Introduction

Functional MRI (fMRI) studies have shown that, in the absence of an explicit task, distributed brain areas are arranged into networks exhibiting correlations in the blood-oxygen-level dependent (BOLD) signal (Biswal et al. 1995). These ‘resting-state’ networks show high consistency across subjects (Damoiseaux et al. 2006) and are identifiable by several characteristics including temporal frequency power predominantly below ~ 0.1 Hz (Cordes et al. 2001; McKeown et al. 2003). Resting-state networks are spatially similar to the neural networks activated by task performance (Smith et al. 2009). They are of pathophysiological significance to diverse brain diseases and are increasingly used to guide diagnostic and therapeutic decisions, for example in neurosurgical treatment planning for patients with epilepsy (Bettus et al. 2010; Boerwinkle et al. 2020; Leuthardt et al. 2018).

However, applications of resting-state fMRI in the clinical setting are limited by the high sensitivity of this technique to head motion (Power et al. 2012). Non-pharmacological strategies to minimise motion include distraction methods (e.g., audio-visual entertainment) (Anderson et al. 2013), pre-scan training with simulated (‘mock’) MRI environments (de Bie et al. 2010), encouraging subjects to fall asleep naturally (Fransson et al. 2011), and real-time prospective motion correction techniques (Maclaren et al. 2013). However, these approaches are of limited benefit in young or intellectually disabled patients who are unable to follow instructions or comply with entering the MRI scanner.

Lennox-Gastaut syndrome (LGS) is one such patient group posing significant challenges to MRI acquisition. LGS is a severe childhood-onset epilepsy associated with frequent seizures, intellectual disability, and complex behavioural, psychiatric, and physical co-morbidities (Archer et al. 2014). MRI is considered critically important in the diagnostic and treatment pathway for LGS: in particular, MRI-guided surgery can control seizures in some cases (Kang et al. 2018; Warren et al. 2017b).

General anaesthesia is routinely administered when acquiring clinically necessary MRI in LGS and other patient groups with similar disabilities (Malviya et al. 2000). For anatomical MRI sequences (e.g., T1- or T2-weighted scans), the choice of anaesthetic regimen is typically guided by the goals of limiting motion while maintaining patient safety and comfort, with propofol and sevoflurane being two commonly used agents. However, in the context of fMRI, an additional goal is the minimisation of anaesthesia’s influence on neuronal activity and neurovascular coupling underlying BOLD signal fluctuations (Petrinovic et al. 2016). While numerous studies have investigated the effects of anaesthesia on BOLD signals in animal models (Hutchison et al. 2010, 2011; Lu et al. 2012; Petrinovic et al. 2016; Vincent et al. 2007), comparatively little is known about the optimal anaesthesia regimens for fMRI in humans (for a review, see Hudetz 2012).

The combination of isoflurane and remifentanil is one regimen with several anaesthetic properties that may be well-suited to fMRI in young patients with epilepsy. Isoflurane is a volatile anaesthetic with rapid onset of action, short recovery time, and low blood solubility (Malviya and Lerman 1990). When combined with remifentanil, an intravenous opioid analgesic (Baker et al. 1997), a synergistic effect is achieved whereby the concentration of isoflurane required to prevent movement and maintain unconsciousness is significantly lowered (Allweiler et al. 2007; Criado et al. 2003; Lang et al. 1996). Combined isoflurane-remifentanil anaesthesia is commonly administered at our hospital for epilepsy neurosurgery (Freeman et al. 2003; Kannan et al. 2016) because low dosages of these agents are known to have limited influence on the frequency and morphology of interictal epileptiform discharges on intracranial EEG (Fiol et al. 1993; Herrick et al. 2002; Watts et al. 1999).

In this study, we aimed to test the feasibility of resting-state fMRI under combined isoflurane-remifentanil anaesthesia in children with severe epilepsy and intellectual disability due to LGS. Group-level independent component analysis (ICA) identified consistent patterns of correlated BOLD signal change across patients. To assess the presence of resting-state networks, we spatially compared ICA results in these anaesthetised patients to network templates available from the ‘Generation R’ study of non-anaesthetised healthy children (Muetzel et al. 2016). We hypothesised that isoflurane-remifentanil anaesthesia would permit observation of resting-state fMRI networks.

