original waveform can
The classification errors are due to the false decisions: the FN
value (False Negative) is the number of PVC beats classified as nor-
be retrieved as a linear combination of the three fractional order
derivations waveforms with parameters ˇi = 3.5, −4.2, 1.7 and
mal beats and the FP value (False Positive) is the number of normal  
beats classified as PVC beats. ˇi = 4.3, −6.3, 2.9 for the normal and PVC beat, respectively.
The accuracy (Accuracy) is the percentage of correctly classified
beats among all the beats considered. 3. Results and discussions
Specificity (Specificity) is defined as the ratio between the num-
bers of correct decisions with respect to the number of beats In this study, we have used real ECG signals from the MIT–BIH
proposed at the input of the classifier when all the tested beats Arrhythmia Database [39]. The database contains 48 records and
are normal. the signals were sampled at 360 Hz. We have used only the MLII
Sensitivity (Sensitivity) is defined as the ratio between the num- electrode signal of the MIT–BIH Arrhythmia Database. In the fol-
bers of correct decisions with respect to the number of beats lowing sections, we present and discuss the results obtained using
proposed at the input of the classifier when all the tested beats the proposed methods.
are PVC. The beats are classified in two groups: Normal beats group and
PVC group.
2.4. The proposed evaluation scheme The beats were selected from 43 ECG records, which correspond
to the following files: 100, 101, 103, 106, 109, 111, 112, 113, 114,
This section is dedicated to the FLP modeling ability to correctly 115, 116, 117, 118, 119, 121, 122, 123, 124, 200, 201, 202, 203, 205,
represent ECG Normal and PVC beats and discriminate them by 207, 208, 209, 210, 212, 213, 214, 215, 217, 219, 220, 221, 222, 223,
comparison with LP modeling. An ECG preprocessing step and the 228, 230, 231, 232, 233 and 234.
fractional operator definition are first detailed, before presenting The two groups are selected according to the AAMI standards.
the proposed evaluation algorithm. Table 2a shows the annotations of the MIT/BIH database, the
selected groups and the total number of used beats.
2.4.1. QRS complex extraction preprocessing
ECG modeling needs previously to properly detect the QRS 3.1. FLP modeling versus LP modeling
occurrences and also to extract the exact QRS duration of each beat.
We have used for this task the ecgpuwave software available with In this section, we compare the modeling quality between FLP
the physiobank database in [39]. The QRS detector in ecgpuwave modeling and LP modeling in terms of SER computed between the
software is based on the famous algorithm of Pan and Tompkins original waves and the reconstructed waves using Eq. (13). The
46 M.L. Talbi, P. Ravier / Biomedical Signal Processing and Control 23 (2016) 42–51

1 1
Original Original
p = -0.15 0.8 p = -0.15
0.8
p = -0.3 p = -0.3
p = -0.45

Normalized
Normalized 0.6 0.6 p = -0.45

0.4 0.4

0.2 0.2

0 0
0 20 40 60 80 0 20 40 60 80 100 120
Sample Sample

Fig. 1. The effect of the three used fractional integral operators on the QRS complex in the case of normal beat (left) and PVC beat (right).

Table 1
Steps of the algorithm used for the comparison between FLP modeling and LP modeling.

Step The FLP modeling The LP modeling

0. Preprocessing Extraction and normalization of the QRS complex


1. Modeling Computation of three fractional integration Linear prediction coefficients
operators Dpi on the QRS waveform with estimation of the QRS complex
values −0.15, −0.30 and −0.45; resolution of using Burg algorithm with
Eq. (12) for the FLP coefficients estimation order 3
2. Evaluation of the modeling quality a. Reconstruction QRS complex reconstruction using the QRS complex reconstruction
estimated FLP coefficients using the estimated LP
coefficients
b. Computing the Computation of the SER value of the two
SER models and comparison of the values of the
two models in the case of PVC and normal beat
2 . Evaluation of the classification a. Classification Learning and testing phases based on FLP Learning and testing phases
performance coefficients of ECG beats for a two classes based on LP coefficients of ECG
problem beats for a two classes problem
b. Computing Compute the three indicators and compare the
Accuracy, performance of the classifiers
Specificity,
Sensitivity

