Primary or idiopathic generalized epilepsy (IGE) accounts for 15–30% of all epilepsy cases (Jallon & Latour, 2005; Szaflarski et al., 2010b). Although much progress has been made in discovering the etiology of IGE, the neuronal causes of treatment resistance remain poorly understood. The treatment armamentarium against IGE is mostly pharmacologic, and a combination of one or more drugs will control seizures in >80% of patients (Faught, 2004). Valproic acid (valproate, VPA) in particular is predictive of drug-resistance, with <10% of patients who fail an adequate dose of VPA achieving seizure control with any combination of other drugs (French et al., 2004; Holland et al., 2010; Szaflarski et al., 2010b). Drug-resistant patients may undergo vagus nerve stimulation (Labar et al., 1999; Ng & Devinsky, 2004; Kostov et al., 2007) or a ketogenic diet (Groomes et al., 2011), but these therapies are better at reducing seizure frequency than they are at achieving seizure control. Surgery is contraindicated in IGE due to presumed genetic etiology and generalized seizure onset (Engel, 1993; Upchurch & Stern, 2006). Therefore, intractable IGE is characterized by uncontrolled seizures despite treatment with syndrome-appropriate antiepileptic drugs (AEDs).
The generalized spike and wave discharge (GSWD) is a hallmark of IGE, and its observation with electroencephalography (EEG) during “staring spells” is pathognomonic for the absence seizure (Yenjun et al., 2003). GSWDs are thought to arise, in part, from abnormal neuronal circuits (Contreras et al., 1996; Moeller et al., 2008a). Magnetic resonance imaging (MRI), especially diffusion tensor imaging (DTI) of white matter tracts, allows for noninvasive investigation of these circuits in humans. However, because the brains of patients with IGE are grossly normal (Opeskin et al., 2000), structural MRI may not be sufficiently sensitive to detect abnormal connectivity. Using functional MRI (fMRI) to study intrinsic, or resting-state brain connectivity provides a measure of these circuits consistent with but independent of measures of structural connectivity (Greicius et al., 2009). There is evidence that this approach is more sensitive than DTI at detecting abnormal connectivity in IGE (Zhang et al., 2011). Previous studies of IGE and its subpopulations have found altered resting-state functional connectivity in the default mode network (DMN) (Luo et al., 2011; McGill et al., 2012), dorsal attention network (Killory et al., 2011), and between hemispheres (Bai et al., 2011).
Simultaneous EEG/fMRI allows for temporal alignment of the fMRI hemodynamic response with electrographic features such as GSWDs detected on EEG (Gotman et al., 2006). These multimodal data are useful in the investigation of functional connectivity because analysis of the fMRI signal can be restricted to periods of normal EEG, preventing contamination of the resting-state by GSWD. The opposite is also possible, and several EEG/fMRI studies have reported on fMRI activation and deactivation associated with GSWDs in IGE (Aghakhani et al., 2004; Gotman et al., 2005; Laufs et al., 2006; Moeller et al., 2008b; Tyvaert et al., 2009; Szaflarski et al., 2010a). Deactivation occurs specifically in brain regions that coincide with the DMN (Aghakhani et al., 2004; Gotman et al., 2005; Laufs et al., 2006; Moeller et al., 2008b), a resting-state network thought to support consciousness (Raichle et al., 2001; Fox et al., 2005). It is probably not a coincidence that deactivation of the DMN during GSWD coincides with transient loss of consciousness during absence seizures (Danielson et al., 2011).
Two recent resting-state studies have found that DMN connectivity is lower in patients with IGE than in healthy controls; furthermore, in these studies, connectivity was negatively correlated with disease duration (Luo et al., 2011; McGill et al., 2012). Although highly relevant, these existing DMN studies were limited with respect to equipment and experimental population. One study with 15 IGE patients was fMRI-only and thus could not exclude or control for the effects of GSWDs on the resting-state (McGill et al., 2012). A separate 12-patient EEG/fMRI study restricted analysis to periods of normal EEG, but all patients in this study had uncontrolled seizures at the time of scanning, preventing the authors from evaluating the effects of disease severity on DMN connectivity (Luo et al., 2011). Both studies used seed-based voxel correlation to assess DMN connectivity, and we are not aware of any studies that have used independent component analysis (ICA) for this purpose. ICA and seed-based voxel correlation are both tools for detecting functional connectivity, but ICA offers superior separation of structured noise from resting-state fluctuations (Damoiseaux et al., 2006; Ma et al., 2007). Therefore, in this resting-state EEG/fMRI study we investigated DMN connectivity in a heterogenous population of 60 IGE patients and 38 healthy controls using ICA and performed confirmatory analysis using seed-based voxel correlation. Furthermore, we examined the effect of VPA resistance and uncontrolled seizures on DMN connectivity. We hypothesized that treatment-resistant IGE has a neuronal basis and is thus linked to an observable decrease in DMN connectivity, a putative neuronal marker for treatment-resistant IGE, when compared to treatment-responsive IGE patients and/or healthy controls.
