ECG Signal Analysis by Pattern Comparison

R Bousseljot, D Kreiseler
Physikalisch-Technische Bundesanstalt, Berlin, Germany

Abstract
A new technique for the ECG interpretation by wave form recognition is introduced. The basis for the pattern matching algorithm without feature extraction is un ECG wave form database with ECG signals and patient diagnosis information. The recognition takes place by comparing the wave forms of 12 leads of the examined ECG with the same leads of different ECGs from the database. The similarity of the compared ECG beats is calculated with the cross correlation function. By means of the algorithm, we select ECGs of the database which are the most similar to the examined ECG considering all of the individual leads. Using the available database information, the diagnosis which corresponds to the most similar ECGs of the database is then identified as the diagnosis for the examined ECG. The new method has been tested with 249 original ECGs, most of them from patients with myocardial injurction or from healthy persons.

2.

Hypothesis

The technique is based on the hypothesis that those ECGs whose signal waveforms are in best agreement in their leads can in all probability be assigned to the same diagnosis. As this assumption is at the basis of medical experience, it should be possible to obtain the diagnosis for an unknown ECG from a comparison between the signal waveforms of this ECG with unknown diagnosis and an ECG of known diagnosis stored in a database. The technique thus starts from the assumption that the waveforms of the quasi-periodical signal sections of the leads of the 12-channel ECG represent a "finger print" of the ECG in question.

3.

Procedure

The method developed for ECG signal analysis by pattern comparison with the aid of an ECG database consists of four succe,sive steps:

1.

Introduction

3.1.

Signal conditioning

For the computer-aided analysis of ECGs, procedures are at present used which, among other things, measure the waveforms of the ECG to extract features from the signals for use as parameters [l].With the aid of sets of features and on the basis of decisions allowing for the applicable rules or using neural networks [ 2 ] , classifications and defined diagnostic statements are made. This decisionmaking takes place automatically without the doctor exerting any influence. The present publication presents a novel technique which is not based on the measurement and extraction of individual features of the ECG. By comparison of the signal patterns of a 12-channel ECG of unknown diagnosis with signal waveforms of an ECG database, those ECG are searched in the database whose waveforms are most similar to those of the unknown ECG.

The first step ensures the computational preparation of the derived and digitized 12-channel ECG. The analysis relates to beat variance, drift, signal offset, extra beats, etc. To realize this step, signal processing methods can be used which often are already implemented in computer-aided ECG equipment. According to the periodicity of the ECG signal, a representative ECG beat is selected isochronous over all leads. This selection can be made automatically or by interaction at the display using a line cursor taking in all leads. The ECG signal patterns thus provided are the prerequisite to the comparison of the waveforms with the ECG of the database.

3.2.

Pattern correlation

The assessment of the similarity of the ECG signal patterns of equal ECG leads is carried out by calculation of

0276-6547/98 $10.00 0 1998 IEEE

349

Computers in Cardiology 1998 Vol25

correlation functions whose function values are formed from the correlation coefficients K: N 1 N

In dependence on the temporal shift of the signal window with the time series x,,(unknown ECG) and y,, (reference ECG) - with 1 5 n 2 N . N = number of function values - the correlation functions form maxima according to the periodicity of the ECG if the patterns are similar. Figs. 1 and 2 show two examples of correlation functions in the case of similarity of the ECG beats as well as in the case of strong differences of the waveforms from one another. According to the periodicity of the ECG signals, the correlation functions are also periodical. The different maxima are due to the variance between the ECG beats. The examples show that it is permissible for ECG without disrhythmias or strong anomalies between the beats to limit the correlation analysis to only one beat typical of the ECG in question.
1
Kit)

In pattern correlation, the unknown ECG is compared with the ECGs of the database in all 12 leads. The result of each reference ECG comparison is a 12-dimensional vector whose elements give the correlation values for the respective lead. The range of values corresponds to the definition range of the correlation coefficient K with -1 I K I +l.For the further considerations, only the positive correlations are taken into account. The correlation measure 100% (K=l) thus means identity of the patterns. A correlation statement towards K=O indicates completely different signal patterns in the waveforms of the lead considered.

3.3.

Classification

p.5
0

It is the aim of classification to summarily assess the agreement of the lead-related correlation results obtained from the comparison of the unknown ECG with any reference ECG. It must be taken into consideration that the result of two ECG pattern comparisons may be multivariable and available as a 12-dimensional vector of the correlation results of the individual leads and the number of comparisons with the reference ECG may be very great. To solve this task, a modified distance technique of the multivariate signal analysis is used. As the method can be applied independantly of the dimension of the vector of the correlation results and thus independantly of the number of leads to be taken into account, the classification step is illustrated by the example of the dimension nz = 3.

-0.5
-1

4
+tis

Fig. 1 Correlation function for very similar signal patterns

x x

- ---"

i
0

Mi

B -

Fig. 3 Graphical interpretation of the multichannel correlation results (dimension nz = 3) Fig. 2 Correlation function with distinct differences of the signal patterns Fig. 3 shows a three-dimensional Cartesian system of coordinates which is formed by the correlation results of three ECG leads. According to the used range of values for the correlation coefficient K, the axial ranges from 0 to 1

350

are obtained. Any comparison of signal patterns of the unknown ECG with those of an ECG from the database leads to a point inside the cube formed for the threedimensional case If the signal patterns are identical in all leads ( K = I), point Po ( l , l , l ) will be obtained. The length of the diagonal in this cube represents the maximum similarity measure in the case of three leads. All other points are associated with vectors whose magnitude are smaller compared with the length of the diagonal and thus indicate less similarity of the signal patterns. When all leads of the 12-channel ECG are allowed for, the dimension m = 12 is obtained. The normalized modulus vector as similarity measure then reads Mi: maxima of the correlations of a lead m: dimension

70
~' .I..IIII._..LX-.X-^I-.

