INDIN 2015
CLASSIFYING TACHYCARDIAS VIA HIGH DIMENSIONAL LINEAR DISCRIMINANT
FUNCTION AND PERCEPTRON WITH MULT-PIECE DOMAIN ACTIVATION FUNCTION
Jing Su, Jun Xiao,
Bingo Wing-Kuen Ling
& Qing Liu
School of Info. Eng.,
G.D.U.T.
Guangzhou, 510006,
China.
Kim-Fung Tsang,
Kwok-Tai Chui &
Haoran Chi
Dept. of Electronic Eng.,
City University of H. K.
Hong Kong, China.
Gerhard P. Hancke Jr.,
Dept. of Computer
Science,
City University of H. K.
Hong Kong, China
Pretoria, South Africa.
Zhangbing Zhou
TELECOM
SudParis, France.
ABSTRACT
This paper proposes a novel method for discriminating the
supraventricular tachycardias and the ventricular
tachycardias via a high dimensional linear discriminant
function and a perceptron with a multi-piece domain
activation function having multi-level functional values. The
algorithm is implemented via the mobile application. First,
the discrete cosine transform is applied to each training
electrocardiogram. Then, these discrete cosine transform
coefficients are scaled down according to their frequency
indices. These scaled discrete cosine transform coefficients
of each electrocardiogram are employed as features for
performing the discrimination. Second, the high order
statistic moments of each feature of the training
electrocardiograms corresponding to the same type of
tachycardias are evaluated. These high order statistic
moments of each feature corresponding to same type of
tachycardias form a vector. Third, the high dimensional
linear discriminant function is employed to minimize the
intraclass separation and maximize the interclass
separation of these statistic moment vectors. In particular,
new vectors are formed by projecting these statistic moment
vectors to the high dimensional linear discriminant function.
Fourth, the principal component analysis is employed to
reduce the dimension of the projected vectors. Finally, a
bank of perceptrons with multi-piece domain activation
functions having multi-level functional values is employed
for performing the discrimination. By using this bank of
perceptrons, the condition for general two class pattern
recognition problems achieving the error free pattern
recognition performance is guaranteed. Computer
numerical simulation results show that our proposed
method is robust and effective.
1. INTRODUCTION
There are two main types of tachycardias. They are the
supraventricular tachcardias and the ventricular tachycardias.
The arterial fibrillation becomes a common type of
supraventricular arrhythmias in elder people [1]. On the
other hand, the ventricular tachycardia is a potentially life
threatening arrhythmia [2].
To classify the supraventricular tachcardias and the
ventricular tachycardias, statistical approaches [3] are
employed. However, the decision rules based on these
statistical approaches are too simple that the discrimination
accuracies are too low for practical applications. On the
other hand, the one dimensional linear discriminant
functions [4] are employed for minimizing the interclass
separations and the intraclass separations of the feature
vectors. However, as the projected values are in a real line,
in general the degrees of freedoms for performing the
discrimination in the real line are too low that the obtained
discrimination accuracies are unacceptable for practical
applications too. To address these issues, two layer neural
networks and support vector machines are employed for
performing the discrimination [5]. However, the effects of
the activation functions of these two layer neural networks
and the kernel functions of the support vector machines on
the discrimination accuracies are unknown. Recently, a
single layer perceptron with the multi-piece domain
activation function having binary level functional values is
proposed for performing general two class pattern
recognition problems [6]. However, in general it is not
guaranteed that the training algorithm would converge.
To address the above issues, this paper proposes to employ a
multi-dimensional linear discriminant function and a bank
of perceptrons with the multi-piece domain activation
functions having multi-level functional values for
performing the discrimination of the supraventricular
tachycardias and the ventricular tachycardias. In our
proposed method, the weights of the perceptron are not
found by the iterative training algorithm. Instead, the design
of these weights are formulated as a linear programming
problem and they are found by the conventional simplex
algorithm. The outline of this paper is as follows. Section 2
presents the proposed discrimination algorithm. Section 3
presents the computer numerical simulation results. Finally,
a conclusion is drawn in Section 4.
2. PROPOSED DISCRIMINATION ALGORITHM
A. Preprocessing
l-)))