Medical Engineering & Physics 34 (2012) 485–497

Contents lists available at SciVerse ScienceDirect

Medical Engineering & Physics

journal homepage: www.elsevier.com/locate/medengphy

Identification of myocardial infarction (MI) using spatio-temporal heart dynamics

Hui Yang a , Satish T.S. Bukkapatnam b,∗ , Trung Le b , Ranga Komanduri c
a
Department of Industrial and Management Systems Engineering, University of South Florida, Tampa, FL 33620-5350, United States
b
School of Industrial Engineering and Management, Oklahoma State University, Stillwater, OK 74078, United States
c
School of Mechanical and Aerospace Engineering, Oklahoma State University, Stillwater, OK 74078, United States

a r t i c l e i n f o a b s t r a c t

Article history: Cardiovascular disorders, such as myocardial infarction (MI) are the leading causes of mortality in the
Received 18 May 2010 world. This paper presents an approach that uses novel spatio-temporal patterns of the vectorcardiogram
Received in revised form 12 May 2011 (VCG) signals for the identification of various types of MI. In contrast to the traditional electrocardiogram
Accepted 17 August 2011
(ECG) approaches, the 3D cardiac VCG signal is partitioned into 8 octants for localized analysis of the
This work is dedicated to the memory of heart’s electrical activities. The proposed method was tested using the PhysioNet PTB database for 368
Professor Komanduri (1942-2011), who MIs and 80 healthy control (HC) recordings, each of which includes 12-lead ECG and 3-lead VCG. Sig-
was the originator of the manuscript. nificant differences are found in the VCG spatial distribution between MI and HC groups. Furthermore,
classification and regression tree (CART) analysis was used to demonstrate that VCG octant features
Keywords: can distinguish MIs from HCs with a sensitivity (accuracy of MI identification) of 97.28% and a specificity
Electrocardiogram (ECG) (accuracy of HC identification) of 95.00%, which is promising compared to the previously reported results
Vectorcardiogram (VCG) using other ECG databases. The results indicate that the present approach provides an effective way for
Myocardial infarction (MI) monitoring, post-processing, and interpretation of ECG data, and hopefully can impact the current cardiac
Spatial octant dynamics diagnostic practice.
Classification and regression tree (CART)
© 2011 IPEM. Published by Elsevier Ltd. All rights reserved.

1. Introduction oftentimes on the memorization of ECG signals for different cardi-

ological disorders [1–6] taught in the medical schools.
Coronary heart diseases have been reported to be responsible
for some 451,326 deaths in 2004 and considered as the leading 2. Background
cause of mortality in America [1]. Myocardial infarction (MI), oth-
erwise known as heart attack, results from an occlusion of the Dower and his colleagues [4–6] conducted pioneering research
coronary artery and insufficient blood supply to the myocardium. on lead transformation, based on Frank’s tank torso model, and
It can manifest in different portions of the heart, i.e., anterior, infe- introduced a linear transformation matrix to convert the 3-lead
rior, posterior, inferior-lateral, anterior-septal, posterior-lateral. VCG signals into 12-lead ECG signals without a significant loss
The triad of MI, namely, ischemia, injury, and necrosis, or any one of clinically useful information regarding heart dynamics. The so-
of the three can occur either in isolation or in combination [2]. called inverse Dower transformation matrix was used to derive the
Ischemia is due, normally, to reduced blood supply; injury indicates 3-lead VCG from the 12-lead ECG. Some even consider Dower trans-
acuteness of infarct; and infarction is the symptom of myocardium formation matrix as “generalized” or “universal” transformation
necrosis [2]. matrix although one finds the need to develop different transfor-
The 12-lead ECG is more commonly used than the 3-lead VCG mation matrices for healthy control (HC) subjects and patients with
because medical doctors have been trained to using them in clini- different cardiological disorders [1–5].
cal diagnosis for more than one hundred years. It has, thus, proven For a given cluster of subjects, e.g., HC subjects in certain age and
its value, time tested, and considered as the Gold Standard. It gender groups, the transformation studies show statistical equiva-
may, however, be noted that a 12-lead ECG contains some redun- lence between the 12-lead ECG and 3-lead VCG signals. However, it
dant information, and even in that, only a small fraction of the is difficult for human beings to visually project a spatial VCG vector
data is used by the physicians based on experience, expertise, and into any specified cardiac measurement angle determined by the
12-lead measurement system, which is the traditional way for the
interpretations of ECG signals. In addition, the absence of tempo-
ral information in the static VCG representation poses extra barriers
for medical doctors to closely relate the 12-lead ECG characteristics
∗ Corresponding author. Tel.: +1 405 744 5900; fax: +1 405 744 7873.
to the 3-lead Frank X, Y, Z VCG signals [1–6].

1350-4533/$ – see front matter © 2011 IPEM. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.medengphy.2011.08.009
486 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497

Fig. 1. (a) Different views of human torso showing various coordinate axes and positions of the electrodes for acquiring VCG signals [32]; (b) Schematic showing the locations
of the eight octants (0–7) for VCG signal analysis.

It may be noted that when it comes to the computer automated knowledge and expertise of spatial decomposition of the heart vec-
analysis of heart monitoring measurements, 12-lead ECG signals tors which is emerging gradually, and is not commonly taught in
will introduce what is commonly known as the curse of dimensional- the medical schools. The conventional 12-lead ECG diagnostic cri-
ity problem, namely, redundant information. Thus, VCG offers some teria are based on mapping the durations and amplitudes of the
positive outcomes for computer processing and analysis because P wave, QRS complex, and T wave extracted from an ECG time
it overcomes not only the information loss from using only one or series to cardiac electrical activities, such as sinoatrial (SA) node
two ECG signals but also the dimensionality problems induced by excitation, atrial depolarization and repolarization, and ventricular
the 12-lead ECG signals. depolarization and repolarization events. Such a sequential division
Conceptually speaking, a 12-lead ECG used widely in clinical of ECG time-series is sensitive to the presence of ectopic waves, arti-
diagnostic applications, views the 3D cardiac electrical activities facts, and noises. Additionally, the presence, location, and extent of
from 12 different perspectives, and a 3-lead vectorcardiogram cardiovascular disorders, i.e., MI, can alter ECG signal patterns in
(VCG) captures the spatiotemporal cardiac electrical dynamics diverse ways. This sheer diversity of patterns poses significant dif-
from the 3 Frank XYZ orthogonal directions, as shown in Fig. 1(a) [3]. ficulties for both computer and human experts to diagnose MI from
The VCG vector loops contain 3D recurring, near-periodic P, QRS, ECG measurements.
and T wave activities distributed among one or more of the eight It may be noted that some of the previous MI diagnostic
octants (0–7). Fig. 1(b) shows the conventional VCG coordinate sys- approaches also considered the T loop morphology [8–10], angles
tem used in the clinical practice with right to left indicating the of QRS, T loops, QRS loops [11–17], recurrent dynamics, QRS and
positive X-direction, superior to inferior in the positive Y-direction, T vector magnitudes [18–20] in the 3-lead VCG signals. Many of
and anterior to posterior in the positive Z-direction. the earlier ECG methods exhibited considerably high sensitivity at
Dower et al. [4–6] showed that a 3-lead VCG can be linearly the expenses of specificity. In 1992, Willems et al. [21], conducted
transformed to a 12-lead ECG without a significant loss of clini- a comprehensive bench marking study of comparing the perfor-
cally useful information pertaining to the heart dynamics. Recent mance of various automated computer programs for MI detection
advances in computer graphics and wireless technologies have along with the average score of eight experienced cardiologists.
renewed interest in the VCG signals that use fewer leads than the The automated MI detection programs include those from Padova,
conventional 12-lead ECG. It surmounts the information loss due Nagoya-Fukuda, IBM Medis, HP (Agilent), Glasgow, GE (Marquette),
to data compaction of the ECG signals as well as the dimensional- Means, Hannover, and Louvaine. The best reported MI detector
ity problems and often redundant information introduced by the recorded 97.1% sensitivity1 and 80.3% specificity2 for the MI identi-
12-lead ECG. It may be noted that geometrically any point in three fication [21] and the average sensitivity and specificity with various
orthogonal planes can be specified by three orthogonal coordinates MI detectors are: 91.65% and 68.38%, respectively.
(here, 3 leads). For example, Dawson et al. [7] presented a statisti- Additional efforts have been made in recent years to improve the
cal approach to transform 3-lead Frank VCG into 12-lead ECG and sensitivity and specificity of MI detection. For example, Heden et al.
vice versa, based on Dower’s lead transformation. They showed [22] used ST segment features for neural network identification of
the statistical affine transform can map the ECG recordings accu- acute MI. They report a sensitivity of about 95% and specificity of
rately for both MI and HC subjects. However, the accuracy issue 86%. Reddy et al. [23] proposed artificial neural networks using the
becomes accentuated under pathological conditions, such as MI, patterns of the QRS complex in leads V2–V4 to detect acute MI. They
natural degradation due to aging, abnormal size variations of the
heart, where dipole assumptions may not hold and VCG may not
capture all salient electrical activities of the heart.
1
It may also be noted that many manufacturers of ECG units do Sensitivity measures the proportion of actual positives which are correctly iden-
not collect signals from all 12 leads but generate them from fewer tified as such (i.e. the percentage of sick people who are identified as having the
condition).
signals. As stated earlier, VCG signals are not as commonly used 2
Specificity measures the proportion of negatives which are correctly identi-
in the medical practice as the 12-lead ECG because, among other fied (i.e. the percentage of healthy subjects who are identified as not having the
reasons, the interpretation of 3D VCG loops requires specialized condition).
 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497 487

