ECG Signal Analysis - ECG Acquisition and Automated Remote Processing - page 24

Suitability of wavelet transforms for non-stationery biomedical signal analysis,
especially ECG, was established by a number of publications in the 1990s. Due to
the multiresolution property of the wavelet transform, it has been used in the
efficient detection of QRS complexes. Many of the reported works are based on
Mallat's and Hwang's approach for singularity detection using local maxima of the
wavelet coefficient signal [ 55 - 57 ]. In [ 55 ], a spline wavelet developed for varying
QRS morphology is proposed. The dyadic wavelet transform (D y WT) was com-
puted using a wavelet which is the first derivative of a smoothing function. It
exhibited local maxima at the QRS occurrences. A continuous wavelet transform
is proposed in [ 56 ] to detect QRS. The algorithm used first-order derivative to
suppresses the noise and baseline drift and high-scale continuous wavelet trans-
form to pick the zero crossing R point produced by differentiator to ease the task of
QRS detection. In the proposed method, at first, a 3-point moving-window inte-
grator is used for low-pass filtering, followed by a first-order derivative to elim-
inate baseline drift and muscle noise. Finally, CWT is used for QRS detection. In
[ 57 ], a CWT implementation is shown using DSP processor for real-time detection
of ECG characteristic points.
Artificial neural networks and other intelligent computational techniques are in
use for ECG analysis, especially arrhythmia monitoring [ 58 ]. In [ 59 ] an adaptive
matched filter is proposed whose coefficients are updated by an ANN. A multilayer
perceptron (MLP) neural network structure is used for an adaptive whitening filter.
The QRS template used for matched filtering is updated by an ANN recognition
algorithm, which provides better adaptation to signal changes. A new method,
employing multichannel adaptive resonance theory (MART) neural network is
described in [ 60 ] for efficient QRS detection. An FIR filter based on Keiser
window is adopted for removal of PLI and BLW from the ECG. The R peaks are
initially detected by 32-point averaging and peak detection method. The RS
segment is approximated by a side of a triangle of variable width for training the
network. The Mart network uses two channels of data for detection of Q and S
points. In [ 61 ], a genetic design for QRS detection is described. At the first stage, a
linear polynomial filter is used for QRS enhancement. In order to operate on small
number of input samples, a genetic algorithm-based optimization of filter coeffi-
cients is performed, which only detect the maxima of the filtered output and
ignores QRS like spikes.
2.3.3 Feature Extraction from ECG Signal
Computerized ECG feature extractions are aimed to delineate the complete
waveforms from the ECG dataset. A detailed survey of different ECG feature
extraction techniques are discussed in [ 62 , 63 ]. In general, the feature extraction
techniques are classified into the following:
(a) Time-domain methods based on morphology and adaptive filtering;