How reliable are fmri EEG studies of epilepsy? A nonparametric approach to analysis validation and optimization

www.elsevier.com/locate/ynimg NeuroImage 24 (2005) 192 199 How reliable are fmri EEG studies of epilepsy? A nonparametric approach to analysis validation and optimization Anthony B. Waites, a Marnie E. Shaw, a Regula S. Briellmann, a Angelo Labate, a David F. Abbott, a and Graeme D. Jackson a,b, * a Brain Research Institute, Austin Health, Heidelberg West, Australia b Department of Medicine and Radiology, The University of Melbourne, Parkville, Australia Received 6 April 2004; revised 18 June 2004; accepted 7 September 2004 Simultaneously acquired functional magnetic resonance imaging (fmri) and electroencephalography (EEG) data hold great promise for localizing the spatial source of epileptiform events detected in the EEG trace. Despite a number of studies applying this method, there has been no independent and systematic validation of the approach. The present study uses a nonparametric method to show that interictal discharges lead to a blood oxygen level dependent (BOLD) response that is significantly different to that obtained by examining random deventst. We also use this approach to examine the optimization of analysis strategy for detecting these BOLD responses. Two patients with frequent epileptiform events and a healthy control were studied. The fmri data for each patient were analyzed using a model derived from the timings of the epileptiform events detected on EEG during fmri scanning. Twenty sets of random pseudoevents were used to generate a null distribution representing the level of chance correlation between the EEG events and fmri data. The same pseudoevents were applied to control data. We demonstrate that it is possible to detect blood oxygen leveldependent (BOLD) changes related to interictal discharges with specific and independent knowledge about the reliability of this activation. Biologically generated events complicate the fmri EEG experiment. Our proposed validation examines whether identified events have an associated BOLD response beyond chance and allows optimization of analysis strategies. This is an important step beyond standard analysis. It informs clinical interpretation because it permits assessment of the reliability of the connection between interictal EEG events and the BOLD response to those events. D 2004 Elsevier Inc. All rights reserved. Keywords: fmri; Electroencephalography; Epilepsy * Corresponding author. Brain Research Institute, Ground Floor, Neurosciences Building, Austin Health, Heidelberg West, Victoria, 3081, Australia. Fax: +61 3 9496 2980. E-mail address: BRI@brain.org.au (G.D. Jackson). Available online on ScienceDirect (www.sciencedirect.com.) Introduction Scalp electroencephalography (EEG) gives good temporal information about interictal epileptiform discharges (IED), while functional magnetic resonance imaging (fmri) can identify blood oxygen level dependent (BOLD)-associated neuronal activity with high spatial resolution (Logothetis, 2002, 2003; Ogawa et al., 1990). EEG that is simultaneous with fmri scanning gives highresolution information spatially and temporally, providing a powerful tool for the study of the neural substrates active during generation of interictal discharges. The simplest way to combine EEG and fmri is to use dspiketriggered fmrit, which simply activates the MRI when an epileptiform event is seen. While this has contributed greatly to our understanding of activated areas during epileptiform activity (Archer et al., 2003a,b,c; Baudewig et al., 2001; Jager et al., 2002; Krakow et al., 1999, 2001a,b; Lazeyras et al., 2000; Seeck et al., 1998; Symms et al., 1999; Warach et al., 1996), it has the disadvantage that it only collects a single acquisition per event, not the whole hemodynamic response. Also, no EEG information is acquired during the time when the fmri image is obtained. Technical developments have enabled simultaneous acquisition of EEG and fmri data (Allen et al., 2000; Bonmassar et al., 2002; Jay et al., 1993; Lemieux et al., 2001). Gradient artefact is eliminated using filters (Hoffmann et al., 2000) or by avoiding measurement during gradient switching (Anami et al., 2003). Pulse artefact can be ameliorated using an adaptive subtraction method (Allen et al., 1998; Bonmassar et al., 2002). The combined effect of these correction methods leads to an EEG signal that is of sufficient quality to reliably identify interictal epileptiform discharges, allowing dspike-related fmri EEGT data sets to be acquired (Benar et al., 2002; Iannetti et al., 2002; Lemieux et al., 2001). Despite the importance of these data, there has been little work done on objectively validating the method. The factors that influence the success of a particular fmri EEG investigation remain poorly understood. For example, the number of spikes required for obtaining a significant BOLD response varies 1053-8119/$ - see front matter D 2004 Elsevier Inc. All rights reserved. doi:10.1016/j.neuroimage.2004.09.005