reported a sensitivity and specificity of 79% and 97%, respectively. enlargements, left ventricular infarction, dorsal and inferior acute
A neural-fuzzy approach for classifying MI suggested by Lu et al. MI, and multiple MIs; VCG showing in a clear way, the special ori-
[24] used ST elevation as the input of the systems. They reported entation and magnitude of the vectors at every moment; greater
that their method correctly identified 89.4% of patients with MI sensitivity in detecting atrial enlargements and greater sensitivity
and 95% of healthy subjects for a validation set of 124 ECG records. in detecting atrial enlargements; greater sensitivity and specificity
Zheng et al. [25] used support vector machines (SVM), Naïve Bayes in the diagnosis of left ventricular enlargement; and greater diag-
(NB), and random forecast (RF) methods to detect MI from 192 lead nostic sensitivity in acute MI.
body surface potential maps. The total classification accuracies of
NB method were reported to be 81.9%, SVM 82.8%, and RF 84.5%.
Matveev et al. [26] obtained signal-averaged QRS and T waves to 3. Octant vectorcardiography
detect acute MI. They reported a classification accuracy of about
90% for identifying between the three groups, namely, healthy, The above cited works were preceded by a not-so-widely
anterior ST elevation MI, and inferior ST Elevation MI. Jayachan- known, yet pioneering work of Laufberger [37–40] of Czechoslo-
dran et al. [27] used the energy-entropy characteristics of wavelet vakia. He published a series of articles (in an East European journal,
decomposed signals to classify between MI and normal subjects. Physiologia Bohemoslovaca) wherein he demonstrated the diag-
Sensitivity and specificity for MI detection were reported to be nostic capabilities of VCG octant information for detecting MI. A
87.5% and 95%, respectively. Arif et al. [28] used k-nearest neigh- series of four papers from 1980 to 1982 focused on analyzing the
bor (KNN) clustering with T wave amplitude, Q wave and ST level distribution of vectorcardiogram in cardioelectrical space by pro-
deviation to achieve classification sensitivity and specificity of 90% jecting the Frank lead system onto a three-dimensional orthogonal
and 99.6% between MI and normal subjects, respectively. coordinate system. The Cartesian coordinate system divides the
Most of the methods proposed in the literature used the mor- 3-dimensional space into eight octants. The QRS complex, repre-
phological features of 12 leads ECG, such as T-wave amplitude, senting the cardiac depolarization process was investigated for
Q wave and ST deviation, QRS complex to detect the MI. Princi- the identification of MI and other pathologies. Specifically, the
pal component analysis (PCA) for feature selection could not be sequence of octants that a QRS loop traverses during depolarization
adequate because the leading eigen direction corresponds to R- (referred to as octant sequence), octant-wise maxima of positive
amplitude related characteristic [29,30], which is not known to be and negative potentials along the X, Y, Z directions (referred to
highly sensitive to MI. Lagerholm et al. [31] used 4–6 Hermite poly- as peak amplitudes), and duration of the QRS complex in various
nomial basis functions to represent the QRS complex, and used the octants (referred to as octant duration) were obtained and taken
coefficients of these basis functions to identify different anoma- as features for MI detection. The variations of these feature values
lies in heart beats including arrhythmias from ECG signals. Also, in the fifth octant were used to detect the onset of a class of MI
Yang et al. [32] used recurrence quantification analysis and neural (the so-called ST segment MI), and to determine the progress of
network (NN) classification of the VCG signals and found testing the coronary arterial diseases following the onset of MI. It may be
accuracies to vary from 92.27 to 98.7% for the MI and 56.25 to 75% noted that the VCG trajectories corresponding to the end portion
for the HC cases. Based on the observed trends for both MI and HC of a QRS loop and the starting portion of a T-loop typically pass
subjects, the lower classification accuracies for the HC cases were through the fifth octant. The presence of ST-elevation MI is marked
attributed to the sparse data available (368 for MI vs. 80 for HC). by a significant increase or decrease in the distribution of the QRS
Yang et al. [33] also conducted nonlinear adaptive wavelet anal- loop portion of a VCG in Octant 5 (left-superior-posterior).
ysis of the ECG signals and found the wavelet representation can Besides the conceptualization of the octant-wise distribution of
provide an effective time, and time-frequency analysis of the ECG VCG (which is referred to as spatiocardiogram [41]), Laufberger also
signals. contributed towards the development of multiple versions of the
VCG signals can monitor a significant portion of spatial and tem- clinical instrumentation system and methods for the acquisition,
poral cardiac electrical activities along three orthogonal planes of recording and visualization of the spatiocardiograms [41]. A typical
the body. However, the absence of temporal resolution in conven- recording instrumentation consists of an electrocardiograph con-
tional static VCG representation has remained a major impediment nected to a magnetic tape recorder with six channels that record
for the medical interpretation and clinical usage of VCG. This is the ECG directly onto a drum to promote image retention, an elec-
especially so since most cardiologists rely on temporal aspects of tronic control panel to adjust the parameters of spatiocardiogram
the 12-lead ECG, such as intervals, durations for the diagnostics of visualization process (e.g., image hold time, number of pictures per
cardiovascular activities. Recently, MI detection using VCG features second), calculate the spatiocardiogram components, and render it
has received some attentions. For example, Bortolan and Christov on an oscilloscope.
[8] used the maximum angle between QRS and T-loop axes, T-axis The spatiocardiogram instrumentation system assembled by
elevation and ratio of maximum to mean T vector magnitude as Laufberger and his associates was rather bulky and complex [41].
features to detect MI. Ghista et al. [34] used the shape charac- This is because that was the state-of-the-art equipment available
teristics of VCG to detect abnormal cardiac conditions including at that time. It also took several years for them to design, assem-
bundle branch block, left ventricular hypertrophy, and MI. Yang’s ble, and operate the instrument satisfactorily. At that time, cathode
[35] approach captures critical spatiotemporal heart dynamics by ray oscilloscopes were just beginning to appear in the market
displaying the real time motion of the VCG cardiac vectors in a 3D place. The instrumentation system Laufberger put together with-
space. Dynamic VCG signal representation has been shown to be out doubt was a significant advance in electrocardiology. However,
effective in enhancing the interpretability of VCG and facilitate in implementation into practice and continuation of this work by
the identification of related spatiotemporal cardiovascular dynam- other researchers was altogether a different question. The inven-
ics. tion of cathode ray oscilloscope, according to Laufberger, has
In a review on the significance of VCG in the cardiological enabled direct observation of the electrical phenomenon of the
diagnosis, Riera et al., stated that VCG could enrich the morpho- heart which heather to been inaccessible. As regards the cathode
logical interpretation of the electrical phenomena of the heart [36]. ray oscilloscope, Laufberger states and we quote, “It would be quite
They cite some 20 advantages of the VCG over ECG that include: justified to state that the oscilloscope has enabled the foundation
improving the onset and offsets of various cardiac events, espe- of the entire field of vectorcardiography and spatiocardiography.”
cially for those relevant for detecting disorders, such as left arterial Today, the same can be said about the introduction of personal
488 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497

Table 1
VCG octant positions and color coding. (For interpretation of the references to color in this table, the reader is referred to the web version of the article.)

