Bayesian multi-modal model comparison: a case study on the generators of the spike and the wave in generalized spike-wave complexes.

Abstract

We present a novel approach to assess the networks involved in the generation of spontaneous pathological brain activity based on multi-modal imaging data. We propose to use probabilistic fMRI-constrained EEG source reconstruction as a complement to EEG-correlated fMRI analysis to disambiguate between networks that co-occur at the fMRI time resolution. The method is based on Bayesian model comparison, where the different models correspond to different combinations of fMRI-activated (or deactivated) cortical clusters. By computing the model evidence (or marginal likelihood) of each and every candidate source space partition, we can infer the most probable set of fMRI regions that has generated a given EEG scalp data window. We illustrate the method using EEG-correlated fMRI data acquired in a patient with ictal generalized spike-wave (GSW) discharges, to examine whether different networks are involved in the generation of the spike and the wave components, respectively. To this effect, we compared a family of 128 EEG source models, based on the combinations of seven regions haemodynamically involved (deactivated) during a prolonged ictal GSW discharge, namely: bilateral precuneus, bilateral medial frontal gyrus, bilateral middle temporal gyrus, and right cuneus. Bayesian model comparison has revealed the most likely model associated with the spike component to consist of a prefrontal region and bilateral temporal-parietal regions and the most likely model associated with the wave component to comprise the same temporal-parietal regions only. The result supports the hypothesis of different neurophysiological mechanisms underlying the generation of the spike versus wave components of GSW discharges.

Full text

Page 1

Bayesian multi-modal model comparison: a
case study on the generators of the spike and
the wave in Generalised Spike-Wave complexes.


Jean Daunizeau1, Anna E. Vaudano2,3 and Louis Lemieux2

1 Wellcome Trust Centre for Neuroimaging, University College of London, United Kingdom
2 Department of Neurology, Policlinico Umberto I°, University of Rome “La Sapienza”, Viale
dell’Università’ 30, 00185 Rome, Italy;

3Department of Clinical and Experimental Epilepsy, UCL Institute of Neurology, Queen
Square, WC1N 3BG, London, UK and MRI Unit, National Society for Epilepsy, Chesham
Lane Chalfont St. Peter SL9 0RJ, Buckinghamshire UK;









Address for Correspondence
Jean Daunizeau
Wellcome Trust Centre for Neuroimaging,
Institute of Neurology, UCL
12 Queen Square, London, UK WC1N 3BG
Tel (44) 207 833 7488
Fax (44) 207 813 1445
Email j.daunizeau@fil.ion.ucl.ac.uk

Key words: generalized spike and wave discharges, absence seizure,
neuroimaging, EEG, fMRI, data fusion, source reconstruction, inverse problem.
* 3. Manuscript
Click here to view linked References

Page 2

Introduction

During the past decade, EEG-correlated fMRI has been used to map
haemodynamic changes correlated with the occurrence of epileptiform
discharges in focal and generalised epilepsies (Salek-Haddadi et al., 2003;
Salek-Haddadi et al., 2006; Laufs 2007; Gotman et al., 2006). In focal
epilepsy, the localising information thus obtained has been found to be
broadly concordant with the location of the sources inferred from other
electro-clinical data and EEG source reconstruction (Lemieux et al., 2001;
Benar et al., 2006). In contrast, generalised epileptiform discharges, such as
generalised spike-wave (GSW) complexes seen on the EEG traces, are
commonly associated with widespread haemodynamic changes in the
neocortex and sub-cortical regions (Laufs et al., 2006; Hamandi et al., 2006).
Given the temporal resolution of fMRI, of the order of a few seconds, one
must assume that these maps are representative of haemodynamic changes
taking place over entire discharges. As a consequence, they are not
informative with respect to the electrophysiological processes that underlie the
different components of such GSW1 complexes. On the other hand, the
problem of localizing these processes from the EEG remains a difficult
challenge due to the widespread nature of the pattern which is thought to
reflect rapidly propagating neural activity over a large part of the cortex. This
is in addition to a fundamental difficulty of the EEG inverse problem namely
that the underlying current sources cannot be estimated uniquely from EEG
scalp measurements without invoking priors or constraints.
EEG/MEG inverse solutions differ in the nature of their priors, which should
ideally be specific to the neuronal phenomenon at hand. For instance,
assuming that the underlying active network consists of a few focal sources
has been used to justify equivalent current dipoles (ECD) localisation methods
(Sherg and Ebersol, 1993). ECD solutions applied to the analysis of focal
epileptic spikes have proven relatively concordant with both fMRI statistical
maps (see e.g. Korjenova 2001) and intracranial electrode recordings (Merlet
and Gotman, 1999). However, this class of inverse solutions can be

1 Within GSW discharges, the spike is thought to reflect neural excitation and the
wave the inhibition (Niedermeyer and Lopes da Silva, chap 13).

