Compartmental multivariate analysis of exercise ECGs for accurate detection of myocardial ischaemia

1 .Introduction EXERCISE ELECTROCARDIOGRAPHY (ECG) is the most widely used method for the non-invasive assessment of myocardial ischaemia and coronary artery disease (CAD). The exercise ECG test has several advantages; it is easily accessible, it is simple to perform and, particularly, it is noninvasive. However, the test accuracy and usefulness has been limited by a moderate diagnostic accuracy (about 750/0) (FROELICHER,1987) and by indeterminate test responses in populations with a low prevalence of CAD (RIFKm and HOOD, 1977; DIAMOND and FORRESTER, 1979). In particular, the generally used ST-segment criterion for ischaemic response is not very specific. In addition, any change in criteria, especially regarding the ST-segment depression during peak exercise, can markedly affect the diagnostic accuracy of the test: either the sensitivity of the specificity will be altered (ZOHMAN and KATrUS, 1977). Therefore, criteria for more effective interpretation of exercise ECG tests remain a very important goal. New criteria have been proposed for the assessment of patients with suspected CAD. The adjustment of the ST-segment response to corresponding changes in the heart rate during exercise has been shown significantly to improve the diagnostic accuracy of the exercise test. This method was based on the linear relationship (ST/HRFirst received 28 January 1992 and in final form 1 March 1993

9 IFMBE: 1994 MBEC

slope) observed between exercise-induced ST-segment depression and heart rate (ELAMIN et al., 1980; 1982; KARDASHet al., 1982). It was concluded that, by means of the ST/HR-slope, it was possible to identify the degree of CAD with absolute accuracy, and the test performance was not affected by cardiac drugs. This promising result was supported, with certain limitations however, in recent studies (AMEISEN e t al., 1985; 1986; KLIGFIELD et al., 1985; 1986; 1987; 1989; OKIN et al., 1985; 1986; 1988; 1990; 1991; SIEVANEN, 1987; 1991; SIEVANENet al., 1991), but was also strongly questioned (Fox, 1982; BALCON et al., 1984; QYUUMIet al., 1984; THWAITES et aL, 1986). Consequently, it would be useful to reduce the effect of factors detracting from the test performance of the ST/HR-slope. Its performance can be improved by modifying the calculation of the ST/HR'-slope and by applying additional independent diagnostic variables. This can be done by computerising data-processing and using multivariate logical decision procedures. The main objective of this study was to design a computerised method for automatic classification of the ECGs recorded during a standardised exercise test. The advantages of computerisation, such as enhanced measurement accuracy and precision, improved quality control and objective use of classification criteria, are obvious (SIMOONSet al., 1981). Therefore, the main issue was whether a new algorithm based on the ST/HR-slope (multivariate st-segment~heart rate analysis (MUSTA)), would significantly enhace the diagnostic accuracy of the exercise test, and whether a significant improvement can

Electrocardiography, Myocardial Contraction and Blood Flow supplement

July 1994

$3

be achieved in the problematic clinical population commonly investigated by the exercise ECG test. In this paper, the development of MUSTA is outlined, its algorithm is described in Prolog, and the results of initial clinical validation are given.

2 Subjects The learning group, where MUSTA was developed, consisted of 47 patients (38 men and nine women), whose mean age was 51 (SD 7, range 36-66) years. These patients had had anginal symptoms and abnormal ECGs in previous tests. Because of suspected ischaemic heart disease, they also underwent a T1-201 SPECT exercise test. In these tests, 37 patients had an abnormal exercise ECG respose. The TI-201 SPECT investigation showed that nine had neither ischaemia nor myocardial infarction (MI), and 20 had signs of previous MIs with associated or separate ischaemia. Myocardial ischaemia with no MIs was diagnosed in 18 patients. Most of the patients continued their normal use of cardiac medication on the day of the examination. The evaluation group consisted of 60 patients (47 men and 13 women), whose mean age was 53 (SD 9, range 2966) years. The selection of the patients for the evaluation of MUSTA was identical to that in its development phase. There were no differences in age and sex distributions between the evaluation and learning groups. Of the 60 patients, 44 had an abnormal exercise ECG response. The T1-201 SPECT investigation showed that 12 patients had no exercise-induced ischaemia, and 25 had signs of previous MI with associated or separate ischaemia. Myocardial ischaemia with no MI was diagnosed in 23 patients. Most of the patients continued their use of cardiac medication on the day of the examination. All patients exercised on an electrically braked bicycle ergometer. The protocol followed standard clinical routine, with an initial work load of 40 or 5 0 W and an increment of 40 W and 50 W every 4 min for women and men, respectively. The amount of work load was individually adjusted when necessary, depending on the patient's physical condition9 An ST-segment depression of 0.1 mV 80 ms after the QRS-offset was considered an abnormal exercise ECG response9 The baseline was drawn between two consecutive PR-segments. An ECG recording was made at the end of each work load and at peak exercise. The computer-assisted manual analysis of individual ST-segment recordings was based on a trimmed average of 20 ECG cycles measured for each lead and at each work load (SIEVA.NEN, 1987). The ECG lead system used was the standard Mason-Likar modification of the 12-lead system. The exercise testing was continued until limiting symptoms occurred. The Tl-201 SPECT imaging and analysis for the assessment of potential cardiac perfusion abnormalities were performed as described by KOSKINEN et al. (KosKINENet al., 1987). The TI-201 (74 MBq) tracer was injected into the antecubital vein via a cannula one minute before the estimated end of exercise. All patients had had a previous exercise test to predict the right time of injection. Immediate and redistribution images were analysed by computer. They were classified into four groups: normal, i.e. no perfusion defect; reversible defect, i.e. myocardial ischaemia; persistent defect, i.e. myocardial infarction; and persistent but partially reversible defect, i.e. a combination of myocardial ischaemia and infarction.

