Bivariate piecewise stationary segmentation; improved pre-treatment for synchronization measures used on non-stationary biological signals

Analysis of synchronization between biological signals can be helpful in characterization of biological functions. Many commonly used measures of synchronicity assume that the signal is stationary. Biomedical signals are however often strongly non stationary. We propose to use a bivariate piecewise stationary pre-segmentation (bPSP) of the signals of interest, before the computation of synchronization measures on biomedical signals to improve the performance of standard synchronization measures. In prior work we have shown how this can be achieved by using the auto-spectrum of either one of the signals under investigation. In this work we show how major improvements of the performance of synchronization measures can be achieved using the cross-spectrum of the signals to detect stationary changes which occur independently in either signal. We show on synthetic as well as on real biological signals (epileptic EEG and uterine EMG) that the proposed bPSP approach increases the accuracy of the measures by making a good tradeoff between the stationarity assumption and the length of the analyzed segments, when compared to the classical windowing method.