|
Bottom-up Approach to Uniform Feature
Extraction in Time and Frequency Domains for Single-lead ECG Signal
T Srikantha, SA Napperb,
H Guc
aCardiac Science
Inc., Irvine, CA-92606, USA
bDepartment of Biomedical Engineering, Louisiana Tech
University, Ruston, LA-71270, USA
cDepartment of Mathematics/Statistics, Louisiana Tech
University, Ruston, LA-71270, USA
Correspondence: T Srikanth, Cardiac Science Inc.,
16931 Millikan Avenue, Irvine, CA-92606, USA.
E-mail: sthiagarajan@cardiacscience.com, phone +1 949 567 9066, fax +1
949 247 4013
 Abstract. Modern ECG classifiers
need both time and frequency domain features from the ECG beats to
improve the accuracy. Accurate calculation of time domain features are
made difficult by the errors related to the detection of T wave end point
and P wave start and end points. The objective of the present work is to
develop a feature extraction methodology from ECG beats suitable for a
wide array of ECG beat classes. The present work overcomes the impact of
possible inaccuracies in time-domain fiducial point estimations with the
spectral domain features. Spectrum is estimated after the subtraction of
a simulated normal beat from the beat under study. A set of features in
both time and frequency domains are calculated. Thirteen different
classes of beats, including atrial flutter, fibrillation and noisy beats,
are studied. Quality of information is indicated by the percentage of
rejected beats. Based on the analysis on 3854 beats belonging to 13 beat
classes, 98.31 % of the total beats provided reasonable feature
estimation.
Keywords: Electrocardiogram; Spectral
Estimation; Fiducial Points; Power Spectral Density; Mean Power
Frequency; Pattern Classification
1. Introduction
The electrocardiogram (ECG) signal has been the major diagnostic tool
for cardiologists and the ECG signal provides almost all information
about the electrical activity of the heart. In clinical practice, a
12-lead ECG provides more information compared with a single lead ECG.
However, any new methodology for automating ECG classification starts
usually on single lead data sets and the procedure is extended to
multi-lead systems, especially for the initial signal pre-processing and
feature extraction steps [Laguna et al., 1994; Trahanias and Skordalakis,
1989]. Combining information from both time and frequency domains
provides an efficient means to overcome the impact of possible errors
associated with wave boundary delineation. Trahanias and Skordalakis,
1989 advocated a bottom-up methodology in delineating the ECG waveforms
into simpler entities and later combining them to form higher ECG
patterns. In the present work, feature extraction is dealt with in a
similar bottom-up manner, with individual focus on sub-systems like
pre-processing, wave detection and wave boundary identification leading
to an overall improved performance.
2. Methods
ECG signal feature extraction in both time and frequency domains
involve standard steps including signal pre-processing, fiducial point
identification, and feature selection.
2.1. Signal Pre-Processing
ECG signal pre-processing for pattern recognition tasks involves
band-pass filtering, detrending and data normalization. In the present
work, pre-processing procedures also need to take into account the signal
characteristics of three different data sets. Such multi-center,
multi-machine data needs to be normalized in terms of amplitude,
analog-to-digital converter (ADC), and frequency characteristics. The
overall schematic is illustrated in Fig. 1.
Figure 1. Block diagram representation of the
overall ECG feature extraction.
Another important aspect of the present study is the need for
annotated databases. Based on the expert-computer interface suggested by
Laguna et al., 1997, a PC and MatlabÒ-based interactive set-up is utilized in the
present work. The interactive interface helps the experts to choose the
location of fiducial points in each beat. Annotations show the location of
the fiducial points indicated by experts. The screen for the selection of
fiducial points and the selected fiducial points in a beat are indicated
in Fig. 2. Such annotated data provide an evaluation set for fiducial
point identification algorithms. The actual fiducial points for R and S
fiducial points are the nearest integer locations in x-axis, satisfying
the extremum conditions. Once the expert clicks at a location for R and S
fiducial points, the algorithm searches for the two nearest integer locations
on either side and finds the location furthest away from baseline. This
method confirms the intended choice of the expert.
Figure 2. (a) & (b) A sample graphical
user interface and the fiducial points identified by an expert on
the screen.