Material and Methods

Patients

Fourteen children with LGS were recruited through The Royal Children’s Hospital in Melbourne, Australia. Following participation, data from three patients were excluded: one due to MRI artefacts and two due to extensive structural brain abnormalities preventing accurate spatial warping of their brains to common template space. The final sample included 11 children (seven females and four males; mean age at study = 9.8 years, range: 2.8–15 years). All patients had severe intellectual disabilities: 10 were non-verbal, and all patients required assistance with activities of daily living (for example, eating and toileting). fMRI data were collected during patients’ clinically-requested MRI sessions. The work was carried out in accordance with The Code of Ethics of the World Medical Association (Declaration of Helsinki) for experiments involving humans. Before recruitment began, the study was approved by the human research ethics committee of The Royal Children’s Hospital, and written informed consent was given by each patient’s legal guardian. The fMRI data collected in these patients were the focus of two of our previous studies (Warren et al. 2017b, 2019).

fMRI under Isoflurane-Remifentanil Anaesthesia

For each patient, up to 30 min (mean = 27.9 min, range = 16.2–30 min) of fMRI was acquired in a 3T Siemens Trio MRI scanner using a T2*-weighted gradient-recalled EPI sequence (TR = 3,200 ms; TE = 40 ms; voxel resolution 3.4 mm × 3.4 mm × 3.4 mm [no gap]; 44 slices with interleaved slice acquisition). A T1-weighted anatomical image was also acquired. Anaesthesia was induced using inhaled sevoflurane in nine patients (inspired concentration up to 8%), intravenous propofol (50 mg) in one patient, and a combination of sevoflurane (8% inspired concentration) and propofol (30 mg) in one patient. Following induction, anaesthesia was maintained in all patients using a combination of inhaled isoflurane (≤ 0.8% end-tidal concentration in all patients; mean = 0.6%, range = 0.3–0.8%) and intravenous remifentanil (≤ 0.1 mcg/kg/min in all patients; mean = 0.06 mcg/kg/min, range = 0.04–0.1 mcg/kg/min). The average delay between anaesthesia induction and commencement of fMRI was 70 min (range = 40–100 min). During this delay, patients were prepared for scanning, including attachment of scalp EEG electrodes and recording of out-of-scanner EEG samples (for description of simultaneous EEG-fMRI analysis of interictal epileptiform discharges in these patients, refer to Warren et al. 2019). Throughout scanning, respiration was maintained using assisted mechanical ventilation, and patients were monitored using electrocardiography, pulse oximetry, end-tidal capnography, and scalp EEG. Patients had no sedative premedication and did not receive neuromuscular blocking agents or nitrous oxide or any other sedative drugs during induction or maintenance of anaesthesia. Patients’ EEGs showed features consistent with light sleep, including sleep transients (spindles and vertex waves), and did not show burst suppression or marked slowing.

fMRI Pre-processing

fMRI data were pre-processed using the fMRIB Software Library version 5.0.8 (FSL)Footnote 1, Advanced Normalisation Tools version 1.9Footnote 2, Freesurfer version 5.3Footnote 3, and iBrain software version 6.0Footnote 4. Each patient’s data were first temporally corrected to match the timing of the first slice acquired in each volume, and then spatially warped to Montreal Neurological Institute (MNI)-152 6th-generation template space (2 mm3 voxels) using (i) a rigid-body transform of each fMRI volume to the middle volume in each dataset; (ii) a rigid-body transform of each subject’s mean fMRI volume to their T1-weighted image, using boundary-based registration; and (iii) a non-linear transform of each patient’s T1-weighted image to MNI space. Individual-level brain masks were created by skull-stripping each patient’s mean fMRI volume using FSL’s brain extraction tool, after which a common group-level brain mask was created by intersecting all individual-level masks. fMRI data were denoised using the Spatially Organized Component Klassifikator (SOCK)Footnote 5, a heuristic and fully-automated classifier that identifies and reduces likely artefactual components from a decomposition of individual-level fMRI data using probabilistic ICA (Bhaganagarapu et al. 2013). We utilised SOCK’s default settings, which are designed to conservatively retain possible neuronal activity even if this means also retaining some potential artefact (Bhaganagarapu et al. 2013). As a final step, spatial smoothing was applied to the SOCK-denoised data using an isotropic Gaussian kernel with full-width-at-half-maximum of 8 mm.

