Event-related functional magnetic resonance imaging (fMRI) is an MRI technique that can be used to detect changes in the Blood Oxygen Level Dependent (BOLD) hemodynamic response to neural activity in response to certain events. The conventional method for detecting event-related responses in fMRI consists of modeling the expected fMRI response to an event by convolving a stimulus presentation time-course with an assumed canonical Haemodynamic Response Function (HRF) and using linear regression to identify voxels with a significant correlation to this expected response (Josephs et al., 1997). One typically assumes that the onset and offset of neuronal activity match stimuli onset and offset, the neuronal activation is sustained at a constant level throughout the stimulus and that evoked fMRI signal changes follow the canonical HRF.

There are instances, however, when these assumptions may not be satisfied. An example is interictal epileptiform discharges (IEDs), which are pathological patterns of activity generated by the brain of patients with epilepsy between seizures (de Curtis et al., 2012). IEDs produce marked and stereotyped trace deviations on electroencephalography (EEG) recordings and can be studied using fMRI by using a simultaneous acquisition of EEG (EEG-fMRI) in order to identify the event timings (Lemieux et al., 2001; Bnar et al., 2002). Studies have shown that the onset of the neuronal activity underlying the EEG discharge may not always coincide with the EEG onset (Bai et al., 2010; Carney et al., 2010; Masterton et al., 2010). For example, Carney et al. (2010) identified changes in BOLD signal which precede the onset of epileptiform activity. In addition, it is also reported that the use of the same HRF in all patients may not be appropriate and that individual-based HRF models provide increases in extent and degree of activation (Masterton et al., 2010; Storti et al., 2013).

To address the above issues, we developed an algorithm, dubbed event-related independent components analysis (eICA), which allowed for less constrained analyses capable of detecting early, non-canonical responses (Masterton et al., 2013a,b). Event-related ICA is a technique that provides an estimate of the underlying components that give rise to the observed event-related fMRI signal changes throughout the brain, and importantly, does not rely upon the specification of an HRF model or predefined Regions of Interest (ROIs). Unlike a standard independent components analysis (ICA), which is applied to the entire fMRI time series, the eICA method is applied only to the event-related time courses at each voxel (an estimate of the event-related signal at each voxel is first obtained by deconvolution of the observed fMRI signal with the observed EEG event timing), which means that only a small number of components are generated that are all explicitly related to the event of interest. Event-related ICA can be applied to data from individual subjects and also to group data using a temporal concatenation approach. We previously demonstrated that the eICA method, when applied to EEG-fMRI data acquired from a group of patients with Benign epilepsy with centro-temporal spikes (BECTS), provided better performance than a standard event-related analysis and a linear deconvolution approach, with a better detection rate in single-subject analyses (73 vs. 53%) and only event-related ICA finding significant group-level activation (Masterton et al., 2013b).

A key element of the eICA is the initial deconvolution. However, the stability of the deconvolution can be compromised by various sources of structured noise (Biswal et al., 1996; Friston et al., 1996; Glover et al., 2000) present in the fMRI signal. These include rapid and slow head movements, physiological activity (breathing and heartbeat) and potential acquisition artifacts. Data driven techniques, especially ICA, are increasingly being employed to separate signal and noise in conventional fMRI data (Thomas et al., 2002; Kochiyama et al., 2005; McKeown et al., 2006; Perlbarg et al., 2007; Stevens et al., 2007; Calhoun et al., 2008; Tohka et al., 2008; Sui et al., 2009; Beckmann, 2012; Kundu et al., 2012; Bhaganagarapu et al., 2013; Salimi-Khorshidi et al., 2014). However, in the context of EEG-fMRI studies in epilepsy, the interpretation of the results from an application of ICA can be difficult as it may produce more than a hundred different components per subject with the majority of these likely having no relationship to the EEG event of interest (Rodionov et al., 2007; LeVan et al., 2010).

To address this we developed a strategy for the automated objective identification of artifactual components from an ICA, that we have dubbed a Spatially Organized Component Klassificator (SOCK) (Bhaganagarapu et al., 2013). The primary objective of SOCK is to distinguish noise from true activation without rejecting the latter. SOCK automatically classifies ICs into one of two categories; artifact or unlikely artifact. It does so using spatial measures likely to indicate motion, physiological noise, or machine or undetermined noise. SOCK was shown to successfully remove artifactual components, without rejecting true activation in resting state data (Bhaganagarapu et al., 2013). Unlike existing automatic classifier methods which are primarly dependent on training data to inform classification (De Martino et al., 2007; Tohka et al., 2008; Salimi-Khorshidi et al., 2014) or require querying a public database (Sochat et al., 2014), SOCK is a standalone, automated and objective method that does not require the user to train the algorithm. It is able to identify a high proportion of artifact-related ICs without removing components that are likely to be of neuronal origin (Bhaganagarapu et al., 2013).