$4

MBEC

3 Diagnostic method 3.1 General description of MUSTA The objective of the development oF MUSTA was to determine multivariate diagnostic criteria that maximise the diagnostic accuracy in the learning group. The experimental background and the development are described in detail elsewhere (SIEv,g,NEN, 1991). The primary diagnostic variables were the maximum ST/HR-slope (index of myocardial ischaemia) and heart rate (index of myocardial oxygen demand). There was obvious overlap between these variables measured from patients with and without myocardial ischaemia, implying a need for more variables. Additional ST-segment value and shape analysis implemented by heuristic compartmental and specific criteria turned out to be effective. Using these mutally exclusive criteria, it was possible to discriminate accurately between exercise-induced nonischaemic and ischaemic E C G responses. All the criteria were implemented in Turbo-Prolog, a declarative programming language used for solving problems that involve objects and the relationships (predicates) between them. The objects of MUSTA were the diagnostic variables and the patient's proposed cardiac status. The relationship between them was the predicate 'diagnosis' used to represent the interrelationship of the variables by means of logical diagnostic rules. The general form of the predicate 'diagnosis' is the following: diagnosis(Status, STvalue, Maxhr, Slope, Shape, Lead) IF criterion~(STvalue, Maxhr, Slope) AND criterion.(STvalue, Maxhr, Slope) AND CUT. Consequently, the implementation of MUSTA required a data list containing the ST-segment value (ST value), the maximum heart rate (Maxhr), the ST/HR-slope (Slope), the shape of the ST-segment (Shape) and the E C G lead (Lead). These variables were submitted to the diagnostic criteria of MUSTA. The data determine which 'diagnosis' was totally satisfied (all related rulest..n are true) a proposed diagnosis of the patient's status was given. 3.2 Determination of diagnostic variables The primary index of ischaemia, the maximum ST/HR-slope, was determined by analysing the deviations of the ST-segment and the heart rate values in each ECG lead during each work load, and a linear regression line was produced (ST= a*HR + b). The line was defined by at least three last-measured pairs of ST-segment and heart rate values showing statistically significant correlation ( p < 0 . 0 5 and r~>0.95). The steepness a of that line is called the regression ST/HR-slope (SS). When SS was incalculable, the slope was expressed as a terminal ST/HR-gradient (ES), determined according to the two last-measured (ST-segment value, heart rate) pairs. As a general rule of diagnostic significance, the correlation between ST-segment and heart-rate values should be negative, and the value of the ST/HR-slope should be greater than 1.3/.tV (beats min-a) -~. MUSTA considered both ST-segment values measured 60 and 80 ms after the QRS-offset (ST60 and ST80), so that the maximum ST/HR-slope was obtained. The optimal time point for the ST-segment was related to the ST-segment shape (see