Estimation of In-Scanner Head Motion

Mean frame-wise displacement (Power et al. 2012) was used to estimate head motion across fMRI volumes in each patient’s scan. For each instance of volume-to-volume movement, frame-wise displacement is expressed as a scalar value in mm and is derived from the six rigid-body parameters estimated during rigid-body fMRI volume realignment (Power et al. 2012). Mean frame-wise displacement represents the average displacement across all volumes in each scan.

Group Independent Component Analysis

Group-level spatial ICA was performed using the Group ICA for fMRI Toolbox (GIFT) version 4.0aFootnote 6 (Calhoun et al. 2001). The optimal number of independent components (ICs) was 37, as determined using the minimum description length criterion implemented in GIFT software (Li et al. 2007). To avoid prohibitive memory requirements, a two-stage principal component analysis (PCA) was used to reduce data dimensionality prior to ICA: first, PCA retained 100 dimensions from each patient’s dataset (Allen et al. 2011); second, reduced data from all patients were temporally concatenated and PCA was used to retain 37 dimensions using GIFT’s expectation-maximisation algorithm (Calhoun et al. 2001). Spatial ICA was then performed on the PCA-reduced group data using the Infomax algorithm (Bell and Sejnowski 1995) repeated 20 times in Icasso software (Himberg et al. 2004) to ensure reliability of the group-level ICs. Patient-specific spatial maps and timecourses for each IC were then estimated using group information guided ICA softwareFootnote 7 (Du and Fan 2013), which implements a multi-objective optimisation strategy that uses group-level spatial priors (in our case, the 37 group-level ICs estimated in the initial group ICA step described above) to compute individual-level ICs. As a final step, patient-specific ICs were z-transformed to have zero mean and unit variance.

Random-Effects Analysis of Component Spatial Maps

To explore regions within each IC showing significant involvement across the group, z-transformed patient-specific spatial maps for each IC were submitted to random-effects, whole-brain analyses using non-parametric, one-sample t-tests implemented in FSL’s Permutation Analysis of Linear Models (PALM)Footnote 8 software. Permutation was achieved via a sign-flipping procedure assuming symmetrically distributed errors (Winkler et al. 2014). A positive t-contrast (i.e., t > 0) was examined for each IC, and threshold-free-cluster-enhancement (Smith and Nichols 2009) was utilised for statistical inference with significance assessed at a voxel-level family-wise error (FWE)-corrected threshold of p < 0.05. To additionally correct for multiple comparisons across all 37 ICs examined, we used the -corrmod option available in PALM. Anatomical structures participating in each thresholded IC were identified by labelling significant voxels according to their positions within the Harvard-Oxford cortical and sub-cortical brain atlasesFootnote 9.

Spectral Analysis of Component Timecourses

Prior studies show that resting-state networks have temporal characteristics that may distinguish them from non-neuronal sources of BOLD signal change (Allen et al. 2011; Bhaganagarapu et al. 2013; Cordes et al. 2001; McKeown et al. 2003). To assess whether the networks identified here conform to these characteristics, each patient-specific IC timecourse was decomposed into its frequency power spectrum using multi-taper spectral estimation, as implemented in GIFT with Chronux softwareFootnote 10. Prior to calculating power spectra, IC timecourses were first temporally de-trended using default settings in GIFT, involving removal of the constant, slope, sine two cycles, cosine two cycles, sine one cycle, and cosine one cycle. From these power spectra we computed two metrics previously shown to distinguish between ICs of likely neuronal origin from ICs likely arising from noise (Allen et al. 2011): (i) dynamic range, defined as the difference between the power at the peak of the spectrum and the minimum power of frequencies higher than this peak; and (ii) low frequency (LF) to high frequency (HF) power ratio, defined as the ratio of the integral of power below 0.1 Hz to the integral of power between 0.1 Hz and the Nyquist frequency, equivalent to 1/(2*TR), or in our case 1/(2*3.2) = 0.156 Hz. We used a cut-off of 0.1 Hz when calculating LF:HF power ratio because prior studies indicate that resting-state networks typically show strongest BOLD correlations < 0.1 Hz whereas higher frequencies tend to contain greater contributions from non-neuronal sources (Allen et al. 2011; Bhaganagarapu et al. 2013; Cordes et al. 2001; McKeown et al. 2003).