comparison is achieved for both classes Normal and PVC indepen- indicate that 3 parameters are sufficient to reconstruct each beat
dently. Fig. 2 gives illustration examples of an original signal (solid without any loss of quality.
line) compared with its reconstructed version (star line) for nor- To compare the statistical values of the two models in mod-
mal and PVC classes (left). The corresponding instantaneous fitting eling normal and PVC beats, the mean value of the SER and the
errors are also displayed for each case (right). standard deviation ‘std’ of each calculated data were computed.
It is quite visible that the proposed modeling technique, with Table 2b provides the summary statistics for these values for the
only three parameters, can model the QRS complex. These results 88,596 normal beats and 7147 PVC beats tested as mentioned
in Table 1.
Table 2a As can be seen from Table 2b, there is a significant difference
SER values statistics for the quality evaluation of the two models; reported values between the statistical values of the two models. Interestingly, the
are in dB. SER of FLP modeling is higher than that of the LP modeling.
MIT/BIH Description Class Number
Overall, these results indicate that FLP modeling is better than
annotation of beats LP modeling both quantitatively (9.4 dB SER more in average
for FLP modeling in comparison with LP modeling) and quali-
V Premature ventricular Class PVC 7147
contraction tatively (a difference between PVC and other beats of 1.5 dB in
E Ventricular escape beat Class PVC comparison with 8.6 dB for FLP modeling and LP modeling, respec-
N Normal beat Class normal tively).
L Left bundle branch block Class normal
A summary of the main findings which have arisen in this sec-
beat
R Right bundle branch block Class normal
tion clearly shows that FLP is more suitable than LP for ECG beats
beat modeling which means that FLP modeling can be advantageously
e Atrial escape beat Class normal 88,596 used for PVC discrimination.
j Nodal (junctional) escape Class normal
beat
A Atrial premature beat Class normal
a Aberrated atrial premature Class normal Table 2b
beat Annotations of the MIT/BIH database, the allocation of the annotated beats to the
J Nodal (junctional) Class normal groups and the total number of used beats.
premature beat
LP FLP
S Supraventricular Class normal
premature or ectopic beat PVC Other PVC Other

Total 95,743 Mean (SER) ±std (SER) 52.41 ± 2.79 43.77 ± 2.30 58.19 ± 3.64 56.70 ± 2.21
M.L. Talbi, P. Ravier / Biomedical Signal Processing and Control 23 (2016) 42–51 47

Original & Model FLP LP


-4
x 10 0.02
1 5
Original Error
Error
0.8 Model 0.01
0
0
Normalized

0.6
Normal

-5
-0.01
0.4

-10 -0.02
0.2
-0.03
-15 0 20 40 60
0 0 20 40 60
0 20 40 60 80 Sample
Sample
Sample
-4
1 x 10 0.01
4
Original Error
Error
0.8 Model 0.005
2
Normalized

0.6 0
PVC

0
-0.005
0.4

-2 -0.01
0.2

-0.015
0 -4 0 20 40 60 80 100
0 20 40 60 80 100 0 20 40 60 80 100 Sample
Sample Sample

Fig. 2. Original signal (solid line) with reconstructed signal (star line) for normal and PVC QRS complex and their corresponding instantaneous fitting errors (FLP case and LP
case, respectively).