The sampling frequency for data from different sources varies between
200 to 500 Hz. Normalization of sampling frequency to a single value is
vital in developing and testing new signal processing algorithms. In this
study, all signals are converted to a single sampling frequency of 500 Hz
in order to preserve maximum information. Based on numerical values
of the time and frequency domain features as well as visual criteria,
cubic interpolation is chosen based on the simplicity in implementation
and the least distortion [Srikanth et al., 1998]. A low-pass filter with
a cut-off of Fc=100 Hz is applied to both the Massachusetts
Institute of Technology-Beth Israel Hospital (MIT-BIH) data and clinical
data after de-trending. The filter is an IIR filter of Chebyshev-Type 1
of sixth order.
2.2. QRS Complex and Fiducial Point Identification
Direct measurement of the various intervals, (e.g. RR and PR
intervals) requires the knowledge of the locations of the boundaries (the
onsets and ends) of the P, QRS, and T waves. Problems in detecting
fiducial points arise due to non-linearity, non-stationarity, missing
beats and noise. Filtering and signal pre-processing alone may not
provide an ideal solution for tackling these problems. Design of high
performance fiducial point detector requires ideas from mathematical
morphology, geometry, and rhythm information.
In the present work, a combination approach is used
for QRS detection and subsequent R-point detection. Figure 3 indicates
the detailed block diagram of the QRS detector. A combination of two
different schemes is implemented to accommodate all types of QRS
complexes. The major algorithm uses short-term energy calculation using
moving window integrator (MWI). Initial pre-processing for the sub-system
involves high-pass filtering of the signal to highlight the sharp
characteristics of the QRS complex. A Chebyshev type-I IIR filter of
sixth order is used for high-pass filtering the signal at Fc=0.5
Hz.
The basic steps in the MWI are differentiator, squaring algorithm, and
summation. Implementation is based on the method applied on long-term
data by Pan and Tompkins, 1985. The width of the moving window is
determined empirically; taking into consideration that the QRS complex
occupies about one-sixth of a second (»150
ms). In the present work, time-domain fiducial point detection is
implemented with a sampling frequency (Fs) of 250 Hz and other
algorithms including feature calculations and frequency domain spectral
estimations are performed after converting Fs to 500 Hz by
cubic interpolation. For a sampling rate of Fs=250 Hz, a
window length of 40 samples is chosen. MWI output contains information
about both the slope and the width of the QRS complex. The upward
slope of the MWI for each beat is found to correspond to the QRS complex
[Pan and Tompkins, 1985].
Figure 3. Combinational algorithm for QRS
detection.
Amplitude and interval thresholds in the detector include mean
baseline amplitude (after removing QRS complexes), mean amplitude,
maximum and minimum amplitudes of ECG signal and MWI output. The location
of the previous QRS complex is taken into account during the search for
the present QRS complex; i.e., the algorithm skips a blanking
period of 120 ms after detecting a fiducial point. This interval is
decided based on the physiology of the pace maker and pulse propagation
inside the heart. Due to the short data duration of six seconds in the
present study, a learning phase of few initial seconds, as mentioned by
Pan and Tompkins, 1985, is not possible in present implementation. Hence,
the algorithm is not adjustable and is prone to minor errors. Two
remedies are implemented in the present work:
i)
Back-tracking during fiducial point detection to see if all the R
points are detected properly.
ii)
Whenever a beat is missed or the algorithm does not find a beat
for more than 1.5 seconds, second method of QRS detection making use of
maximizing the first difference, is attempted.
The MWI tends to give poor performance for the following cases;
i)
beats dominated by noise.
ii)
data segments with QRS complexes with varying morphologies.
iii)
rare presence of very low amplitude QRS complexes in the middle of
a sequence of normal QRS complexes, and
iv)
beats in which QRS complexes followed by very high T waves.
In such situations, maximizing the first difference provides adequate
performance. If one assumes a simple situation with uniform distribution
of ECG amplitudes, the estimation procedure carries out a nonlinear
transformation of the linearly filtered ECG is given by
| |
 |
(1) |
where, t1 and t2 are positive integers.
This differentiator operation is different from the 5 point
differentiator used in Pan and Tompkins algorithm.
