Estimation of narrowband amplitude and phase from electrophysiology signals for phase-amplitude coupling studies: a comparison of methods

Abstract

Many functional connectivity studies based on electrophysiological measurements, such as electro-and magnetoencephalography (EEG/MEG), start their investigations by extracting a narrowband representation of brain activity time series, and then computing their envelope amplitudes and instantaneous phases, which serve as inputs to subsequent data processing. The two most popular approaches for obtaining these narrowband amplitudes and phases are: bandpass filtering followed by Hilbert transform (we call this the Hilbert approach); and convolution with wavelet kernels (the wavelet approach). In this work, we investigate how these two approaches perform in detecting the phenomenon of phase-amplitude coupling (PAC), whereby the amplitude of a high-frequency signal is driven by the phase of a lowfrequency signal. The comparison of both approaches is carried out by means of simulated brain activity, from which we run receiver operating characteristic (ROC) analyses, and of experimental MEG data from a visuomotor coordination study. The ROC analyses show that both approaches have comparable accuracy, except in the presence of interfering signals with frequencies near the high-frequency band. As for the visuomotor data, the most noticeable impact of the choice of approach was observed when evaluating task-based changes in PAC between the delta (2-5 Hz) and the high-gamma (60-90 Hz) frequency bands, as we were able to identify widespread brain areas with statistically significant effects only with the Hilbert approach. These results provide preliminary evidence of the advantages of the Hilbert approach over the wavelet approach, at least in the context of PAC estimates.

Tensorpac : an open-source Python toolbox for tensor-based Phase-Amplitude Coupling measurement in electrophysiological brain signals

“…Among existing methods, [35] introduced a time lag to the amplitude, while [44] swap amplitude and phase trials and [57] swap time blocks, cut at a random point. The latter method, with only two blocks, has been described as the most conservative strategy to generate the distribution of PAC that can be observed by chance [47]. Finally, this null distribution is then used to perform non-parametric inference or to correct the measure estimated from the data (usually by subtracting the mean and divide by the deviation of this distribution).…”

“…One issue that has been raised is that, since PAC is computed across time, non-stationary signals can cause the appearance of spurious coupling [47]. An illustrative example taken from the same study explains that if there is an induced phase locking of lower frequencies and simultaneously a high frequency power increase, a coupling between them is going to be observed.…”

“…Nevertheless, there is still no consensus on the minimal data length (i.e. the number of cycles) that is required or the most appropriate filtering methods [46,47]. It has also recently been shown that spurious PAC can occur for a variety of reasons that may be difficult to systematically control [48][49][50][51].…”

Tensorpac: An open-source Python toolbox for tensor-based phase-amplitude coupling measurement in electrophysiological brain signals

“…Nevertheless, there is still no consensus on the minimal data length (i.e. the number of cycles) that is required or the most appropriate filtering methods [ 48 , 49 ]. It has also recently been shown that spurious PAC can occur for a variety of reasons that may be difficult to systematically control [ 48 , 50 – 52 ].…”