Electrocardiography, Myocardial Contraction and Blood Flow supplement

July 1994

Table 1 Categorydefinition of the ST-segment shape ST-slope, mV s- ~

ST-segment shape

ST-slope> 5 51>ST-slope> 1 1/> ST-slope > - 1 - 1/> ST-slope

rapidly ascending ascending horizontal down sloping

ST-slope = 50* (ST80- ST60) mV s-

the category definitions in Table 1) and to the heart rate during peak exercise (StEv~,NEN, 1991). Only the ST80 value was used, when the ST-segment was rapidly ascending and the maximum heart rate (Maxhr) was greater than 160 beats min -j. Otherwise, the selection depended on the ST60 or ST80 values during the peak exercise: if the ST-segment was more depressed at 60ms after the QRS-offset than at 80 ms, the ST60 values were used, and, in turn, the ST80 values used were used if the ST-segment was more depressed at that time point. The ECG lead system of MUSTA comprised leads V4, V5, II, aVL in this particular hierarchical order (V4 or V5 are equal). This hierarchy means that if a diagnostic SS was found in either V4 or V5, the search was terminated, and the maximum SS found in these leads was selected for a diagnostic variable. If not found the search was continued to II and, if necessary, to aVL. If SS was not found in any of these leads, then the search was carried out correspondingly for ES. The lead used in the analysis was the ECG lead, where the diagnostic slope was found. If a diagnostic slope in any lead or mode was not found, then the value of the slope was set to zero. Using this reducedlead system, a diagnostic slope was found in 37 of the 38 ischaemic patients. The determination of the diagnostic variables was achieved by providing numerical ECG data in the following form ((lead) (variable) (time point)): IIST60, IIST80, IISS60, IISS80, IIES60, IIES80 aVLST60, aVLST80, aVLSS60, aVLSS80, aVLES60, aVLES80 V4ST60, V4ST80, V4SS60, V4SS80, V4ES60, V4ES80 V5ST60, V5ST80, V5SS60, V5SS80, V5ES60, V5ES80 Maxhr These data were submitted to the logical selection procedure described above, which determines unequivocally the lead (Lead) and the values of two variables (Shape and Slope). Of the remaining two variables (STvalue and Mhxhr) required for the predicate 'diagnosis', ST value was the ST80 during the peak exercise in the selected lead, and Maxhr is self-explanatory. In principle, this procedure considered the ST-segment shape and the hierarchy between regression ST/HR-slope, terminal ST/HR-gradient and ECG leads during the calculation, and provided an unequivocal diagnostic ST/HR-slope for MUSTA, in addition to other variables needed. In other words, it performed a heart-rate and ST-segment shape adjustment of the ST-segment analysis. 3.3 Diagnostic criteria The compartmental multivariate criteria ( I . . X X ) , which gave the best diagnostic accuracy in the learning group (see the results in Section 4.1), were implemented by logically interrelated, partially linear discriminant func-

MBEC

tions. These criteria are given below in the form of the predicate 'diagnosis'.*t I

diagnosis(normal, STvalue, Maxhr, Slope . . . . ) IF STvalue> - 0.1 AND Maxhr> 130 AND Slope < 1.3 AND CUT.

II

diagnosis(normal, STvalue, Maxhr, _, rapidly_ ascending, v4) IF STvalue > 0.2 AND Maxhr> 160 AND CUT.

III

diagnosis(normal, STvalue, Maxhr, Slope, rapidly_ascending, v4) IF STvalue > 0.2 AND Maxhr> 130 AND Slope< (0.07*Maxhr- 3.7) AND CUT.

IV

diagnosis(normal, STvalue, Maxhr, _, rapidly_ ascending, v4) IF ST value > 0.05 AND Maxhr> 170 AND CUT.

V

diagnosis(normal, STvalue, Maxhr, slope, rapidly_ ascending, v4) IF STvalue > 0.05 AND Maxhr> 130 AND Slope < (0.09*Maxhr - 10.4) AND CUT.

VI

diagnosis(normal, STvalue, Maxhr, _, rapidly_ ascending, v4) IF STvalue> - 0.1 AND M a x h r > 180 AND CUT.

VII

diagnosis(normal, STvalue, Maxhr, Slope, rapidlyascending, v4) IF S t v a l u e > - 0 . 1 AND Maxhr > 143 AND Slope < (0.08*Maxhr- 10.1 ) AND CUT.

VIII

diagnosis(normal, STvalue, Maxhr, Slope, ascending, v4) IF STvalue>0.2 AND M a x h r > 130 AND Slope< (0.09*Maxhr- 10.4) AND CUT.

IX

diagnosis(normal, STvalue, Maxhr, Slope, horizontal, v4) IF. STvalue > 0.2 AND M a x h r > 130 AND Slope < (0.09*Maxhr- 10.4) AND CUT.

X

diagnosis(normal, STvalue, Maxhr, Slope, ascending, v4) IF STvalue>0.1 AND M a x h r > 143 AND Slope < (0.08*Maxhr- 10.1 ) AND CUT.

* The syntaxof Prolog is followed. 1"The unitsfor Slope, STvalueand Maxhrare/~V (beats min-1)-1, mV and beat min-1, respectively.

