Application to biomedical signals - Practical Biomedical Signal Analysis Using MATLAB - page 175

separation method. The independent components for a set of signals can be found if
the mixing matrix D is known (Sect. 3.6.2, equation 3.52). We can expect that spe-
cific ICA components will be connected with the noise/artifact sources. Since ICA
components are transformed combinations of leads, in order to recover the informa-
tion about the ECG signals the back projection has to be performed (Sect. 3.6.2).
During this procedure the noise sources may be removed. Namely the components
corresponding to artifact/noise may be set to zero during reconstruction.
However, this procedure is not quite straightforward, since we have to determine
which ICA components are noise. The system based upon kurtosis and variance have
been devised to automatically distinguish noise components [He et al., 2006], but
the quality of results may be dependent on the particular data set. The problem is
complicated, since the ICA mixing/demixing matrix must be tracked over time and
the filter response is constantly evolving. A robust system of separating artifacts from
ECG is still to be found.
4.2.1.3.2 Morphological ECG features Clinical assessment of ECG mostly re-
lies on evaluation of its time domain morphological features such as positions, du-
rations, amplitudes, and slopes of its complexes and segments ( Figure 4.28) . The
morphological feature may be estimated using a sequence of individual heartbeats
or using averaged heartbeat. The first step in the analysis is usually detection of the
QRS complex; it serves as a marker for averaging of heart cycles, for evaluation of
heart rate, for finding the heart axis. Usually the analyzed parameters include ampli-
tudes of Q,R,S,T peaks, depression or elevation of the ST segment, durations of the
QRS complex, QT interval, dispersion of QT (difference between the longest and the
shortest ST).
The algorithms for the determination of ECG morphological time features based
on wave boundaries, positions, amplitudes, polarizations may be found, e.g., in
[Pahlm and Sornmo, 1984, Pan and Tompkins, 1985, Laguna et al., 1994].
For feature extraction from ECG beats, with the aim of beat recognition, several
approaches based on different formalisms have been proposed. Examples of such ap-
proaches include: Fourier transform [Minami et al., 1999], Hermite functions [Lager-
holm et al., 2000], wavelet transform [Senhadji et al., 1995, Yu and Chen, 2007].
An interesting approach to ECG features quantification was proposed by [Chris-
tov et al., 2006]. Each training heartbeat was approximated with a small number
of waveforms taken from a Wavelet Packet dictionary (Symlet 8). Matching pursuit
algorithm was used as the approximation procedure. During the training procedure
each of the five classes of heartbeats was approximated by 10 atoms, which were
then used for the classification procedures. The results showed high classification
accuracy. Matching pursuit was also used in the procedure of fitting the wavelets
from the Daubechies family to the ECG structures, with the aim of obtaining sparse
time-frequency representation of ECG [Pantelopoulos and Bourbakis, 2010].
At present there are also ready to use systems for ECG features extraction avail-
able on the Internet. The software for determination of the characteristic time points
of ECG may be found in PhysioNet ( http://www.physionet.org/physiotools/