A.B. Waites et al. / NeuroImage 24 (2005) 192 199 193 substantially between studies, ranging from 4 at 3 T (Archer et al., 2003a) to about 20 discharges at 1.5 T (Krakow et al., 1999). Many clinical studies using EEG and fmri show that the area of significant BOLD signal change is congruent with the suspected area of seizure focus (Al-Asmi et al., 2003; Jager et al., 2002; Krakow et al., 1999, 2001a; Lazeyras et al., 2000; Patel et al., 1999; Seeck et al., 1998; Symms et al., 1999), but significant BOLD signal changes may be identified in only 40% (Al-Asmi et al., 2003; Krakow et al., 2001a) to 80% (Archer et al., 2003a; Lazeyras et al., 2000) of the subjects. The sensitivity and specificity of using fmri/eeg to detect a BOLD response to a neurally generated epileptiform event have not been formally assessed. If we are to use BOLD data to understand the neural substrates that generate epileptiform (EEG) events, the levels of false positives and false negatives must be known. A high level of false negatives would suggest the method is not sensitive to detecting the true BOLD response associated with interictal epileptiform EEG events. False positives would mean that BOLD responses occur at times when there is no epileptiform EEG activity. In the present study, we present a nonparametric analysis approach that allows objective consideration of the statistical difference between the BOLD response to true IEDs and the BOLD variations occurring by chance. Since this approach assesses the level of a summary statistic for true EEG spike labelings compared to a null distribution generated by randomizing the timings of the spikes, it is independent of the variance structure of a single voxel and can assess how unusual any activation is relative to the total behavior of brain signal in the absence of EEG events. We have two major aims: (i) to objectively determine that spike-related fmri EEG analysis can reliably detect BOLD changes associated with interictal epileptiform discharges. This is a necessary first step in considering the validity of current approaches as well as identifying possible improvements to the methodology. (ii) to apply this nonparametric framework to optimize analysis strategies in order to maximize detection of true spike-related BOLD changes while managing the level of false positives. Materials and methods Subjects Three subjects were selected for this study. All subjects gave written informed consent. The study was approved by the Austin and Repatriation Medical Centre Ethics committee, and conformed to the guidelines of the National Health and Medical Research Council and the Declaration of Helsinki. Patient 1 is an 8-year-old girl with recent onset of childhood absence epilepsy (CAE). At the time of the fmri/eeg scan, she had hundreds of absences per day. The absences were typically associated with eye flickering and loss of awareness, and had a duration of a few seconds. During the 30-min scanning session she showed 10 EEG spike and wave bursts of between 2 and 5 s. This subject was chosen since she showed significant BOLD response to IEDs when analyzed using our standard approach. Patient 2 is a 10-year-old girl with onset of childhood absence epilepsy about 10 months before scanning. Initially, she had many absences per day, but was not put on medication. She was scanned shortly after her first generalized tonic clonic seizure, which occurred while she was playing a video game. She showed 20 IEDs during scanning, as detected on simultaneous EEG, but standard analysis showed no significant BOLD response. Finally, a healthy female control subject was included. She was 24 years old and showed no epileptic discharges on EEG. Image acquisition The fmri studies were performed with a 3-T GE Signa LX whole body scanner (General Electric, Milwaukee, WI) using a standard birdcage quadrature head coil. Structural imaging included T1-weighted 2D spin-echo images acquired with the same geometric orientation and slice thickness as the subsequent functional images. Functional images were acquired as a series of gradient-recalled echo planar imaging (GR-EPI) volumes (TR/TE = 3600/40 ms, flip angle = 608, 22 oblique slices 4-mm thick + 1-mm gap, 24 cm FOV, 128 128 matrix), with one imaging volume being acquired every 3 s. The first four acquisitions were discarded in order to allow the tissue magnetization to reach equilibrium. EEG acquisition Eighteen nonmetallic scalp electrodes with carbon fiber leads were fixed to the scalp in the conventional d10 20T EEG format, with two chest electrodes to record ECG signal. Electrodes were