Computational Intelligence in Cardiology and Heart Disease Diagnosis - Computational Intelligence in Biomedical Engineering - page 179

4.6.1.3 Transform and Classification Algorithms

Signal transforms have also been used to detect peaks in the ECG. The wavelet
transform is a popular method that can be used to capture time and frequency
characteristics of the signal and is closely related to the filter bank method.
Most of the wavelet methods applied to ECG peak detection are derived from
the Mallat and Hwang (Mallat, 1999) algorithm for singularity detection and
classification. Peak detection is accomplished by using local maxima of the
wavelet coecients and computing the singularity degree, that is, the local
Lipschitz regularity, α , estimated from

Wf (2 j +1 ,n j +1 ) Wf (2 j ,n j ) α j =ln | |− ln | | (4.14)

Heart rate variability (HRV) algorithms examine the intervals between
heartbeats to detect potential pathologies and have, in the past, used Poincare
plots for two-dimensional visualization of the classes. Brennan et al. (2001a)
later showed that these Poincare plots extracted only linear information from
the intervals and neglected nonlinear information, thus, motivating the use
of nonlinear CI techniques. They also proposed a network oscillator interpre-
tation for the information depicted by the Poincare plots (Brennan et al.,
2002) as a transform method to extract HRV information. The authors later
investigated signal-processing algorithms, specifically, the integral pulse fre-
quency modulation (IPFM) model (Brennan et al., 2001b) to modulate the
heart intervals into digital pulses and a spectrum count method to interpret
the results.
ANNs have also been used to detect the QRS complex, the more common
networks being the MLP, RBF networks, and the learning vector quantiza-
tion (LVQ) networks. In QRS detection, NNs have been applied as nonlinear
predictors (Hu et al., 1993; Xue et al., 1992), where the objective is to predict
the value of the X ( n + 1) sample based on previous samples. The reversed
logic is used where the network is trained to recognize the nonexistence of the
QRS waveform well. Portions of the waveform that contain the QRS will cause
large network output errors, e ( n ), and can be used to infer the location of the
QRS. Suzuki (1995) developed a self-organizing QRS-wave-recognition system
for ECGs using an ART2 (adaptive resonance theory) network. This network
was a self-organizing NN system consisting of a preprocessor, an ART2 net-
work and a recognizer. The preprocessor was used to detect R points in the
ECG and segment it into cardiac cycles. The input to the ART2 network
was one cardiac cycle from which it was able to approximately locate one
or both the Q and S points. The recognizer established search regions for
the Q and S points where the QRS wave was thought to exist. The average
recognition error of this system was reported to be less than 1 ms for Q and
S points.
Genetic algorithms have also been used to design optimal polynomial filters
for the processing of ECG waveforms (Poli et al., 1995). Three types of filters
have been applied: quasilinear filters with consecutive samples, quasilinear