Combining MCG and ECG in Classifying Myocardial Infarction

Juha Nousiainen^a, Jaakko Malmivuo^a, O Sakari Oja^b

^aRagnar Granit Institute, Tampere University of Technology, Tampere, Finland
^bTampere Univeristy Hospital, Tampere, Finland

Correspondence: J Nousiainen, Ragnar Granit Institute, Tampere University of Technology, P.O. Box 692, FIN-33101 Tampere, Finland.
E-mail: juha.nousiainen@tut.fi, phone +358 3 3115 2594, fax +358 3 3115 2162

1.  Introduction

For already 40 years, one interesting topic in biomagnetism has been the clinical applications of magnetocardiography (MCG). Although a large number of studies about the ability of different MCG methods to provide new information on the heart diseases has been published, the comparisons of the diagnostic powers of MCG and ECG and thus the amount of additional diagnostic information provided by MCG have been an object of minor interest.

This paper focuses on the information contents of the MCG and ECG to classify old myocardial infarctions. The correct classification rate is used as a measure of the information content that is evaluated with a statistical approach. Comparative analysis requires that both the MCG and ECG must be modeled and recorded with methods that are equivalent in complexity. This is fulfilled by using a dipolar heart model for both methods and by applying vector recording for both signals, the unipositional VMCG lead system and the Frank VECG system. In addition to the comparison of diagnostic information content of the VMCG and VECG, this study shows how the application of MCG combined with ECG can increase the amount of clinical information and thus improve the diagnostic performance of the methods.

2.  Material and Methods

The material of the study consisted of 313 subjects including 152 normal, healthy subjects (85 male and 67 females, age 54 ± 11 years), 90 patients with old inferior myocardial infarction (IMI), (73 male and 17 females, age 59 ± 10 years) and 71 patients with old anteroseptal myocardial infarction (AMI), (59 male and 12 female, age 59 ± 5 years).

The clinical diagnosis of myocardial infarction was based on a history of chest pain, a significant release of creatine kinase from the heart (CK>300 U/l and CK-B>10 %) and characteristic changes in the ECG in the acute phase of the infarction during the hospital period. In 106 patients, localization of the infarction was based on one or more positive findings in the following ECG-independent diagnostic tests: (a) a rest perfusion defect in clinical cardiac isotope tomography (Tl-201 SPECT), (b) a hypo- or akinetic wall of the heart detected in echocardiography, or (c) an epicardial scar in the myocardium found during a coronary by-pass operation. In the remaining 55 patients, localization of the infarction was solely based on the diagnostic clinical 12-lead ECG changes measured in the acute phase of the infarction during the initial period of hospitalization.

The normal material consisted of healthy subjects without a history of medical treatment or chest pain or other possibly heart-related symptoms, normal arterial blood pressure and normal heart sounds.

From all subjects a similar set of recordings was made: the recording of the VECG with the Frank lead system and the VMCG with the unipositional lead system [Malmivuo and Plonsey, 1995] provided three orthogonal components (signals) of the electric and magnetic heart dipoles, respectively. In this study, a coordinate system was used where the positive X axis is directed anteriorly, Y axis to the left and Z axis cranially in distinction to that recommended by the AHA. Consequently, similar data processing and analysis was performed for all six signals, including time-averaging to improve the S/N ratio, time-normalizing of the QRS and ST-T waves to compensate for the effect of interindividual variability in the wave duration on the instantaneous wave amplitudes, and parameter extraction to perform the statistical analysis [Oja, 1993]. From the time-normalized waves, 84 QRS and 63 ST-T amplitude measures were extracted. In addition to these, maximum and minimum wave amplitudes, time intervals and durations as well as the QRS-duration were derived from the signals.

The statistical analysis of the data was performed with the BMDP Statistical Software. The group mean VECG and VMCG were computed separately for the three groups. Discriminant indexes [Kornreich, 1989] were computed for each bigroup comparison by subtracting the group mean values of the groups and dividing resulting differences by the corresponding pooled standard deviations.

The correct classification rate of the methods was evaluated with stepwise linear discriminant analysis (LDA) by finding the linear combination of parameters (a classification function) that best predicts the group to which each individual case belongs. The classification results were assessed with the jackknife method that leaves each case in turn out of the computation of the classification function and then uses the function to classify the case omitted. The McNemar test of symmetry was used to test the significance of the observed differences in the classification.

3.  Results

3.1. Bigroup Comparisons

Figure 1 shows as an example the bigroup comparison between normal cases and patients with IMI. In the upper panel, the group mean QRS complexes and ST-T waves in the X, Y and Z leads are given for VMCG and VECG. The lower panel shows the corresponding discriminant indexes. In the VMCG, best discriminating features appear at the end of the T wave of the Z lead and at the beginning of the QRS of the Y lead. VECG provides highest discrimination during the T wave of the Z and Y leads and also during the first half of the Y lead QRS. Similar analysis was performed for the two other bigroup comparisons.

Figure 1.   Upper panel: Averaged time-normalized VMCG and VECG wave forms in X, Y and Z leads for normal subjects (solid lines) and patients with IMI (broken lines). VMCG amplitudes are in mAm² and VECG amplitudes in mV. Lower panel: Corresponding discriminant index (DI) for the comparisons above. Time scale is for the QRS complex in percentages of the total QRS duration (0-100%) and for the ST-T wave in percentages of the total ST-T wave (100-200%).