twisted in pairs then woven in chains and taken straight out of the head end of the scanner bore. A head-box with fiber optic coupling transmitted the EEG signal out of the MR room for display and recording in real time. Simultaneous acquisition of EEG and fmri leads to artefact in the EEG due to induced currents caused by gradient switching. This artefact is eliminated at our institute using custom-built hardware and software (patent pending) and postacquisition filtering (Allen et al., 2000). Image analysis All preprocessing and analysis were performed using SPM99 (Wellcome Department of Cognitive Neurology, http://www.fil. ion.ucl.ac.uk/spm) and ibrainr (Abbott and Jackson, 2001). Preprocessing fmri data were preprocessed prior to analysis. This involved motion correction, where a rigid body transformation was applied to each image volume, using the set of six parameters (three translations and three rotations) that minimized the squared difference between each image and the first. The volumes were then slice-timing corrected, nonlinearly transformed into a space approximating the SPM standard space, which is based on a 152 brain average produced by the Montreal Neurological Institute. Finally, the data were then spatially smoothed by convolution with an 8-mm isotropic Gaussian kernel. Spike-related analysis EEG signals were examined by two trained epileptologists (AL, RB). From the EEG trace IED events (spikes, spike waves, slow spike waves), other physiological EEG events (slowing, sleep transients), and EEG artefacts (motion artefacts, electrode artefacts) were identified. Of these, only IED events were considered to be of interest in detecting the site generating IED. These events were convolved with the hemodynamic response function (HRF)

194 A.B. Waites et al. / NeuroImage 24 (2005) 192 199 chosen. Analysis of the data to identify voxels correlated with this spike-based basis function was performed using the general linear model, as implemented in SPM99. In SPM99, corrections for multiple comparisons exploit Gaussian random field theory (Worsley et al., 1996). Analysis strategies For each subject, a total of eight analysis strategies were used, as is summarized in Table 1. The different analysis strategies differed with respect to two variables; (i) the combination of function(s) used to model EEG events and (ii) which EEG events should be modeled. For the analyses referred to henceforth as A1 A4, a standard canonical hemodynamic response function (HRF) and the associated temporal derivative were used to model EEG events. For the analyses referred to as B1 B4, a flexible model was used. The flexible model, comprised of three gamma functions, was able to model signal changes of unexpected shape whereas the canonical HRF is more sensitive to signal changes that conform to a specific shape. In each case, four alternative approaches were used to model artefactual sources of variance. In the first (A1, B1), only IED events were modeled. In the second approach (A2, B2), IED events were modeled as events of interest and all other EEG events modeled as events of no interest. Those events judged as physiological EEG events were convolved with the relevant HRF, and those judged artefact related were not. In the third analysis approach (A3, B3), scans temporally aligned with EEG artefact identified from EEG were discarded from the analysis. Finally, the fourth approach (A4, B4) involved discarding scans associated with both EEG artefact and other EEG events from the analysis. Permutation test In order to judge the significance of any results for each subject and analysis strategy, we used a permutation test (Holmes et al., 1996; Nichols and Holmes, 2002). Since permutations of the scans would change the temporal autocorrelations of the fmri data, we approached the problem by permuting the labels relating to the spikes (Raz et al., 2003), which in this case involves generating a random list of the spike timings. As described in Nichols and Holmes (2002), permutation tests exploit the fact that if there is no difference between experimental conditions (the null hypothesis), then applying labels of an experimental condition (such as rest, or EEG spike) to an observation is arbitrary. Under the null hypothesis, the significance of a measurement can then be assessed by comparison with the distribution of values obtained when the labels are permuted. This nonparametric approach involves fewer assumptions than a standard parametric analysis; in fact, all that is required is that the distributions of values in each condition have the same shape or are symmetric (Nichols and Holmes, 2002). For each analysis strategy (A1...A4, B1...B4) and subject, we generated 20 random relabelings of the set of spike timings. These Table 1 Summary of analysis strategies Scans included (design matrix) Canonical HRF All scans (model spikes only) A1 B1 All scans (model all EEG events) A2 B2 Remove artefact scans A3 B3 Remove all nonspike EEG scans A4 B4 Gamma basis functions 20 sets of bpseudoeventsq were analyzed, together with the true spikes. In each case, the maximum T statistic within the brain was noted. The 20 random relabelings thus yielded a distribution of 20 maximum T values against which the true maximum T value could be compared and assessed for significance. Results Validity of the permutation test In order for the permutation test to be valid for generating the null distribution of no BOLD response to the IED, we require that the distribution is symmetric (Nichols and Holmes, 2002). In Fig. 1 the distribution of T scores for each of the 80 randomizations (20 relabelings 4 approaches) is presented for each of three subjects. A similar distribution is seen for each subject. In each case, the distribution is symmetric and approximately normal in form (for all subjects, jsj b 0.31, jkj b 1.14, Jarque Bera test does not reject the assumption of normality). Patient 2 and the control show similar distributions of maximum T scores. Patient 1 shows a somewhat broader distribution, possibly influenced by the large BOLD signal changes associated with this subject s IED. If the random events lie close to the true spikes, one would expect higher T values, resulting in a longer tail at high T values, and in other cases, the increased variance associated with spikes would decrease the significance of false positives, leading to a lower mean of the distribution of maximum T scores. Is the analysis of spike-related fmri EEG reliable? The main result with regard to our first aim is the finding that patient 1 showed much stronger maximum T values when analyzed with true spikes, suggesting that there is variance associated with those spikes that is much greater than variance detected by chance (due to autocorrelations). This result is seen in Fig. 2, which presents a summary of results of the three subjects (top row in red patient 1, middle row in blue patient 2, and bottom row in green, control subject) for four of the eight possible analysis strategies (A1, A2, A3, A4). It can be seen in Fig. 2A that for patient 1, the true spike labelings (dark color) lead to maximum T scores significantly above the mean of the random labelings (light color) (5.83 8.18 SDs above the mean). In each case the maximum T scores are also above the line representing a significance of P b 0.05, corrected for multiple comparisons. The mean of the random relabelings lies below the 0.05 corrected level. To examine the reliability of the observed spatial pattern of activation between analysis strategies, the detected BOLD activation associated with true spikes is presented in Fig. 3 together with the most significant (highest maximum T value) of the relabelings for each subject and analysis strategy. Fig. 3A is the set of results for patient 1, Fig. 3B those for patient 2, and Fig. 3C for the control subject. For the patients, the top panels represent true spike-related BOLD change, and the bottom panels the changes associated with the most significant random relabeling. Left to right panels display the four analysis strategies A1 A4. The most important feature to note is that in Fig. 3A the pattern of activation associated with the relabeling (lower panels) is not equivalent to the pattern of true activations. Patient 1 shows reproducible true activations for each of the analysis strategies, whereas random labelings show different

A.B. Waites et al. / NeuroImage 24 (2005) 192 199 195 Fig. 1. The frequency distribution of maximum T scores obtained using 20 sets of random spike labelings for each of the three subjects. As indicated in the legend, red shows the distribution of patient 1, blue that of patient 2, and green the distribution of the control. activation patterns. In Fig. 3B, we see that patient 2 shows true activation with low significance, but that the pattern of activation is relatively reproducible across the different strategies. In the lower panels of Fig. 3B together with Fig. 3C, we see once again that the random labelings lead to a pattern that varies as a function of analysis strategy. A final point to note is that in all subjects, analysis strategies A1 and A2 show very similar patterns of activation. Is the standard canonical analysis optimal? Patient 1 showed a significant response independent of analysis strategy. The results of patient 2 show a contrasting situation (Fig. 2B). In this case, there are no significant differences between true and random labels for any of the four strategies presented. This is despite the fact that this subject had 20 IEDs during her scanning session. This suggests that either there is no significant BOLD response