80

90 -r1%

10'0

_"I~-_I~",-_^I."_.xII*l'~~.."".."~~~~~~~~~~~~

Fig. 4 Detail of an r-b diagram showing the distribution of the pattern correlations of an unknown ECG

3.4
11 = 7

Evaluation of the classification results

JTW
i=l

Jm

Fig. 5 shows as an example the distribution of the correlation results, including the diagnostic information from the database.

A second value should allow for the uniformity of the correlations of the leads. This is possible by indicating the angle cp or the distance b:
/

NORM MI A Ahdl V ASMl + ALMl

0

0.1 5
0.1

The value b = 0 means here that equal correlation values were calculated for all leads. This does not allow conclusions to be drawn for the similarity of the signal patterns of all leads. It is, however, possible, for example, to take in the influence of noise-affected signals, or other technical disturbances in a lead, on the overall result. The two values r and b can be calculated for any comparison between an ECG of unknown diagnosis with the reference ECG. In the following, the value r will be given in %. The assignment of the known diagnoses of the reference ECGs to the distributions according to Fig. 4 allows the diagnosis of the unknown ECG to be estimated.

0.05

n
85 90 95 -r(%

100

Fig. 5 Assignment of the pattern correlations according to Fig. 4 to the diagnoses of the reference ECGs (detail) The correlations of the patterns of the ECG of unknown diagnosis with themselves furnish in all leads a 100% similarity (r = loo%, b = 0). In the range r > 85%, some correlation results indicate database ECGs with a great similarity of the signal patterns in all leads. These ECGs stem from patients with inferior myocardial infarctions (IMI). In the example given, the IMI diagnosis can therefore also be inferred for the unknown ECG.

351

The evaluation of the classification results according to Fig. 5 can be supported as follows: clustering of the distributions taking into consideration of selected leads or of their combinations for the correlation process inserting of the ECG signal patterns representation of the discrete correlation results for every lead inserting of detailed information from the database about findings, anamnesis, anonymized patient information, medications, technical parameters, etc. These interactive possibilities are intended to enable the doctor - in contrast to conventional methods of computeraided ECG interpretation - to separately evaluate the results including the patient information available. The result of the signal pattern comparison is not a diagnostic statement which can be accepted or rejected but a probability statement on the basis of similar cases of an ECG database. Accordingly, the quality of this database as regards the extent and the validation is a decisive prerequisite for the function of the technique presented.

5.

Advantages

The presented method for computer-aided interpretation of ECGs by signal pattern comparison with ECG databases offers the following advantages: no feature extraction for the ECG no limitation of the diagnostic statements inclusion of rare diseases by specialized ECG databases possibility of extending the databases as bases of knowledge without changes of the algorithm inclusion of further patient information from the database for making diagnoses lead-related inclusion of ECG patterns robustness with respect to disturbances or signal failures in a lead learning ability by inclusion of the results of current patient examinations cost advantages by use of existing PCs local availability of the ECG database on CD or use of centralized databases via modem, ISDN or network.

References
[ l ] Willems J.L., Abreu-Lima C., Arnaud P., van Bemmel J.H., Brohet C., Degani R., Denis B., Gehring J., Graham I., van Herpen G., Machade H., Macfarlane P.W., Michaelis J., Moulopoulos S.D., P. Rubel P. The diagnostic performance of computer programs for the interpretation of electrocardiograms New England Journal of Medicine, vol. 325, 1767-1804, 1991 [2] Macfarlane P.W. Recent developments in computer analysis of ECGs, Clinical Physiology, vol. 12 (1 992), 3 13-7 [3] Bousseljot R., .Aufbau der EKG-Datenbank CARDIODAT der PTB (Setting-up of the ECG database CARDIODAT of PTB). ,26th Annual Meeting of the Deutsche Gesellschaft fur Biomedizinische Technik, Biomedizinische Technik, vol. 39, Supplement 1 (1994), 250-251

4. Results
The operation of the method presented was checked with the aid of the CARDIODAT signal pattern database of the Physikalisch-Technische Bundesanstalt [3]. Within the scope of initial investigations, 249 ECGs of patients with myocardial infarctions and normal persons were allowed for. For all types of investigation the distributions in the rb diagram was calculated and evaluated in accordance with the known diagnoses of the reference ECGs. For the evaluation those five ECGs were taken into account for each diagnostic group which, compared with the respective unknown ECG, furnished the best correlations. Pooling of the five best results is made by unweighted averaging. The following table shows the sensitivities and specificities reached. In the anterior myocardial infarct group, 75 ECGs of patients with the AMI, ASMI and ALMI diagnoses were pooled. The inferior myocardial infarct group considers 121 ECGs with the diagnoses IMI, ILMI, IPMI, IPLMI. The normal persons group contains 53 test persons. Finding anterior MI inferior MI normal person

Authors addresses

1
I

Sensitivity 89.33%
89,26%

67,92%

Specificity 93.68%
84,38Yo

Dr. Ralf Bousseljot Dr. Dieter Kreiseler Physikalisch-TechnischeBundesanstalt Labor 8.21 Abbestrasse 2- 12 10587 Berlin, Germany e-mail: rbousseljot@biodat.berlin.ptb.de e-mail: dkreiseler@biodat.berlin.ptb.de

96,43%

352