- Top of page
- Supporting Information
All patients with IGE evaluated in the Cincinnati Epilepsy Center were approached for participation and enrolled prospectively irrespective of their seizure control, medication status, or demographic characteristics. Enrolled subjects fulfilled all criteria for the diagnosis of IGE based on published criteria (Commission on Classification & Terminology of the International League Against Epilepsy, 1989). As in our previous studies, diagnosis and treatment were directed by an epilepsy specialist (Szaflarski et al., 2008). Patients with a history of absence seizures received 24-h ambulatory EEG to document seizure-freedom. Response to VPA was defined as seizure freedom lasting at least 3 months during treatment with VPA with at least one VPA level documented as therapeutic. Patients who did not achieve seizure control for 3 months while receiving VPA with consistently therapeutic VPA levels were considered VPA-resistant. Patients who were not treated with VPA or whose treatment with VPA lasted <3 months due to intolerable side effects were categorized as VPA-unknown. Patients who had one or more seizures in the 3 months leading up to scanning were categorized as having uncontrolled seizures independent of their VPA status. Most (n = 9) patients with uncontrolled seizures also satisfied our criteria for VPA resistance, but some did not because they had never tried VPA (n = 3) or because they were VPA-responsive but were not being treated in the months leading up to scanning (n = 4); see Table 1.
Table 1. Subject demographics
| ||Male||Female||Age (years ± SD)||Duration (years ± SD)||No. current drugs||No. failed drugs|
|Epilepsy (All)||25||35||31.5 ± 11.7||15.5 ± 12.0||1.47||1.95|
|Epilepsy (Seizures−)||17||27||32.2 ± 11.3||15.1 ± 11.0||1.25||1.45|
|Epilepsy (Seizures+)||8||8||29.4 ± 13.1||16.5 ± 14.6||2.06||3.40|
|Epilepsy (VPA+)||14||14||27.5 ± 6.7||13.2 ± 7.3||1.18||1.54|
|Epilepsy (VPA−)||7||6||35.6 ± 15.7||23.7 ± 17.7||2.31||3.67|
|Epilepsy (VPA?)||4||15||34.4 ± 13.2||13.2 ± 10.9||1.32||1.47|
|Epilepsy (JME)||14||16||28.1 ± 9.2||13.5 ± 9.4||1.40||2.07|
|Epilepsy (other IGE)||11||19||34.8 ± 13.2||17.4 ± 13.9||1.53||1.83|
All of the 89 epilepsy patients and 40 healthy controls participated in the study after providing written informed consent for a protocol approved by the institutional review board of the University of Cincinnati. Each epilepsy patient underwent 1–3 consecutive 20-min EEG/fMRI scans, and each healthy control subject underwent 1–2 consecutive scans. All patients and 20 control participants listened to self-selected music during scanning to increase comfort and compliance. The effect of music-listening on resting-state data is discussed elsewhere (Kay et al., 2012) and was included as a covariate in analysis. Eleven epilepsy patients did not complete the scanning procedure due to claustrophobia (n = 3), metallic artifacts (n = 1), or not wanting to continue the procedure (n = 7). After excluding 25 low-quality scans (23 epilepsy, two control) and 37 epilepsy scans with abnormal EEG (i.e., GSWD), resting-state analysis was carried out on 189 scans from 60 patients with idiopathic generalized epilepsy (IGE) (122 scans) and 38 healthy controls (67 scans); see Table 1.