The set of values, z(n) in equation (1) is thus the rectified,
maximum amplitude difference between two samples. In the present
study, a first order difference equation is used with t1 =T=1 and a variable t2, dependent on the moving
window width and the previous RR interval. The location of the
maximum difference is taken as the fiducial point location. This
detection process provides good fiducial point estimation when the
regular MWI algorithm fails to locate the exact fiducial point in the
above mentioned situations.
2.3 Overall Estimation of Fiducial Points
Fiducial points detected in the present work include P wave on- and
off-set points (P1 and P2), QRS complex on- and
off-set points (Q start, Q point, R point, S point and J point) and T
wave on- and off-set points (T1 and T2). The
detection is performed in three stages, namely signal smoothing, wave
detection, and wave boundary identification for P and T waves. The J
point is defined as the junction between a QRS complex and its ST
segment. A negative wave at the onset of the QRS complex is defined
as Q wave and the valley is defined as Q point. Occasionally, Q
wave may be absent. The junction of the baseline and Q wave is
defined as Q start point. The negative deflection following an R wave is
called an S wave [Wagner, 1994].
Signal Smoothing
The input to the signal smoothing is the original ECG signal in the
frequency range 0.1-100 Hz for the MIT-BIH database and 0.05-100 Hz for
the clinical database used in this study. Clinical
database bandwidth was chosen to increase the accuracy of ST-segment
information. Signal smoothing is performed before detecting P and T
waves and after R wave detection. A simple causal smoothing makes
use of a summation operation for the few previous points. Equation
(2) exactly indicates such a causal relation.
| |
y(n)=(x(n)+x(n-1)+x(n-2)+x(n-3))/4 |
(2) |
The smoothing operation is a low-pass filter operation and eliminates
noisy and turbulent components. Detection of both peaks and edges
of the component waves becomes easier after smoothing.
Component Wave Detection
The algorithm below indicates the step-by-step identification of
fiducial points, starting from QRS detection. The algorithm makes use
of some information from previous beats (after the first beat).
Step 1: Single lead QRS detection, using
the MWI algorithm, is done. When the QRS complex is missed for more than
1.5 seconds, a search is performed using the first-order differentiation and
maximizing the output.
Step 2: Calculation of tentative boundaries of
QRS complex, using the bottom and top edges of the rising arm of the MWI
output, and search for the R point. In the case of using maximum of
the first difference, approximate the boundaries. If QRS
characteristics differ, go to fiducial point calculation for ventricular
beats, i.e., alternate steps 3a - 6a.
Step 3: Searching along the right side of
the detected R point for an S point, based on duration, zero-crossing of
slope, and angle between the lines formed by successive points. The
formula for angle between two lines, given their start and end points is
provided by
| |
q(i) = atan((m1+m2)/(1-m1*m2))*(180/pi)
|
(3) |
where, m1= ecg(i)-ecg(i-1) and m2=ecg(i+1)-ecg(i).
Here, {…ecg(i-1),ecg(i),ecg(i+1),….} indicate successive digitized
ECG sample points.
Step 3a: Here, presence of premature
ventricular beat is assumed and a large negative Q wave (Qr) is detected
from the tentative QRS boundary calculations.
Step 4: The detection of a J point is
also based on the angle and direction of the slope change. The
changes are small compared to S wave detection.
Step 4a: Choosing S point as the point of
baseline crossing. The J point location varies with the location of
S point.
Step 5: Searching for T wave peak based
on the definition of peak, after smoothing the curve. A smooth peak
shows a maximum with at least two descending points on either side as
well as a sign change in slope at the peak. Another constraint for
the peak should be that it should be above an adjustable amplitude
threshold, based on R peak amplitude. Local baseline and the
amplitude difference with respect to baseline are used and, hence, even
inverted T waves are detected easily.
Step 5a: T wave detection using simple
peak calculation after the J point and before the next beat.
Step 6: Proceeding with the P wave
detection and the process is similar to T wave peak detection, except
that the search is in the other direction of QRS complex. A peak is
defined as a local maximum where the sign of the derivative changes and
smooth descent occurs on either side. With this step, the first
stage of wave detection gets completed for the normal beat. The
absence of a P or T peak is also noted. Cross-checks are introduced
for the negative wave detection.