Spatial Comparison with ‘Generation R’ Resting-State Network Templates

Using FSL’s fslcc functionFootnote 11, we calculated pairwise spatial cross-correlations between the 37 mean IC spatial maps observed here and 25 resting-state network templates from the ‘Generation R’ fMRI study of 536 non-anaesthetised healthy children aged 6–10 years (Muetzel et al. 2016). The Generation R templates were estimated using a group-level ICA approach comparable to the one we used, and are available to download in standard MNI template spaceFootnote 12. Prior to calculating spatial cross-correlations, the Generation R maps were first spatially masked using the group-level brain mask created for our study (see Sect. "fMRI Pre-processing" above). Anatomical structures involved in the Generation R maps are listed in the original paper (Muetzel et al. 2016). Each IC in our study was labelled according to the resting-state network label of the Generation R map showing peak spatial cross-correlation.

Results

In-Scanner Head Motion

Minimal head motion was observed throughout fMRI acquisition. The average mean frame-wise displacement across patients was 0.2 mm (± 1 SD = 0.3 mm; range = 0.03–1.2 mm), indicating that head movement was usually below the native resolution of a single fMRI voxel (3.4 mm3).

Group-Level Spatial Components

For each of the 37 IC spatial maps, brain areas showing significant involvement across patients are displayed in Fig. 1a–d. Anatomical regions participating in each result are listed in Table 1. All ICs showed specific areas of significant group-level involvement. Numerous resting-state networks commonly described in non-anaesthetised healthy subjects were identifiable, such as the default-mode (IC07), sensorimotor (IC03), and frontoparietal (IC34, IC36) networks. While some components showed a clear left- or right-hemispheric emphasis (for example, IC04 and IC05), most were bilaterally represented, both cortically (for example, IC07) and subcortically (for example, IC06). A small number were restricted to either cortical (IC03, IC11, and IC16) or subcortical (IC 01) structures only; however, the majority were spatially distributed over both cortical and subcortical sites. Of the 34 components with subcortical involvement, the most commonly involved areas were the thalamus (27 ICs), hippocampus (23 ICs), putamen (23 ICs), pallidum (20 ICs), brainstem (21 ICs), amygdala (17 ICs), caudate (17 ICs), cerebellum (16 ICs), and nucleus accumbens (13 ICs).

Fig. 1
figure 1figure 1

Axial views of 37 independent component (IC) spatial maps from the group-level independent component analysis. Warm-coloured areas within each spatial map indicate regions showing a significant difference from zero across the group (p < 0.05, corrected for family-wise error after threshold-free-cluster-enhancement). Spatial maps are coloured using a − log10 transformation of voxel-level p-values, where hotter (i.e., closer to yellow) colours indicate more significant areas. The z coordinates at the base of the image indicate axial positions (mm) in Montreal Neurological Institute (MNI)152 6th-generation space. Anatomical regions involved in each IC are listed in Table 1

Full size image
Table 1 Cortical and sub-cortical regions involved in each of the 37 independent component (IC) spatial maps shown in Fig. 1a–d (IC numbers in the leftmost column match the IC spatial maps in Fig. 1a–d). For each IC in the children with LGS, peak spatial cross-correlation (r) with ICs from the ‘Generation R’ (Gen. R) dataset is reported, along with the functional label of the Generation R IC showing peak cross-correlation (as described in the original publication; see Muetzel et al. 2016)
Full size table

Spectral Characteristics of Component Timecourses

For each IC timecourse, the mean (across patients) dynamic range is plotted against the mean LF:HF power ratio in Fig. 2. A strongly positive linear relationship was evident between these two metrics (Pearson’s correlation coefficient = 0.8). Example power spectra (and corresponding spatial maps) are shown in Fig. 3 for two components with high values of both metrics (IC07, IC37), and two components with low values (IC01, IC10). Components with high values tended show more tightly clustered spatial maps resembling resting-state networks (IC07, IC37), whereas components with lower values tended to have spatial features suggestive of substantial non-neuronal noise (Allen et al. 2011; Bhaganagarapu et al. 2013; Cordes et al. 2001; McKeown et al. 2003), such as voxels of peak significance located in the ventricles (IC01, IC33) or cerebral arteries (IC10, IC32).

