Do we need spikes?

Ideally, patients who undergo an EEG-fMRI study show clearly identifiable and classifiable interictal epileptic activity occurring with a temporal spacing that will give optimal statistical power in an event-related general linear model. In reality, spiking patterns are different. In 63 patients with focal epilepsy, Afraim Salek-Haddadi found a tendency for more spikes, less motion, more runs of IEDs and less EEG background abnormality in cases showing any IED-correlated BOLD signal changes compared to those not showing any (Salek-Haddadi, Diehl et al. 2006). In his series, 42% of the patients analysed and which were pre-selected based on spiking on a recent routine EEG did not have any interictal activity in 35 min. In our series of patients with focal cortical dysplasia (FCD), only about half of the subjects considered for epilepsy surgery exhibited interictal epileptic discharges during two or three twenty minute sessions of EEG-fMRI acquisition (Thornton, Vulliemoz et al. 2011) (see Error: Reference source not found). This raises the question whether “spike-less” EEG-fMRI data can be

analysed such that useful clinical information can be obtained. One solution can be to build a GLM based on information extracted from the ongoing “background” EEG. A proof of

principle case report was published in MRI (Laufs, Hamandi et al. 2006) (see 7.2.1): We studied a patient with refractory focal epilepsy using continuous EEG-correlated fMRI.

Seizures were characterized by head turning to the left and clonic jerking of the left arm, suggesting a right frontal epileptogenic region. Interictal EEG showed occasional runs of independent nonlateralized slow activity in the delta band with right frontocentral

dominance and had no lateralizing value. Ictal scalp EEG had no lateralizing value. Ictal scalp EEG suggested right-sided central slow activity preceding some seizures. Structural 3-T MRI showed no abnormality. There was no clear epileptiform abnormality during simultaneous EEG–fMRI. I modelled asymmetrical EEG delta activity at 1–3 Hz near frontocentral electrode positions. Significant blood oxygen level-dependent (BOLD) signal changes in the right superior frontal gyrus correlated with right frontal oscillations at 1–3 Hz but not at 4–7 Hz and with neither of the two frequency bands when derived from contralateral or posterior electrode positions, which served as controls. Motor fMRI activations with a finger-tapping paradigm were asymmetrical: they were more anterior for the left hand compared with the right and were near the aforementioned EEG-correlated signal changes. A right frontocentral perirolandic seizure onset was identified with a subdural grid recording, and electric

stimulation of the adjacent contact produced motor responses in the left arm and after discharges. The fMRI localization of the left hand motor and the detected BOLD activation associated with modeled slow activity suggest a role for localization of the epileptogenic region with EEG–fMRI even in the absence of clear interictal discharges.

An alternative approach to “spike-less”, or data with too few events are data driven,

hypothesis-free methods. Being based on the fMRI data with homogeneous resolution across the entire brain, these methods might be sensitive to epileptic activity in brain regions to which scalp EEG is insensitive, e.g. subcortical and other surface-distant areas. After

evaluation of a temporal clustering method showing false positive results most likely due to subject motion (Hamandi, Salek Haddadi et al. 2005), in our group, we decided not to pursue the approach originally suggested by (Morgan, Price et al. 2004). The clustering method is based on the temporal profile of BOLD signal intensity values which are thresholded based on observer-defined criteria defined yielding “activation” maps. An alternative method is

independent component analysis, which separates a multivariate signal into additive subcomponents supposing the mutual statistical independence of the source signals. The main advantage of this method is that it represents the original functional time series as a set of independent components (IC), which may separate meaningful neurophysiological sources and artefacts. However, the lack of a prior hypothesis and the potentially large number of IC generated render interpretation of the results difficult (Beckmann and Smith 2004). An automated characterization technique has been introduced and implemented to reduce the number of meaningful IC requiring interpretation (De Martino, Gentile et al.

2007). In this method, the classification of patterns as BOLD-like relies on a set of spatial and temporal characteristics derived from data acquired in normal healthy subjects. Given our findings of mainly canonical haemodynamic responses also in the context of epilepsy, we consider this classifier suitable for our data. For evaluation of the technique, we applied ICA to EEG-fMRI data sets of patients in whom interictal spikes were present on the EEG and GLM-based BOLD signal localized the epileptic foci to regions also identified with other techniques (EEG, structural MRI, or surgical outcome) (Rodionov, De Martino et al. 2007).

Concordance between the ICA and GLM-derived results was assessed based on spatio-temporal criteria. In each patient, one of the IC satisfied corresponded to the IED-based GLM result. The remaining IC were consistent with BOLD patterns of spontaneous brain activity and may include epileptic activity that was not evident on the scalp EEG. We found that ICA of fMRI is capable of revealing areas of epileptic activity in patients with focal epilepsy and may be useful for the analysis of EEG–fMRI data in which abnormalities are not apparent on scalp EEG.

Another situation when surface EEG may not sufficiently reflect epileptic activity is during seizures. We evaluated the technique in this context and compared it to a GLM-based approach and the gold standard of intracranial EEG recordings (Thornton, Rodionov et al.

2010) (see Error: Reference source not found). Nine of 83 patients with focal epilepsy undergoing pre-surgical evaluation had seizures during EEG-fMRI and were analysed using three approaches, two based on the general linear model (GLM) and one using independent component analysis (ICA). The canonical GLM analysis revealed significant BOLD signal changes associated with seizures on EEG in 7/9 patients, concordant with the seizure onset zone in 4/7. The GLM analysis revealed changes in BOLD signal corresponding with the results of the canonical analysis in two patients. ICA revealed components spatially

concordant with the seizure onset zone in all patients (8/9 confirmed by intracranial EEG). In conclusion, ictal EEG-fMRI visualises plausible seizure related haemodynamic changes. The GLM approach to analysing EEG-fMRI data reveals localised BOLD signal changes concordant with the ictal onset zone when scalp EEG reflects seizure onset. ICA provides additional information when scalp EEG does not accurately reflect seizures and may give insight into ictal haemodynamics.

If interictal epileptic discharges are present on the scalp EEG, their accurate classification, or grouping, into distinct and reproducible event classes is vital for the successful creation of regressors in a GLM-based analysis. If paroxysms of different spatial origins are erroneously modelled as a single event type, this will be detrimental to the sensitivity of the analysis. We assessed whether an automated spike classification procedure can increase the model’s sensitivity and additionally, automated detection procedures can serve to increase IED identification on the EEG in the first place (Liston, De Munck et al. 2006): For patients with multiple IED generators, sensitivity to IED-correlated BOLD signal changes can be improved when the fMRI analysis model distinguishes between IEDs of differing morphology and field.

In an attempt to reduce the subjectivity of visual IED classification, we implemented a semi-automated system, based on the spatio-temporal clustering of EEG events. We illustrated the technique's usefulness using EEG-fMRI data from a subject with focal epilepsy in whom 202 IEDs were visually identified and then clustered semi-automatically into four clusters. Each cluster of IEDs was modelled separately for the purpose of fMRI analysis. This revealed IED-correlated BOLD activations in distinct regions corresponding to three different IED

categories. In a second step, Signal Space Projection (SSP) was used to project the scalp EEG onto the dipoles corresponding to each IED cluster. This resulted in 123 previously

unrecognised IEDs, the inclusion of which into the GLM increased the experimental efficiency as reflected by significant BOLD activations. We also showed that the detection of extra IEDs is robust in the face of fluctuations in the set of visually detected IEDs. We concluded that automated IED classification can result in more objective fMRI models of IEDs and

significantly increased sensitivity.

7 Results of original work

7.1 Mapping of ongoing physiological EEG information identifies different brain states in