3.2. PVC discrimination using k nearest neighbor classifier Table 4 presents the results obtained for the arrhythmic beat
classification by k nearest neighbor classifier in the case of FLP and
The results in the previous section indicate that the estimated LP coefficients.
coefficients of FLP modeling are more suitable to precisely repre- It can be seen from the data in Table 4 that the statistical per-
sent the two classes than those of the LP modeling. formances of the classifier in the case of FLP modeling are better
This is confirmed by Fig. 3 representing the scatter plot of the than those of the LP modeling in the majority of the tested records.
two classes: we can see that the two classes in the case of FLP mod- Comparing the two confusion matrices (Table 5), it can be also
eling are clearly separated compared to those of the LP modeling. seen that from 7147 PVC beats, 6071 beats are correctly classi-
This illustrates the potential of FLP modeling in discriminating fied in the case of FLP modeling; on the other hand, only 3333
both populations compared to LP modeling. In this section we thus PVC beats are correctly classified for LP modeling. This corresponds
investigate the usefulness of the FLP coefficients for a discrimina- to sensitivity percentages of 46.6% for LP modeling compared to
tion of the PVC arrhythmia. 85% obtained for FLP modeling (Table 6) which is an impor-
For that purpose, we used a simple k nearest neighbor classifier tant issue because PVC beats detection should not be missed for
to separate the two classes. Two classification experiences were clinicians.
elaborated with the same conditions. In the first experience, the A summary of the statistical performances of the k nearest
classifier is fed by the three LP coefficients. In the second one, the neighbor classifier for the two models is given in Table 6. The results
classifier is fed by the three FLP coefficients. reported in this table indicate that there was a significant differ-
The training set is randomly selected from two records; the first ence between the statistical performances of the two models, FLP
110 normal beats are selected from record ‘112’, and 110 PVC beats coefficients based classification performs better.
are taken from record ‘119’ (Table 3).
A summary of the results obtained for arrhythmic beat classifi-
3.3. Evaluation of the FLP modeling via classifiers performances
cation in the case of the FLP model and the LP model are shown in
Table 4.
In this section, we study the influence of the used classifier type
on the performances statistics of the classification task (Accuracy,
Specificity and Sensitivity).
Table 3
Numbers of training and test beats.
Our motivation is to study the robustness of the results with
respect to any classifier type. This study will help us choosing the
Class Normal (N) Premature ventricular Total most appropriate classifier when applying FLP modeling to ECG
contraction beat (V)
signals, if any.
Training beats 110 110 220 To that aim, we have chosen four categories of classifiers as fol-
Test beats 88,596 7147 95,743
lows: 4 belong to the Statistical Density based classifiers category; 3
48 M.L. Talbi, P. Ravier / Biomedical Signal Processing and Control 23 (2016) 42–51

Fig. 3. The scatter plot of some records, The blue stars (*) indicate normal ECG beats, and the red crosses (×) are abnormal (PVC) beats. (For interpretation of the references
to color in this figure legend, the reader is referred to the web version of this article.)

belong to the neural networks category; 2 are statistical-geometric Documentation about the classifiers description and implementa-
based classifiers and 1 is a syntactic based classifier as shown in tion can be found in [44].
Table 7. We have used the prtools toolbox [43] for the imple- In order to be able to compare statistics performances of the
mentation of classification techniques for classification purpose. chosen classifiers, we keep the same conditions that we set in
M.L. Talbi, P. Ravier / Biomedical Signal Processing and Control 23 (2016) 42–51 49

Table 4
Results for arrhythmic beat classification by k nearest neighbor classifier.

