ECG Analysis Systems
I.Interpretation of the 12-lead ECG
ECG interpretation starts with feature extraction, which has two parts as shown in the figure. The goals
of this process are (1) waveform recognition to identify the waves in the ECG including the P and T
waves and the QRS complex, and (2) measurement to quantify a set of amplitudes and time durations
that is to be used to drive the interpretation process. Since the computer cannot analyze the ECG
waveform image directly like the human eye-brain system, we must provide a relevant set of numbers
on which it can operate.
The first step in waveform recognition is to identify all the beats using a QRS detection algorithm .
Second, the similar beats in each channel are time-aligned and an average (or median) beat is produced
for each of the 12 ECG Analysis Systems 267 leads . These 12 average beats are analyzed to identify
additional waves and other features of the ECG, and a set of measurements is then made and assembled
into a matrix. These measurements are analyzed by subsequent processes .
Computerized Interpretation of the ECG:
There are two basic approaches for computerized interpretation of the ECG.The one used in modern
commercial instrumentation is based on decision logic . A computer program mimics the human
expert’s decision process using a rule-based expert system. The second approach views ECG
interpretation as a pattern classification problem and applies a multivariate statistical pattern
recognition method to solve it . Figure shows the complete procedure for interpretation of the ECG.
The feature extraction process produces a set of numbers called the measurement matrix. These
numbers are the inputs to a decision logic or statistical process that drives an interpretation process
which assigns words to describe the condition of the patient.
The decision logic approach is based on a set of rules that operate on the measurement matrix derived
from the ECG. The rules are assembled in a computer program as a large set of logical IF-THEN
statements.
The rules are usually developed based on knowledge from human experts. The pathway through the set
of IF-THEN statements ultimately leads to one or more interpretive statements that are printed in the
final report. Unfortunately, it is well known that a group of cardiologists typically interpret the same set
of ECGs with less than 80 percent agreement. In fact, if the same ECGs are presented to one
cardiologist at different times, the physician typically has less than 80 percent agreement with his/her
previous readings. Thus, a decision logic program is only as good as the physician or group of
physicians who participate in developing the knowledge base. One advantage of the decision logic
approach is that its results and the decision process can easily be followed by a human expert.
However, since its decision rules are elicited indirectly from human experts rather than from the data, it
is likely that such a system will never be improved enough to outperform human experts. Unlike
human experts, the rule-based classifier is unable to make use of the waveforms directly. Thus, its
capability is further limited to looking at numbers that are extracted from the waveforms that may
include some measurement error. Also, with such an approach, it is very difficult to make minor
adjustments to one or few rules so that it can be customized to a particular group of patients.
For the multivariate statistical pattern recognition approach to ECG interpretation, each decision is
made directly from the data; hence this approach is largely free from human influence. Decisions are
made based on the probabilities of numbers in the ECG measurement matrix being within certain
statistical ranges based on the known probabilities of these numbers for a large set of patients. Since
this technique is dependent directly on the data and not on the knowledge of human experts, it is
theoretically possible to develop an interpretive system that could perform better than the best
physician. However, unlike the decision logic approach, which can produce an explanation of how the
decision is reached, there is no logic to follow in this approach, so it is not possible to present to a
human expert how the algorithm made its final interpretation. This is the major reason that this
technique has not been adopted in commercial instrumentation. In clinical practice, physicians overread
and correct computerized ECG interpretive reports. If similar waveforms are analyzed subsequently,
the computer software makes the same diagnostic error over and over. Although it is desirable for an
ECG interpretation system to “learn” from its mistakes, there is no current commercial system that
improves its performance by analyzing its errors. Figure shows the final summary provided to the
clinician by an interpretive ECG machine . The machine has classified this patient as “Normal” with
normal “Sinus rhythm.”