EEG acquisition and processing
Acquisition and processing of EEG data simultaneous with fMRI was carried out as described previously (Espay et al., 2008; Szaflarski et al., 2010b; Kay et al., 2012). Briefly, subjects were fitted with an MRI-compatible EEG cap with electrodes arranged according to the international 10/20 system (Compumedics U.S.A., Ltd., El Paso, TX, U.S.A.). Sixty-four channels of data, including an ECG channel, were recorded at 10 kHz concurrent with fMRI using an MRI-compatible system. Time marks generated by the scanner at the onset of each volume acquisition were used to reduce gradient-related artifacts via an average artifact subtraction method (Allen et al., 2000). Heartbeat timings generated from the electrocardiography (ECG) channel were used to reduce the ballistocardiographic artifact via a linear spatial filtering method (Lagerlund et al., 1997; ). All EEG data were reviewed by a board certified epilepsy specialist (JPS). Based on this review, 37 scans containing GSWDs from 25 patients (14 male, 11 female, age mean years ± SD = 29.1 ± 12.0 years) were excluded from further analysis.
MRI acquisition and processing
Acquisition of MRI and fMRI data was carried out as described previously (DiFrancesco et al., 2008; Szaflarski et al., 2010b; Kay et al., 2012) on a four Tesla, 61.5-cm bore Varian Unity INOVA system (Varian, Inc., Palo Alto, CA, U.S.A.) equipped with a standard head coil. T1-weighted structural images were acquired for use as an anatomic reference. A modified driven equilibrium Fourier transform (MDEFT) method (Uğurbil et al., 1993; Duewell et al., 1996) was used with an 1,100-msec inversion delay, 256 × 196 × 196 mm field of view, 256 × 196 × 196 voxel matrix, 22 degree flip angle, and repetition time/echo time (TR/TE) = 13.1/6.0 msec. T2*-weighted echo-planar functional images with blood oxygenation level-dependent (BOLD) contrast were acquired with an 256 × 256 mm field of view, 64 × 64 voxel matrix, 90 degree flip angle, and 5-mm slice thickness in axial orientation without gap. Four hundred volumes consisting of 30 slices each and TR/acquisition time = 3,000/2,000 msec were collected during each scan.
Data were reconstructed and corrected for geometric distortion and Nyquist ghosting with the aid of multiecho reference scans (MERS) (Schmithorst et al., 2001). Functional scans underwent slice timing correction, motion correction (Jenkinson et al., 2002), rigid-body registration to a high-resolution anatomic scan (Jenkinson & Smith, 2001), and nonlinear registration (Andersson et al., 2007a,b) to a Montreal Neurological Institute (MNI)152 standard using the FMRIB software library (FSL) (Smith et al., 2004). Data were spatially blurred with a gaussian kernel of full width at half maximum (FWHM) = 6 mm using analysis of functional neuroimages (AFNI) (Cox, 1996). Because residual motion has been shown to have an artifactual effect on resting-state connectivity (Power et al., 2012; Van Dijk et al., 2012) data were low-pass filtered at 0.1 Hz (Cordes et al., 2001), and the six rigid-body motion parameters were regressed out of the data using AFNI in addition to the motion correction performed using FSL. Unfortunately, data on physiologic regressors such as pulse and breathing were unavailable.
The quality of functional to anatomic registration was measured using the mutual information cost function (Jenkinson & Smith, 2001). Twelve functional scans with an outlying cost indicating unsatisfactory registration were excluded from the study. The quality of motion correction was measured using the normalized correlation ratio cost function (Jenkinson & Smith, 2001) of each time point to the reference volume. Thirteen additional scans with outlying costs indicating excessive motion were excluded from the study.
Independent component analysis
Group ICA (Calhoun et al., 2001; Schmithorst & Holland, 2004) of all scans followed by dual regression (Filippini et al., 2009) was carried out using the group ICA of fMRI toolbox (GIFT) (Calhoun et al., 2009). To make analysis of the large number of scans in the study computationally feasible, principle component analysis (PCA) reduction was applied in two steps. Each scan was reduced to 75 temporal components, the number suggested by the GIFT software, prior to temporal concatenation, and the concatenated result was reduced to 50 temporal components prior to ICA. The number of components used in the final reduction was chosen empirically (Schmithorst, 2005) because the number obtained using Bayesian information criterion (BIC) estimation (Calhoun et al., 2001) was very large. This process yielded 50 independent components.
Each independent component is thought to represent either a resting-state network, such as the DMN, or an artifact, such as head motion (Damoiseaux et al., 2006). The most spatially comprehensive DMN component was identified visually and confirmed (Greicius et al., 2004) via DMN template distributed with GIFT. Component voxels with intensities >99% of the robust range were used as an DMN region of interest (ROI), where the robust range was defined as the 2nd–98th percentiles of voxel intensities. Voxels with a 50% or greater probability of being white matter, based on the Johns Hopkins University (JHU) white matter atlas distributed with FSL, were excluded from the DMN ROI (Wakana et al., 2007; Hua et al., 2008).