Fig. 2
figure 2

Spectral characteristics of 37 independent component (IC) timecourses from the group-level independent component analysis. For each IC timecourse, the average (across patients) low frequency to high frequency power ratio is plotted on the x axis, and the average (across patients) dynamic range is plotted on the y axis. Each IC is labelled using a number inside (or adjacent to, for ICs with closely overlapping values) a green circle. This number matches the number of the IC spatial map displayed in Fig. 1 and described anatomically in Table 1

Full size image
Fig. 3
figure 3

Frequency power spectra of four example independent component (IC) timecourses and their associated spatial maps. The IC number matches Figs. 1 and 2, and Table 1. The upper two rows display example ICs with spectral features suggestive of a neuronal origin of fMRI signal change, including high dynamic range and high low frequency to high frequency power ratio (as per values displayed in Fig. 2), whereas the lower two rows display example ICs with spectral features suggestive of non-neuronal noise, including low values of the same metrics. The solid pink line in each plot (right) indicates the mean (across patients) frequency power, and the surrounding shaded purple area indicates ± 1 standard error of the mean. The spatial maps (left) are thresholded using the strategy described in Fig. 1, and are coloured using a − log10 transformation of voxel-level family-wise error-corrected p-values. The z and y coordinates shown below each map indicate axial and coronal positions (mm), respectively, in Montreal Neurological Institute (MNI)152 6th-generation space

Full size image

Spatial Comparison with ‘Generation R’ Resting-State Network Templates

For each IC map, peak spatial cross-correlation with the Generation R maps is shown in Table 1, along with the resting-state network label of the Generation R map with peak cross-correlation. Overall, ICs extracted from anaesthetised children with LGS were well-represented by maps in the Generation R dataset (average peak cross-correlation across all ICs = 0.5; range = 0.2–0.8). Six example ICs with very strong similarity (r ≥ 0.7) to a corresponding map in the Generation R dataset are displayed in Fig. 4. These included a variety of resting-state networks associated with diverse functional roles, including the sensorimotor (IC03), motor (IC02), anterior visual (IC24), cerebellar (IC09), default-mode (IC07), and parietal (IC16) resting-state networks. Of interest, timecourses of these six ICs tended to have higher-than-average dynamic range and LF:HF power ratio (Fig. 2).

Fig. 4
figure 4

Six example independent component (IC) spatial maps in the anaesthetised children showing high spatial correlation with a corresponding IC from the ‘Generation R’ study of non-anaesthetised healthy children. Maps for the anaesthetised children are thresholded using the strategy described in Fig. 1, and are coloured using a − log10 transformation of voxel-level family-wise error-corrected p values. The Generation R maps are thresholded at z = 3.09 (corresponding to p < 0.001) as in the original paper (Muetzel et al. 2016). The IC number provided for each spatial map in the left column matches Figs. 1 and 2, and Table 1. The label provided for each spatial map in the right column (sensorimotor, motor, etcetera) matches the label used in the original paper (Muetzel et al. 2016) and the label used in Table 1. The z, y, and x coordinates below each spatial map indicate axial, coronal, and sagittal positions (mm) in Montreal Neurological Institute (MNI)152 6th-generation space, respectively

Full size image

Discussion

Our study demonstrates the feasibility of resting-state fMRI in children with severe epilepsy and intellectual disability scanned under combined isoflurane-remifentanil anaesthesia. Group-level ICA showed that resting-state networks commonly studied in non-anaesthetised healthy children (Muetzel et al. 2016) could be detected in the anaesthetised patients. ICA timecourses associated with these networks showed spectral characteristics consistent with a likely neuronal origin of BOLD signal fluctuations (Allen et al. 2011; Bhaganagarapu et al. 2013; Cordes et al. 2001; McKeown et al. 2003), including high dynamic range and frequency power predominantly below 0.1 Hz. Additionally, components with spatial and spectral features suggestive of non-neuronal noise (e.g., physiological artefacts) could be readily distinguished from resting-state network-related BOLD changes.

Previous human studies have observed resting-state networks under other anaesthetic and sedative agents, including sevoflurane (Peltier et al. 2005), propofol (Bisdas et al. 2016; Boveroux et al. 2010; Liu et al. 2017), and midazolam (Greicius et al. 2008). The majority of these studies were performed in healthy adults, with few studies reported in young children or clinical populations. In our study we extend these findings in three ways: we show that resting-state networks are observable (i) under anaesthesia maintained using combined isoflurane with remifentanil; (ii) in anaesthetised young children; and (iii) in patients with severe epilepsy and intellectual disability due to LGS.