Record Normal PVC FN FP Sensitivity% Specificity% Accuracy%

LP FLP LP FLP LP FLP LP FLP LP FLP

100 2272 1 1 0 446 1 0 100 80.37 99.96 80.33 99.96


101 1862 0 1 0 745 0 – – 59.99 100 59.96 100
103 2084 0 1 0 2038 0 – – 2.207 100 2.206 100
106 1507 520 461 209 174 0 11.52 59.88 88.45 100 68.69 89.69
109 2491 38 36 26 1434 1784 7.69 33.33 42.43 28.38 41.9 28.46
111 2123 1 2 0 0 45 0 100 100 97.88 99.91 97.88
112 2538 0 1 0 20 0 – – 99.21 100 99.17 100
113 1795 0 1 0 34 0 – – 98.11 100 98.05 100
114 1831 43 29 0 331 0 34.09 100 81.92 100 80.8 100
115 1953 0 1 0 1736 0 – – 11.11 100 11.11 100
116 2302 109 64 2 604 17 41.82 98.18 73.76 99.26 72.31 99.21
117 1534 0 1 0 25 0 – – 98.37 100 98.31 100
118 2261 16 7 12 1521 58 58.82 29.41 32.73 97.43 32.92 96.93
119 1542 444 3 0 1480 0 99.33 100 4.021 100 25.36 100
121 1861 1 2 0 1 24 0 100 99.95 98.71 99.84 98.71
122 2475 0 1 0 718 0 – – 70.99 100 70.96 100
123 1514 3 4 2 1508 0 0 50 0.40 100 0.40 99.87
124 1562 47 33 1 1525 1 31.25 97.92 2.369 99.94 3.23 99.88
200 1774 826 5 0 227 0 99.4 100 87.2 100 91.08 100
201 1752 198 199 117 3 68 0 41.21 99.83 96.12 89.65 90.52
202 2116 19 20 15 310 9 0 25 85.35 99.57 84.55 98.88
203 2526 438 351 42 79 459 20.05 90.43 96.87 81.83 85.5 83.1
205 2574 71 29 6 2180 2 59.72 91.67 15.31 99.92 16.52 99.7
207 1649 210 210 5 0 1561 0.47 97.63 100 5.337 88.71 15.81
208 1588 992 881 0 559 15 11.28 100 64.8 99.06 44.21 99.42
209 3003 1 2 0 1848 0 0 100 38.46 100 38.44 100
210 2443 194 194 17 60 3 0.51 91.28 97.54 99.88 90.37 99.24
212 2748 0 1 0 1075 8 – – 60.88 99.71 60.86 99.71
213 2668 220 5 3 2666 0 97.74 98.64 0.075 100 7.55 99.9
214 2001 256 97 100 369 9 62.26 61.09 81.56 99.55 79.36 95.17
215 3197 164 136 117 182 2 17.58 29.09 94.31 99.94 90.54 96.46
217 245 162 113 10 13 0 30.67 93.87 94.69 100 69.12 97.55
219 2087 64 11 13 1972 1 83.08 80 5.51 99.95 7.85 99.35
220 2047 0 1 0 732 0 – – 64.24 100 64.21 100
221 2029 396 397 2 242 0 0 99.5 88.07 100 73.66 99.92
222 2269 0 1 0 12 7 – – 99.47 99.69 99.43 99.69
223 2100 473 298 350 1596 9 37.13 26.16 24 99.57 26.42 86.05
228 1691 362 180 19 0 15 50.41 94.77 100 99.11 91.24 98.34
230 2253 1 1 0 2223 0 50 100 1.33 100 1.375 100
231 1564 2 2 0 1556 0 33.33 100 0.51 100 0.5743 100
232 1779 0 1 0 55 4 – – 96.91 99.78 96.85 99.78
233 2236 829 26 8 2086 1 96.87 99.04 6.708 99.96 31.12 99.71
234 2749 3 4 0 2365 0 0 100 13.97 100 13.95 100

the previous section experiences. Therefore, the data described in From Table 7, we can also see that all the values of sensitivities
Table 1 and the performance statistics previously defined in Eqs. in the case of LP modeling are lower than 62%, despite some good
(14) to (16) are similarly used in those experiences. A summary of values for the accuracy and for the specificity. This indicates that
the statistics performances of the chosen classifiers are presented more than 40% PVC beats are misclassified. This result is inaccept-
in Table 7. able for clinicians and proves the interest of our method since the
Table 7 shows that whatever the classifier, the performances sensitivity value with FLP modeling can reach up to 88% depending
statistics obtained for FLP modeling are better than those obtained on the classifier type (the lowest being 47%).
for LP modeling, except in one case where the specificity is 0.5% A summary of the results of the performances statistics are set
better for LP compared to FLP (neural network classifier trained out in Table 8.
by back-propagation algorithm). The best results are obtained for Table 8 compares the summary statistics of the performances of
normal densities based linear classifier’ and fishers classifiers. the 10 used classifiers.
Table 8 shows that the averaged statistical performances of FLP
modeling are better than those of LP modeling. However, only the
Table 5 standard deviation of the sensitivity percentage presents an excep-
Confusion matrix in the FLP and LP modeling cases.
tion. This is probably due to the quadratic classifier and radial basis
Class Class (N) Class (V) Totals neural network classifier that present the poorest sensitivity values
Confusion matrix in the FLP case
which are rather far from all the other sensitivity values.
Class (N) 84,493 4103 88,596
Class (V) 1076 6071 7147
Table 6
Totals 85,569 10,174 95,743 Results for arrhythmic beat classification (PVC beats).
Confusion matrix in the LP case
Class (N) 51,846 36,750 88,596 Accuracy% Specificity% Sensitivity%
Class (V) 3814 3333 7147
LP case 57.6 58.5 46.6
Totals 55,660 40,083 95,743 FLP case 95.0 95.0 85.0
50 M.L. Talbi, P. Ravier / Biomedical Signal Processing and Control 23 (2016) 42–51

Table 7
Summary the statistics performances of the chosen classifiers.