Octant X Y Z Binary code Location (X, Y, Z) Color code

0 − − − (000) Right-superior-anterior Black

1 − − + (001) Right-superior-posterior Blue
2 − + − (010) Right-inferior-anterior Gray
3 − + + (011) Right-inferior-posterior Cyan
4 + − − (100) Left-anterior-superior Magenta
5 + − + (101) Left-superior-posterior Green
6 + + − (110) Left-inferior-anterior Orange
7 + + + (111) Left-inferior-posterior Red

computers (PCs). The advent of inexpensive and powerful PCs and VCG are located into one or more of the eight octants (labeled 0–7)
communication platforms, visualization tools, and software, such for the spatiotemporal decomposition of cardiovascular electrical
as Matlab, the entire system has become rather compact, simple to activities. For ease in the spatial octant identification, a binary sys-
operate, and analyze data. tem is introduced in this investigation to designate each octant, i.e.,
Despite such an innovative concept, further work and transla- binary code 0 is used to represent the negative directions in the X-,
tion into practice of Laufberger’s pioneering work, conducted some Y-, and Z-axes and binary code 1 is used for the positive directions
30 years ago, had been somewhat limited, if any, so far. This prob- of the X-, Y-, and Z-axes as shown in Table 1. For example, if the
ably was due to bulky, complex, and expensive system not easily octant lies in the (−, +, −) X, Y, and Z directions, binary coding for
accessible to many researchers. This may also be the reason why this will be (010) and the resulting octant number is 2 in the dec-
this work was not followed by other researchers as evidenced by imal numeral system (binary coding: 0 × 22 + 1 × 21 + 0 × 20 = 2). It
the absence of research publications in this area since then. may be noted that octant numbers in Table 1 range from 0 through
Besides technological limitations, non-availability of sufficient 7 instead of the conventional 1 through 8 numbering system. This
number of systematically examined and clinically diagnosed was chosen to take advantage of expressing the octant numbers in
subjects appear to have impeded the validation studies neces- the binary code. Fig. 1(b) shows that all the octants with even num-
sary for broader acceptance of the spatiocardiography concepts. bering are in the anterior locations and those with odd numbering
However, his approach to quantitatively describe the VCG in three- are in the posterior locations. Octants 0 and 2 are located in the
dimensional space in terms of octants holds significant parallels right anterior sides and octants 4 and 6 in the left anterior sides of
with the present work where we demonstrate that the octant fea- the heart. This binary octant system is simple in concept and does
tures can be effective for MI detection. It may please be noted that not require memorization. Hence, it is more convenient for users
we have developed the reported spatiotemporal approach inde- to locate the spatial positions of the octants even if the VCG coordi-
pendently without being aware of Laufberger’s pioneering work. nate system is rotated. The VCG signals in all eight octants are also
It was during the manuscript review process the authors became color coded (see Table 1) for easy identification.
aware of this pioneering work and hence have incorporated an
account of his work in this paper for completion and for appropriate
recognition. 4.1. ECG database

Some 448 VCG recordings (368 MIs and 80 HCs) available in

4. Research approach the PhysioNet PTB Database [42,43] were analyzed in this investi-
gation. Each recording contains 15 simultaneous heart monitoring
In the present investigation, we used the spatial octant analysis signals, namely, the conventional 12-lead ECG and the 3-lead Frank
of 3-lead VCG signals for monitoring the cardiac electrical activities (XYZ) VCG. The signals were digitized at 1 kHz sampling rate with a
and further diagnostic applications. Instead of inspecting the ECG 16-bit resolution over a range of ±16.384 mV. The 80 HC record-
leads one at a time in the time domain; the 3D cardiac vectors from ings were acquired from 54 healthy subjects and the 368 MI

Fig. 2. ECG ensembles extracted in the time domain for the computation of conventional interval features including RR interval average (RRavg), RR interval standard
deviation (RRstd), QT interval average (QTavg), QT interval standard deviation (QTstd), Q magnitude (Qmag), ST segment elevation (STelev), and ST segment integration area
(STitg) [44].
 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497 489

Table 2
Summary of 48 features in three groups used in this study. Group 1 (conventional ECG features), Group 2 (VCG vector features), and Group 3 (VCG octant features), with the
corresponding D-statistic values obtained by applying Kolmogorov–Smirnov (KS) test.

Group No. Feature Full description KSSTAT (D)

1 RR avg Average of RR intervals 33.0%

2 RR std Standard deviation of RR intervals 40.8%
Features

3 QT avg Average of QT intervals 13.3%

4 QT std Standard deviation of QT intervals 19.3%
ECG

5 ST elev ST segment elevation 35.5%

I
6 ST Itg ST segment integration area 11.6%
7 Q amp Q wave amplitude (from leads II&v3) 58.0%

8 Q mag Q-vector magnitude 33.3%

9 T mag T-vector magnitude 58.0%
Features

10 R mag R-vector magnitude 43.2%

Vector
VCG

11 Tang T-vector angle 20.3%

II 12 Rang R-vector angle 14.9%
13 RTang Angle between R and T-vector 46.4%

14 R pos R-vector octant position 17.3%

15 T pos T-vector octant position 61.6%
16 Q pos Q-vector octant position 11.7%
17 Oct0AvgN Average vector magnitude in Octant 0 67.8%
18 Oct0MaxN Maximal vector magnitude in Octant 0 69.3%
19 Oct0Ratio Ratio of VCG components in Octant 0 9.2%
20 Oct0VarN Vector magnitude variance in Octant 0 64.7%
21 Oct1AvgN Average vector magnitude in Octant 1 39.5%
22 Oct1MaxN Maximal vector magnitude in Octant 1 29.4%
23 Oct1Ratio Ratio of VCG components in Octant 1 45.7%
24 Oct1VarN Vector magnitude variance in Octant 1 26.4%
25 Oct2AvgN Average vector magnitude in Octant 2 61.7%
26 Oct2MaxN Maximal vector magnitude in Octant 2 55.7%
27 Oct2Ratio Ratio of VCG components in Octant 2 36.0%
VCG Octant Features

28 Oct2VarN Vector magnitude variance in Octant 2 59.7%

29 Oct3AvgN Average vector magnitude in Octant 3 10.4%
30 Oct3MaxN Maximal vector magnitude in Octant 3 48.0%
31 Oct3Ratio Ratio of VCG components in Octant 3 22.3%
32 Oct3VarN Vector magnitude variance in Octant 3 30.7%
33 Oct4AvgN Average vector magnitude in Octant 4 30.7%
34 Oct4MaxN Maximal vector magnitude in Octant 4 42.7%
III 35 Oct4Ratio Ratio of VCG components in Octant 4 29.8%
36 Oct4VarN Vector magnitude variance in Octant 4 12.1%
37 Oct5AvgN Average vector magnitude in Octant 5 29.8%
38 Oct5MaxN Maximal vector magnitude in Octant 5 37.1%
39 Oct5Ratio Ratio of VCG components in Octant 5 39.3%
40 Oct5VarN Vector magnitude variance in Octant 5 17.4%
41 Oct6AvgN Average vector magnitude in Octant 6 43.3%
42 Oct6MaxN Maximal vector magnitude in Octant 6 29.5%
43 Oct6Ratio Ratio of VCG components in Octant 6 47.3%
44 Oct6VarN Vector magnitude variance in Octant 6 35.8%
45 Oct7AvgN Average vector magnitude in Octant 7 36.6%
46 Oct7MaxN Maximal vector magnitude in Octant 7 53.9%
47 Oct7Ratio Ratio of VCG components in Octant 7 37.4%
48 Oct7VarN Vector magnitude variance in Octant 7 45.9%

The rows with bold values are features with large KSTAT values (Table 3). They have been selected for MI identification through CART classification.