Electrocardiography, Myocardial Contraction and Blood Flow supplement

July 1994

$5

Xl

diagnosis(normal, STvalue, Maxhr, Slope, horizontal, v4) IF STvatue >0.1 AND Maxhr> 143 AND Slope< (0.08*Maxhr- 10.1) AND CUT.

XII

diagnosis(normal, STvalue, Maxhr, Slope, ascending, v4) IF STvalue > 0.0 AND Maxhr> 155 AND Slope < (0.08*Maxhr- 11.1 ) AND CUT.

XIII

diagnosis(normal, STvalue, Maxhr, Slope, hori9 zontal, v4) IF STvalue > 0.0 AND Maxhr > 155 AND Slope < (0.08*Maxhr- 11.1) AND CUT.

XIV

diagnosis(normal, STvalue, Maxhr, Slope, rapidly_ascending, ii)IF STvalue > 0.0 AND Maxhr> 155 AND Slope < (0.08*Maxhr- 11.1) AND CUT.

XV

diagnosis(normal, STvalue, Maxhr, Slope, ascending, ii)IF STvalue > 0.0 AND Maxhr> 155 AND Slope < (0.08*Maxhr- 11.1) AND CUT.

XVI

diagnosis(normal, STvalue, Maxhr, Slope, horizontal, ii) IF STvalue > 0.0 AND Maxhr> 155 AND Slope < (0.08*Maxhr- 11.1) AND CUT.

XVII

diagnosis(normal, STvalue, Maxhr, Slope, rapidly_ascending, avl) IF STvatue > 0.0 AND Maxhr> 143 AND Slope< (0.08*Maxhr- 10.1) AND CUT.

XVIII

diagnosis(normal, STvalue, Maxhr, Slope, ascending, avl) IF STvalue > 0.0 AND Maxhr> 143 AND Slope < (0.08*Maxhr- 10.1) AND CUT.

XIX

diagnosis(normal, STvalue, Maxhr, Slope, horizontal, avl) IF STvalue > 0.0 AND Maxhr> 143 AND Slope< (0.08*Maxhr- 10.1) AND CUT.

XX

diagnosis(probable_myocardial_ischaemia . . . . . . . . . . ) IF CUT.

To summarise, the diagnostic process of MUSTA is as follows: First, the variables are determined; the ST/HR-slope and the ECG lead are determined according to the procedure given in Section 3.2; the maximum heart rate is measured at peak exercise, the ST-segment deviation is measured at 80 ms after the QRS-offset at peak exercise, and its shape in the selected lead is determined

$6

MBEC

according to Table 1. A criterion ( I . . X X ) that is totally true is then searched; if such a criterion is found the proposed status is 'normal', otherwise, the status is 'probable myocardial ischaemia' (criterion XX). Consequently, the selection of the criterion and the ST/HR-slope partition used in each case depends on individual exercise ECG response.

4 Performance characteristics of MUSTA The diagnostic ~tccuracy (the fraction of correct diagnoses), the sensitivity (the fraction of correctly diagnosed ischaemic patients) and the specificity (the fraction of correctly diagnosed normal subjects) of the ST/HR-analyses (ST60/HR, ST80/HR, MUSTA) and the standard exercise ECG based on clincal interpretation by the supervising physician were assessed in relation to the computer-aided visual interpretation of the TI-201 SPECT images. Comparison of the performance characteristics was performed by McNemars's modification of the zZ-test for paired proportions and with Yates' continuity correction for unpaired cases. The diagnostic criterion for a positive test in the exercise ECG was an ST-depression greater than 0.1 mV in any lead (aVR and V1 excluded). The diagnostic criterion for a positive test in the ST60/HR- and ST80/HR-analyses was the maximum regression ST/HR-slope greater than 1.3/~V (beats min-l) -1 in any lead (aVR and V1 excluded). Incalculable slopes appearing in these analyses were considered negative test results. The diagnostic criteria of MUSTA were those given above.

4.1 Initial performance The initial performance characteristics were determined in the learning group (38 ischaemic and nine normal patients). The diagnostic accuracy of the exercise ECG analysis was 77%, with a sensitivity of 84% and a specificity of 44%. For the ST60/HR-analysis, the corresponding values were 70%, 79% and 33%. The ST80/HR-analysis performed less well; the diagnostic accuracy was 60%, the sensitivity was 69%, and the specificity was 25%. The sensitivity and specificity of MUSTA were 100% and 89%, indicating diagnostic compatibility between TI-201 SPECT and the present method. The diagnostic accuracy was thus 95%. With regard to T1-201 SPECT, MUSTA performed significantly better compared with standard exercise ECG (p