Classifier Approach Accuracy% Specificity% Sensitivity%

FLP LP FLP LP FLP LP

Normal densities based quadratic classifier Statistical-density 94.0 93.18 98.0 97.87 47.0 35.09
k-Nearest neighbor classifier Statistical-density 95.0 57.63 95.0 58.51 85.0 46.63
Normal densities based linear classifier Statistical-density 91.0 87.04 91.0 89.13 88.0 61.14
Parzen classifier Statistical-density 95.0 56.86 95.0 57.64 84.0 47.13
Neural network classifier trained by back-propagation algorithm Neural network 80.0 78.98 81.0 81.54 75.0 47.33
Neural network trained by Levenberg–Marquardt rule Neural network 96.0 82.70 98.0 85.87 74.0 43.37
Radial basis neural network Neural network 95.0 86.81 98.0 90.89 59.0 36.35
Support vector machine Statistical-geometric 95.0 26.39 95.0 24.67 85.0 47.60
Fisher classifier Statistical-geometric 91.0 87.04 91.0 89.13 88.0 61.14
Binary decision tree classifier Syntactical 94.0 51.71 95.0 52.24 85.0 45.09

Table 8 References
Summary statistics of the performances of the 10 tested classifiers.

Accuracy (m ± std %) Specificity (m ± std %) Sensitivity (m ± std %) [1] S.V. Madihally, Principles of Biomedical Engineering, Artech House, Norwood,
MA, 2010.
FLP 92.6 ± 4.74 93.7 ± 5.14 77 ± 13.82 [2] L. Khansa, A. Naït-Ali, Parametrical modeling of a premature ventricular con-
LP 70.83 ± 21.6 72.75 ± 23.35 47.09 ± 8.64 traction ECG beat: comparison with the normal case, Comput. Biol. Med. 37
(2007) 1–7.
[3] S. Edla, N. Kovvali, A. Papandreou-Suppappola, Electrocardiogram signal mod-
eling with adaptive parameter estimation using sequential bayesian methods,
IEEE Trans. Signal Process. 62 (2014) 2667–2680.
Taken together, these results suggest that the FLP coefficients
[4] H. Baali, R. Akmeliawati, M.J.E. Salami, A. Khorshidtalab, E. Lim, ECG parametric
are best suited to model the QRS complex. modeling based on signal dependent orthogonal transform, IEEE Signal Process.
Lett. 21 (2014) 1293–1297.
[5] A. Koski, Modeling ECG signals with hidden Markov models, Artif. Intell. Med.
4. Conclusion 8 (1996) 453–471.
[6] R.V. Andreão, B. Dorizzi, J. Boudy, ECG signal analysis through hidden Markov
models, IEEE Trans. Biomed. Eng. 53 (2006) 1541–1549.
The present study was designed to demonstrate the effective- [7] P. Laguna, I.R. Janz, S. Olmos, N.V. Thakor, H. Rix, P. Caminal, Adaptive estimation
ness of the FLP technique on the ECG signal modeling. One of the of QRS complex wave features of ECG signal by the Hermite model, Med. Biol.
Eng. Comput. 34 (1996) 58–68.
most significant findings that emerge from this study is that FLP
[8] O. Sayadi, M.B. Shamsollahi, G.D. Clifford, Robust detection of premature ven-
modeling is more suitable for ECG signal modeling than LP mod- tricular contractions using a wave-based Bayesian framework, IEEE Trans.
eling. The results of this investigation show that three coefficients Biomed. Eng. 57 (2010) 353–362.
are largely sufficient to characterize each ECG beat. Consequently, [9] O. Sayadi, M.B. Shamsollahi, ECG denoising and compression using a mod-
ified extended Kalman filter structure, IEEE Trans. Biomed. Eng. 55 (2008)
these coefficients are reliable for the PVC arrhythmia discrimina- 2240–2248.
tion. In terms of quantity, the proposed method was applied to [10] E.K. Roonizi, R. Sameni, Morphological modeling of cardiac sig-
PVC arrhythmia discrimination and the best accuracy of 96% was nals based on signal decomposition, Comput. Biol. Med. 43 (2013)
1453–1461.
achieved. [11] A. Sandryhaila, S. Saba, M. Püschel, J. Kovacevic, Efficient compression of QRS
The present study confirms previous studies in FLP physiological complexes using Hermite expansion, IEEE Trans. Signal Process. 60 (2012)
signal modeling, particularly in speech and EEG signals. Similarly 947–955.