recordings from 148 patients. Within the MI recordings, there Group-II VCG vector features (features 8–13): Vector magnitudes
were 89 inferior, 56 inferior-lateral, 19 inferior-posterior-lateral, of the Q-wave, R-wave, T-wave vectors, angle between
1 inferior-posterior, 47 anterior, 43 anterior-lateral, 77 anterior- R and T vector, azimuth R vector angle, and azimuth T
septal, 2 anterior-septal-lateral, 3 lateral, 4 posterior, 5 posterior- vector angle (see Fig. 3);
lateral and 22 unknown site cases. The VCG recordings were Group-III VCG octant features (features 14–48): Octant location of
typically of ∼2 min duration and all the signals were recorded for the Q-wave, R-wave, T-wave vectors, percentage of the
at least 30 s. vector points in an octant, and maximum, average, and
variance of vector magnitude distributions in each octant
(see Fig. 3).
4.2. Feature extraction

To characterize and distinguish MIs from HCs, the following

4.2.1. ECG features
three groups of features, namely, conventional ECG features, VCG
Traditional ECG features, such as RR and QT intervals are highly
vector features, and VCG octant features are extracted (see features
sensitive to the heart rate variability, respiration, artifacts, and
1–48 in Table 2).
power-line interferences [45]. In this investigation, we removed
the effects of these extraneous factors using a wavelet filter cus-
Group-I Conventional ECG features (features 1–7): Mean and tomized to the ECG signals [35]. We determined the mean and the
standard deviation of RR interval, QT interval, ST eleva- standard deviation of RR and QT intervals, as well as the extent
tion/depression, and Q wave amplitude (see Fig. 2) [44]; of ST elevation and Q amplitude through the extracted heart beat
490 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497

Fig. 3. VCG trajectory and its 2D projection plots with maximal Q, R, T vectors for a representative HC subjects. (a) VCG trajectory in the 3D coordinate system; (b) projection
plot in Vy and Vz plane; (c) projection plot in Vx and Vz plane; (d) projection plot in Vx and Vy plane. Arrows indicate the directions of the resultant maximal Q, R, T vectors.

cycles over time as shown in Fig. 2. We have used two features it was shown that QRS and T loops are close to each other in the
to capture ST elevation. The first, ST elev uses wavelet filter to HC cases, while they are far from each other in some cardiovascular
determine the locations of the J point (location where QRS com- diseases [8].
plex joins the ST segment, which is taken as the onset point of
the ST segment) and the onset point of the T wave. This fea-
4.2.3. Octant features
ture estimates the deviation of a line connecting these two points
As stated earlier, Laufberger propounded the ideas of spati-
from the local iso-electric line. The second ST itg, computes the
ocardiography [40] where he pointed out the potential of the
ST segment integration area, which is determined by computing
octant-specific distribution of VCG signals for MI detection. Coin-
the areas from the J point to J + 80 ms, if the heart rate (HR) is less
cidentally, the octant features, including the location and vector
than 100 beats/min (or J + 72 ms, if 100 < HR < 110; or J + 64 ms, if
magnitude distribution we had used in the present work hold close
110 < HR < 120; or J + 60 ms, if HR > 120) [46]. Since MI often causes
similarities to the octant locations of QRS loop, and amplitudes of
ST elevation/depression in the 12-lead ECG signals, the drifts from
QRS loop in various octants that Laufberger had used for MI diag-
isoelectric line, therefore, will also result in an increase in the ECG
nostics [40]. The extraction and application of these features for MI
ST segment area. The Q amp feature is estimated from the ampli-
diagnosis are detailed in the following sections.
tude of the Q wave as measured from the lead V3. The initial
As illustrated in Fig. 4, VCG signals for healthy control (HC)
investigations indicate that the Q amp estimated from V3 has the
exhibit significant spatial distribution differences from MI. The car-
best signal content, and the values showed significant difference
diac vectors in each octant are identified by a single color (see
depending on whether the recording was from HC vs. MI subject.
Table 1). For example, the red color segments (Octant 7) have differ-
ent magnitudes and percentages to the entire trajectory for HC and
MI (see Fig. 4). Therefore, a group of VCG octant features is investi-
4.2.2. VCG features
gated for a high separation probability of MI from HC subjects. The
Recent studies by the present authors on dynamic VCG
T vector octant position will locate those T-wave inversions and
representation gave some indications regarding the intrinsic rela-
abnormal cases. R vector octant location indicates the heart’s elec-
tionships between cardiovascular diseases and VCG octant-wide
trical axis. Cardiac vectors’ magnitude and proportion in each of the
features [47,48]. In this investigation, a group of VCG vector features
8 octants are designed for the cardiac vector changing directions
was used to capture the MI pathological patterns. For example, the
and strengths in different local areas.
Q vector magnitude (Qmag) is used to facilitate in the identification
of the Q-wave infarction, T vector magnitude (Tmag) is related to
the enlargement of the T-wave amplitude, and the R vector mag- 4.2.4. Feature selection and Kolmogorov–Smirnov D statistic
nitude (Rmag) provides the size of the ventricular chambers and For each VCG signal recording, 48 features in three groups (con-
proximity of chest electrodes to the ventricular chamber. As shown ventional ECG interval features (7), VCG vector features (6), and
in Fig. 3, the Q, R, T vectors are from the isoelectric point in the VCG octant features (35)) were analyzed. Such a high-dimensional
VCG to the farther most point in the respective Q, R, and T loops. In feature space, however, is unsuitable for classification in terms
addition, the angle between R and T vectors is computed to provide of computational efficiency. Consequently, a statistical goodness-
more information regarding the myocardial infarction (MI) because of-fit hypothesis test (known as Kolmogorov–Smirnov test or
 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497 491

Fig. 4. VCG trajectories are color coded for difference octant numbers (see Table 1
for details) in the 3D coordinate system, (a) HC – PTB database recording No. 511; (b)
MI – PTB database recording No. 187. It can be seen that VCG trajectory is distributed Fig. 5. Illustration of the split of tree node t into left and right child nodes tL , tR with
in more than one octant, and the spatial distributions are vastly dissimilar between probability pL and pR .
MI and HC. (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of the article.)
the goodness of the split is defined to be the decrease in impurity
defined as follows:
KS test [49]) was performed to examine the differences in both i(s, t) = i(t) − pL i(tL ) − pR i(tR ) (3)
location and shape of the cumulative distribution functions of the
48 features between HC and MI subjects in the PTB database. The where the impurity function i(t) is a function of the splitting class
Kolmogorov–Smirnov statistic is given by probabilities p(1|t), p(2|t), . . ., p(J|t).
The Gini diversity index criterion was commonly adopted for
D = sup|EnMI (x) − EnHC (x)| (1) the impurity function in the form:
x ⎛ ⎞2
   
which is the maximum difference between HC and MI feature i(t)= p(j|t)p(i|t) = ⎝ p(j|t)⎠ − p2 (j|t) = 1 − p2 (j|t)
empirical distribution functions developed using samples of fea-
j=
/ i j j j
ture values x (realized from a random feature variable X), and
n (4)
En (x) = (1/n) i=1 I(Xi < x), and the asymptotic p-value is calcu-
lated as Therefore, the best split s* is selected by solving the following opti-
 √ 2  mization problem:
0.11
p = exp −2 × ñ + 0.12 + √ × D2 (2) i(s∗ , t) = arg maxs i(s, t) − arg maxs (i(t) − pL i(tL ) − pR i(tR ))
ñ
⎡ ⎤
where ñ = (nMI × nHC )/(nMI + nHC ).   
= arg maxs ⎣− p2 (j|t)+pL p2 (j|tL ) + pR p2 (j|tR )⎦
Here, the p-value is the probability that any variation in the input
feature will not cause a statistically significant change in the output j j j

