Application to biomedical signals - Practical Biomedical Signal Analysis Using MATLAB - page 178

untypical propagation of the depolarization vector or distinctly weaker amplitude
of depolarization from particular fragments of the cardiac muscle. It was reported
that high resolution vectorcardiography analysis permits a fast, non-invasive, and
cheap confirmation of myocardial infarction and determination of its localization in
persons whose standard and exercise ECG as well as echocardiography do not reveal
changes. It also provides the recognition of changes in the depolarization amplitude
related to, e.g., the effect of a drug on the cardiac muscle in ischemia [Krzyminiewski
et al., 1999]. At present the on-line vectorcardiography is available and may be used
for patient monitoring in hospital [Lundin et al., 1994].
4.2.1.3.4 Statistical methods and models for ECG analysis Statistical methods
and models are applied for ECG segmentation, feature extraction, and quantification.
For ECG analysis probabilistic models may be used, in particular hidden Markov
models (HMM) (Sect. 2.2.1) were applied for the signal segmentation. The first step
is to associate each state in the model with a particular feature of the ECG, namely
to connect the individual hidden states with structures such as: P wave, QRS com-
plex, etc. The next step involves training of HMM, usually by supervised learning.
The important aspect is the choice of observation model to be used for capturing the
statistical characteristics of the signal samples from each hidden state. As the ob-
servation models, Gaussian density, autoregressive model, or wavelets may be used.
The latter two methods were reported as performing better [Hughes, 2006].
For ECG quantification and identification of the features helpful for distinction of
pathological changes in ECG, the analysis based on statistical procedures such as
PCA and models such as AR may be applied without making explicit reference to
time-amplitude features of ECG structures.
AR model technique was proposed to classify different types of cardiac arrhyth-
mias [Ge et al., 2007]. Preprocessing involved removal of the noise including respi-
ration, baseline drift, and power line interference. The data window of 1.2 s encom-
passing R peak was used. AR model of order 4 was fitted to the data. Four model
coefficients and two parameters characterizing noise level were used as input param-
eters to classification procedure. For classification, quadratic discriminant function
was used. The performance of classification averaged over 20 runs (different training
and testing data sets), showing specificity and sensitivity above 90% for each of six
classes of arrhythmias. These results demonstrated the usefulness of the AR model
for quantification of ECG. The model may also be used for ECG data compression
and in the portable telemedicine ECG systems, since the algorithms are relatively
simpleandworkfast.
Wavelet transform was used in analysis of high-resolution ECG for risk evalua-
tion in tachycardia [Lewandowski et al., 2000]. The time-frequency maps were ob-
tained by modified Morlet wavelet ( Figure 4.31) . For quantification of late potentials
(appearing after QRS complex) the index called irregularity factor, quantifying the
variation of energy, was proposed.
For ECG feature extraction and reduction of redundancy, the orthonormal function
model may be used. It is based on the Karhunen-Loeve transform, which provides