Computational Intelligence - February 2014 - 62

Table 4 Recognition rate for some
competitive algorithms and proposed
one (chaos theory + bpnn).

MeThod

Fiducial RecogniTion
deTecTion RaTe

PRD

Yes

$ 75%

Yes

$ 79%

WDM

$ 85%

HigH-oRDeR
LegeNDRe
PoLYNoMiaLs

$ 88%

CHaos
No
THeoRY+BPNN

$ 90%

and the activation function for the neural network must be made. In this
research, six characteristic parameters
were acquired from the tested subjects
(19) that demand 6 nodes for the input
layer, for the output layer, there are 20
nodes corresponding to 19 classes and 1
undefined class. The number of nodes
for the hidden layer was determined via
experimental trials to compromise the
acceptable identification accuracy and
computational economy. There are 20
and 5 neurons in the first and second
hidden layers respectively. One may also
refer to [44, 45] for determining an
appropriate network structure. The
training of the resilient back propagation
algorithm was applied for the neural
network training. The adaptation learning function was of the gradient descent
type as depicted in Section III. The performance function is defined by the
normalized mean square error. The activation function of the output layer is the
linear transfer function, and the activation functions of the hidden layers are
log-sigmoid. The increase and the
decrease factor were set to be f + = 1.2
and f - = 0.5 respectively.
Referring to Hashemi's experimental
results [46], the percentage of number of
training and testing data sets were 90%
and 10% respectively. There were 648
patterns of the 19 subjects collected after
conducting data outliers. In which, 580
patterns were selected for network
training and 68 patterns for subsequent
testing. Each subject was possessed of
about 30 patterns for training and 3 or
more patterns for testing. After training,
the testing patterns were utilized for

identification performance. The experimental results show that 62 out of 68
testing patterns were accurately classified
and 6 patterns were misidentified. The
recognition rate is about 91%.
As mentioned before the previous
research focusing on the recognition rate
of ECG-based biometrics were not conducted under the same conditions. To
show robustness of the proposed
approach, Table 4 compares the recognition rate of the proposed method with
some competitive algorithms under the
same conditions such as tested subjects
(19 subjects), data source (collected from
laboratory), electrode orientation (Lead I)
etc. When the percentage root-meansquare deviation (PRD), cross-correlation
(CC) and wavelet distance measurement
(WDM) were applied to recognized person, as seen, the recognition rate of fiducial detection is lower than non-fiducial
detection. They resulted in lower recognition rates because ECG biometric
based with fiducial point detection is
inherently flawed as reported in [27],
there is no standard definitions where the
ECG feature wave boundaries lie and
relied heavily on the detection of the
PQRS signature. These misclassifications
occurred because of not prioritizing the
ECG features and occurrence of abnormal heart beats. The classification methodology proposed based on chaotic features explains ECG signal behavior and
its relation to time series information; the
discriminate advantages of the method
lies in its feature to quantify the integral
dynamic characterizations of ECG.
Therefore misclassification can be
avoided in some extents.
V. Conclusions and Future Works

Discriminable advantages of the ECGbased recognition can be summarized
as follows:
1) Uniqueness of ECG signal makes it
hard to be duplicated.
2) Individual identification based on multimodal biometrics is becoming a trend
for biometric identification system.
ECG signal embed bountiful individual information, some of which might
be used for the identification purpose,
remain to be discovered.

IEEE ComputatIonal IntEllIgEnCE magazInE | FEbruary 2014

3) ECG signal varies from person to
person that is difficult to be learned
or duplicated.
We have utilized ECG signals of the
lead-I measurement to conduct individual identification. The chaos theory
and time-domain analyses were utilized
to analyze the ECG signals. In the past,
ECG analyses based on the chaos theory were only used to differentiate
between the healthy people and the
people with heart diseases [12-17]. It
was shown here that it is possible to
advance to the purpose of individual
identification. The features of ECG
were extracted by applying Lyapunov
exponents and correlation dimension,
and a back propagation neural network
was utilized as the individual identification machine after training. The experimental results show the proposed
method for ECG-based personal identification is possible.
In the published works, the values
of the Lyapunov exponent and correlation dimension were mostly utilized to
differentiate between normal people
and the people with heart diseases
whose values of the cardiac are smaller
than the normal people. Wavelet analysis is also a popular method used to
analyze ECG signals. However, it needs
to analyze the signals within the same
intervals. If the intervals are not the
same, the parameters related to the
analysis exhibit significant variances.
The presented approach waives the
weakness as described.
In the current stage of this research,
the recognition process is conducted
off-line and effectiveness of the proposed method was tested on normal rest
subjects. Investigation of the on-line
operation to demonstrate possibility of
the proposed verification systems in
practical applications such as the tested
subjects are under stressed or after
exercise is undergoing. The preliminary
results obtained show promising potential of the present approach.
Acknowledgment

This research was sponsored by National
Science Council, Taiwan, R.O.C. under
the Grant NSC-99-2221-E-005-066.

Table of Contents for the Digital Edition of Computational Intelligence - February 2014