to IEDs for this subject, or that the analysis strategies presented are not optimal for their detection. To explore the possibility of optimizing analysis strategy, Fig. 4 considers the differences between all of the eight different analysis strategies. Fig. 4A looks at the significance (measured as the number of SDs above the mean random maximum statistic that the true spikes reach) as a function of model, for all eight models discussed above. Results for patient 1 are shown in red and show a high level of significance independent of modeling approach. There appears to be a trend toward the second strategy (A2, B2, modeling motion and artefact as regressors of no interest) yielding the highest rate of detection of activation associated with true EEG events, as well as a trend toward the strategy of fitting the HRF with basis functions (B1...B4) yielding higher true positives than using a canonical HRF (A1...A4). As mentioned above, the spikes of patient 2 (data shown in blue) are not significant for any of the approaches using a canonical HRF. Considering the use of a flexible model using three gamma functions, we see that approaches B3 and B4 reach significance. Fig. 4B looks at the rate of false positives in each of the strategies used, measured as the number of relabelings (of 20) where one or more voxels reach the threshold of P b 0.05 (corrected for multiple comparisons). By chance one would expect one false positive on an average per set of 20 labelings. There were more false positives than expected for both patient 1 (mean = 2.62) and the control (mean = 1.5). Furthermore, false positives were more frequently observed with the approach using basis functions (B1 B4). Discussion In this study, we have established an analysis framework to assess the reliability of the BOLD response to IED (spikes) detected in simultaneously acquired EEG signal. We have used this framework to validate fmri EEG results and assessed different analysis strategies according to their level of true and false positives. The application of the nonparametric permutation test to fmri EEG analysis appears to be valid, since the distribution of relabeled maximum T values is symmetric. The approach we present here is to use the fmri data of a patient to generate the null distribution of variance relevant to that subject. We were initially concerned that the presence of true spikes in patient data may influence this null distribution and bias subsequent analysis. The distributions we observed in patients and controls are very similar, and all are normally distributed, indicating that the variance associated with the true spikes does not bias the distribution in a manner that would invalidate this approach. The present study has a limitation that should be considered. The nonparametric analysis approach described here is not automated within the SPM99 software and is thus time consuming and labor intensive, which is why we limited ourselves to 20 permutations per analysis strategy. This is a low number of realizations of the null distribution, but we feel it is adequate for the present purposes since each distribution proved to be normal in form, allowing us to judge the significance of true spike labels relative to the null distribution using a Student s t test. Given the utility of the method as demonstrated in this study, we believe it

196 A.B. Waites et al. / NeuroImage 24 (2005) 192 199 motion, or other physiological changes, for example). Motion in association with an IED is actually quite common, but since the hemodynamic response is delayed by approximately 6 s, it is less likely to be the cause of the detected signal change. Further, the distribution of activation for patient 1 was, in our experience, biological and unlike typical motion patterns seen in other fmri studies, supporting the finding that the variance is due to IEDs. This independent validation is important, since there are a number of reasons to question fmri EEG results. Standard analysis of simultaneously acquired fmri EEG involves looking for correlation between BOLD and EEG events significant above some threshold, typically 0.05 corrected for multiple comparisons. We have shown that a healthy control subject can show significant bactivationsq for some randomly assigned events, which demonstrates the ease with which false positive bactivationsq can arise from the standard approach. Indeed, in two of three subjects, a corrected level of significance was reached more often than expected by chance. This may reflect the fact that the model is nonspecific when only few (10 20) spikes are identified. The present framework provides a solution, providing a test to ensure that there is a significant difference between true and random labelings for each subject studied. It may even be possible to use a permutation