[12] R. Dubois, P. Roussel, M. Vaglio, F. Extramiana, F. Badilini, P. Maison-Blanche,
the study has successfully demonstrated that FLP modeling can be G. Dreyfus, Efficient modeling of ECG waves for morphology tracking, Comput.
a good alternative to LP modeling. The main advantage of FLP for Cardiol. 36 (2009) 313–316.
detection of PVC in ECG signals is its ability in better achieving [13] S. Edla, N. Kovvali, A. Papandreou-Suppappola, Sequential Markov Chain Monte
Carlo filter with simultaneous model selection for electrocardiogram signal
this task by comparison with a classical LP method, with a reduced
modeling, in: 34th Annual International Conference of the IEEE EMBS, 28
number of parameters (only three in the studied case). The obtained August–1 September, 2012, San Diego, California USA, 2012.
results can also be improved by tuning the fractional integral values [14] P.E. McSharry, G.D. Clifford, L. Tarassenko, L.A. Smith, A dynamical model for
generating synthetic electrocardiogram signals, IEEE Trans. Biomed. Eng. 50
constituting a new degree of freedom for the FLP method compared
(2003) 289–294.
to the LP method. The limitations of FLP modeling are the absence [15] S. Edla, N. Kovvali, A. Papandreou-Suppappola, Electrocardiogram Signal Mod-
of efficient rules or algorithms for estimating the fractional integral eling using Interacting Multiple Models, in: ASILOMAR, 6–9, November, Pacific
values. The choice of the optimal number of coefficients is also a Grove, CA, 2011.
[16] J.C. Nunes, A. Nait Ali, Hilbert transform based ECG modeling, Biomed. Eng. 39
limitation which similarly exists for LP method. (2005) 133–137.
The influence of some parameters on the presented results can [17] M. Lagerholm, C. Peterson, G. Braccini, L. Edenbrandt, L. Sörnmo, Clustering
be investigated in future studies. The truncation order of the Mac ECG complexes using Hermite functions and self-organizing maps, IEEE Trans.
Biomed. Eng. 47 (2000) 838–848.
Laurin series expansion, the fractional integration order and the [18] R. Sameni, M.B. Shamsollahi, C. Jutten, M. Babaie-Zadeh, Filtering Noisy ECG
fractional integration values are parameters of the method which signals using the extended Kalman filter based on a modified dynamic ECG
variation is interesting to study with the purpose of classification model, Comput. Cardiol. 32 (2005) 1017–1020.
[19] K.P. Lin, W.H. Chang, QRS feature extraction using linear prediction, IEEE Trans.
statistics improvement, also taking into account computation time Biomed. Eng. 36 (1989) 1050–1055.
limitations. Moreover, a number of possible future studies using [20] D. Ge, N. Srinivasan, S.M. Krishnan, Cardiac arrhythmia classification using
the same experimental set up can be carried out. A natural pro- autoregressive modeling, Biomed. Eng. Online 1 (1) (2002) 1–5, Article number
5.
gression of this work is to design a system that can classify several
[21] C.A. Monje, B.M. Vinagre, V. Feliu, Y.Q. Chen, Tuning and auto-tuning of frac-
arrhythmic beats. tional order controllers for industry applications, Control Eng. Pract. 16 (2008)
The success encountered by FLP modeling for detection of PVC in 798–812.
[22] B. Mathieu, P. Melchior, A. Oustaloup, C. Ceyral, Fractional differentiation for
ECG signals encourages future works in other fields that need para-
edge detection, Signal Process. 83 (2003) 2421–2432.
metric modeling of data, like ECG data compression or synthetic [23] R.L. Magin, Fractional calculus models of complex dynamics in biological tis-
ECG data generation, in the cardiac domain among other. sues, Comput. Math. Appl. 59 (2010) 1586–1593.
M.L. Talbi, P. Ravier / Biomedical Signal Processing and Control 23 (2016) 42–51 51