II. ST-SEGMENT ANALYZER
The ST-segment represents the period of the ECG just after depolarization, the QRS complex, and just
before repolarization, the T wave. Changes in the ST-segment of the ECG may indicate that there is a
deficiency in the blood supply to the heart muscle. Thus, it is important to be able to make
measurements of the ST-segment. Figure shows an ECG with several features marked. The analysis
begins by detecting the QRS waveform. Any efficient technique can be implemented to do this. The R
wave peak is then established by searching the interval corresponding to 60 ms before and after the
QRS detection mark, for a point of maximal value. The Q wave is the first inflection point prior to the
R wave. This inflection point is recognized by a change in the sign of slope, zero slope, or a significant
change in slope. The three-point difference derivative method is used to calculate the slope. If the ECG
signal is noisy, a low-pass digital filter is applied to smooth the data before calculating the slope. The
isoelectric line of the ECG must be located and measured. This is done by searching between the P and
Q waves for a 30-ms interval of near-zero slope. In order to determine the QRS duration, the S point is
located as the first inflection point after the R wave using the same strategy as for the Q wave.
Measurements of the QRS duration, R-peak magnitude relative to the isoelectric line, and the RR
interval are then obtained.
The J point is the first inflection point after the S point, or may be the S point itself in certain ECG
waveforms. The onset of the T wave, defined as the T point, is found by first locating the T-wave peak
which is the maximal absolute value, relative to the isoelectric line, between J + 80 ms and R + 400 ms.
The onset of the T wave, the T point, is then found by looking for a 35-ms period on the R side of the T
wave, which has values within one sample unit of each other. The T point is among the most difficult
features to identify. If this point is not detected, it is assumed to be J + 120 ms. Having identified
various ECG features, ST-segment measurements are made using a windowed search method. Two
boundaries, the J + 20 ms and the T point, define the window limits. The point of maximal depression
or elevation in the window is then identified. ST-segment levels can be expressed as the absolute
change relative to the isoelectric line.
The ST slope is defined as the amplitude difference between the ST-segment point and the T point
divided by the corresponding time interval. The ST area is calculated by summing all sample values
between the J and T points after subtracting the isoelectric-line value from each point. An ST index is
calculated as the sum of the STsegment level and one-tenth of the ST slope.
III.PORTABLE ARRHYTHMIA MONITOR :
There is a great deal of interest these days in home monitoring of patients, particularly due to cost
considerations. If the same diagnostic information can be obtained from an ambulatory patient as can
be found in the hospital, it is clearly more cost effective to do the monitoring in the home.
Technological evolution has led to a high-performance computing capacity that is manifested in such
devices as compact, lap-sized versions of the personal computer. Such battery-powered systems
provide the ability to do computational tasks in the home or elsewhere that were previously possible
only with larger, nonportable, line-powered computers.
Holter recording :
Figure shows the Holter approach, which is to record the ECG of a patient for a full day on magnetic
tape. This recording and its subsequent return to a laboratory for playback and analysis restricts the
timely reporting of suspected arrhythmias to the physician. The results of a Holter recording session are
typically not known by the physician for several days.
Portable arrhythmia monitor hardware design :
The intelligent portable arrhythmia monitor will capture the ECG during suspected abnormal periods,
and immediately send selected temporal epochs back to a central hospital site through the voice-grade
telephone network. This approach should provide a significant diagnostic edge to the cardiologist, who
will be able to make judgments and institute therapeutic interventions in a much more timely fashion
than is possible today. In addition, the clinician will be able to monitor the results of the therapy and
modify it as needed, another factor not possible to do in a timely fashion with the tape recorder
approach .The hardware design of a portable arrhythmia monitor is quite straightforward, dependent
only on the battery-operable, large-scale-integrated circuit components available in the marketplace.
The primary semiconductor technology available for battery-operated designs is CMOS.
Block diagram of portable arrhythmia monitor consisting of microprocessor is shown above. Analog
amplifiers do the front-end ECG amplification and signal conditioning. An analog-to-digital (A/D)
converter integrated circuit changes the analog ECG to the digital signal representation needed by the
microprocessor. ROM memory holds the program that directs the performance of all the functions of
the instrument, and RAM memory stores the captured ECG signal. Input/output (I/O) ports interface
audio and visual displays and switch interactions in the device. A modem circuit provides for
communication with a remote computer so that captured ECG signals can be transmitted back to a
central site .