In this paper, we extend the eICA procedure by utilizing SOCK to automatically de-noise fMRI data from an EEG-fMRI acquisition prior to running eICA. As we are de-noising the entire fMRI time series prior to the eICA, we hypothesize that this approach has the potential to be more sensitive than eICA alone. The use of an automated de-noising procedure like SOCK in the context of eICA is a novel methodology and to our knowledge has not been investigated previously. We demonstrate the effectiveness of SOCK by comparing the extent of activation obtained in a standard eICA analysis of EEG-fMRI data with and without the use of SOCK for 14 patients with rolandic epilepsy who exhibited stereotypical IEDs arising from a focus in the rolandic fissure.

In this paper we have demonstrated the superiority of SOCK+eICA compared with eICA alone for mapping functional brain activity associated with epileptic spikes. It has previously been demonstrated that that eICA is superior to conventional event-related analyses when the BOLD response does not closely match the canonical haemodynamic response function (HRF) (Masterton et al., 2013b). Taken together, our results suggest that SOCK+eICA should replace eICA alone as the preferred method for such analyses.

The centro-temporal spikes of Rolandic epilepsy served as a good test case for our analysis methodology, as it is known that the BOLD response does not well match the standard HRF. This is in part due to neuronal activity associated with the spikes being detectable with fMRI before activity becomes sufficiently widespread and synchronized to manifest as a spike visible on the EEG, and in part due to a larger post-spike undershoot (Masterton et al., 2010, 2013b). The eICA procedure provides a less constrained approach than a GLM incorporating a conventional HRF model, however this flexibility comes at the cost of lower power (increased susceptibility to noise). Whilst the event-related nature of the eICA approach provides a much stronger constraint than conventional ICA on the full fMRI time series, the eICA method is still susceptible to noise, particularly in the initial deconvolution step. Temporally non-stationary noise would be expected to increase the heterogeneity of the raw signal response associated with events, making deconvolution more challenging, and spatial non-stationary of the noise would be expected to increase heterogeneity of the derived event-related responses across voxels. This would then deleteriously affect the performance of the subsequent ICA of the event-related responses. Thus using a procedure which removes a substantial quantity of noise from the input data may improve the end result. The SOCK procedure is a standalone, automated and objective method which is able to remove a substantial fraction of noise without removing biologically interesting signal (Bhaganagarapu et al., 2013). The results of the present study indicate in practice the improvement can be substantial when SOCK is used to de-noise fMRI data prior to eICA.

Applying SOCK+eICA to the existing EEG-fMRI of our BECTS cohort has improved the confidence in the initial results of Masterton et al. (2010) and Masterton et al. (2013b) (i.e., that the centro-temporal spikes arise from low in the ipsilateral post-central gyrus, extending along its opercular surface into the lateral fissure), with two of the previously negative-result individuals now showing activity in this region. There now remain just two individuals with negative results: These two subjects (3 and 6 in Table 1) along with subject 14 registered the smallest number of events compared to all other subjects. The lower the event count, the less power one has to detect an effect (Flanagan et al., 2009).

In these BECTS subjects the epileptic spikes were unilateral during the imaging session. Homologous regions of cortex are connected via fibers projecting through the corpus callosum and typically inhibit neural activity in the contralateral hemisphere. The new SOCK+eICA group analysis reveals a more complex response than previously evident, with initial ipsilateral activity, followed by a more extensive bilateral pattern of deactivation (i.e., the time-course of the deactivation component displays a later rise and peak compared to the ipsilateral-only component). We would interpret the new finding as distinguishing the activation of the ipsilateral cortex during epileptiform events, and a later bilateral decrease in activity in response to this activation.

We have demonstrated a novel application of our ICA classifier, SOCK, to de-noise fMRI prior to an event-related ICA in patients with rolandic epilepsy. The procedure outlined in this paper harnesses the advantage of both techniques: (1) SOCK de-noises fMRI in an objective and automated manner utilizing the entire fMRI time-series. (2) eICA utilizes the EEG information to derive event-related responses which are input into an ICA, thus constraining the final eICA decomposition to a small number of components time-locked to the events of interest. The use of SOCK increased power to detect activity of interest in both individual and group analyses.