response. The p-value should be as low as possible for the input (5)
feature to be significantly discriminating. Features of interest are
The step is repeated till a terminal node t is reached (optimization
the ones with the highest D-statistic values and p-values below
at each step) such that no significant decrease in impurity measure
0.05.
i(s, t) is possible.
However, the tree can grow to a significant size with numer-
4.3. Classification and regression tree (CART) model ous levels. Some insignificant nodes and sub-branches need to
be pruned so as to reduce the high complexity and prevent
The CART approach [50] was used in this investigation for the “overfitting” problems. The optimal tree size is achieved through
classification of MI cases from these extracted features. It is easy cost-complexity function R˛ (T) as shown in Eq. (4), which includes
to interpret and quite effective in creating clinical decision rules both tree complexity and misclassification error.
that combine medical knowledge-base with data analysis. The
arg minT R˛ (T ) = R∗ (T ) + ˛(T ) (6)
CART classification model utilizes tree-like structures to assign an
explanatory variable based on a set of feature quantifiers, x1 , x2 , . . ., where ˛(T) is the complexity measure of tree T, and is the R* (T)
xM − 1 , and xM . For example, taking the feature space  to be the set misclassification error of tree T. The present investigation uses k-
of all possible values of (x1 , . . ., xM ), and letting C = {1, 2, . . ., J} be fold cross validation method3 to obtain the test sample estimate of
the set of possible explanatory variables. The classification model R* (T). The learning space  is divided into k subgroups  1 ,  2 , . . .,
is a function with domain  and range C, and it corresponds to a  k of equal sizes. k-fold cross validation estimate is the propor-
partition of  into disjoint sets, A1 , A2 , . . ., AJ , such that the pre- tion of cases in the learning space  that are misclassified by the
dicted class is j, if x ∈ Aj , where x = (x1 , . . ., xM ). The training sample classifier constructed from the subsample  ∩ ¯ k . This estimate of
consists of data (x1 , j1 ), (x2 , j2 ), . . ., (xN , jN ) on N cases where xn ∈ , misclassification error is computed as in Eq. (7):
and jn ∈ {1, 2, . . ., J}. The learning space is denoted by  = {(x1 , j1 ),
1 
(x2 , j2 ), . . ., (xN , jN )}. R∗ (T ) ≈ Rts (T ) = (d(k) (xn ) =
/ jn ) (7)
Nk
In the CART model, each parent node is optimally partitioned (xn ,jn ) ∈ k
so as to achieve homogenous descendant nodes. Let us denote that
the class probability p(j|t), j = 1, . . ., J, to be the proportion of the
cases xn ∈ t belonging to class j, so that p(1|t) + p(2|t) + · · · + p(J|t) = 1. 3
k-fold cross validation method involves randomly partitioning the data into k
For any tree node t, the candidate split s divides learning space  partitions, and using k − 1 of the partitions for training and the remaining one for
into left and right child nodes tL , tR such that a proportion pL of the testing. Such a partitioning, followed by training and testing is performed k times
cases in t go into tL and a proportion pR go into tR (see Fig. 5). Then and the final performance metrics (e.g., misclassification errors) are reported.
492 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497

Fig. 6. Kolmogorov–Smirnov or KS statistic variations for the 48 features. KSSTAT indicates the maximal feature distribution differences between HC and MI samples. Details
of the features and corresponding index are given in Table 2.

where d(k) (xn ) is the classifying function from tree T, Nk is the num- 0.05. The relatively high D-statistic values and low p-values imply
ber of cases in  k and jn ∈ {1, 2, . . ., J}. that the nine feature subsets incorporate much of the discriminat-
In this study each parent node is optimally partitioned so as to ing information of the total 48 features. Consequently, the selected
achieve homogenous descendant nodes with less correlation. The nine features were used to build the classification and regression
use of k-fold cross validation helps in surmounting the “overfitting tree (CART) model [50] for the identification of MI vs. HC recordings
problem,” and can thereby lead to generalized, optimal, and simpler from the PhysioNet PTB database.
tree structure for the MI and HC classifications.
5.2. Classification
5. Results
The CART model [50] uses decision rules learnt from the pattern
(clustering) of features in the training data set to classify MI and
5.1. Feature analysis
HC from the PTB database. The CART approach facilitates the inter-
pretation and is quite effective in creating clinical decision rules
As stated in the foregoing, for each 15-lead signal recording,
that combine medical knowledge-base with the data analysis. Each
48 features in three groups (conventional ECG interval features
parent node is optimally portioned so as to achieve homogeneous
(7), VCG vector features (6), and VCG octant features (35)) were
descendant nodes with less correlation. To prevent “overfitting”
analyzed. Such a high-dimensional feature space, however, is
problems, k-fold cross validation [50] is used to find a generalized,
unsuitable for classification in terms of computational efficiency.
optimal, and simpler tree structure for the MI and HC classifica-
Consequently, a statistical goodness-of-fit hypothesis test, known
tions. In this case, a subset of 90% of the data (332/368 MIs and
as Kolmogorov–Smirnov test or KS test [49], was performed to
72/80 HCs) is randomly selected as the training dataset, and the
examine the differences in both location and shape of the cumu-
lative distribution functions of the 48 features between HC and MI
subjects in the PTB database.
Fig. 6 shows the variations of D-statistic values obtained from
the application of the KS test for all the 48 features (see also,
Table 2). It can be noted that the following 9 VCG vector and octant
features have high KSSTAT values (D > 0.5) (see Table 3). They are: T
vector magnitude (T mag – 58%), T vector position (T pos – 61.6%),
average vector magnitude in Octant 0 (Oct0AvgN – 67.8%), maxi-
mal vector magnitude in Octant 0 (Oct0MaxN – 69.3%), variance of
vector magnitude in Octant 0 (Oct0VarN – 64.7%), average vector
magnitude in Octant 2 (Oct2AvgN – 61.7%), maximal vector magni-
tude in Octant 2 (Oct2MaxN – 55.7%), variance of vector magnitudes
in Octant 2 (Oct2VarN – 59.7%), and maximal vector magnitude in
Oct 7 (Oct7MaxN – 53.9%). We did not use any of the ECG features
in our classifier in order to investigate the classification accuracies
obtainable with VCG and VCG octant features compared to those
from ECG waveforms. Eight out of the nine features are VCG octant
features and the remaining one is a VCG vector feature. In effect,
they capture the VCG activity in Octant 0, Octant 2, and along the T
loop. Fig. 7. Optimal CART tree structure for the MI classification in the PTB database.
Table 3 provides the KS test statistics, as well as the mean and This tree creates if–then rules for the classification of MI and HC using four octant
features (Oct0MaxN, T pos, Oct2AvgN, Oct0VarN). The tree branches in the red dash
the standard deviation of the nine important subset features that
line enclosed area are pruned with k-fold cross validation method to prevent “over-
can effectively discriminate between MI and HC recordings in the fitting”. (For interpretation of the references to color in this figure legend, the reader
PTB database. The p-values for all the nine features were well below is referred to the web version of the article.)
 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497 493

Table 3
Statistical analysis (Kolmogorov–Smirnov test or KS test) of 9 key VCG and VCG octant features used for classification, out of the 48 total features extracted.

Key feature no. Feature p-value KSSTAT (D) MI HC

Mean Std Mean Std

9 T mag 2.70E−20 58.0% 0.25 0.14 0.40 0.14

15 T pos 8.71E−23 61.6% 3.46 2.26 5.88 0.72
17 Oct0AvgN 1.48E−27 67.8% 0.09 0.001 0.12 0.00
18 Oct0MaxN 9.63E−29 69.3% 0.38 0.02 0.54 0.00
20 Oct0VarN 4.30E−25 64.7% 0.005 0.00 0.007 0.00
25 Oct2AvgN 7.26E−23 61.7% 0.11 0.00 0.15 0.00
26 Oct2MaxN 9.99E−19 55.7% 0.45 0.07 0.62 0.09
28 Oct2VarN 1.83E−21 59.7% 0.01 0.00 0.02 0.00
46 Oct7MaxN 1.44E−17 53.9% 1.48 0.17 2.05 0.23