Portable arrhythmia monitor software design :
Unfortunately, the frequent hardware changes lead to equally frequent software redesign. The software
for an instrument is very hardware dependent. Each new microprocessor has its own unique machine
language. Thus, we need to rewrite the same programs for different processors, and thereby waste
considerable programming time. This software problem has led to explore higher-level languages that
are transportable from one kind of microprocessor to another. C language is now available for this type
of real-time instrumentation application. Its primary advantages are (1) a programming level low
enough to achieve the requisite machine control, and (2) transportability from one type of
microprocessor to another. Provided that we follow a few software design rules, a real-time program
written in the C language for one type of microprocessor can be easily reconfigured to run on a
different one. Although it is not perfect, the C language considerably reduces the programming time
necessary to rewrite the software when changing microprocessors.
For a portable arrhythmia monitor, the two major software design tasks are (1) QRS detection
and (2) arrhythmia analysis
QRS detection algorithm :
The various techniques that are used to implement a QRS detector are linear digital filters, nonlinear
transformations, decision processes, and template matching . Typically two or more of these techniques
are combined together in a detector algorithm .The most common approach in contemporary
commercial ECG instrumentation is based on template matching. A model of the normal QRS complex,
called a template, is extracted from the ECG during a learning period on a particular patient. This
template is compared with the subsequent incoming real-time ECG to look for a possible match, using
a mathematical criterion for goodness of fit. A close enough match to the template represents a detected
QRS complex. If a waveform comes along that does not match but is a suspected abnormal QRS
complex, it is treated as a separate template, and future suspected QRS complexes are compared with
it. But it requires considerable memory for saving the templates and significant computational power
for matching the templates to the real-time signal. Digital Filters can also be used
Arrhythmia analysis :
From the QRS detector, the QRS duration and the RR intervals are determined. The ECG signal is then
classified based on the QRS duration and the RR interval. Figure below is a conceptual drawing of an
arrhythmia analysis algorithm based on the two parameters, RR interval and QRS duration . In this
two-parameter mapping, we establish a region called normal by permitting the algorithm to first learn
on a set of eight QRS complexes defined by a clinician as having normal rhythm and morphology for
the specific patient. This learning process establishes the initial center of the normal region in the two-
dimensional mapping space.
Boundaries of all the other regions in the map, except for region “0”, are computed as percentages of
the location of the center of the normal region. Region “0” has fixed boundaries based on physiological
limits. Any point mapped into region “0” is considered to be noise because it falls outside what we
normally expect to be the physiological limits of the smallest possible RR interval or QRS duration .An
abnormality such as tachycardia causes clusters of beats to fall in region “1” which represents very
short RR intervals. Bradycardia beats fall in region “6”. Typically, abnormalities must be classified by
considering sequences of beats. For example, a premature ventricular contraction with a full
compensatory pause would be characterized by a short RR interval coupled with a long QRS duration,
followed by a long RR interval coupled with a normal QRS duration. This would be manifested as a
sequence of two points on the map, the first in region “3” and the second in region “5”. Thus,
arrhythmia analysis consists of analyzing the ways in which the beats fall onto the mapping space.
The center of the normal region is continuously updated, based on the average RR interval of the eight
most-recent beats classified as normal. This approach permits the normal region to move in the two-
dimensional space with normal changes in heart rate that occur with exercise and other physiological
changes. The boundaries of other regions are modified beat-by-beat, since they are based on the
location of the normal region. Thus, this algorithm adapts to normal changes in heart rate. The
classification of the waveforms can be made by noting the regions in which successive beats fall.
Figure lists some of the algorithms to detect different arrhythmias. The technique described is an
efficient method for extracting RR interval and QRS duration information from an ECG signal. Based
on the acquired information, different arrhythmias can be classified.