Step 6a: If the P wave is absent, P wave
start and end points correspond to the same point. The points are
positioned after a previous T point and before the start of present beat
(Q start point).
Wave Boundary Identification
An important finding of the CSE study is that the algorithms tend to
locate the end points of the T wave significantly earlier than human
experts do [Willems et al., 1990]. The end points of the P
or T waves are defined in terms of average baseline amplitude in the
present work. Baseline amplitude is approximated as the mean value
of the beat after removing the QRS complex, and this value tends to be
closer to actual baseline. The end points (on- and off-sets) tend to
merge smoothly with the baseline, and this fact is made use of in the
present detection of end points. A derivative value of around zero
occurs nearer to end points, compared with the rising and falling arms of
P and T waves. Sometimes, a derivative sign change occurs, but it
is not a reliable indicator. Hence, a combination of criteria based
on curvature, slope and merging with baseline are used for endpoint
detection for both P and T waves.
A fiducial point is assumed for cases like premature
ventricular contraction (PVC) that have no visible P waves.
Inversion of waveform causes another problem. Inversion of P and T
waves usually does not matter with respect to baseline amplitude, and end
points usually return back to baseline level. However, waveform
asymmetry in the component waves causes problems for algorithms,
especially the asymmetry between ascending and descending arms of the
waves. Approximate locations in such cases are decided based on the
length of the other arm, previous beat intervals, and the conflicts with
the next fiducial point, if available. Location of J point and the
calculation ST segment slope are not accurate under MIT-BIH frequency
bandwidth of (0.1-100) Hz. Hence, the values of ST segment slope
are considered zero for MIT-BIH beats and calculated only for the
clinical ECG beats with a bandwidth of (0.05-100) Hz.
2.4. Feature Selection
In ECG interpretation, a feature set of size 10-20
values, consisting of intervals and amplitudes and quantities derived
from them, is found to provide as much information as an entire beat of
raw signal [Borovsky and Zywietz, 1980]. The end-users, physicians
and cardiologists, have also been trained to think in terms of features
with the last four decades of computerization effort. Another
advantage of feature extraction occurs with respect to noise in observed
beats. The impact of noise is relatively easier to reduce in the
initial fiducial point detection than in the final classification step.
Another major gain from the feature extraction algorithms is the
elimination of non-stationarity characteristics, before the
classification stage. The problem changes from one of stochastic
non-stationary process to one involving normally distributed random
variables. A compromise has to be reached in choosing parameters
based on familiarity, input space, computational load and fit into the
overall scheme.
ECG Features Calculated in Time Domain
Time-domain features calculated in the present work
are PP interval (T1), mean base-line (T2), QRS shape and sign (T3), ratio
of areas under QRS and T waves (T4), JT1 interval (T5), ST
segment slope (T6), PR interval (T7), T wave inverted or upright (T8), TP
interval (T9), P wave duration (T10), ratio of upward versus downward
slopes for QRS complex (T11), ratio of R amplitude and Q amplitudes
(T12), QT interval (T13), noise reduction after smoothing (T14), number
of zero-crossings (T15), and previous TP interval (T16). Baseline
amplitude is taken as the mean amplitude in the TP interval. Time
domain features essentially follow the standard definitions derived from
Wagner, 1994, Greenhut et al., 1989 and Willems et al., 1990. Noise
reduction after smoothing (T14) indicates the reduction in variance of
the signal after the signal smoothing. This feature is an effective
parameter for noisy beats, which show a very high reduction in variance
after smoothing compared to regular beats. The other non-regular
feature is the number of zero crossings in a beat which points towards
the presence of atrial fibrillation and atrial flutter beats. The
zero crossings are calculated with respect to the estimated baseline.
Frequency Domain Features
Applications related to ECG signals in frequency domain have used both
Fourier-transform based estimation methods and parametric methods
[Berbari et al, 1990; Murthy and Niranjan, 1992]. In their
comparison of four different techniques for recognition of ventricular
fibrillation and atrial flutter, Clayton et al., 1993, advocated the
higher amount of information provided by the signal spectra, compared to
time domain methods and methods based on auto-correlation
calculations. They have suggested the importance of combining
spectral algorithms with an efficient QRS detector and activating the
spectral domain detection only if no QRS complex can be detected.
