Improving Ischemia Diagnosis With Synthesized ECG Leads JJ Viik, JA Hyttinen, RM Lehtinen, JA Malmivuo Ragnar Granit Institute, Tampere University of Technology, Tampere, Finland
Abstract The aim of this study is to improve the diagnosis of ischemic heart diseases by developing new synthesized leads. The synthesized leads were constructed in learning population (60 CAD and 60 low prevalence of heart disease patients) with mathematical optimization. The optimized parameter was T value of the Student's t test, Three synthesized leads were constructed as a combination of the three (I, 11, V2),five (1, 11, V2, V5, V6) and eight (I,II and VI-V6) leads. In the control population (100 CAD and 100 non-CAD) the synthesized leads had 80.3%, 79.7% and 78.1% ROC-areas using three, five and eight leads, respectively. ROC-areas of the standard leads ranged from 52.8% (lead Ill) to 80.9% (lead 1). The results of this study indicated that the detection of the ischemia using exercise ECG can be improved by developing the synthesized leads. Additionally, the developed method would give opportunity to more effective localization of the ischemia. The usefulness of the method and the synthesized leads need further studies.
1.
Introduction
An increased diagnostic accuracy of the exercise ECG can be achieved by improving the signal processing or developing the ECG leads. In signal processing the classically powerful methods, filtering, baseline correction, averaging, etc., are already being utilized. Development of ECG leads has been studied by changing the location of the electrodes or recording the body surface potential maps. Combining the leads using the mathematical optimization has not been conducted. All the potential information derivable from ECG signals has yet to be exploited. Due to the location of the electrode and the body's inhomogeneities, each of the standard leads has individual sensitivity in detection of the myocardial activation. Arising fkom the different sensitivities of the leads, the merging of the signals' information should improve the diagnosis of the diseases. In other words the merging of the standard leads should yield more accurate leads for the detection of the coronary artery disease (CAD). Due to the different sensitivities of the standard leads, they 02764547195 $4.00 0 1995 IEEE
713
detect the electrical activation (and the injury current) of the myocardium with different intensity [I 1. Using the combination of the standard leads it should be possible to construct new leads that have the desired spatial properties for the detection of a heart disease (e.g. CAD). The main idea of the new leads is to gather the spatial information of the standard leads providing more accurate representation in CAD diagnasis. The objective of this study was to develop a mathematical optimization method for the construction of new synthesized ECG leads. The leads in this study were derived by optimizing coefficients for the linear combination of the 12-lead ECG. The end-exercise ST depression of the specific standard leads was utilized to provide maximal discrimination between the patients with and without CAD.
2.
Method and material
2.1.
Method
The basis of the construction of the synthesized leads utilized the standard 12-leads. Each synthesized lead was constructed by combining the known standard leads with characteristic coefficients. The idea for the construction of the synthesized leads is illustrated in equation 1.
Sx
=
a.I
+
b-N + ...
+
h.V6
(1)
Sx is synthesized lead constructed with x number of standard leads, a to h are coefficients for each independent standard lead. The determination of the coefficients was made by mathematical optimization in a database, which had the parameters and the status of the patients, The optimized variable was T value of the Student's t test. The optimization of the coefficients was performed with Newton's iteration method. The iteration method searches the maximal discrimination between the diseased group and non-diseased group maximizing the difference of the averages and the standard errors of the standard deviations of the desired predictor. The synthesized leads were constructed using eight independent leads, two limb leads (I and 11) and six chest leads (Vl-V6). All eight leads can be used simultaneously in the development of synthesized leads. On the other hand we can select a few of these leads for the optimizing procedure. Three Computers in Cardiology 1995
groups: the learning population with 60 CAD patients and 60 patients with a low likelihood of CAD and the control population with 100 CAD patients and 100 patients with a low likelihood of CAD. The partition of the patients to the groups was made randomly. The characteristics of the learning and control populations are presented in Table 1. The CAD group and non-CAD group of control population did not include any patients who had been used in the creation of the synthesized leads, in learning population.