test approach to evaluate the null distribution for the particular data set, then use an appropriate threshold (perhaps defined as two standard deviations above the mean of the nonparametric null distribution) to control false positives and give more confidence in the reliability of the subsequent activation maps (obtained using a standard parametric approach). The fact that patient 1 showed a consistent and significant difference between true IED events and random pseudoevents also provides an independent proof of the validity of the fmri EEG technique. The impact of analysis approach on detection of spike-related BOLD changes Fig. 2. A summary of the findings for the three subjects. The maximum T score of the true spikes (dark color) is compared to the mean and standard deviation of the maximum T score of the 20 random relabelings (light color). Results are presented for four of the eight possible analysis strategies (A1, A2, A3, A4). In the upper panel (A), the results are presented for patient 1, with light red representing the maximum T score for random events and dark red that for true spikes, in B results for patient 2 (light blue for random labels, dark blue for true labels), and in C the results of the 20 relabelings for the control subject (light green), who obviously has no true spikes for comparison. will be useful to automate the analysis to permit larger numbers of realizations. How reliable is spike-related fmri EEG analysis? It is possible to reliably detect variance that is correlated with IEDs, detected in simultaneously acquired EEG signal, at a level that is significantly above other sources of variance. For a subject with a strong BOLD response to IEDs, that response can be reliably detected. In Fig. 1, we saw that for such a patient all analysis strategies were able to distinguish between true and random labelings of IED events within the data. Of course, this does not prove that the variance is due to the IEDs (it could be The results presented here demonstrate that our nonparametric framework can help develop an optimized strategy for detecting spike-related BOLD changes. Patient 2 showed a significant activation for two of the eight analysis strategies, despite the fact that no BOLD response was detected when using the standard analysis approach. This finding indicates that for at least a subgroup of subjects, choosing the optimal analysis strategy may be a prerequisite for identification of significant BOLD changes in response to interictal discharges. This is particularly important if we are trying to understand the strength and reliability of the coupling between neurally generated interictal epileptiform discharges and the BOLD response. For example, if we do not detect a BOLD response to a definite IED (Krakow et al., 2001a), is this a biological phenomenon, or merely the sensitivity of our method of detecting the BOLD response? There are several further points worth discussing about the different analysis strategies. Considering the rate of detection associates with true EEG events (Fig. 4A), we saw that fitting the HRF with basis functions was more sensitive than using a canonical HRF. This is not surprising, since basis functions represent a more flexible model, so that BOLD responses that deviate from the canonical form would be better detected using basis functions. Unfortunately, the flexible HRF model was also more likely to incorrectly model other variance sources as false

A.B. Waites et al. / NeuroImage 24 (2005) 192 199 197 Fig. 3. Statistical parametric maps of spike-related activation (upper panels) and the most significant of the 20 random relabeled events (lower panels) for four of the analysis strategies (A1 A4), for the three subjects: patient 1 (A), patient 2 (B), and the random events only for the healthy control (C). Four axial slices are shown per strategy. The activation is overlayed on the mean EPI image for each subject. Positive and negative responses are shown in red yellow and blue green, respectively. positives. Further, modeling motion and other EEG artefact as regressors of no interest was more sensitive than removing scans affected by these sources of variance. The fact that subjects vary in the level of significance in their BOLD response should not be surprising. Significance relates to noise sources in the fmri data, which may include physiological noise and motion, but will also depend on the shape of the HRF of each subject, which has been shown to vary across subjects (Benar et al., 2002). It is not known if the HRF varies over time. In a recent study using spike-triggered fmri, Krakow et al. (2001b) assessed individually the 43 spikes that occurred in a patient with focal epilepsy. About a third of these spikes were associated with a significant BOLD signal change. This may be due to instability in the HRF, but may also have been a result of increased motion as the study progressed. Finally, the approach of Salek-Haddadi et al. (2003) in modeling the BOLD response to a clinical seizure using a cosine basis function set showed a very strong response with a period of around 100 s. Although the parametric map appeared biologically plausible, such an approach is also very sensitive to patient motion, which may occur in concert with the seizure. Similarly, Benar et al. (2002) showed that one may measure the actual signal time course of spike-related BOLD response. The limitation of this approach is that one must choose a region of interest. If this is

