Computational Intelligence in Cardiology and Heart Disease Diagnosis - Computational Intelligence in Biomedical Engineering - page 192

F4

{U k }

First
comparison
detection

{

η k }

Class
credits

Initialization
classes

{ µ k }

[inic x ]

{T 1K }
channel
block 1

{T 1K }
channel
block 1

•• •

Selection of
class to
initialize

{E 1 } d 1 {E 1j }

{E 1j } d

{act 1k } {act 1k }

inic

Orientation

{re k }

Initialization
detection

FIGURE 4.30
A multichannel adaptive resonance theory network (MART). (Proposed by
Barro, S., Fernandez-Delgado, Vila-Sobrino, J., Regueiro, C., and Sanchez, E.,
IEEE Eng. Med. Biol. Mag., 17, 45-55, 1998. c

IEEE.)

genetic algorithm a variable node layer during training. When Fourier trans-
form coecients of the QRS complex in the R-wave region were used as input
features, comparisons with an MLP network showed improved training e-
ciency using GARCE. An adaptive NN as seen in Figure 4.30 was used by
Barro et al. (1998) to deal with a larger range of patients. Their network
was based on the ART model (Carpenter and Grossberg, 1987; Carpenter
et al., 1991) with segment inputs of the ECG signal. Template matching was
used to detect new QRS waveforms and a difference function was employed
to calculate the global difference and classify the waveforms.

4.7.2 Support Vector Machines
SVMs are a recent classifier formulation (early 1990s) and are only beginning
to find use in ECG classification. The advantage of using SVMs is that good
performance can still be obtained using smaller datasets. A general SVM
recognition system is similar to that of an NN system with the difference that
the NN classifier is replaced by an SVM classifier as seen in Figure 4.31.
Rojo-Alvarez et al. (2002a,b) initially hypothesized that ventricular EGM
onset could be used to discriminate supraventricular tachycardias and VTs.