Data from 22 healthy subjects (11 females) served as control in this study (age 35 ± 12 years), and data from 27 epilepsy patients [13 classified as left TLE (age 32 ± 9 years), 14 with right (age 33 ± 11 years); 16 females], who were selected from a larger dataset with the inclusion criteria of TLE based on electro-clinical information-informed expert classification (Table 1).

In this study, we assumed the presence of an “epileptic process” possibly reaching beyond epochs of interictal epileptic activity visible on scalp EEG. Accordingly, the EEG was not used for the detection of epileptiform EEG activity but to identify fMRI epochs of wakefulness in order to increase the sensitivity and specificity of our results (12).

EEG via a cap (modified Brain Cap MR, Easycap, Herrsching, Germany) was recorded continuously (sampling rate 5 kHz, low pass filter 1 kHz) during fMRI acquisition with an MRI-compatible EEG system (BrainAmp MR⁺ and Brain Vision Analyzer; Brain Products GmbH, Gilching, Germany) yielding two 20-min sessions consisting of 404 T2*-weighted single-shot gradient-echo echo-planar images (EPIs; echo time/repetition time, 30/3000 ms; flip angle, 90°; 43 2.5 mm interleaved slices; FOV, 24 cm × 24 cm; matrix 64 × 64) acquired continuously on a 3T Signa Excite HDX MRI scanner (General Electric, Milwaukee, WI, USA).

MRI and pulse artifact correction were performed based on the average artifact subtraction (AAS) method (13, 14) as implemented in Vision Analyzer 2 (BrainProducts, Germany) resulting in EEG with a sampling rate of 250 Hz. EEG was re-referenced to common average. Sleep stages were scored manually by an expert according to the AASM criteria (15). Epochs other than wakefulness were erased from the analysis (by excluding the corresponding BOLD time courses from both the variance and dynamic functional connectivity analyses). Epochs of wakefulness shorter than 2 min were not included in the analysis.

Using Statistical Parametric Mapping (SPM 8, http://www.fil.ion.ucl.ac.uk/spm), EPI data were realigned, normalized (MNI space), and spatially smoothed (Gaussian kernel, 5 mm³ full width at half maximum). The time course of the average signal at the ventricles (CSF as given by FSL’s Ventricle Mask standard image) as well as motion-induced noise was regressed out. We did not remove the global brain signal to avoid the issue of induced anti-correlations (16). Also, since there are direct electrophysiological correlates of the global resting-state signal (17), we believe that its removal is arbitrary and results in a loss of information. fMRI data was band pass filtered in the range 0.01–0.1 Hz using a sixth order Butterworth filter as described previously (8). In all cases, both 20 min fMRI sessions were analyzed and then the results were averaged within every subject (to guarantee the assumption of statistical independence) prior to statistical testing.

Subjects displayed similar degrees of head displacement across groups. Controls: 0.06 ± 0.02 mm, Left TLE: 0.08 ± 0.04 mm, Right TLE: 0.08 ± 0.05 mm. No significant differences (p > 0.05) between groups were found. As an additional control, we reproduced the results after erasing volumes associated with high ( >0.3) head displacement amplitude (see Appendix).

Both maps of temporal variability and of dynamical functional connectivity variability are compared between groups by means of mass univariate t-tests as implemented in the SPM8 software. All maps are reported at a level of p < 0.001 uncorrected with only clusters passing a threshold of p < 0.05 FWE corrected being shown.

The first observation that in our group of TLE patients, we found the highest BOLD signal variance in the anterior temporal lobe including, of note, the hippocampus supports our hypothesis of the presence of paroxysmal interictal activity resulting in regionally increased BOLD signal variance. It is well-established that local field potentials (LFP) explain the largest portion of this variance (25) and that interictal epileptic spikes in the anterior temporal lobe are reflected in LFP changes (26). We hence propose interictal epileptic activity (“epileptic process”) as the most likely explanation for the high anterior temporal lobe variance common to our TLE group. Kobayashi and colleagues as well as ourselves previously identified hippocampal BOLD signal increases at the group level in TLE, when correlating by means of a general linear model (GLM) interictal scalp EEG activity with the BOLD signal (27, 28). For these GLM-based EEG-fMRI studies, the occurrence of a number of interictal spikes was essential, whereas in the present approach, we on purpose designed our analysis to be independent of scalp EEG interictal activity. We did this for two reasons: (1) the sensitivity of scalp EEG in TLE is compromised with respect to hippocampal interictal epileptic activity (3, 29), allowing the creation of a scalp EEG-based GLM with limited sensitivity only and (2) we sought to identify an fMRI marker for paroxysmal activity specific to the epileptic condition independently of the simultaneously recorded EEG. Such paroxysmal activity in addition to “true” interictal spikes (as can be demonstrated with intracranial EEG) might include other types of epileptic (neuronal) activity leading to bursts of energy consuming processes, e.g., owing to damaged cells in the hippocampus, alterations in the glial milieu, or pathology of blood flow parameters (30). The exclusive selection of epochs during which the patients and control subjects, respectively, were awake ensured increased sensitivity when comparing resting-state BOLD signal properties: it is well known that the functional architecture of the brain changes significantly depending on the level of wakefulness (12, 31) including the variance of the BOLD signal in different brain regions (11, 32), but also graph theoretical network measures (33, 34), and variance in the brain’s functional connectivity structure (8).