Back-projection is the process by which a group component is translated onto a single scan. The back-projection of the group DMN component onto each scan was computed using dual regression (Filippini et al., 2009), a technique that normalizes component intensity across scans so as to allow for direct statistical comparisons between scans and between experimental groups. Back-projected component intensities of scans from the same subject were averaged voxelwise to obtain one map of DMN connectivity for each subject in the study. These per-subject connectivity maps were subsequently used in higher levels of analysis as described below.
Seed-based voxel correlation
Seed-based voxel regression was done using AFNI's 3dDeconvolve and 3dREMLfit tools (Cox, 1996). A spherical seed region with a radius of 4 mm centered at MNI coordinates x = 2, y = −58, and z = 24, in the posterior cingulate cortex (PCC), was selected based on a previously reported seed region (McGill et al., 2012). For each subject, the average fMRI time course within the seed region was used as the regressor of interest. Each subject's fMRI time course was regressed voxelwise against the subject's seed-region time course. The t-values of the corresponding regression coefficients at each voxel were used as each subject's connectivity map. 3dREMLfit was used to achieve prewhitening via an autoregressive (AR) model. No global signal regressor was used, as inclusion of a global signal regressor has been shown to introduce an unwanted bias (Weissenbacher et al., 2009).
The average of all subjects' connectivity maps underwent a voxelwise one-sample t-test, and the resultant t-values were used as a group DMN connectivity map. Voxels in the group DMN connectivity map with intensities >99% of the robust range were used as an DMN ROI, where the robust range was defined as the 2nd–98th percentiles of voxel intensities. As above, voxels with a 50% or greater probability of being white matter, based on the JHU white matter atlas distributed with FSL, were excluded from the DMN ROI (Wakana et al., 2007; Hua et al., 2008).
High-level analysis was performed using R (R Development Core Team, 2011; Whitcher et al., 2011). Each analysis was repeated separately with data from ICA and dual regression and with data from seed-based voxel correlation. Voxelwise analysis using a two-sample t-test was carried out on subjects' DMN connectivity maps to compare patients with uncontrolled seizures to healthy controls; age (Sambataro et al., 2010) and music-listening (Kay et al., 2012) were included as coregressors. Voxelwise regression using a linear model with duration of epilepsy as the explanatory variable was carried out on DMN connectivity maps of patients with uncontrolled seizures. Resultant t-maps were masked with a dilated DMN ROI and corrected for multiple comparisons with cluster-based thresholding using AFNI.
Region of interest analysis
For ROI analysis, each subject's DMN connectivity map was summarized by averaging voxels within the group DMN ROI. The resultant value was used as a measure of DMN connectivity for each subject. Linear models were used to assess the effect of epilepsy, duration of epilepsy, VPA resistance, and uncontrolled seizures on connectivity.
- Top of page
- Supporting Information
In this study we examined the effect of resistance to VPA and the effect of uncontrolled seizures on resting-state DMN functional connectivity. We found that VPA resistance and uncontrolled seizures are associated with a greater reduction in DMN connectivity than epilepsy alone. These findings confirm that the DMN (Luo et al., 2011; McGill et al., 2012) is among the resting-state features (Bai et al., 2011; Killory et al., 2011) that differ between IGE patients and healthy controls.
Previous resting-state fMRI studies have found that DMN connectivity is reduced in IGE patients versus healthy control subjects, and that this reduction is correlated with duration of disease (Luo et al., 2011; McGill et al., 2012). However, we are not aware of any previous studies that specifically investigated the effect of treatment resistance on connectivity in IGE. Our results suggest that reduced DMN connectivity in patients with IGE as a whole is due predominantly to the presence of treatment-resistant patients within the study population. When treatment-resistant and treatment-responsive patients are considered separately, treatment resistance is associated with significantly lower connectivity whereas, depending on the technique used, it may not be possible to distinguish between treatment-responsive patients and healthy controls; see Fig. 2 and Tables 2–4. These findings are consistent with a neurologic contribution to the etiology of treatment resistance and suggest that reduced DMN connectivity may be useful as a biomarker for treatment resistance.
It is notable that DMN connectivity declines with duration of epilepsy (Luo et al., 2011; McGill et al., 2012), and that this diminishing effect is significantly more pronounced in treatment-resistant patients (Fig. 3 and Tables 5 and 6). This finding suggests a cumulative effect of epilepsy on the brain and, if replicated, might weigh in favor of a clinical decision to treat aggressively. It would be interesting to learn if uncontrolled seizures have a uniform effect on connectivity or if seizure or GSWD frequency is negatively correlated with DMN connectivity. Lack of detailed information on seizure load (number of seizures per unit of time) or GSWD frequency during 24 h monitoring did not allow for these analyses in this study.
