Signal-averaged electrocardiography (SAECG) today usually denotes a widely used technique of acquiring ECG signals with high accuracy at low noise using coherent averaging of non-ectopic QRS complexes, relatively high sampling rates (typically 1000 Hz or more) and high-precision analog-to-digital conversion (12-16 bit). This technique is mostly applied in the search for pathologic late potentials (LPs) - higher-frequency low-amplitude signals in the terminal QRS portion, proven to be indicators of increased risk of developing malignant ventricular tachycardia (VT) in patients after a myocardial infarction (MI) (BERBARI et al., 1989, FLOWERS et al., 1991).

Different methods have been developed to obtain relevant information from the SAECG in recent years. Time domain analysis of the high-pass filtered SAECG is the most widely used approach. SIMSON (1981) developed a technique which is based on bidirectional filtering and is still most widely used. A task force committee of the ESC, AHA, ACC (BREITHARDT et al., 1991) set the basic guidelines for time domain analysis. Also frequency domain analysis techniques emerged, originally based on estimating the Fourier spectra of a windowed time interval covering the terminal QRS part and the ST-segment (CAIN and colleagues, 1984,1985,1989, PIERCE et al., 1989). Spectral area ratios were often used to characterize the proportion between high- and low-frequency components in the time interval selected (BUCKINGHAM, et al.). The next step was to utilize the short-time Fourier Transform (STFT) in order to analyse the SAECG signal in the form of a "moving spectrum", where each "instantaneous" frequency distribution is gained by first multiplying the analyzed time interval in the vicinity of the time moment of interest by a window (usually of Backman-Harris type), and then conducting a Fast Fourier Transform (FFT).

Time-domain and frequency-domain analysis in separate did not yield sufficient diagnostic power. Since LP's are of undefined and/or possibly varying frequency and are superimposed on a relatively large-amplitude but slow changing R- and S- electrocardiographic waves, time-domain analysis alone is not able to accurately detect these pathologic oscillations, although high-pass filtering is normally used prior to measurement. The main problems appear to be the selection of a steep and linear phase filter causing little or no ringing in the QRS-area being examined, while preserving signal shape. Bidirectional Butterworth IIR filters may strongly influence signal morphology, whereas alternative FIR filters are difficult to optimize - a low number of taps results in a poor frequency response, while a large number of taps increases filter ringing and obstructs precise delineation of the QRS-complex (GRAMATIKOV,1993). Filtering in several frequency bands has been applied to check for LP's of various spectral contents, but that has not been accepted enthusiastically by most pf the physicians. Besides, no universal measurement rules have been established so far for any frequency band, to extract diagnostically reliable time domain parameters. Latter depend considerably on the detection of the QRS endpoint, which in many cases is ambiguous.

On the other hand, traditional frequency domain analysis suffers from several drawbacks : it fails to provide time localization of signal singularities characterized by high-frequency components and hence information on the precise incidence of LP's is lost. Hump-like windows are largely being applied in order to reduce edge discontinuities, but they tend to widen narrow spectral components, since a multiplication of a signal by a window function in time dimain is equivalent to convolution of the Fourier transform of the signal and the window function. In an effort to widen the window and thus increase frequency resolution, many investigators have extended the analysed epoch toward the T-wave, covering the entire ST-segment. Latter however revealed no significant diagnostic value. Furthermore, the low-amplitude window edges strongly suppress the ECG-signal there, the most interesting QRS end-area being often almost masked out. And last but not least, the regularity of the frequency of the oscillatory potentials in the terminal QRS also plays an important role in spectral analysis. MACHAC et al.(1988) showed that a very regular frequency has most of its energy concentrated in a single spectral peak, whereas a lack of periodicity results in energy being distributed over a wide range of frequencies. Obviously these two cases will be reflected differently by any spectral area ratio chosen. A time- frequency representation would help avoid this misleading effect, since a signal component of relatively constant frequency might occasionally dominate an epoch being subject to the current spectral analysis, but the situation would most probably change with time, since there is no electrophysiological background for a persisting narrow-spectrum component when dealing with depolarization fronts.

Apparently, combined time-frequency analysis would be required to increase predictive accuracy of SAECG

Literature

Gramatikov, B., Georgiev, I.   Wavelets as an Alternative to STFT in signal-averaged electrocardiography.    Medical and Biological Engineering and Computing, Vol.33, No.3, May 1995, pp. 482-487.

Gramatikov, B.   Detection of late potentials in the signal-averaged ECG - combining time and frequency domain analysis.   Medical and Biological Engineering and Computing, Vol.31, No.4, July 1993, pp.333-339.