We note that a very similar result (increased variance of BOLD signals in bilateral hippocampi) was recently found for subjects under the influence of a psychedelic substance (psilocybin) (35). Furthermore, a related result is discussed in the context of REM sleep (36), and in the psilocybin study the intensity of the “dream-like” experience, which was rated subjectively by the subjects, correlated positively with the magnitude of the hippocampal BOLD signal fluctuations. It can be speculated that subjective alterations in the conscious awareness of the patients – more subtle than full-blown seizures, but which can escalate into the former and also into auras – are shared in TLE and in the other altered states of consciousness mentioned above.

Increased variance of dynamic functional connectivity in TLE compared to healthy controls indicates region pairs, which exhibit stronger fluctuations of the correlation of their BOLD time series in the epileptic condition than in the healthy. While the measure does not assess an absolute difference in connectivity between the two cohorts, regions are identified, which link and unlink typically more in TLE than in the control group. Our study design motivates the hypothesis that the “epileptic process” is responsible for this coupling and uncoupling, possibly in the form of the interictal epileptiform activity in the hippocampus, which leads to increased functional connectivity variance. Further analyses should reveal whether such epileptic activity weakens or strengthens the functional connection between regions. It may even be that “erratic” hippocampal coupling (and uncoupling) with cortical regions disrupts the physiological interplay of the latter with yet further parts of the brain – including subcortical and that this eventually results in the observed dysfunctionality.

Van Paesschen and colleagues used single photon emission computed tomography (SPECT) to study patients with hippocampal sclerosis (HS) and observed both ictal hyper- and hypoperfusion in (lateralized) regions partly overlapping with those we describe here in our study, which also includes many cases of HS. Hyperperfusion was described in temporal lobe, middle frontal and central regions, and in the frontal lobes and the precuneus hypoperfusion was found (37). It hence appears that both a “gain” as well as a “loss of function” in different brain regions is linked with ictal dysfunctioning. For example, Chassagnon and colleagues studied patients with mesial TLE and found ictal–interictal hypoperfusion in the posterior cingulate and prefrontal regions, which might be interpreted as a loss of function in the sense of impaired consciousness (precuneus) and executive functioning (frontal regions) (38).

Looking at the particular brain regions in TLE with increased connectivity variance, we speculate that (1) the hippocampal activity interferes with language (39) and memory function (40), both interictally and ictally and that (2) the increased dynamic connectivity to the precuneus and frontal cortex is ictally associated with impaired consciousness (4, 41) and executive functioning (42). We also propose that (3) the increased dynamics in functional connectivity between the hippocampus and the sensorimotor cortices might pave the way for ictal sensory and motor dysfunction and – probably tightly related to the supplementary motor area – in particular motor automatisms (43). The superior frontal gyrus was also described relevant for introspection (44).

Our study does not offer any objective clues as to what the changes we describe in the TLE group interictally have to do with ictal changes in brain function. While it is well described that language and memory impairment are present interictally (45), sensorimotor dysfunction does not obviously occur interictally. One explanation could be that a qualitative difference between what is called interictal and ictal activity may not exist as such, but rather a quantitative one: Binnie a good decade ago pointed out that if only we tested carefully enough, transitory cognitive impairment could be related to “interictal” activity in many individuals (46). In our own EEG-fMRI study looking at BOLD signal changes related to formally interictal activity, we interpreted signal changes in regions of the so-called default mode network to explain reduced consciousness during dyscognitive seizures. However, we selected for the study cohort patients with very frequent interictal discharges on the EEG increasing the sensitivity of our discharge-correlated GLM analysis (27). With this in mind and taking Binnie’s idea forward, it is conceivable that with so-called interictal activity vastly the same set of brain regions (network) is recruited as is ictally with the difference that when behavioral alterations become obvious they are called seizures and hence define “ictal activity.” Of course, some additional features distinguish seizure from interictal activity going beyond “duration” alone but include spreading of epileptic activity. Still, such spreading of activity might occur along interictally pre-existing paths (47). Supporting our speculation further, in TLE, structural changes have been shown to progress over time and memory function was more closely related to structural hippocampal changes than the overt seizure frequency: the group of Bernasconi found neocortical thinning in TLE progressing over time in bilateral frontal (sensorimotor) and temporal (hippocampal, entorhinal, temporo-polar) regions – overlapping with the regions we report here (48), and Pacagnella and colleagues most recently presented data proposing that memory impairment is more influenced by hippocampal damage than by seizure frequency (40). In our limited cohort, we failed to identify a correlation between epilepsy duration and BOLD signal variance or dynamic functional connectivity (analysis not shown).