These results build on our recent study in this patient group using simultaneous EEG-fMRI, where we found that BOLD activation during interictal epileptiform discharges in LGS was detectable under isoflurane-remifentanil anaesthesia (Warren et al. 2019). The present study extends these findings by showing that resting-state network activity (i.e., BOLD fluctuations occurring beyond the time of discharges alone and referable to intrinsic brain functions) is also observable under this regimen.

Our observations may be relevant to emerging clinical applications of resting-state fMRI. For example, spatial mapping of networks located in eloquent cortex (such as the sensorimotor and visual networks displayed in Fig. 4) may assist with neurosurgical planning, particularly in patients with cognitive or behavioural impairments who cannot tolerate awake or non-anaesthetised scanning, and in those with distorted cortical anatomy in whom functional mapping with resting-state fMRI can assist with defining safer surgical targets (Boerwinkle et al. 2020; Leuthardt et al. 2018).

Resting-State Networks under Isoflurane-Remifentanil Anaesthesia

Previous animal studies have explored the effects of isoflurane or remifentanil separately on BOLD patterns. The spatial distributions of numerous resting-state networks described in healthy non-anaesthetised humans (including default-mode and frontoparietal networks) are observable under isoflurane anaesthesia in rodents (Hutchison et al. 2010) and macaques (Hutchison et al. 2013), and in rats anaesthetised using combined isoflurane with dexmedetomidine (Lu et al. 2012). When administered in low dosages similar to that provided in our study, effects of isoflurane on some physiological parameters underlying the BOLD response appear to be minimal: for example, in humans (Newman et al. 1986) and animal models (Hoffman et al. 1991; Todd and Drummond 1984), cerebral blood flow during isoflurane anaesthesia at dosages ≤ 1 MAC (minimum alveolar concentration required to prevent movement) does not significantly differ from the awake state. Additionally, deoxyhemoglobin-weighted optical imaging studies in isoflurane-anaesthetised macaques (Shtoyerman et al. 2000) and felines (Fukuda et al. 2004) have found that the shape of the hemodynamic response to visual stimulation is similar to that seen in awake animals, although the response amplitude is reduced.

Fewer studies have explored the effects of remifentanil during fMRI. One study in macaques found persistence of within-network dynamic functional connectivity during anaesthesia maintained using remifentanil in combination with mivacurium chloride (Hindriks et al. 2016). Additionally, two human studies of patients undergoing fMRI in an intra-operative setting found that combined remifentanil and propofol anaesthesia did not suppress resting-state networks (Bisdas et al. 2016; Roder et al. 2016). Remifentanil is an ultra-short-acting µ-opioid agonist with an elimination half-life < 10 min (Baker et al. 1997) and is thus thought to have limited influence on neurovascular coupling in brain areas beyond those with high opioid receptor densities, such as pain processing circuits (Hindriks et al. 2016; Wise et al. 2002). In humans undergoing craniotomy, remifentanil dosage concentrations similar to that provided in our study (≤ 0.1 mcg/kg/min) appear to have minimal effects on cerebral blood flow, vasodilation, and spectral properties of brain activity measured by EEG (Baker et al. 1997; Engelhard et al. 2004).

However, deeper stages of anaesthesia can have significant impacts on resting-state networks. Using ICA of fMRI data in macaques, Hutchison et al. (2014) examined dose-dependent effects across six isoflurane levels ranging from 1 to 2.75% concentration in medical air. Dosages below 1.5% were associated with stable patterns of ICA-derived resting-state networks, whereas breakdowns in functional connectivity were observed at dosages exceeding 1.75%, including loss of inter-hemispheric connectivity patterns for networks that are normally expressed bilaterally. Although we did not explore specific dose-dependent effects here, our results using ≤ 0.8% isoflurane did not appear to abolish inter-hemispheric connectivity of bilateral networks (Fig. 4), suggesting that stable patterns of connectivity seen by Hutchison et al. (2014) in macaques under low-level isoflurane are also preserved at similarly low dosages in humans.

The combination of isoflurane and remifentanil may have additional advantages for fMRI in patients with epilepsy specifically. Previous electrophysiological studies show that low dosages of these agents have minimal influence on the occurrence of epileptiform discharges (Fiol et al. 1993; Herrick et al. 2002; Watts et al. 1999), suggesting that this anaesthetic regimen is well-suited to performing simultaneous EEG-fMRI studies in epilepsy, where capturing discharges with similar morphology to those seen during patients’ routine out-of-scanner EEG recordings is typically desired. A similar preservation of discharges has been described in intracranial EEG studies of epilepsy patients anaesthetised using other agents, including dexmedetomidine (Oda et al. 2007; Talke et al. 2007).