This concept is vital in the current work and eliminates the necessity
for excessive computation.
Recent trend points to more direct application of frequency domain
analysis related to individual waves in ECG beats. Langberg et al.,
1998, presented a new technique to analyze atrial fibrillation. The
major steps in the procedure are band-pass-filtering, subtraction of QRST
segment, and Fourier transformation. The study suggests that the
frequency analysis of the surface ECG may be a useful means to augment
clinical decision-making. Shkurovich et al., 1998 also
performed similar beat-by-beat analyses in the frequency domain using
cancellation techniques.
Power spectral estimation in the current scheme uses a combination of
modeling and subtraction of individual wave components. In the
present work, two different types of spectral estimations are
performed. The first estimation is based on the original beat with
no subtraction of individual components. Burg's algorithm is used to
estimate the power spectral density with order P=32 based on
auto-regressive modeling [Kay, 1992].
A combination approach involving simulation and subtraction of
individual components in a beat is applied to generate the second
spectral estimation. Each beat is modeled as a linear combination
of the normal sinus rhythm component and a rhythm disturbance component.
| |
beat(t) = nsr(t)+disturb(t) |
(4) |
where,
beat(t)= an observed beat
nsr(t) = normal
sinus rhythm beat of matching timings and
disturb(t) =
disturbance in the beat
Disturbance in the beat, disturb(t), is used to
estimate the spectrum instead of the observed beat. The normal
sinus rhythm component is simulated based on amplitude and interval
calculations of each subject's observed ECG beat. A normal sinus
rhythm beat corresponding to an observed PP interval is simulated using
MatlabÒ-Simulink
algorithms. A linear superposition of P waves, QRS complex and T
waves with standard timings and with amplitude matching are used to
simulate waveforms. Peter Bonadonna, [1998], in his article on
understanding QT/QTc measurements, provides a QT chart and a QT duration
graph showing the variation of QT interval with respect to RR interval values.
Standard variations in time intervals like QT interval, PR interval and
TP intervals are derived from literature [Bonadonna, 1998; Wagner, 1994].
The simulation model recognizes that major interval
changes occur more in base-line intervals compared to wave
duration. Hence, changes in heart rate and PP intervals have
maximum impact on TP interval, followed by PR interval and ST segments,
in that order. The resulting difference, disturb(t), is modeled
using an auto-regressive (AR) model using Burg's algorithm. An auto
regressive model of order P=32 is used for modeling and spectral
estimation. The number of points chosen for the estimation is 1024
in order to provide adequate resolution [Srikanth et al., 1999; Voss et
al., 1992].
Selected Features in Frequency Domain
Selected frequency domain features include
1. Mean power frequency of the original beat
(mpf1) (F1): This feature indicates the distribution of
power across the frequency band of (0-125) Hz. The formula for
calculating mpf1 is given in equation (5).
| |
 |
(5) |
where, P(f(i)) indicates the power spectral
densities (PSD) at frequency f(i). The actual resolution of the
spectrum was dependent on number of points chosen to estimate the
spectrum based on AR model. In the present case, resolution is 0.25
(~250/1024) Hz.
2. PSD ratio for the band (0.5-6) Hz for the
original beat (psd1) (F2): Indicates the fraction of the power in (0.5-6)
Hz band. The calculation is shown in (6).
| |
 |
(6) |
3. PSD ratio for the band (6-12) Hz for the
original beat (psd2) (F3): Indicates the fraction of the power in (6-12)
Hz band and is calculated similarly as (6).
4. PSD ratio for the band (12-18) Hz for the
original beat (psd3) (F4): Indicates the fraction of the power in (12-18)
Hz band and is calculated similarly as (6).
5. Mean power frequency of (original-simulated)
beat (mpf2) (F5): The formula for calculation is similar to (5); only
difference being P(f(i)) replaced by P1(f(i))
where P1(f(i))=power spectrum of disturb(t) from
equation (4).