Page 3

misleading in presence of spatially extended sources (see e.g. Kobayashi et
al., 2005). In contradistinction, distributed linear (DL) methods aim at
estimating the amplitude of a predefined highly dense ensemble of dipoles,
typically spread over the cortical sheet (Dale and Sereno, 1993). Usually,
additional spatial and/or temporal constraints are used to finesse the under-
determination of DL inverse solutions (see e.g. Daunizeau et al., 2007b). In
this context, fMRI-derived spatial priors are thought to significantly improve
the spatial resolution of DL inverse solutions, particularly if the inverse
technique is able to account for the potential mismatch between EEG and
fMRI sources (see e.g. Daunizeau et al., 2008 for a comprehensive review of
EEG/fMRI information fusion).
Recently, such probabilistic DL source reconstruction methods using spatial
priors derived from fMRI statistical parametric maps (SPMs) have been
developed (see e.g. Daunizeau et al., 2005; Sato et al., 2005; Daunizeau et
al., 2007a). Most of these techniques fall into a Bayesian framework, which
provides an estimation of the current sources (posterior probability maps or
PPMs, see e.g. Friston et al., 2007) and allows for generic model comparison,
through the computation of the model evidence/marginal likelihood (see e.g.
Trujillo-Barreto et al., 2005 or Mattout et al., 2006).
Source reconstruction using DL and ECD models from high-density EEG have
already been applied to GSW, suggesting a focal frontal origin for the spike
and broader frontal generator for the wave (Holmes et al., 2004; Tucker et al.,
2007), broadly in line with earlier work (Lemieux and Blume 1983;
Niedermeyer and Lopes da Silva, 2004). This encouraging preliminary result,
along with previous work on both EEG-correlated fMRI data analysis and
fMRI-constrained EEG source reconstruction, has led us to develop a
principled probabilistic technique dedicated to identifying the respective
networks involved in the respective generation of the spike and wave
components of GSW complexes from EEG and fMRI data.
In Daunizeau et al., 2005, we have proposed a Bayesian model comparison
scheme for assessing the relevance of fMRI-derived spatial priors in
probabilistic EEG source reconstruction. In Grova et al., 2008, we have
applied this method in the context of focal interictal spike localization, using
EEG-correlated fMRI statistical parametric maps (SPMs).

Page 4

In this work, we systematize and extend this method to cope with multi-
regions EEG-correlated fMRI SPMs, by means of the Expectation-
Maximization (EM) algorithm (see e.g. Friston et al., 2008). The method is
designed to take advantage of the spatial resolution of fMRI and the temporal
resolution of EEG, and consists of the three following steps:
(i) Deriving the static EEG-correlated fMRI SPM. Due to the low
temporal resolution of fMRI compared to EEG, the resulting map of
activations can be taken to reflect multiple aspects of the EEG
events of interest such as the generators of the spike and the wave
for the specific case of GSW.
(ii) Building the cortical source space partitions that are associated with
every combination of the activated regions.
(iii) Using Bayesian model comparison to identify the most likely source
space partition with respect to the spike and the wave component of
the EEG scalp measurement of GSW complexes, respectively.
We demonstrate the potential of this method in an analysis of multi-modal
EEG-fMRI data acquired in a patient affected by idiopathic generalized
epilepsy (IGE) with frequent absence seizures.

Page 5

Methods

Patient clinical history
We studied a right-handed 23 year-old man (written informed consent and
ethics committee approval obtained) affected by juvenile absence epilepsy
(JAE; onset age 12y) with frequent absence seizures (2-3 episodes per week)
and rare (fewer that one per year) generalized tonic clonic seizures. He was
born by a caesarean section three weeks preterm; he was well at birth and his
developmental milestones were within normal limits. He has no history of
febrile convulsions, brain injury or other risk factors for the development of
epilepsy. His father’s mother was diagnosed with temporal lobe epilepsy at
the age of 58.
Neurological examination and morphological MRI scans (at 1.5T and 3T) were
normal.
Previous EEG recordings showed a normal background interrupted by 2.5-3
per second generalized spike-and-wave activity with anterior predominance,
facilitated by hyperventilation. There was no response to photic stimulation.
The patient was treated with AED, in mono-therapy or association
(Ethosuximide, Lamotrigine, Levetiracetam and Valproate) without achieving
complete control of the absences.
At the time of our investigations, the patient was taking Levetiracetam
2500mg/day and Valproate 1000mg/day. The EEG showed frequent
spontaneous and hyperventilation-related generalized spike-and-wave (GSW)
discharges lasting between <1 and 20 s; clinically, the longest discharges
(more than 15 seconds) were accompanied by psychomotor arrest and eyelid
blinking.

Simultaneous EEG-fMRI acquisition
The head was immobilized using a vacuum cushion. Thirty-two channels of
EEG were recorded using the MR-compatible BrainCap electrode cap and
recording system (Brainproducts, Munich, Germany; cap: Falk Minow
Services, Herrsching-Breitbrunn, Germany), along with bipolar
electrocardiogram and MR scanner synchronisation signal (Krakow et al.,
2000). Four hundred and four T2*-weighted single-shot gradient-echo