EVALUATION OF A NEW SIGNAL PROCESSING APPROACH OF HIGH-RESOLUTION BSPM

Gy. Kozmann^1,2, K. Haraszti¹, L. Gerencsér ³and Zs. Vágó^3
1University of Veszprém, Veszprém,
²Research Institute for Technical Physics and Materials Science, and ³Computer and Automation Institute,
Budapest, Hungary

Abstract: As a part of our effort in the development of a high-resolution body surface potential mapping system, the utility of the simultaneous perturbation stochastic approximation method, SPSA, was tested. The method was used for the high-precision classification of QRS patterns, prior to synchronized averaging. According to this pilot study, the SPSA outperforms the correlation method usually applied for similar problems. The classes separated by the SPSA were characterized by the averaged ECGs and by the averaged prediction error functions. According to our preliminary interpretation the new approach sensitive enough to separate cycles different in localized propagation velocities.

INTRODUCTION

According to our previous modeling results, propagation of depolarizing wave-fronts is largely deterministic, resulting in piece-wise smooth, body surface ECG signals. However, even in normal hearts there are times when ECG exhibits (usually) small-amplitude sudden unexpected changes (notches and slurs). The small-amplitude signs of discontinuous propagation occurs at times of epicardial or endocardial breakthrough, collision of activation waves (the most common “characteristic events”, CEs). Furthermore, discontinuous propagation is also occurring at the interface of regions with altered activation (or activation propagation) properties, consequently, the pattern of small details of surface signals provide diagnostic information on spatial (or spatio-temporal) bioelectrical tissue properties of the myocardium, not accessible by other methods [1,2]. When the “relative-curvature” of the activation wave and the boundary of the region with altered propagation properties is small, the “jump” is large, while changes experienced at the time of an extended activation wave collision against geometrically small “islands” of unexcitable tissues should be £ a few mV [3]. As the information carrier is hidden in the sub-noise level details of the surface ECG pattern, the need for a high precision, high resolution body surface mapping (HR BSPM) seems to be justified. From a diagnostic point of view, the detection of localized defects (discontinuities) in tissue electrical properties due to ischemic heart disease or fibrosis might have a significant impact on the care and treatment of patients.

Recovery the body surface signs of discontinuous propagation can be revealed by a simple linear predictive algorithm (LPA) and a decision making procedure [1]. However, before applying the LPA, the improvement of the signal-to-noise ratio is an essential requirement. In case of similar low noise signal requirements, usually the time-aligned averaging procedure is applied. In this case at first the subsequent heart cycles are classified according to their QRS pattern, subsequently the time-aligned averaging is performed for the majority class. The classification is usually based on the use of the correlation method. Because the correlation method is not sensitive enough on the low-amplitude small details of the ECG pattern, in this study, the performance of a new classification approach, called simultaneous perturbation stochastic approximation (SPSA) method, was tested and compared with the common method.

METHODS

The correlation method and the SPSA method was used on the very same test data for the classification of QRS patterns.

The correlation method is well known e.g. from the late potential analysis [4].

The SPSA method represents each QRS cycle by a point in a multidimensional space. In our experiment representative segments of the leads were selected by a time-window around the point of steepest negative gradient. By this windowing n sample points were taken from each heart cycle and from all the three leads, i.e. each cycle was represented by 3n sample points (in our example n=66). Subsequently, the method covers the measured ECG pulses with spheres (with definite radius) in the 3n dimensional space. The cycles falling outside of the sphere with a predefined dimension, form members of new clusters. (The SPSA method does not use the Euclidean distance definition in contrast to the classical classification methods like cluster-analysis, vector-quantization, correlation coefficient.)

The number of classes is not predefined; but the maximal deviation within a class is kept fix. Mathematically the SPSA method concludes in a non-smooth optimalization problem, which is not a trivial exercise in a high dimensional space [5].

Human test measurements of 7 healthy subjects were taken by the orthogonal lead system suggested by Simson for late potential studies. QRS patterns were classified, subsequently the difference in the two kinds of classifications were analysed in terms of the averaged QRS signal and in terms of the linear prediction error signals [1].The prediction error was computed by the equation:

e_i(t) = φ_i(t) - 2 φ_i (t-Δt) + φ_i (t-2Δt).          where:

φ_i(t)  : potential at the ith lead at the moment t,

Δt : time step,

e_i(t) : error at the ith body surface ECG, at the moment t.

The number of cardiac cycles considered was always >100. Residual noise level was <1 mV.

RESULTS AND DISCUSSION

In Fig.1 a typical example is plotted on the correlation coefficients (CC) of the subsequent cycles with a reference QRS pattern. According to this plot the similarity of the subsequent cycles is high for normal subjects (CC>0.99). SPSA classified the very same cycles in two classes, the members one of the classes is marked by *.

Figure 1: Example on the QRS similarity levels of a normal subject, in terms of CC. SPSA formed a second cluster from the cycles marked by *.

Averaged signals of the classes of the SPSA classification, revealed small but significant differences both in the ECG and especially in the e_i(t) prediction error signals. According to the example shown in Fig. 2, the timing of certain peaks did show ms order of shifts, while the location of other peaks were stable. Similarly, changes significant to the residual noise level (left from the vertical dotted line) were observed in certain error-peak amplitudes, while others were stable.

Figre 2: Example on the averaged prediction error functions computed from the two classes separated by the SPSA method. Before the Q_onset (dotted line) the residual noise is illustrated. In the QRS interval observe the difference in timing and amplitude of certain signal peaks (e.g. peaks marked by arrows of Dt and Da) and the stability of others.

Based on our previous modeling studies, these type of changes might be attributed to slight and local temporal changes in the propagation velocity of activation waves in certain parts of the myocardium.

Though the number of cases in this study was rather limited, based on these preliminary results we expect, that the rather computation intensive SPSA signal processing step may contribute to the elaboration of a clinically meaningful new HR BSPM procedure.

Acknowledgements: This study was supported by the National Research Found grants T30747, T33085, and by the NKFP grant 2/052/2001 of the Ministry of Education, Hungary.

REFERENCES

[1] Gy. Kozmann, Zs. Cserjés, I. Préda, “Manifestation of characteristic events of ventricular activation in body surface potential field” In „Electrocardiology ‘83” pp. 2o4-2o7, Excerpta Medica, Amsterdam, 1984.

[2] Zs. Cserjés, Gy. Kozmann, M. Tysler, et al., “Noninvasive spatio-temporal detection of epicardial breakthroughs and septal wave collision: Results of a model study” Jpn. Heart J., 35, Suppl. pp. 93-94, (1994).

[3] F Greensite, “Some imaging parameters of the oblique dipole layer cardiac generator derivable from body surface electrical potentials” IEEE Trans. Biomed. Eng. vol. BME-39, pp. 159-167, 1992.

[4] G. Breithardt, M. Cain, N. El-Scharif et al., “Standards for analysis of ventricular late potentials using high resolution signal-averaged electrocardiography” Eur. Heart J., vol. 12, pp.473-480, 1991.

[5] L. Gerencsér, Gy. Kozmann, Zs. Vágó, K. Haraszti, “The use of the SPSA method in ECG analysis.” IEEE Trans. on Biomed. Eng. (accepted)