remaining 10% (36/368 MIs and 8/80 HCs) is used for validation recordings with Oct0MaxN less than 0.5187 are MIs as can be seen
(10-fold cross validation). The tree size and structure are varied from the one-dimensional histogram plot in Fig. 9(a). However, the
to test cross-validation errors of the CART model. Optimal tree determination of all four feature boundary values are optimally
structure for MI and HC classifications in the PTB database is shown experimented and validated in the 4-dimensional space, instead
in Fig. 7. of lower dimensional projections. In this stage of Rule 1, 28 out of
It may be noted that the classification boundaries are deter- 368 MI subjects and 2 out of 80 HC subjects are incorrectly clas-
mined based on localized regression that forms part of the CART sified. The final misclassification rates in Rule 2 are 10 out of 368
model [50]. The tree branches in the dash line enclosed area (in MI subjects and 4 out of 80 HC subjects. Detailed performances are
red) (see Fig. 7) are pruned to minimize cross-validation errors. obtained as follows:
Two simple, if-then, rules using four VCG octant features (top four
highest D values) (see Table 3) are generated from the CART model - CART Result 1: % MI = 100 (1–28/368) = 92.39%; % HC = 100
as follows: (1–2/80) = 97.50%;
- CART Result 2: % MI = 100 (1–10/368) = 97.28%; % HC = 100
- CART Rule 1: If Oct0MaxN < 0.5187 or T pos < 5.5, then the record- (1–4/80) = 95.00%.
ing is MI;
- CART Rule 2: For the rest of the recordings after Rule 1, if The four feature boundary values are obtained from CART model
Oct2AvgN < 0.1233 or Oct0VarN > 0.0144, then the recording is training process, and the convergence of these boundary values is
MI; shown through random subsampling methods, as shown in Fig. 10.
The CART classification tests with random subsampling methods
The classification boundaries (the numbers in CART Rules 1 aim to show the generality and effectiveness of the VCG octant
and 2) are obtained by a process of optimally separating the MI features by randomly selecting training datasets. The optimal tree
and HC clusters in the 4-dimensional space, namely, Oct0MaxN, structure from Fig. 7 is retained but the training dataset size is
T pos, Oct2AvgN, and Oct0VarN. It may be noted that choice of the varied from 20% to 90%. For example, the training subset is ran-
features of the four highest D values (from KS statistic) is a com- domly selected at 20% from both MI and HC groups, i.e., 74/368 MIs
promise between accuracy and computational time required. Fig. 8 and 16/80 HCs. Rest of the database (80%) is used for testing pur-
shows the distribution of the feature values in a 3-dimensional pose. The random training dataset selection and classification are
space formed with the axes of Oct0MaxN, T pos, and Oct2AvgN. replicated 200 times to collect the performance statistics.
Alongside Fig. 9 provides the histograms showing how the val- It may be noted from Figs. 10(a) to (d) that feature boundary
ues of the features are distributed in the HC vs. MI samples. For values in the CART models vary because of random selection of the
example, the Oct0MaxN boundary 0.5187 will guarantee that all the training datasets. However, the boundary variations demonstrate a
trend of convergence to fixed numbers with increase in the training
percentage from 20% to 90% (feature boundary values: Oct0MaxN
– 0.5187, T pos – 5.5, Oct2AvgN – 0.1233 and Oct0VarN – 0.0144).
Fig. 10(e) shows the median of CART classification model sensitivity
is higher than 95% for all 8 training sizes (red line in the middle
of the box represents the median). Fig. 10(f) shows the median of
specificity to be higher than 90%. These results indicate that the
feature boundary values used in the optimal CART model in Fig. 7
can lead to robust CART model to identify MI.
Next, we used the ECG features alone (without the VCG vector or
octant features) to develop a CART classifier. The conventional ECG
features were selected based on their relevance to the clinical diag-
nosis of MI. The CART classifier indicated that Q amp followed by
ST elev, RR avg and RR std were the most discriminating features.
The classifier reported a comparably high sensitivities (∼97%) for
detecting MI but the specificity tends to be somewhat low (∼65%)
for detecting HC. This result affirms the assertions of the previ-
ous researchers on the low specificity of ECG based classifiers for
Fig. 8. Distribution of MI (368) and HC (80) subjects in 3D feature space (Oct0MaxN, MI identification. High specificity is desirable towards minimizing
T octant position and Oct2AvgN). The HC subjects are found to converge in the
excessive false positives. In fact, one of the issues the clinicians have
green circle area, while MI subjects spread over the feature space and mostly outside
the green enclosed area. (For interpretation of the references to color in this figure with computerized automated diagnostic systems is their high lev-
legend, the reader is referred to the web version of the article.) els of false positives, which have lowered the trust the caregivers
494 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497

place on the computer programs. The proposed VCG octant features

can potentially offer superior candidate features to improve clinical
detection of MI.

6. Discussion and conclusions

The diagnostic tests of myocardial infarction, commonly used

in clinical practice, include electrocardiogram, computer imaging,
blood enzyme, chest X-ray, cardiac catheterization etc. But com-
puter imaging and X-ray approaches are not only very expensive
but also not always available. In addition, medical professionals
are strictly required to perform invasive methods, such as blood
enzyme tests and cardiac catheterization. The ECG or VCG signals
tracks the real time dynamics of cardiac functioning, and have been
used widely for the diagnostic of myocardial infarction. However,
it is often a difficult task for cardiologists to visually inspect the
long-term ECG signals for warning signs when continuous ECG
monitoring generates an enormous torrent of data. Therefore, it
is imperative to develop an effective automated computer system
that can classify the underlying cardiac pathological patterns with
high sensitivity and specificity.
It is generally known that Q-wave (significant for necrosis),
ST segment (elevation for injury), and T-wave (inverted T wave
for ischemia) play a significant role in the identification of differ-
ent stages of MI. But conventional ECG intervals and previously
proposed VCG features are somewhat inadequate to capture the
essential characteristics of MI (see Fig. 6). The newly proposed VCG
spatial octant features and CART analysis can achieve high sensi-
tivity and specificity using two simple rules and four VCG octant
features. It may also be noted that Rule 1 can detect 92.39% of MI
and 97.50% of HC with only two features, namely, Oct0MaxN and
T vector octant position. In ongoing activities, we intend to inves-
tigate how these features are quantitatively related to the heart
dynamics.
Fig. 8 shows the distribution of HC (368) and MI (80) recordings
from the PTB database characterized with three features, namely,
Oct0MaxN, T pos, and Oct2AvgN. It can be seen that these three
octant features can effectively discriminate HC from MI subjects
since the healthy controls are found to cluster in the enclosed area
marked green (HC cluster) while MI patients spread mostly out-
side that green area. Note that these features capture the variations
in the anterior side as well as the variation in the repolarization
activities. These activities are known to be highly affected by MI.
Histogram plots of three features, namely, Oct0MaxN, T pos,
and Oct2AvgN for MI (368) and HC (80) samples are shown in
Fig. 9(a)–(c). It may be noted that Oct0MaxN has significant distri-
bution differences between MI (0.38 ± 0.02) and HC (0.54 ± 0.0002)
with a p-value of <0.005 in Fig. 10(a). In addition, the majority
(76/80) of HCs’ T vector positions are also found to cluster in
Octant 6 (right-inferior-anterior), while MIs spread over all the
eight octants (see Fig. 10(b)). Oct2AvgN also shows >60% dissim-
ilarity of the statistical distribution between HC and MI subjects
(see Table 3).
Both Octant 0 and Octant 2 are located in the right anterior
side. Therefore, MI subjects are statistically shown to have not only
the spatial shift of ventricular repolarization directions (T pos),
but also abnormal electrical activities in the right anterior posi-
Fig. 9. Histogram plots of (a) the maximal vector magnitude in octant 0 – Oct0MaxN,
tions of the body. This can be understood based on how cardiac
(b) the maximal T vector octant position – T Pos and (c) average vector magnitude
in octant 2 – Oct2AvgN between HC (green) and MI (red) samples. The majority electrophysiological activity is affected during MI. The onset of
(76/80) of T Pos for HC subjects was also found to cluster in Octant 6 while that for MI creates electrically nonconductive areas about the damaged
MIs spread over multiple octants. (For interpretation of the references to color in or dead heart muscle tissues. Most of the MI recordings in this
this figure legend, the reader is referred to the web version of the article.) research are related to the ventricular MI. During the ventricle
depolarization, the membrane of normal cardiac tissue are perme-
able to sodium ions allowing large numbers of positively charged
sodium ions to diffuse into the tissue cells [51]. However, the dead
 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497 495