[24] M.S. Couceiro, N.M.F. Ferreira, J.A. Tenreiro Machado, Application of fractional [35] S. Zhou, J. Cao, Y. Chen, Genetic algorithm-based identification of fractional-
algorithms in the control of a robotic bird, Commun. Nonlinear Sci. Numer. order systems, Entropy 15 (2013) 1624–1642.
Simul. 15 (2010) 895–910. [36] P.P. Vaidyanathan, The Theory of Linear Prediction, Morgan & Claypool, San
[25] K. Assaleh, W.M. Ahmad, Modeling of speech signals using fractional calculus, Rafael, CA, 2008.
in: Proc. Ninth International Symposium on Signal Processing and Its Applica- [37] S.M. Kay, Modern Spectral Estimation: Theory and Application, Prentice Hall,
tions, 12–15 February 2007, Sharjah, UAE, 2007, pp. 1–4. Englewood Cliffs, NJ, 1988.
[26] V. Joshi, R.B. Pachori, A. Vijesh, Classification of ictal and seizure-free EEG signals [38] A.K. Jain, R.P.W. Duin, J. Mao, Statistical pattern recognition: a review, IEEE
using fractional linear prediction, Biomed. Signal Process. Control 9 (2014) 1–5. Trans. Pattern Anal. Mach. Intell. 22 (2000) 4–37.
[27] H. Sheng, Y.Q. Chen, T.S. Qiu, Fractional Processes and Fractional-Order Signal [39] R. Mark, G. Moody, MIT-BIH Arrhythmia Database Directory, Mass Inst. Technol,
Processing, Springer, London, 2012. Cambridge, MA, 1988.
[28] K.B. Oldham, J. Spanier, The Fractional Calculus, Academic Press, New York, NY, [40] J. Pan, W.J. Tompkins, A real-time QRS detection algorithm, IEEE Trans. Biomed.
1974. Eng. 32 (1985) 230–236.
[29] K.S. Miller, B. Ross, An Introduction to the Fractional Calculus and Fractional [41] P. Laguna, New Electrocardiographic Signal Processing Techniques: Applica-
Differential Equations, Wiley-Interscience, Hoboken, NJ, 1993. tion to Long-term Records, Science Faculty, University of Zaragoza, 1990 (Ph.D.
[30] M.D. Ortigueira, Fractional Calculus for Scientists and Engineers, Springer, Dor- Dissertation).
drecht/Heidelberg/London/New York, 2011. [42] P. Laguna, R. Jane, P. Caminal, Automatic detection of wave boundaries in mul-
[31] S. Das, Functional Fractional Calculus for System Identification and Controls, tilead ECG signals, Comput. Biomed. Res. 27 (1994) 45–60.
Springer, Berlin/Heidelberg/New York, 2008. [43] R.P.W. Duin, P. Juszczak, P. Paclik, E. Pekalska, D. de Ridder, D.M.J. Tax, S. Verza-
[32] T. Kaczorek, Selected Problems of Fractional Systems Theory, Springer, kov, PRTools4.1, A Matlab Toolbox for Pattern Recognition, Delft University of
Berlin/Heidelberg, 2011. Technology, 2007.
[33] Y.Q. Chen, K.L. Moore, Discretization schemes for fractional-order differentia- [44] F. van der Heijden, R.P.W. Duin, D. de Ridder, D.M.J. Tax, Classification, in:
tors and integrators, IEEE Trans. Circuits Syst. I: Fundam. Theory Appl. 49 (2002) Parameter Estimation and State Estimation: An Engineering Approach Using
363–367. MATLAB, John Wiley & Sons, Hoboken, NJ, 2004.
[34] D.P.M.O. Valério, Ninteger v. 2.3 Fractional Control Toolbox for MatLab, User and
Programmer Manual, Universidade técnica de lisboa instituto superior técnico,
2005.

You might also like