(I, II and VZ), five (I, 11,V2, V5, and V6) and eight (I 1, 1, and V1-V6) independent leads was utilized for defining the effective number of the leads in the construction of the synthesized leads. The synthesized leads were named S3, S 5 and S8, respectively. Conventional ECG parameter, ST-segment was selected for indicating ischemic changes in exercise ECG. ST-segment deviation hom baseline was measured to the nearest 10 pV at 60 ms after QRS-offset (ST60) in each 12 standard leads. Due to the precision of measurement the values of the synthesized leads were also determined to an accuracy of 10 pV. The capabilities of the leads were studied by using the Receiver Operating Characteristics (ROC) analysis. The larger area under the ROC curve, the better overall discriminative capacity of the classifier [2]. The analysis of the area under the ROC curve allows the comparison of the classifier without the k e d partition value (cutpoint) [3]. This is essential especially for the developed leads, because they are not in clinical usage and thus they do not have any clinically validated cutpoints.
2.2.
Table 1. Characteristics of learning and control population Learning population Control population Characteristic
CAD Non-CAD CAD Non-CAD (n=60) (n=60) (n=100) (n=100)
Age (years) 55 (7) Gender 4911 1 Medication p blocker (y/n) 4911 1 Ca antag. (yln) 20140
Study population
The whole study population was collected at Tampere University Hospital during 1990 and 1991 [4]. All patients underwent the bicycle ergometer exercise ECG test. The initial workload and the increment workloads every 4 minutes were 50 W for men and 40 W for women. The lead system used was the Mason-Likar modification of the standard 12-lead system. Exercise tests were sign or symptom limited maximal tests using recommended criteria for termination. The ECG recordings were made with a commercial Siemens Sicard 440s ECG recorder. The ECG data was filtered and averaged before storing to the hard disk. All the cases were referred by a physician for a routine exercise electrocardiogram. Coronary angiography had been used as a gold standard to prove the presence of CAD. Each of the CAD patient's coronary angiography was referred by a physician. The maximum time between the exercise test and coronary angiography was set at 180 days. The patients who where treated with coronary angioplasty or surgical operations within that time were excluded. The aiterion used for CAD was more than 50% coronary artery stenosis at least in one major coronary artery. The criteria for the reference group (nonCAD) were no history of any heart disease, normal resting ECG, no anginal type chest pain, and no cardiac medication. This group can be considered to be a population with a low likelihood (p < 0.005) of CAD [ 5 ] .Furthermore, the reference group included five patients who had no significant CAD according to coronary angiography and eight patients who had no exercise induced myocardial ischemia according to ME31 SPECT. This study population was divided to two populations; the learning population consisting 120 patients and the control population consisting 200 patient. Both learning and control population was divided to the CAD group and the reference
47 (12)** 2913 1**
54 (8) 76/24
48 (12)** 50/50**
0/60** 0/60**
82/18 35/65
5/95** 2/98**
Digitalis (yln) 2/58 0160 2/98 1/99 Nitrate (yln) 39/21 0/60** 74/26 3/97** ACP (yln) 31/29 0/60** 45/55 4/96** Maximum HR 126 (23) 165 (19)** 123 (22) 160 (21)** 27/33 0/60** MI ( Y W 43/57 0/60** Vessels (11213) 20117123 34/29/37 Continuous data are mean values (SD) CAD = Coronary Artery Disease, LAD = Left Anterior Descending, LCX = Left Circumflex, RCA = Right Coronary Artery, AP = Anginal Chest Pain, MI = Myocardial Infarction, HR = Heart Rate ** p < 0.01 between CAD and non-CAD groups
3.
Results
The results of the optimization in the learning group are illustrated in Figure 1. The T values of the synthesized leads were better than T values of the standard leads. Increasing the number of standard leads improved the T value (i.e. discrimination between the groups). In the control population synthesized leads had 80.3%, 79.7% and 78.1% ROC-areas using S3, S5 and S8 leads, respectively. ROC-areas of the standard leads ranged from 52.8% (lead III) to 80.9%(lead I). Each of the chest leads, V4, V5 and V6 had approximately 76% ROC-area. The results are illustrated in Figure 2. More than a 70% ROC-area was achieved in I, aVR and V3 - V6 of standard leads and in every synthesized ischemia leads. To achieve at least 50% ROC-area standard leads aVR and V1 were inverted.