We did not find any significantly lateralized or side-dependent results. Instead, significant differences between healthy subjects and controls were usually bilateral. This might be due to a lack of sensitivity of our analysis, which was not optimized for this purpose (balancing of handedness, type of left- and right-sided pathologies, and EEG abnormalities). We hence do not discuss our lateralized results. In general, however, on first sight, a lack of lateralized findings is surprising taking into account clinical practice and surgical success with unilateral resections and the efforts spent with non- and invasive video-EEG telemetry and imaging to identify in which temporal lobe (hippocampus) the epileptogenic zone is located (49). Although it is clinically not a contested issue that it is relevant indeed to operate on the correct side of the brain, we are not aware of any systematic review of epilepsy surgery cases in which – for whatever reason – retrospectively the wrong side was operated upon. It is likely that the vastly symmetrical results we present reflect secondary bilateral “network” effects of a lateralized primary cause.

It is equally interesting that we found exclusively variance increases – and not any decreases – of BOLD signal amplitude and hippocampal functional connectivity in patients with TLE compared to control subjects. However, we would like to note that this does not rule out the possibility of decreased absolute functional inter-regional connectivity in the epilepsy cohort, as this differs from the variance of functional connectivity, which we report in our study and the interpretation of which is discussed above. Still, a relationship between the two measures might exist because many of the regions in which we found altered dynamic functional connectivity were also reported by Haneef and colleagues. They compared TLE patients to controls and found that the classical measure of static hippocampal functional connectivity was greater to the bilateral temporal lobes, insula, fornix, frontal poles, angular gyrus, basal ganglia, thalamus, and cerebellum. They found reduced connectivity with the occipital pole, calcarine, lingual, precuneus, sensorimotor cortex, and parts of insula and frontal lobes as well as medial frontal areas (50). We report results of a “Standard Seed Correlation Analysis” of our data in the Appendix. Because we analyzed wakefulness epochs exclusively, comparability to other data remains limited.

We are aware of the many factors influencing resting-state – and any other – fMRI studies (51) but not aware of any study formally assessing the order of relevance of the numerous confounding effects. We do know, though, from our own data (12, 31) that sleep alters the neuronal resting-state brain architecture significantly; and motion is known to introduce BOLD signal changes of several magnitudes the size of those commonly induced neuronally (52). In comparison, e.g., effects of sex and age are less pronounced both in terms of extent and intensity, and only optimized analysis methods will reveal such differences (53, 54). Nevertheless, we tried to match both gender and age as much as possible between the examined cohorts. We regressed rigid body motion from the data and in an additional analysis accounting for motion-induced variance in a very conservative way showed that our results were robust w.r.t. motion-induced variance (see “Reproduction of Results Erasing High Amplitude Head Movements” in the Appendix; Figures 5 and 6).

Regarding confounding fluctuations in wakefulness, to our knowledge, this is the first “wakeful rest” functional connectivity study controlling for and restricting the analysis to true, EEG-defined, awake epochs only and in addition accounting for non-stationarity of the functional connectivity even within the awake state. We advocate both as a desirable standard. To demonstrate independence of our results to the length of the sliding window used in the dynamic functional connectivity analysis, we performed an additional analysis with a different (shorter) window length (see “Robustness against Changes in Sliding Window Length” in the Appendix, Figure 7).

We did not make a strong point about left TLE vs. right TLE comparisons, because we did not observe significant group differences. Evidently, absence of proof is not proof of absence but may be due to patient heterogeneity. Such is almost inevitable in patient studies like this one caused by a variety of factors, such as the range of dyscognitive seizure semiology (not any two patients have the same seizures), seizure frequency, and the potential occurrence of additional generalized tonic clonic seizures, type and location – or by current radiological standards even absence – of structural pathology visible on MRI, IQ, duration of epilepsy, and medication – just to name a few. Although epilepsy syndrome diagnosis relied on multidisciplinary experts reviewing extensive electro-clinical information, we cannot rule out classification errors. But differences within the TLE cohort between left and right TLE patients will not affect our main positive findings, i.e., the group differences between TLE and healthy controls. However, we note that any medication taken by the patients if systematically leading to, e.g., alterations of consciousness could have biased our results; but there is no evidence to date supporting this possibility for the specific patterns we observed.

In addition to these mere technical issues, like in any resting-state study, result interpretation conceptionally is limited given the absence of a task. We based the justification of the study design on the fact that interictal epileptic activity occurs spontaneously at rest as indicated by neuronal discharges measurable with EEG, although we assumed a more general “epileptic process” based on the well-established general clinical observation of interictal cognitive compromitation in TLE. In an attempt to tie our interpretation of our observations more closely to the results, we regressed IQ against BOLD signal variance and dynamic functional connectivity. We report this analysis only in “Correlations Between BOLD Signal Variance/Variance of Dynamic Connectivity Time Series with the Left Hippocampus and VIQ/PIQ” in the Appendix as it needs to be considered of anecdotal character owing to a lack of statistical significance. Causes of the latter include those discussed above.