Computational Intelligence - February 2014 - 57

the presence of noise will corrupt the
desired signal, and make the feature
extraction and classification less accurate
or even discriminable. A de-noising procedure is necessary to minimize the negative effects of the noise. The chaotic
ECG extractor pursues key features
including six parameters include Lyapunov exponents spectrum (m 1, m 2, m 3, m 4),
correlation dimension (D), and RMS
level (Vrms) as depicted above. A threshold
is then set to remove outliers that are not
appropriate for training and classification.
All the features of ECG are removed
while it exceeds 1.5 times of the standard
deviation. The threshold is dependent on
every subject's feature centroid, which has
been summarized in Table XVI in the
supplementary materials [42]. Finally, a
neural network is applied as a classifier for
the features classification.
A. Outlier

When the value is not within in the
normal range, the value is called outlier.
It is not easy to distinguish the cause for
the outlier. If the outlier is not handled
properly, errors will be induced in the
analysis. The standard deviation of signal
distribution is used as a threshold for
dealing with the outlier. If a data set of
characteristic parameters x i ! R for
i = 1, f, N whose average is
x avg = 1
N

/ xi .

(11)

i =1

The standard deviation is
v

1
N

/ ^x i - x avgh2 ,

(12)

i =1

where N is the length of the data set
and x avg is the average of the data set.

First Active Sensor
Electrode

Second Active
Sensor Electrode

Bio-Potential
Sensor

ESD
Protection Circuits

This shows the discrete degree for the
data. The standard normal distribution is
illustrated in Fig. 3.
B. Neural Network

The back propagation neural network
(BPNN) is used here for classification
[43]. The operational process involves
two stages. The stages are the forward
pass and backward pass respectively.
The resilient back propagation
(RPROP) algorithm is one of the fastest weight training methods [43]. The
algorithm is a local adaptive learning
scheme. The bad influence of the partial derivative's size on the weight
updating step can be eliminated with
this method.
Six characteristic parameters are used
as the inputs for networking training:

-3v -2v -v

Figure 3 Normal distribution of measured data.
Table 2 Age, height and weight of
nineteen subjects joining the
experiment.
Sex

age HeigHT WeigHT
(Yr) (cm)
(kg)

Subj. A

FemAle 25

153

Subj. b

mAle

172

Subj. C

mAle

175

Subj. D

mAle

173

Subj. e

mAle

170

Subj. F

mAle

166

Subj. G

FemAle 22

152

Each subject refers to one output.

Subj. H

FemAle 17

158

o = 6o 1 o 2 o 3 o 4 o 5 o 6 o 7 o 8 g o p@,
(14)

Subj. I

mAle

173

Subj. j

mAle

175

Subj. K

mAle

180

Subj. l

FemAle 19

x = 6x 1 x 2 x 3 x 4 x 5 x 6@
= 6m 1 m 2 m 3 m 4 D 2 Vrms@T . (13)
T

where the last output o p refers to as
the undefined class. The details of
neural network training are similar to
the popular artificial neural networks
based on the supervised lear ning
mechanism for the connection weight
establishment and thus are omitted
for briefness.
If the network was adequately
trained, the training patterns should
correspond to the designate targets.
Next, the testing patterns are added for
testing. If testing patterns cannot correspond to the correct target, the condition means the result which is fault to

Bio-Signal
Measurement

Buffer/Balanced
Circuit

Analog
Filter/Amplifier Unit

156

175

Subj. N

FemAle 32

166

Subj. O

mAle

173

Subj. P

FemAle 23

155

Subj. Q

FemAle 27

160

Subj. R

FemAle 27

155

Subj. S

mAle

177

Subj. m mAle

do classification. One observes the testing result to evidence whether the
method is suitable for the personal
identification or not.

Associative
Processing Unit

Signal
Processing Unit

External Input
Device

Display Device

Negative Feedback Difference
Common Mode Signal
Figure 4 Structure of the patented portable instrument eT-600.

February 2014 | Ieee ComputatIonal IntellIgenCe magazIne

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