In simultaneous EEG-correlated fMRI of epilepsy patients, the prediction of the epilepsy-related BOLD response may be compromised by the presence of signal fluctuations of non-neuronal origin, commonly referred to as physiological noise. Several methods have already been reported to model and remove those fluctuations, usually relying on the simultaneous recording of cardiac and respiratory signals. Here, we propose to use the ECG-derived respiratory (EDR) signal estimated by Empirical Mode Decomposition (EMD) as a surrogate of the respiratory signal, in typical EEG-fMRI studies where respiratory data is unavailable but ECG is recorded. We evaluate the impact of physiological noise modeling and correction using the proposed approach, on EEG-correlated fMRI mapping of epileptic networks.Data: Six epilepsy patients with drug-refractory focal epilepsy were studied on a 3T Siemens Verio MRI system using an MR-compatible 32-channel EEG system including one ECG channel (Brain Products). Resting-state fMRI data was obtained using whole-brain 2D-EPI (TR/TE=2500ms/50ms) with 3.5×3.5×3.0 mm³ resolution. EEG data were subjected to MR-induced artefact correction and band-pass filtering (1-45 Hz); for ECG a 4-45 Hz band-pass filter was used.

EDR Estimation: The EDR signal was estimated by EMD, which retrieves Intrinsic Mode Functions (IMF) from the data¹. Since respiration is known to modulate the ECG, it is expected that one of the IMFs reflects the respiratory waveform. Therefore, the EDR signal was defined as the IMF whose FFT yielded the highest ratio between the power within (0.2–0.4Hz) and outside (0–5Hz) the respiratory frequency band.

Pre-processing: BOLD-fMRI data was first brain-extracted and cardiac- and respiratory-related periodic fluctuations were regressed out by slice-wise RETROICOR, followed by motion and slice-timing correction. The associated terms were successively added up to the 4^th order, if the variance explained (VE) was statistically significant different from 0 (p<0.05)². Following the same rationale, the following sets of explanatory variables were subsequently added: 1) respiratory volume (RV) and heart rate (HR) convolved with the respiration (RRF) and cardiac response functions (CRF), respectively^3,4, and subsequently shifted by a time lag optimized for each subject; 2) average CSF and WM BOLD time courses; and 3) estimated motion parameters. Finally, high-pass temporal filtering (cut-off=100s) and a Gaussian spatial smoothing (FWHM=8mm) were applied. For comparison, BOLD-fMRI data was also subjected to standard pre-processing (motion and slice-timing correction, high-pass temporal filtering and spatial smoothing), without physiological noise correction.

Epileptic network mapping: Epileptic activity was recorded in 3/6 patients, which were used to assess the impact of the proposed physiological noise correction methodology. After pre-processing, the EEG were analysed as follows: ICA decomposition, selection of an epilepsy-related IC⁵, extraction of the root mean square frequency (RMSF) by Morlet Wavelet time-frequency decomposition⁶, convolution with the canonical haemodynamic response function, and downsampling to the fMRI sampling rate. EEG-correlated fMRI analysis was then performed by fitting a general linear model to the BOLD data using the EEG-IC-RMSF as regressor, and cluster thresholding the resulting Z-score maps using voxel Z>2.3 and cluster p<0.05. The sensitivity was quantified as the number of grey matter voxels (N_GM) in the clusters, and the specificity as the average Pearson correlation between the time-course of the voxel with highest Z-score and all other grey matter voxels in the cluster (r_GM).

An example of the estimated respiratory signal is shown in Fig. 1, exhibiting the typical frequency spectrum. The group average VE values by the different terms of the physiological noise model are shown in Fig. 2. Although significant contributions were found for all orders of the cardiac RETROICOR term, as expected, this was only observed for the second order of the respiratory term. This can be explained by the fact that the extraction of the respiratory phase takes into account the amplitude of the respiratory signal, but EMD only guarantees an accurate estimation of the respiratory rate, not its amplitude. In contrast, the optimal RV model significantly contributed to the physiological noise model (0.9±0.4%), as well as all the remaining model terms.

In Fig. 3, the epileptic networks obtained by EEG-correlated fMRI analysis for three representative datasets are shown, with and without physiological noise correction. Clear improvements can be observed, in terms of increased specificity (Fig. 3A), as well as sensitivity (Fig. 3B). Ultimately, Fig. 3C illustrates a case where the epileptic activity was successfully mapped only upon physiological noise removal.

We show that it is possible to model both cardiac- and respiratory-related BOLD fluctuations resorting to the ECG signal exclusively, with significant impact in the sensitivity and specificity of EEG-correlated fMRI mapping of epileptic networks.