6. Mean power frequency of (0.5-10) Hz band
(mpf3) (F6): This feature indicates the distribution of power in the
spectrum of disturb(t) in the frequency band (0.5-10) Hz. The
formula for calculation is given in equation (7).
| |
 |
(7) |
7. Mean power frequency of (10-20) Hz band
(mpf4) (F7): This feature indicates the distribution of power in the
spectrum of disturb(t) in the frequency band (10-20) Hz. The
formula for calculation is similar as in equation (7).
8. Mean power frequency of (20-30) Hz band
(mpf5) (F8): This feature indicates the distribution of power in the
spectrum of disturb(t) in the frequency band (20-30) Hz. The
formula for calculation is similar as in equation (7).
9. PSD ratio for (0.5-6) Hz band for (original
beat-simulated beat)(psd4) (F9): This feature indicates the distribution
of power in the spectrum of disturb(t) in the frequency band
(20-30) Hz. The formula for calculation is similar as in equation
(7).
10. PSD ratio for (6-12) Hz band for (original
beat-simulated beat)(psd5) (F10): This feature indicates the fraction of
power in (6-12) Hz for the disturb(t). The formula is given in
equation (8).
| |
 |
(8) |
11. PSD ratio for (12-18) Hz band for
(original beat-simulated beat)(psd6) (F11): This feature indicates the
fraction of power in (12-18) Hz for disturb(t). The formula is
similar to equation (8).
Parameters F9, F10 and F11 provide information about the impact of QRS
complex relative to rest of the beat. The information is amplified
by the subtraction of a simulated beat from the original beat.
Standard definitions related to spectral estimation are found elsewhere
[Kay, 1992].
3. Data Set and Results
Electrocardiogram data used in the present work consists of 430 data
segments from standard MIT-BIH database and 65 data segments from routine
clinical ECG data recorded from Kilpauk Medical College Hospital,
Chennai, India; and Madras Medical College Hospital, Chennai,
India. All chosen data are derived from limb lead II or modified
limb lead II (ML II). Modified limb leads are similar to lead II,
except that the electrode positions are placed on the chest and hip
locations closest to the limbs [Moody GB, 1992]. Data
characteristics in time and frequency domains for ML II are same as limb
lead II. At least two independent physicians and/or cardiologists
annotated each beat in all the data segments chosen for the study from
both sources. ECG segments varied in duration from six to ten
seconds.
Table 1.
Statistics on data set
3.1. Results of QRS Detection
Table.2 presents the results from QRS detection, tested with the
overall set of 3854 beats selected from both MIT-BIH database and
clinical data. QRS complex identification was successful even in
some cases where the rest of the beats are noisy.
Table 2.
Statistics on QRS detection
3.2. Results of Overall Fiducial Point Estimations
Performance of overall fiducial point estimation methods is dependent
on QRS detection efficiency. Figures 5-12, illustrate a variety of
problems associated with both QRS detection and overall fiducial point
identification. The exact quantity providing the numerical accuracy
of fiducial point detection algorithms is the number of discarded beats,
due to abnormal value in features (Table 2.). Total number of
discarded beats constitutes just 1.5% of the total of 3854 beats.
An accuracy of 98.5% is quite excellent, considering the choice of beat
classes include atrial flutter, atrial fibrillation, ventricular beats
and extremely noisy beats. In the present work, the accuracy of
detected fiducial points is decided by an ability to estimate feature
values for further classification. This criterion is not as strict
as the exact location criteria suggested by others [Laguna et al., 1994;
Greenhut et al., 1989], but an attempt is made to incorporate the
expert feedback on exact locations.
Figure 5. Errors in T wave identification for
atrial premature beats.
Figure 6. Fiducial point detection in a simple
data segment, with no complexities.
Figure 7. A case of signal corrupted by large
noise -- one of the discarded segments.
Figure 8. Fiducial point detection in PVC beat
in the middle of Left Bundle Branch Block beats.
Figure 9. Fiducial point detection in Atrial
Fibrillation beats - approximate Pstart(P1) location is
enough.
Figure 10. Fiducial point identification in a
noisy segment.
Figure 11. Fiducial point detection in a
simple right bundle branch block segment.
Figure 12. Fiducial point detection in atrial
fibrillation with inverted T waves.
