Schrödinger filtering: a precise EEG despiking technique for EEG-fMRI gradient artifact

Gabriel B Benigno,Ravi S Menon,Hacene Serrai

doi:10.1016/j.neuroimage.2020.117525

Abstract

In EEG data acquired in the presence of fMRI, gradient-related spike artifacts contaminate the signal following the common preprocessing step of average artifact subtraction. Spike artifacts compromise EEG data quality since they overlap with the EEG signal in frequency, thereby confounding frequency-based inferences on activity. As well, spike artifacts can inflate or deflate correlations among time series, thereby confounding inferences on functional connectivity.We present Schrödinger filtering, which uses the Schrödinger equation to decompose the spike-containing input. The basis functions of the decomposition are localized and pulse-shaped, and selectively capture the various input peaks, with the spike components clustered at the beginning of the spectrum. Schrödinger filtering automatically subtracts the spike components from the data.On real and simulated data, we show that Schrödinger filtering (1) simultaneously accomplishes high spike removal and high signal preservation without affecting evoked activity, and (2) reduces spurious pairwise correlations in spontaneous activity. In these regards, Schrödinger filtering was significantly better than three other despiking techniques: median filtering, amplitude thresholding, and wavelet denoising.These results encourage the use of Schrödinger filtering in future EEG-fMRI pipelines, as well as in other spike-related applications (e.g., fMRI motion artifact removal or action potential extraction).

Highlights

Simultaneous acquisition of electroencephalography and functional magnetic resonance imaging (EEG-fMRI) has been popular for over two decades (Abreu et al, 2018; Gotman et al, 2006; Huang-Hellinger et al, 1995; Ives et al, 2001; Lemieux et al, 2001) due to each modality’s complementary strengths
The success of Schrödinger filtering with high-spike artifact datasets like the FMRIB dataset and the corresponding simulated datasets encourages the use of EEG-fMRI in scenarios where EEG-MRI clock synchronization or low subject motion is not possible, and enables a greater volume of retained EEG-fMRI data in general thanks to Schrödinger filtering’s spike-cleaning capabilities
Since despiking by Schrödinger filtering robustly removes spike artifacts while preserving signal, and since spike artifacts overlap in frequency with the EEG bands, Schrödinger filtering relaxes the need to aggressively low-pass filter the data, which is presently done to remove artifacts present in the beta and gamma bands

Summary

Introduction

Simultaneous acquisition of electroencephalography and functional magnetic resonance imaging (EEG-fMRI) has been popular for over two decades (Abreu et al, 2018; Gotman et al, 2006; Huang-Hellinger et al, 1995; Ives et al, 2001; Lemieux et al, 2001) due to each modality’s complementary strengths. EEG, using scalp electrodes, measures changes in electric potential associated with ion fluxes from action potentials and postsynaptic potentials at the cortical surface (Beres, 2017). As the electric potential passes through the head’s tissues before reaching the scalp electrodes, it gets blurred in both space and time to the extent that spatial resolution is poor while temporal resolution is high enough to study the classical frequency bands ranging up to the order of 100 Hz (Burle et al, 2015; Niedermeyer, 2005). In fMRI, a time series of images is acquired, with the voxel size ranging from microns to centimeters (Glover, 2011). The hemodynamic response is on the order of 1 s, giving fMRI poor temporal resolution regardless of how fast each image is acquired (Glover, 2011). The relatively high temporal resolution and low spatial resolution of EEG complements the relatively high spatial resolution and low temporal resolution of fMRI, motivating the use of EEG-fMRI

Methods

Results

Conclusion