714
Standard leads
Synthesized lads
1
ROC-areas
9.0 8.0
Standard leads
Synthesized leads
7.0
6.0
-2 5.0 E 4.0 3.0 2.0
1.0
I
0.0
Figure 1. T values of the standard leads and optimized synthesized leads in the separation of the learning population.
4.
Discussion
The mathematical optimization method for the construction of the new ECG leads were introduced in this study. This optimization procedure was utilized for the development of the synthesized leads. The new leads were constructed for the detection of the general, non-located ischemia. Optimization of the synthesized leads produced the greater discrimination (i.e. higher T value) between CAD and non-CAD groups in the learning population. Improvements were noteworthy when comparing the new leads to the standard leads except to the lead I. The best lead of the standard leads in the overall ischemia detection was lead I. Apparently, the sensitivity of the lead I is most uniform to those regions of the heart in which the myocardial ischemia most frequently occurs. The poorest standard leads in ischemia detection were leads 111 and V1. The ROC-areas of these leads barely achieved 50%, indicating unsatisfactorydiagnostic capacity. The chest leads V4, V5 and V6 have quite high and equal ROC-areas (75.3%, 76.3% and 76.5%, respectively), but the superior diagnostic capacity was in lead I. The results of the standard leads are in accordance with previously performed study t61. In general it can be noted that every synthesized lead had good diagnostic capacity; in the control population the areas under the ROC curves were near 80%. The number of the leads used in the construction was important in the optimization. The synthesized leads, which were constructed using a large number of standard leads did not achieve the better diagnostic capacity: the ROC-area of the lead S3 was higher than ROC-areas in leads S5 and S8. The result alludes that the optimization method fits extremely well to the learning group when more variables are available. This indicates that more patients should be included to the learning population when constructing the synthesized lead with more than three standard leads. The new method for improving ischemia diagnosis was introduced in this paper. Here the method was used in the construction of new leads for the detection of ischemia in general. However it may be practically impossible to construct a lead that detect all localized injury sources. The developed
I "
I
II
Ill aVR aVL aVF V1
VZ
V3
V4
V5
V6
53
S5
S8
1
Figure 2. Values of ROC-areas in the control group for all standard leads and constructed synthesized leads.
method should be employed for the localization of the ischemia. In the construction of the regional synthesized leads it might be useful to utilize a different number of the standard leads for dlfferent regions. The number of leads depends on the size of the region; it is obvious that for the detection of the larger region more leads are required. For this purpose a larger and more selective learning population including the evidence of the location of the coronary artery disease is required. 'The usefulness of the method and the synthesized leads need further studies.
References Hyttinen J: Development of regional aimed ECG leads especiallyfor myocardial ischemia diagnosis [Thesis]. Tampere University of Technology, 1994. 162 p. Hanley JA, McNeil BJ: The Meaning and Use of the Area under a Receiver Operating Characteristic (ROC) Curve. Radiology 1982;143:29-36. Metz CE: Basic principles of ROC analysis. Sem Nucl Med VI11 1978;4:283-298. Lehtinen R, Sievanen H, Uusitalo A, Niemela K, Turjanmaa V, Malmivuo J: Performance Characteristics of Various Exercise ECG Classifiers in Different Clinical Populations. J Electrocardiol 1994;27:11-22. Diamond GA, Forrester JS: Analysis of probability as an aid in the clinical diagnosis of coronary artery disease. N Eng J Med 1979;300:1350-1358. Viik J, Lehtinen R, and Malmivuo J: Capability of the Single ECG Leads of the 12-leadSystem to Discriminate Patients With CAD and Without CAD - ROC-analysis Approach. XXIInd International Congress on Electrocardiology. Nijmegen, The Netherlands 1995. Address for correspondence Jari Viik Ragnar Granit Institute Tampere University of Technology P.O.Box 692 FIN-33101 Tampere, Finland Tel+358-3 1-3162158 Fax +358-31-3162162 Internet:
[email protected]
715