Fiducial point detection algorithms are evaluated for the entire data
segments and also for individual beats. Figure 6 shows a clean ECG data
segment with no detection errors. Possible errors in fiducial point
detection are shown in Figures 5,7, and 8-12. Presence of an
ectopic beat usually complicates the fiducial point detection as seen in
Figures 8, 9 and 12. Inverted T waves are detected with no errors
as indicated in Figure 12. Only three data segments out of initial
498 segments were discarded due to entirely bad detection, and a sample
is shown in Figure 7. Errors in other segments vary from no errors to
errors in multiple beats. In cases of atrial fibrillation, atrial
flutter and ventricular beat, approximate detection of start and end
points of a beat are enough to estimate features in the frequency
domain. Accuracy of fiducial point detection in these beat classes
is not well discussed in the literature and in the present work,
frequency domain features play a prominent role in characterizing these
beats.
4. Discussion
The first advantage of present approach is the inclusion of fiducial
point detection and feature extraction systems in implementation and the
second one is the ability to include complementary information from
frequency domain, a derivation of the first advantage. A third
advantage may be the inclusion of noisy beats in implementation of every
sub-system. Finally inclusion of the re-sampling procedure provides
a machine-independent implementation to include data acquired at
different conditions.
Fiducial point extraction and QRS detection algorithms form a major
part of the present implementation. A high accuracy in both steps
is the norm in last few years [Laguna et al., 1994; Trahanias and
Skordalakis, 1989]. A performance of nearly 99% is not uncommon in many
new approaches utilizing wavelet transforms and neural networks etc
[Li et al., 1995]. Pan and Tompkins, 1986, achieved a near
perfect performance in QRS detection and continuously improved upon it
[Xue et al., 1992]. Most of these studies were performed with
standard databases, especially MIT-BIH database. Accuracy of
commercial algorithms has been quite good in the detection of QRS and
fiducial point detection algorithms. However, there have been minor
problems related to wave boundary estimations, and the need for accuracy
increases depending on the nature of information sought.
Interestingly, detection of fiducial points for beats belonging to
fibrillation and flutter classes have not been attempted on a large scale
nor discussed much statistically. Even manual definitions of exact
fiducial points are difficult in such cases. Usually, those beats
are detected early after QRS detection [Sornmo and Pahlm, 1984].
Discussions on the performance for noisy beats have been minimal in the
literature.
In the present work, accuracy of the algorithms has been quite high
for short duration signals. Noisy beats, paced beats, premature
ventricular beats, atrial flutter and atrial fibrillation beats
constitute around 30% of the data set in present work. Performance
needed for fiducial point detection algorithms for such beats are kept
minimal and two major criteria in accepting a beat for further
classification tasks are:
i)
Ability to detect the QRS complex and to approximate the beat
boundaries.
ii)
Ability to provide acceptable feature values in time domain.
The gold standard is comparison with the annotated locations by the
experts and also the statistics of the evaluated features. Clear
outliers in the value of features are rejected. Only three data
segments out of 498 got rejected based on the above two criteria, though
the overall rejection rate for beats is around 1.5%, at the input stage
of the neural network based pattern classifier described in the companion
paper [Srikanth et al.,
2002]. Inclusion of difficult classes and noisy beats
seems to be the reason for this rejection rate.
5. Conclusions
Syntactic rule-based systems provide high accuracy in the separation of
beats and in the location of fiducial points. Beat delineation is
obviously a linearly separable problem and hence, precise syntactic
algorithm satisfies the role in separation of individual beats.
Frequency domain information adds to the information in time domain, in
select beat classes where the accuracy of time domain features is
suspect. The feature extraction procedure discussed in the present
article provides an ideal setting for verifying the nature of useful
information provided by the selected time and frequency domain features
using a classifier. Multi-lead ECGs should play a significant role
in further reduction in fiducial point location, as the fiducial point
detection is attempted in two or more channels, especially in cases of independent
noise.
Acknowledgements
The authors wish to acknowledge Dr. R.W. Schubert for useful
discussions on ECG beat modeling, Dr. S. A. Jones for discussions on ECG
beat modeling and spectral estimations and Ms. Nandhini for the
documentation help.
References
Berbari EJ, Albert
DE, Lander P. Spectral estimation of the
electrocardiogram, in Electrocardiography Past and Future. Annals
of The New York Academy of Sciences. Coumel P,Garfein OB, Editors. New
York, 1990, 197-208.