Fig. 10. CART classification model performance using randomly selected training datasets. The CART model structure is retained as shown in Fig. 6, and uses randomly
selected training datasets in the specified percentage (X-axis) to train the model for the estimation of feature boundary value variations. The boundary variations of (a)
Oct0MaxN, (b) T pos, (c) Oct2AvgN, and (d) Oct0VarN are shown to converge with the increase of training dataset percentage. The classification results from testing datasets
also demonstrate sensitivity (e) and specificity (f) in a high range despite the fluctuations of training datasets. For the specified percentage, random selection of the training
datasets and classification of testing datasets are repeated for 200 times to provide statistical performance evaluations. Red line in the middle of box represents the median,
the blue box shows the lower quartile and upper quartile of data distributions, and black dash lines represent the most extreme values within 0.5 times the interquartile
range. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)

tissue cells are unable to open the permeable channel to allow the post-processing, and interpretation by medical practitioners. The
diffusion of positive ions. The loss of positivity in the infarcted proposed CART model yields a sensitivity of 97.28% and a speci-
area after the depolarization is responsible for the characteristic ficity of 95.00% for the MI identification with four VCG octant
Q waves as well as QRS loops in the VCG leads. These Q waves in features. The random CART classification experiments with differ-
MI patients are negative and last longer than those in the normal ent percentages of training data size also shows high sensitivity
objects. The QRS complex leaks into Octant 0 (−Vx − Vy − Vz) and (median > 95%) and specificity (median > 90%) (see Fig. 8), which
Octant 2 (−Vx + Vy − Vz), referring to the right anterior positions of further shows the generality and robustness of the CART model
the body, and therefore the features from these octants tend to be and the VCG octant features. Comparison with CART models based
MI-sensitive. on ECG features indicates that the use of proposed octant features
The novel VCG octant features are shown to yield much more can lead to improvements in the specificity of detecting MI. It is
statistical differences than previous ECG features, as shown in Fig. 6. anticipated that this method can effectively be applied for the iden-
Additional pathological information is obtained in the 3D VCG spa- tification of other cardiovascular disorders, and make a significant
tial dynamics than in the 12-lead ECG signals [52]. It will also be impact on the current ECG practice.
more convenient to extract VCG octant features with automated New and more accurate techniques are evolving fast (due to
computer methods than inspecting the 12-lead ECG traces and the availability of fast, inexpensive computers and associated soft-
memorizing the ECG interval rules for HS and various MI conditions. ware), such as electrocardiographic body surface mapping (BSM)
Thus, the approach proposed in this investigation provides for the detection of transient MI. BSM is a technique that samples
a new set of features towards improving ECG monitoring, multiple points around the thorax (80 electrode vest) to provide
496 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497

more comprehensive electrocardiographic dataset than conven- cardiogram in patients with inferior myocardial infarction. Jpn Circ J 1983;47,
tional 12-lead ECG [53]. It may be noted that while both BSM and 860–860.
[21] Willems J, Abreu-Lima C, Arnaud P, van Bemmel JH, Brohet C, Degani R, et al.
VCG can capture spatiotemporal variations of the cardiac electri- The diagnostic performance of computer programs for the interpretation of
cal activities of the heart, this is not the case with ECG. The extent electrocardiograms. N Engl J Med 1991;325:1767–73.
to which higher special resolution achieved in identifying various [22] Heden B, Ohlin H, Rittner R, Edenbrandt L. Acute myocardial infarction
detected in the 12-lead ECG by artificial neural networks. Circulation
MIs vs. the diagnostic capability and complexity of using 60 or more 1997;96:1798–802.
leads in BSM, vs. the 12-lead ECG, vs. the 3-lead Frank X, Y, Z is yet [23] Reddy MRS, Edenbrandt L, Svensson J, Haisty WK, Pahlm O. Neural network
to be determined. versus electrocardiographer and conventional computer criteria in diagnos-
ing anterior infarct from the ECG. In: Proceedings of computers in cardiology
conference. 1992. p. 667–70.
[24] Lu HL, Ong K, Chia P. An automated ECG classification system based on a neuro-
Acknowledgments
fuzzy system. In: Proceedings of computers in cardiology conference. 2000. p.
387–90.
The authors would like to thank the National Science Foundation [25] Zheng H, Wang H, Nugent CD, Finlay DD. Supervised classification models to
(CMMI-0729552 and CMMI-0830023) as well as OSU/Tulsa Center detect the presence of old myocardial infarction in body surface potential maps.
In: Proceedings of computers in cardiology conference. 2006. p. 265–8.
for Health Sciences (CHS) (Dr. J. Hess and Dr. B. Benjamin) for the [26] Matveev M, Krasteva V, Naydenov S, Donova T. Possibilities of signal-averaged
support of the research. One of the authors (RK) also wishes to thank orthogonal and vector electrocardiography for locating and size evalua-
A.H. Nelson Jr. Endowed Chair in Engineering for additional finan- tion of acute myocardial infarction with ST-elevation. Anatolian J Cardiol
2007;7(1):193–7.
cial support. Thanks are also due to Mrs. Sharon Green for excellent [27] Jayachandran E, Joseph KP, Acharya UR. Analysis of myocardial infarction using
editorial assistance. discrete wavelet transform. J Med Syst 2010;34:985–92.
[28] Arif M, Malagore I, Afsar F. Detection and localization of myocardial infarction
using k-nearest neighbor classifier. J Med Syst 2010:1–11. Online FirstTM .
Conflict of interest statement [29] Polat K, Günes S. Detection of ECG Arrhythmia using a differential expert sys-
tem approach based on principal component analysis and least square support
None declared.
vector machine. Appl Math Comput 2007;186:898–906.
[30] Castells F, Laguna P, Sornmo L, Bollmann A, Roig JM. Principal component
analysis in ECG signal processing. EURASIP J Adv Signal Process 2007;2007:
References 98–119.
[31] Lagerholm M, Peterson C, Braccini G, Edenbrandt L, Sornmo L. Clustering
[1] Heart disease and stroke statistics – 2009 update. American Heart Association; ECG complexes using Hermite functions and self-organizing maps. IEEE Trans
2009. Biomed Eng 2000;47:838–48.
[2] Dubin D. Rapid interpretation of EKG’s: an interactive course. Cover Publishing [32] Yang H, Bukkapatnam STS, Malshe M, Komanduri R. Recurrence quantification
Company; 2000. analysis and principal components in the detection of myocardial infarction
[3] Malmivuo J, Plonsey R. Bioelectromagnetism: principles and applications of from vectorcardiogram signals. In: Proceedings of the 3rd INFORMS workshop
bioelectric and biomagnetic fields. USA: Oxford University Press; 1995. on data mining and health informatics (DM-HI 2008). 2008.
[4] Dower GE, Machado HB. XYZ data interpreted by a 12-lead computer program [33] Yang H, Bukkapatnam STS, Komanduri R. Nonlinear adaptive wavelet analysis
using the derived electrocardiogram. J Electrocardiol 1979;12:249–61. of electrocardiogram signals. Phys Rev E 2007;76:026214.
[5] Dower GE, Machado HB, Osborne JA. On deriving the electrocardiogram from [34] Ghista D, Acharya U, Nagenthiran T. Frontal plane vectorcardiograms: the-
vectorcardiographic leads. Clin Cardiol 1980;3:87–95. ory and graphics visualization of cardiac health status. J Med Syst 2010;34:
[6] Dower GE, Yakush A, Nazzal SB, Jutzy RV, Ruiz CE. Deriving the 12-lead 445–58.
electrocardiogram from four (EASI) electrodes. J Electrocardiol 1988;21: [35] Yang H. Nonlinear stochastic modeling and analysis of cardiovascular system
s182–7. dynamics – diagnostic and prognostic applications. Ph.D. Thesis (Advisers: S.
[7] Dawson D, Yang H, Malshe M, Bukkapatnam STS, Komanduri R, Benjamin Bukkapatnam and R. Komanduri), Oklahoma State University, Stillwater, OK;
B. Linear affine transformations between 3-lead (Frank XYZ leads) vector- 2009.
cardiogram and 12-lead electrocardiogram signals. J Electrocardiol 2009;42: [36] Riera ARP, Uchida AH, Filho CF, Meneghini A, Ferrerira C, Schapacknik E, et al.
622–30. Significance of VCG in the cardiological diagnosis of the 21st century. Clin
[8] Bortolan G, Christov I. Myocardial infarction and ischemia characterization Cardiol 2007;30:319–23.
from T-loop morphology in VCG. In: Proceedings of computers in cardiology [37] Laufberger V. Octant vectorcardiography. Physiol Bohemoslov
conference. 2001. p. 633–6. 1980;29:481–94.
[9] Lemire D, Pharand C, Rajaonah J, Dube B, LeBlanc AR. Wavelet time entropy, [38] Laufberger V. Octant vectorcardiography. Physiol Bohemoslov
T wave morphology and myocardial ischemia. IEEE Trans Biomed Eng 1981;30:481–95.
2000;47:967–70. [39] Laufberger V. Octant vectorcardiography and automatic diagnosis of coronary
[10] Rubulis A, Jensen J, Lundahl G, Tapanainen J, Wecke L, Bergfeldt L. T vector and artery disease. Physiol Bohemoslov 1982;31:485–95.
loop characteristics in coronary artery disease and during acute ischemia. Heart [40] Laufberger V. Octant vectorcardiography – the evaluation by peaks. Physiol
Rhythm 2004;1:317–25. Bohemoslov 1982;31:1–9.
[11] Starr JW, Wagner GS, Behar VS, Walston II A, Greenfield Jr JC. Vectorcardio- [41] Laufberger V. Spatiocardiography: textbook and atlas (translated by J.H. Cort).
graphic criteria for the diagnosis of inferior myocardial infarction. Circulation Springfield, IL: C.C. Thomas Publisher; 1965.
1974;49:829–36. [42] Goldberger A, Amaral L, Glass L, Hausdorff J, Ivanov P, Mark R, et al. Phys-
[12] Yamazaki T, Froelicher VF, Myers J, Chun S, Wang P. Spatial QRS-T angle predicts ioBank, PhysioToolkit, and PhysioNet: components of a new research resource
cardiac death in a clinical population. Heart Rhythm 2005;2:73–8. for complex physiologic signals. Circulation 2000;23:e215–20.
[13] Cowdery CD, Wagner GS, Starr JW, Rogers G, Greenfield JC. New vectorcardio- [43] Moody GB, Koch H, Steinhoff U. The PhysioNet/computers in cardiology
graphic criteria for diagnosing right ventricular hypertrophy in mitral-stenosis challenge 2006: QT interval measurement. In: Proceedings of computers in
– comparison with electrocardiographic criteria. Circulation 1980;62: cardiology conference. 2006. p. 313–6.
1026–32. [44] Yanowitz FG. ECG intervals and waves-KH. Available:
[14] Hugenholtz PG, Levine HD, Whipple GH. A clinical appraisal of vector- http://library.med.utah.edu/kw/ecg/mml/ecg 533.html.
cardiogram in myocardial infarction: I. Cube system. Circulation 1961;24: [45] Bukkapatnam STS, Komanduri R, Yang H, Rao P, Lih W-C, Malshe M, et al. Clas-
808–24. sification of atrial fibrillation episodes from sparse electrocardiogram data. J
[15] Chou TC, Masangka MP, Young R, Conway GF, Helm RA. Simple quantitative Electrocardiol 2008;41:292–9.
vectorcardiographic criteria for diagnosis of right ventricular hypertrophy. Cir- [46] Jager F, Taddei A, Moody GB, Emdin M, Antolic G, Dorn R, et al. Long-term
culation 1973;48:1262–7. ST database: a reference for the development and evaluation of automated
[16] De Torbal A, Kors JA, Van Herpen G, Meij S, Nelwan S, Simoons ML, et al. The ischaemia detectors and for the study of the dynamics of myocardial ischaemia.
electrical T-axis and the spatial QRS-T angle are independent predictors of long- Med Biol Eng Comput 2003;41:172–83.
term mortality in patients admitted with acute ischemic chest pain. Cardiology [47] Lemay M, Vesin JM, Jacquemet V, Forclaz A, Kappenberger L, Van Oosterom
2004;101:199–207. A. Spatial dynamics of atrial activity assessed by the vectorcardiogram: from
[17] Wolff L. The vectorcardiographic diagnosis of myocardial infarction. Chest sinus rhythm to atrial fibrillation. EP Eurospace 2007;9:vi109–18.
1955;27:263–81. [48] Pérez Riera AR, Uchida AH, Filho CF, Meneghini A, Ferreira C, Schapacknik E,
[18] Carson W, Tseng YZ. Maximal spatial ST-vector patterns in patients with acute et al. Significance of vectorcardiogram in the cardiological diagnosis of the 21st
anteroseptal myocardial infarction. Int J Cardiol 1994;43:165–73. century. Clin Cardiol 2007;30:319–23.
[19] Burch GE, Abildskov JA, Cronvich JA. Studies of the spatial vectorcardiogram in [49] Marsaglia G, Tsang W, Wang J. Evaluating Kolmogorov’s distribution. J Stat
normal man. Circulation 1953;7:558–72. Softw 2003;8(18):1–4.
[20] Uehata A, Kurita A, Takase B, Mizuno K, Isozima K, Satomura K, et al. Incidences [50] Breiman L, Friedman J, Stone C, Olshen RA. Classification and regression trees.
of coronary-artery stenosis and mean QRS vector on the frontal plane vector- Chapman & Hall; 1993.
 H. Yang et al. / Medical Engineering & Physics 34 (2012) 485–497 497

