ECG Signal Analysis - ECG Acquisition and Automated Remote Processing - page 20

filter is designed where the filter weights are optimized by an unbiased linear
ANN. Here, the network weights related to each wavelet sub-band are computed
by steepest deepest algorithm. A new technique with reduced number of hidden
layers and less computational time by using An MADALINE (Multiple ADAptive
LINear Element) structure is reported to eliminate white, artifact sand muscular
noise in [ 25 ].
Wavelet-based multiresolution analysis is one of the popular methods for ECG
denoising in recent times, due to its excellent time frequency resolution properties.
Using discrete wavelet transform (DWT), the signal can be resolved in different
frequency bands from which a combination of different coefficients can be utilized
for identification noise components and eliminating them, in principle, to get a
clean signal. In [ 26 , 27 ] the authors propose a dyadic DWT decomposition using
Daubechie's wavelet and discard appropriate coefficients in the decomposition
layer and reconstruct the signal to get a fairly accurate results. The methods
provide a quick and easy removal method for BW and muscle noise. Wavelet-
based denoising can effectively reject the noisy components. In wavelet-based
thresholding methods, the coefficients after wavelet decomposition are adjusted by
use of a threshold that can either be set as 'hard' or 'soft.' In 'hard' thresholding,
the coefficient values below the threshold are set to zero, thus, eliminating the
noise associated with those components of the signal. In 'soft' thresholding, the
values below the threshold are reduced by the same magnitude from their original
values. The processed coefficients are then reconstructed back to the time domain
to get a noise-free signal. In [ 28 ], a non-linear denoising approach was proposed
by applying soft and hard thresholding methods, in which thresholds were chosen
using four different methods, viz., 'Stein's unbiased risk estimate' (SURE) [ 29 ],
heuristic SURE (HEUR-SURE), fixed threshold (FIX-THRES), and MINIMAX
[ 30 ]. An application of wavelet thresholding method for EMG noise removal is
described in [ 31 ], and the results are compared with Vapnik-Chervonenkis (VC)
learning theory, which is related to statistical learning theory and to empirical
processes. Some more applications of wavelet-based denoising are available [ 32 -
34 ].
EMD proposed by Huang et al. is one of the fully data-driven techniques used
for non-linear and non-stationery signals and does not require a priory knowledge
of the signal. The EMD process defines a signal into a sum of intrinsic mode
functions (IMF), with equal number of extrema and zero crossing with its enve-
lope, symmetric with respect to zero. The signal is reconstructed by eliminating
the IMFs which correspond to frequency components, similar to DWT approach
[ 35 ]. In [ 36 ], the authors also deal with the problem that arises due to truncation of
R peaks resulting out of direct elimination of IMFs during reconstruction by
introducing a 'peak correction method.'