Resting-State Networks and Intellectual Disability

Resting-state networks were spatially similar to networks previously described in non-anaesthetised, cognitively typical children from the Generation R study (Muetzel et al. 2016). High similarity was observed in networks normally associated with a range of cognitive and sensory functions, including sensorimotor, default-mode, and parietal networks (Fig. 4). Persistence of these networks in children with severe intellectual disability appears somewhat at odds with the hypothesis that the emergence of resting-state networks occurs in concert with functional skill acquisition during neurodevelopment (Fair et al. 2008; Supekar et al. 2009; Uddin et al. 2010). Contrary to this view, the majority of children with LGS in our study experienced seizure onset in the first year of life, and subsequently failed to reach numerous developmental milestones, including language acquisition.

Several mechanisms may explain the persistence of resting-state networks in these extreme cases of atypical neurodevelopment. One possibility is that the spatial topography of resting-state networks emerges at an early stage of development, prior to epilepsy onset. This hypothesis is supported by recent fMRI studies showing that key elements of many resting-state networks, including the sensorimotor and default-mode networks, are already observable in the brains of human foetuses in utero (Seshamani et al. 2016) and infants born pre-term (Doria et al. 2010). These findings suggest that the core arrangements of some networks may be established prior to birth: this may confer a degree of resilience to pathologies manifesting postnatally, such as seizures and LGS.

A second possibility is that the topographies of resting-state networks continue to evolve following epilepsy onset, but their developmental trajectories are decoupled from functional skill acquisition, instead supporting abnormal expression of the epileptic process—in other words, the “hardware” of resting-state networks is co-opted to sustain a pro-epileptogenic state (Archer et al. 2014). In the context of severe early-onset epilepsies like LGS, this perhaps implies that the functional labels normally used to describe resting-state networks (for example, ‘executive-control’, ‘default-mode’, or ‘dorsal-attention’) may be misnomers, at least to the extent that the spatial arrangement of an affected network may not reflect consolidation of the associated functional skill (for example, executive function, language, or visuospatial attention), but rather a set of brain regions that are functionally related by their shared involvement in a specific seizure type or cognitive/behavioural co-morbidity. In support of this hypothesis, EEG-fMRI studies in childhood epilepsy have shown that BOLD patterns during epileptic activity often closely mirror resting-state networks. For example, generalised spike-and-wave discharges in childhood absence epilepsy are associated with BOLD deactivation in a set of brain regions closely resembling the default-mode network, and activation within areas of the anterior-salience network (Carney et al. 2010). Additionally, bursts of generalised paroxysmal fast activity in children with LGS are expressed via BOLD activation in the frontoparietal executive-control and default-mode networks (Warren et al. 2017b; Warren et al. 2019).

Importantly, our observation of resting-state networks in anaesthetised children with LGS does not imply these networks are unaffected by anaesthesia or epilepsy. We have previously shown that functional connectivity strength among resting-state networks in non-anaesthetised patients with LGS is different to healthy controls (Warren et al. 2017a; Warren et al. 2016). Anaesthetic agents also influence the strength of functional connectivity among networks, even at very low dosages (Hudetz 2012). These results, taken together with the present study, suggest that anaesthesia and epilepsy may have greater influence on the temporal interactions of resting-state networks than on their spatial arrangements per se. Compatible with this notion is the finding that healthy ageing appears to have greater impact on connectivity strength than on networks’ spatial distributions (Doria et al. 2010).

Conclusions

We demonstrate the technical feasibility of resting-state fMRI in children with LGS scanned under isoflurane-remifentanil anaesthesia, highlighting that the emerging clinical applications of resting-state fMRI (for example, neurosurgical planning) may be extendable to the wide variety of paediatric patient groups in whom awake or non-anaesthetised scanning is impractical. Resting-state networks found under anaesthesia were spatially similar to networks previously described in non-anaesthetised healthy children (Muetzel et al. 2016). Observation of these networks in children with severe epilepsy and intellectual disability provides evidence for the hypothesis that resting-state networks are an intrinsic feature of brain organisation that may persist despite neurodevelopmental failure to acquire the specific functional skills that are normally ascribed to the spatial arrangements of these networks.