The eight best discriminating parameters of each bigroup comparison selected by the LDA are listed in Table 1. Figure 2 depicts how the correct classification rate improves in two classification problems (normal/IMI and normal/AMI) when the number of parameters included in LDA model is increased in the order shown in Table 1. The best correct classification rates that are also given in Table 1 show that VECG and VMCG have similar ability to discriminate between normal and two infarction groups. However, VECG performed better than VMCG in the AMI-IMI classification.

Table 1.    Best discriminating parameters in the three bigroup classifications selected by LDA, and the best correct classification rates (CCR) after the eight parameters in the classification model. Abbreviations: X, Y and Z stand for the X, Y and Z leads, respectively. Q and T stand for QRS complex and ST-T wave. The numbers stand for the amplitudes at the normalized time instant of QRS complex or ST-T wave. Subscript min and max stand for the minimum and maximum wave amplitude and min-t and max-t stand for the time to the minimum and maximum amplitude from the beginning of the QRS complex.

When both the VECG and VMCG parameters were entered simultaneously to the LDA, an improvement in the correct classification rates was achieved. This is shown in Figure 2 that gives the correct classification rate for the combined VEMCG analysis, too. The best discriminating features were those underlined in Table 1. In both classifications (normal/IMI and normal/AMI) the two best discriminating parameters were electrical and the third and fourth were magnetic. The best classification rate of 95.5 % for the normal/IMI classification was achieved when seven parameters were included into the classification model. The corresponding best value for normal/AMI classification was 91.3 % with eight parameters in the model. When these values were compared to the best classification rates of VECG given in Table 1, statistically significant (p=0.019) improvement was achieved in the normal/IMI classification but not in the normal/AMI classification. The frequency table of the best classification results of normal/IMI classification is given in Table 2.

Normal / IMI	Normal / AMI

Figure 2.   Correct classification rates (in percentages) as a function of number of parameters (from 1 to 8) used in LDA. Left graph is for the normal/IMI and right graph for the normal/AMI classification.

3.2. Threegroup Classification

The correct classification rate of the threegroup comparison is depicted in Figure 3 with increasing number of the classifying parameters. The best combined VEMCG CCR was 82.7 % compared to 79.2 % in VECG and 76.6 % in VMCG. These improvements were not statistically significant. LDA selected two VMCG parameters (fourth and sixth) and four VECG parameters.

Table 2.   Frequency table of the best classification results (in percentages) of normal / IMI between VECG and VEMCG.

Figure 3.   Correct classification rates (in percentages) of the threegroup classification as a function of number of parameters (from 1 to 8) used in LDA.

4.  Discussion and Conclusions

This study was focused on one side of clinical information carried by ECG and MCG, the classification between different disease groups. Although its limitations in number of patient cases and disease groups, the study clearly demonstrated that MCG can bring valuable diagnostic information in addition to the ECG. The correct classification rate increased in all classification problems, when VMCG parameters were used in combination with VECG parameters (combined VEMCG) in the classification function, although the diagnostic performance of VMCG alone was similar to that of VECG. This is explained by the fact that the patient sets classified correctly by VECG and VMCG were approximately equal in size but not identical by patient cases.

The worse quality of the MCG signal and higher interindividual variability of MCG parameters compared to those of ECG directly reflected as lower discriminating indexes of MCG than ECG. This explains why in the LDA of the combined VEMCG analysis electrical parameters were preferred to magnetic ones.

This study also demonstrated, how different abnormal conditions of the heart can create both redundant and complementary diagnostic information in VECG and VMCG. In the combined threegroup VEMCG classification, the best discriminating MCG features (Z T80, Z Q10 and Z Q15) were omitted from the classification function showing redundancy with the VECG Z T80 and X Q20 parameters, respectively. Clearly the very initial QRS complex of the VMCG Y lead (Y Q5) and the terminal phase of the QRS of the VMCG X lead (X Q80-85) contain complementary information independent of the VECG.

The amount of new information provided by the VEMCG can be assessed against the information provided by the combination of the 12-lead and Frank VECG methods. Willems et al. [1987] tested the performance of this combination to discriminate seven patient groups. They found that it did not improve the correct classification rate. The reason for this is that of these 15 leads only the three orthogonal ones are independent and the other 12 leads are redundant.

References

Kornreich F, Montague TJ, Rautaharju PM, Kavadias M and Horacek MB. Multigoup Diagnosis of the Standard 12-lead Electrocardiogram J. Electrocardiology 22 (Supplement): 141151, 1989.

Malmivuo J, Plonsey R. Bioelectromagnetism: Principles and Application of Bioelectric and Biomagnetic Fields. Oxford University Press, New York, 1995.

Oja, OS. Vector magnetocardiogram in myocardial disorders. Thesis, University of Tampere, Finland, 1993

Willems JL, Lesaffre E and Pardaens J.Comparison of the classification ability of the electrocardiogram and vectorcardiogram. Am. J. Cardiol. 59: 119-124, 1987.