ECG Signal Analysis - ECG Acquisition and Automated Remote Processing - page 25

(b) Time-frequency analysis using wavelet and other non-linear transforms;
(c) PCA, neural networks, and other methods.
Computerized interpretation of electrocardiogram is started in the early 1960
with the introduction of digital computers in the USA by Caceres et al. In [ 64 ], a
tele-transmission and computerized processing is described. The data acquisition
from the patient was done in a portable ECG machine, coupled with a patient
coder and FM tape recorder by a technician. The incoming ECG was recorded and
processed by an IBM 360/50 computer. The system showed fairly accurate result
acceptable to cardiologists. Later on, in the 1970s, microprocessor standalone units
were in use for automated interpretation. In [ 65 ], use of a microcomputer-based 3-
channel ECG analysis system for P wave, arrhythmia, and axis analysis is
described. In [ 66 ], a time-domain morphology and gradient-based algorithm is
presented. The algorithm is based on a combination of extrema detection and slope
information, with the use of adaptive thresholding to achieve the extraction of 11
number of time instances. These 11 time signatures are onset and offset points of
waves and wave peaks. After the initial QRS detection, the onset and offset of QRS
are detected, by considering them as points of maximum or minimum electrical
activity around them. From the S-offset T, T-onset, and T-off set are detected by a
zonal search of appropriate width imposed with slope-based criteria. To detect the
wave peaks in 'double humped' cases, adaptive threshold values are used to
enhance the detection accuracy. A statistical method based on comparison of
relative magnitudes of ECG samples and their slope in time domain is described in
[ 67 ]. In [ 68 ], an investigation on detecting the boundaries of P and T wave is
carried out using the LMS algorithm. The authors propose an adjustment of the
adaptation constant imposed with an extreme condition to determine the end of T
wave. A morphology-based ECG heartbeat classifier is reported in [ 69 , 70 ]. From a
training data using MIT-BIH arrhythmia dataset, a set of 12 features are selected
based on ECG morphology, heartbeat intervals, and RR intervals. After the pre-
processing for noise elimination, heartbeats are segmented followed by fiducial
point detection. Linear discriminants (LDs) were used to operate on heartbeat
interval features and segmented ECG morphology features. A similar study [ 71 ]
used Kohonen self-organizing maps (SOM) and linear vector quantization algo-
rithms to operate on ANSI/AAMI EC57 standard data.
Use of wavelet transform for ECG characterization is described [ 72 - 74 ]. In
[ 75 ], the QRS detection is performed by modulus maxima of wavelet transform.
The modulus maxima of a normal biphasic QRS are also biphasic. Hence, an
adjustable threshold is used to select the QRS in the scale of 21-24 in wavelet
transform. The QRS-onset and QRS-offset are selected at first modulus maxima
pair. The onset and offset of P and T waves are searched in the scale of 2 3 . The P
wave searched in a 200-ms window ahead of Q-onset, and for T wave, the same
procedure is adopted after S-offset. The algorithm is implemented on a TMS
320C25 DSP which accesses 12-lead ECG at a 16-bit resolution and sampling of