Identification of epileptic brain states by dynamic functional connectivity analysis of simultaneous EEG-fMRI: a dictionary learning approach

Abstract

Most fMRI studies of the brain's intrinsic functional connectivity (FC) have assumed that this is static; however, it is now clear that it changes over time. This is particularly relevant in epilepsy, which is characterized by a continuous interchange between epileptic and normal brain states associated with the occurrence of epileptic activity. Interestingly, recurrent states of dynamic FC (dFC) have been found in fMRI data using unsupervised learning techniques, assuming either their sparse or non-sparse combination. Here, we propose an l₁-norm regularized dictionary learning (l₁-DL) approach for dFC state estimation, which allows an intermediate and flexible degree of sparsity in time, and demonstrate its application in the identification of epilepsy-related dFC states using simultaneous EEG-fMRI data. With this l₁-DL approach, we aim to accommodate a potentially varying degree of sparsity upon the interchange between epileptic and non-epileptic dFC states. The simultaneous recording of the EEG is used to extract time courses representative of epileptic activity, which are incorporated into the fMRI dFC state analysis to inform the selection of epilepsy-related dFC states. We found that the proposed l₁-DL method performed best at identifying epilepsy-related dFC states, when compared with two alternative methods of extreme sparsity (k-means clustering, maximum; and principal component analysis, minimum), as well as an l₀-norm regularization framework (l₀-DL), with a fixed amount of temporal sparsity. We further showed that epilepsy-related dFC states provide novel insights into the dynamics of epileptic networks, which go beyond the information provided by more conventional EEG-correlated fMRI analysis, and which were concordant with the clinical profile of each patient. In addition to its application in epilepsy, our study provides a new dFC state identification method of potential relevance for studying brain functional connectivity dynamics in general.

Figure 1

Schematic diagram of the processing pipeline (example from patient P3). After the estimation of dFC using a sliding-window Pearson correlation approach: (A) dFC states are identified using DL (A₁), k-means clustering (A₂) and PCA (A₃), ranging the number of states k to be estimated, and the regularization parameters λ in DL. The dFC states estimated using each method for k = 4 are shown. (B) The respective non-sparse weight time-courses (dashed gray or black traces) are correlated with the EEG-PSI metric (red trace, representative of the epileptic activity), and the state yielding the highest correlation ρ is identified. The optimal k, and λ for DL, are determined for each patient and method as those that maximize ρ.

Figure 2

Results of the dFC analyses for patient P1 when using the l₁-DL_PCA method. (A) Bar plot depicting the contribution (dictionary weight) of the various dFC states (color-coded as in B) over time windows, superimposed with the EEG-PSI metric time course (black trace); the black dashed line represents 90–95% of the EEG-PSI maximum value. All values are normalized between 0 and 1, for visualization purposes. (B) Estimated dFC states, ordered by statistical significance and correlation between the respective non-sparse weight time-courses and EEG-PSI (titles are color-coded as in (A)). Statistically significant dFC states are highlighted by a solid black square; epilepsy-related dFC states are highlighted by a dashed black square (and indicated with * in the title). All matrices are normalized between −1 and 1. (C) Epileptic network obtained by standard EEG-correlated fMRI analysis: Z score map of BOLD changes significantly associated with the EEG-PSI, superimposed on the patient’s structural image, shown for four illustrative axial slices. (D) Epilepsy-related dFC states with connectivity strengths averaged across the 14 groups of AAL regions (left and right frontal, limbic, occipital, parietal and temporal lobes, thalamus and other subcortical areas). For each group present in the epileptic network shown in (C) (vertical axis), the average connectivity strength with all other groups is shown (horizontal axis).

Figure 3

Results of the dFC analyses for patient P2 when using the l₁-DL_PCA method.

Figure 4

Results of the dFC analyses for patient P3 when using the l₁-DL_PCA method.

Figure 5

Results of the dFC analyses for patient P4 when using the l₁-DL_PCA method.

Figure 6

Results of the dFC analyses for patient P5 when using the l₁-DL_PCA method.

Figure 7

Results of the dFC analyses for patient P6 when using the l₁-DL_PCA method.

Figure 8

Results of the dFC analyses for patient P7 when using the l₁-DL_PCA method.

Figure 9

Results of the dFC analyses for patient P8 when using the l₁-DL_PCA method.