198 A.B. Waites et al. / NeuroImage 24 (2005) 192 199 Acknowledgments We gratefully acknowledge Matthew Harvey, Claire Jennings, and Richard Masterton for assistance in acquiring and analyzing the data for this study. This work was supported by the NHMRC (via program grant number 144105 and career development award ID 260050), Neurosciences Victoria, and the Brain Imaging Research Foundation. References Fig. 4. Relation of positives associated with true EEG events (A) and false positives (B) to the analysis strategy for all of the eight possible analysis strategies considered. In A, a comparison is made between the maximum T value seen for true spikes and distribution of randomly relabeled events. The vertical scale is the number of standard deviations above the mean random maximum T value that the true maximum T reaches. The grey-shaded area represents no significant difference between true and random events. In B, the number of false positives detected from the 20 random relabelings is presented for each subject and analysis strategy. In both figures, patient 1 is shown in red, patient 2 in blue, and control in green. selected using the activation detected with a standard canonical HRF, then one expects the time courses measured to be similar in form to the input model. Thus, their method is unlikely to detect other hemodynamic response shapes that are undetected by the standard analysis but may nonetheless exist and be of clinical significance. Our data suggest an analysis approach using basis function modeling in a nonparametric framework, with removal of EEG-artefact-related scans, could be effectively used to overcome these limitations. Our study independently validates the application of eventrelated fmri analysis methods to simultaneous fmri EEG. It has shown that the approach can adequately detect spike-related BOLD changes above the level of randomly occurring changes, but only in subjects with an adequate BOLD response. It has also shown that the nonparametric framework provides a suitable basis for the optimization of analysis strategies. The nonparametric framework could be used to address issues such as artefact removal, as well as investigating the number of scans and the number of spikes required for a successful study. Abbott, D., Jackson, G.D., 2001. ibrainr-software for analysis and visualisation of functional MR images. NeuroImage 13, 59. Al-Asmi, A., Benar, C.G., Gross, D.W., Khani, Y.A., Andermann, F., Pike, B., Dubeau, F., Gotman, J., 2003. fmri activation in continuous and spike-triggered EEG fmri studies of epileptic spikes. Epilepsia 44 (10), 1328 1339. Allen, P.J., Polizzi, G., Krakow, K., Fish, D.R., Lemieux, L., 1998. Identification of EEG events in the MR scanner: the problem of pulse artifact and a method for its subtraction. NeuroImage 8 (3), 229 239. Allen, P.J., Josephs, O., Turner, R., 2000. A method for removing imaging artifact from continuous EEG recorded during functional MRI. Neuro- Image 12 (2), 230 239. Anami, K., Mori, T., Tanaka, F., Kawagoe, Y., Okamoto, J., Yarita, M., Ohnishi, T., Yumoto, M., Matsuda, H., Saitoh, O., 2003. Stepping stone sampling for retrieving artifact-free electroencephalogram during functional magnetic resonance imaging. NeuroImage 19 (2 Pt. 1), 281 295. Archer, J.S., Abbott, D.F., Waites, A.B., Jackson, G., 2003a. fmri deactivation of the posterior cingulate during generalised spike and wave. NeuroImage 20, 1915 1929. Archer, J.S., Briellmann, R.S., Abbott, D.F., Syngeniotis, A., Wellard, R.M., Jackson, G.D., 2003b. Benign epilepsy with centro-temporal spikes: spike triggered fmri shows somato-sensory cortex activity. Epilepsia 44, 200 204. Archer, J.S., Briellmann, R.S., Syngeniotis, A., Abbott, D.F., Jackson, G.D., 2003c. Spike triggered fmri in reading epilepsy: involvement of left frontal cortex working memory areas. Neurology 60, 415 421. Baudewig, J., Bittermann, H.J., Paulus, W., Frahm, J., 2001. Simultaneous EEG and functional MRI of epileptic activity: a case report. Clin. Neurophysiol. 