[51] Guyton AC, Hall JE. Textbook of medical physiology. 11th ed. Saunders/Elsevier; Trung Le was born in Da Nang City, Vietnam. He is cur-
2005. rently in the doctoral program in the School of Industrial
[52] Edenbrand L, Pahlm O, Lyttkens K, Albrechtsson U. Vectorcardiogram more Engineering and Management at Oklahoma State Uni-
sensitive than 12-lead ECG in the detection of inferior myocardial infarction. versity as a Vietnam Education Foundation fellow. His
Clin Physiol Funct Imaging 1990;10:551–9. research interests are sensor based modeling on human
[53] Robinson M, Curzen N. Electrocardiographic body surface mapping: potential cardiovascular system, real-time monitoring and non-
tool for the detection of transient myocardial ischemia in the 21st century. Ann linear dynamics analysis with focus on diagnosis and
Noninvasive Electrocardiol 2009;41(2):201–10. prediction in the evolution of physiological processes. He
is a member of IEEE (Institute of Electrical and Electronics
Hui Yang was born in Nanjing, Jiangsu Province, P.R. Engineers), INFORMS (Institute for Operations Research
China. He received his Ph.D. degree from the School of and Management Science), and IIE (Institute of Industrial
Industrial Engineering and Management at Oklahoma Engineers).
State University. He is an Assistant Professor in the Depart-
ment of Industrial and Management Systems Engineering Ranga Komanduri received B.E. (Mechanical) and M.E.
at the University of South Florida. His research interests (Heat Power) from Osmania University, Hyderabad, India,
are sensor based computational modeling and analy- and Ph.D. and D. Eng. from Monash University, Melbourne,
sis of complex systems with special focus on nonlinear Australia. He is a Regents Professor and the A. H. Nel-
stochastic dynamics, and the resulting chaos, multifractal, son, Jr. Chair in Engineering in the School of Mechanical
self-organization, long range dependence behaviors. He is and Aerospace Engineering of Oklahoma State University,
a member of IEEE (Institute of Electrical and Electronics Stillwater, OK. Dr. Komanduri coauthored some 230 pub-
Engineers), INFORMS (Institute for Operations Research lications and holds 22 U.S. Patents. He is a Fellow of ASME,
and Management Science), and IIE (Institute of Industrial Fellow of SME, and Fellow of CIRP. He is the recipient
Engineers). of several prominent awards in the areas of manufac-
turing and materials, including SME Albert M. Sargent
Satish T.S. Bukkapatnam was born in Tirupati, India. Progress Award; ASME Int. William T. Ennor Manufac-
He received his M.S. and Ph.D. degrees in Industrial and turing Technology Award; ASME Charles Russ Richards
Manufacturing Engineering from Pennsylvania State Uni- Memorial Award; ASME Blackall Machine Tool and Gage Award; F. W. Taylor Medal
versity, University Park. He is an AT&T Professor in the of CIRP; OSU’s Eminent Faculty Award; OSU Regents Distinguished Research Award;
School of Industrial Engineering and Management, Okla- OSU President’s Distinguished Service Award; OSU Halliburton Outstanding Fac-
homa State University. His research is in sensor-based ulty Award’; OSU Phoenix Outstanding Graduate Faculty Award; and Distinguished
modeling for monitoring quality and performance of man- Honorary Professor, ITT/Kanpur, India.
ufacturing machines and processes, and other real-world
complex systems, including human cardiovascular sys-
tem. This approach is based on augmenting the statistical
and intelligent systems foundations of modern sensor
and monitoring technologies with nonlinear dynamics
principles. His research has led to over 100 refereed pub-
lications. Dr. Bukkapatnam was a recipient of three best paper awards, Alpha Pi
Mu/Omega Rho Outstanding Teacher of the Year in Industrial Systems Engineering,
and the SME Dougherty Outstanding Young Manufacturing Engineer Award.