This study investigated a cross section of IGE that included newly diagnosed patients as well as patients who had lived with epilepsy for >50 years. As such, it was not possible to directly examine whether reduced DMN connectivity is an outcome of treatment-resistance or a predictor of it. We came close to this question by including treatment-resistance and duration of disease in the same model (Tables 3 and 4). Using ICA, we found that the effect of uncontrolled seizures on connectivity remains significant even after controlling for the effect of duration. Although the effect of VPA resistance was significant on its own, it did not reach our threshold for significance (α = 0.05) after controlling for duration. This could have been due to the small number of VPA-resistant patients in the study (n = 13) or to a confounding correlation between VPA resistance and duration. VPA-resistant patients were observed to be older (by 8.1 ± 3.5 years, p < 0.05) and to have had epilepsy for a longer period of time (by 10.5 ± 3.9 years, p < 0.01) than VPA-responders. Although including duration in the model of connectivity versus VPA response was appropriate (F = 7.85, p = 0.008), its collinearity with VPA resistance precipitated a drop in statistical power from 0.96 to an unsatisfactory level of 0.60. A prospective study of newly diagnosed drug-naive IGE patients would help resolve this confound and establish whether DMN connectivity can predict treatment resistance. At least one such study (Moeller et al., 2008b) has demonstrated deactivation of DMN regions during GSWD in a drug-naive population.
We had not expected to observe a correlation between VPA status and age or duration, and no such correlation was observed for uncontrolled versus controlled seizures (p > 0.4). We postulate that, due to the less favorable adverse effects profile of VPA compared to newer drugs, nowadays clinicians are postponing treatment with VPA until other treatment options have been exhausted. This bias was especially evident in women of childbearing age (χ2 = 4.86, p = 0.053); fortunately, we were able to recruit a similar number of men (n = 21) and women (n = 20) who had tried VPA. Response to VPA among those who had tried it was also similar between men (14/21) and women (14/20; Table 1).
Drugs represent another possible confound in our study, as these neuromodulatory agents could plausibly modulate the DMN independently of their effects on epilepsy (e.g., Szaflarski & Allendorfer, 2012). Unfortunately, the large number of different AEDs combined with the high proportion of patients on AED polytherapy precludes modeling every AED effect separately. We instead considered the number of current drugs (taken at the time of scanning) and the number of previously failed drugs (Table 1). As expected, patients who were taking more (p = 0.023) or had failed more (p < 0.001) medications were more likely to have uncontrolled seizures. On the other hand, number of current drugs was a poor predictor in the model of connectivity versus uncontrolled seizures and duration (F = 1.51, p = 0.224). This lack of correlation is unsurprising because there is no reason to expect that all drugs would have the same effect (i.e., increasing or decreasing) on connectivity. We estimate that an additional 54 epilepsy patients would be needed to retain the statistical power of the original model and conclusively refute the possibility that observed differences in DMN connectivity were due to drug effects rather than treatment resistance. Again, a prospective study would help to resolve these confounds.
To our knowledge, this is the first study of DMN connectivity in IGE to use ICA and dual regression. We performed confirmatory analysis using the more common technique of seed-based voxel correlation. These two techniques yielded similar findings. Both showed greater changes in posterior than anterior regions. The PCC is considered the “hub” of the DMN (Greicius et al., 2003), and so may have been more severely affected by IGE. In addition, posterior DMN regions are larger than anterior ones and thus may have been favored by cluster-based thresholding; ancillary ROI analysis was not affected by this factor. Seed-based voxel correlation showed decreased connectivity in the thalamus associated with uncontrolled IGE, whereas ICA did not show thalamus to be part of the DMN. The interpretation of this discrepancy is unclear because thalamus is not classically considered part of the DMN (Raichle et al., 2001). Nevertheless, thalamus is functionally and structurally connected to DMN regions in cortex (Greicius et al., 2003, 2009; Zhang et al., 2008) and has been observed to modulate DMN connectivity (Jones et al., 2011). Reduced DMN connectivity in thalamus could arise due to generalized spike and wave discharges (GSWDs) produced by corticothalamic circuits (Contreras et al., 1996; Moeller et al., 2008a).