THE PERFORMANCE OF INFORMATION-THEORETIC CRITERIA IN DETECTING THE NUMBER OF INDEPENDENT SIGNALS IN MULTILEAD ECGS

Three different methods to detect the number of independent signals in multilead ECGs were evaluated by using different ECG realizations with known specifications for signal and noise: a threshold method (TM), the minimum description length (MDL) and Akaike's information criterion (AIC). The fundamental assumption in this kind of signal processing is that both the signal and the noise stem from independent stochastic generators. The consequence is that the detection of the number of signals is only possible if the noise is white, or if the noise properties have been specified. The evaluation was performed with respect to the QRS complex of individual multilead ECG simulations and to the entire ensemble. In the simulated ECGs the number of independent signals was fixed (eight). It was found that, out of the three methods studied, the performance of MDL was the best, especially when the number of available observations in the noise (used to estimate the noise specifications) was moderate.