Bonadonna P. Understanding the QT/QTc measurement. Monroe Community
College. http://www.monroecc.edu/depts/pstc/parasqt.htm.
Rochester, New York, 1998.
Borovsky D, Zywietz Chr. Feature extraction from a short - term
record, in Optimization of Computer ECG Processing. Wolf HK,
Macfarlane FW, Editors. North Holland Publishing Company, IFIP, 1980,
109-116.
Clayton RH, Murray A, Campbell RWF, Comparison of four
techniques for recognition of ventricular fibrillation from the surface
ECG. Medical & Biological Engineering & Computing,
31: 111-117, 1993.
Greenhut SE, Chadi BH, Lee JW, Jenkins JM, Nicklas JM. An
algorithm for the quantification of ST-T segment variability. Computers
and Biomedical Research, 22: 339-348, 1989.
Kay SM. Modern Spectral Estimation. Prentice-Hall Signal
Processing Series, NJ, 1992.
Laguna P, Jane R, Caminal
P. Automatic detection of wave boundaries in multilead
ECG signals: Validation with CSE database. Computers and
Biomedical Research, 27: 45-60, 1994.
Laguna P, Mark RG, Goldberg A, Moody GB. A database for evaluation
of algorithms for measurement of QT and other waveform intervals in the
ECG. In Proceedings of Computers in Cardiology, 1997, 673-676.
Langberg JL, Burtle JC, McTeague KK. Spectral analysis of
electrocardiogram predicts recurrence of atrial fibrillation after
cardioversion. Journal of Electrocardiology, 31(Supplement):
80-84, 1998.
Moody GB. MIT-BIH Arrhythmia Database Directory, Harvard
University-MIT Division of Health Sciences and Technology, 1992.
Murthy ISN, Niranjan UC. Component wave delineation of ECG by
filtering in the Fourier domain, Medical & Biological Engineering
& Computing, 30: 169-175, 1992.
PahlmO, Sornmo L. Software QRS detection in ambulatory
monitoring -- a review. Medical & Biological Engineering &
Computing, 22: 289- 297, 1984.
Pan J, Tompkins WJ. A Real Time QRS Detection Algorithm. IEEE
Transactions on Biomedical Engineering, 32: 230-235, 1985.
Shkurovich S, Sahakian AV, Swiryn S. Detection of atrial
activity from high-voltage leads of implantable ventricular fibrillators,
using a cancellation technique. IEEE Transactions on Biomedical
Engineering, 45: 229-234, 1998.
Srikanth T, Napper SA, Gu H. A combination approach to frequency
domain analysis of electrocardiogram signals, In Proceedings of
IEEE-EMBS/BMES Joint Meeting, Atlanta, 1999.
Srikanth T, Napper SA, Gu H. Assessment of re-sampling
techniques for electrocardiogram signals for feature extraction,
statistical and neural network applications. In Proceedings of Computers
in Cardiology, Cleveland, 1998, 537-540.
Srikanth T, Napper SA, Gu H, A Bottom-Up
Design of Neural Network Based ECG Beat and Rule-based Rhythm Classifier.
International Journal of Bioelectromagnetism, 4, 1, 2002.
Thakor NV, Webster JG, Tompkins WJ. Estimation of QRS complex
power spectra for design of a QRS Filter. IEEE Transactions on
Biomedical Engineering, 31(11): 702-706, 1984.
Trahanias P, Skordalakis E. Bottom-up approach to the ECG
pattern-recognition problem. Medical & Biological Engineering
& Computing, 27: 221-229, 1989.
Voss A, Kurths J, Fiehring H. Frequency domain analysis of
highly amplified ECG on the basis of maximum entropy spectral estimation.
Medical & Biological Engineering & Computing, 30: 277-282,
1992.
Wagner GS. Marriot's practical electrocardiography. Ninth
Edition, Williams & Wilkins, 1994.
Willems JL, Arnaud P, van Bemmel JH, Degani R, Macfarlane PW, Zywietz
Chr. Common standards for quantitative electrocardiography : goals
and main results. Methods of Information in Medicine, 29:
263-271, 1990.
| 
Home
Current Issue
Table of Contents