112 (7), 1196 1200. Benar, C.G., Gross, D.W., Wang, Y., Petre, V., Pike, B., Dubeau, F., Gotman, J., 2002. The BOLD response to interictal epileptiform discharges. NeuroImage 17 (3), 1182 1192. Bonmassar, G., Purdon, P.L., Jaaskelainen, I.P., Chiappa, K., Solo, V., Brown, E.N., Belliveau, J.W., 2002. Motion and ballistocardiogram artifact removal for interleaved recording of EEG and EPs during MRI. NeuroImage 16 (4), 1127 1141. Hoffmann, A., Jager, L., Werhahn, K.J., Jaschke, M., Noachtar, S., Reiser, M., 2000. Electroencephalography during functional echo-planar imaging: detection of epileptic spikes using post-processing methods. Magn. Reson. Med. 44 (5), 791 798. Holmes, A.P., Blair, R.C., Watson, J.D., Ford, I., 1996. Nonparametric analysis of statistic images from functional mapping experiments. J. Cereb. Blood Flow Metab. 16 (1), 7 22. Iannetti, G.D., Di Bonaventura, C., Pantano, P., Giallonardo, A.T., Romanelli, P.L., Bozzao, L., Manfredi, M., Ricci, G.B., 2002. fmri/ EEG in paroxysmal activity elicited by elimination of central vision and fixation. Neurology 58 (6), 976 979. Jager, L., Werhahn, K.J., Hoffmann, A., Berthold, S., Scholz, V., Weber, J., Noachtar, S., Reiser, M., 2002. Focal epileptiform activity in the brain: detection with spike-related functional MR imaging Preliminary results. Radiology 223 (3), 860 869. Jay, V., Becker, L.E., Otsubo, H., Hwang, P.A., Hoffman, H.J., Harwood- Nash, D., 1993. Pathology of temporal lobectomy for refractory

A.B. Waites et al. / NeuroImage 24 (2005) 192 199 199 seizures in children. Review of 20 cases including some unique malformative lesions. J. Neurosurg. 79, 53 61. Krakow, K., Woermann, F.G., Symms, M.R., Allen, P.J., Lemieux, L., Barker, G.J., Duncan, J.S., Fish, D.R., 1999. EEG-triggered functional MRI of interictal epileptiform activity in patients with partial seizures. Brain 122 (Pt. 9), 1679 1688. Krakow, K., Lemieux, L., Messina, D., Scott, C.A., Symms, M.R., Duncan, J.S., Fish, D.R., 2001a. Spatio-temporal imaging of focal interictal epileptiform activity using EEG-triggered functional MRI. Epileptic Disord. 3 (2), 67 74. Krakow, K., Messina, D., Lemieux, L., Duncan, J.S., Fish, D.R., 2001b. Functional MRI activation of individual interictal epileptiform spikes. NeuroImage 13 (3), 502 505. Lazeyras, F., Blanke, O., Perrig, S., Zimine, I., Golay, X., Delavelle, J., Michel, C.M., de Tribolet, N., Villemure, J.G., Seeck, M., 2000. EEGtriggered functional MRI in patients with pharmacoresistant epilepsy. J. Magn. Reson. Imaging 12 (1), 177 185. Lemieux, L., Salek-Haddadi, A., Josephs, O., Allen, P., Toms, N., Scott, C., Krakow, K., Turner, R., Fish, D.R., 2001. Event-related fmri with simultaneous and continuous EEG: description of the method and initial case report. NeuroImage 14 (3), 780 787. Logothetis, N.K., 2002. The neural basis of the blood-oxygen-leveldependent functional magnetic resonance imaging signal. Philos. Trans. R. Soc. Lond., B Biol. Sci. 357 (1424), 1003 1037. Logothetis, N.K., 2003. The underpinnings of the BOLD functional magnetic resonance imaging signal. J. Neurosci. 23 (10), 3963 3971. Nichols, T.E., Holmes, A.P., 2002. Nonparametric permutation tests for functional neuroimaging: a primer with examples. Hum. Brain Mapp. 15 (1), 1 25. Ogawa, S., Lee, T.M., Kay, A.R., Tank, D.W., 1990. Brain magnetic resonance imaging with contrast dependent on blood oxygenation. Proc. Natl. Acad. Sci. 87, 9868 9872. Patel, M.R., Blum, A., Pearlman, J.D., Yousuf, N., Ives, J.R., Saeteng, S., Schomer, D.L., Edelman, R.R., 1999. Echo-planar functional MR imaging of epilepsy with concurrent EEG monitoring. Am. J. Neuroradiol. 20 (10), 1916 1919. Raz, J., Zheng, H., Ombao, H., Turetsky, B., 2003. Statistical tests for fmri based on experimental randomization. NeuroImage 19 (2 Pt. 1), 226 232. Salek-Haddadi, A., Lemieux, L., Merschhemke, M., Friston, K.J., Duncan, J.S., Fish, D.R., 2003. Functional magnetic resonance imaging of human absence seizures. Ann. Neurol. 53 (5), 663 667. Seeck, M., Lazeyras, F., Michel, C.M., Blanke, O., Gericke, C.A., Ives, J., Delavelle, J., Golay, X., Haenggeli, C.A., de Tribolet, N., et al., 1998. Non-invasive epileptic focus localization using EEG-triggered functional MRI and electromagnetic tomography. Electroencephalogr. Clin. Neurophysiol. 106 (6), 508 512. Symms, M.R., Allen, P.J., Woermann, F.G., Polizzi, G., Krakow, K., Barker, G.J., Fish, D.R., Duncan, J.S., 1999. Reproducible localization of interictal epileptiform discharges using EEG-triggered fmri. Phys. Med. Biol. 44 (7), N161 N168. Warach, S., Ives, J.R., Schlaug, G., Patel, M.R., Darby, D.G., Thangaraj, V., Edelman, R.R., Schomer, D.L., 1996. EEG-triggered echo-planar functional MRI in epilepsy. Neurology 47 (1), 89 93. Worsley, K.J., Marrett, S., Neelin, P., Vandal, A.C., Friston, K.J., Evans, A.C., 1996. A unified statistical approach for determining significant signals in images of cerebral activation. Hum. Brain Mapp. (4), 58 73.