Altered physiological brain variation in drug-resistant epilepsy

2.1 Participants

The study sample consisted of 10 patients with DRE (age 34.5 ± 10.9 years, seven females, Table 1) and 10 age- and sex-matched healthy controls (HC, age 34.4 ± 11.3 years, seven females). We included also another control group for verification purposes (HC_2nd, 10 subjects, age 23.7 ± 2.2 years, eight males). The patients were recruited from the outpatient department of neurology at Oulu University Hospital. The patients were diagnosed with focal epilepsy with ongoing seizure activity despite the use of two to four antiepileptic drugs. The patients either had been or were to be evaluated for the possibility of epilepsy surgery or electrical stimulation treatments. The mean age at epilepsy onset was 24.2 ± 12.8 years. Subjects were scanned between 2012 and 2016.

Written informed consent was obtained from all of the participants according to the Declaration of Helsinki. The Ethics Committee of the Northern Ostrobothnia Hospital District, Finland, approved the research protocol.

2.2 Multimodal data acquisition

Multimodal imaging was carried out using our Hepta-scan concept, in which fast fMRI data was imaged simultaneously and synchronously with EEG, noninvasive blood pressure (NIBP), and near-infrared spectroscopy (NIRS), and anesthesia monitor data (Korhonen et al., 2014). We also utilized the MRI scanner′s own respiratory bellows and peripheral capillary oxygen saturation meters. All data were synchronized with the scanner trigger-pulse. Anesthesia monitor data were used for verification purposes.

2.3 Functional MR imaging

All subjects were scanned using a Siemens MAGNETOM Skyra 3T MRI scanner (Siemens Healthcare GmbH, Germany) with a 32-channel head coil. We utilized a MREG sequence (Assländer et al., 2013; Lee, Zahneisen, Hugger, LeVan, & Hennig, 2013; Zahneisen et al., 2012), which includes a three-dimensional (3D) single-shot undersampled spiral trajectory with the following sequence parameters: repetition time (TR) = 100 ms, echo time (TE) = 36 ms, field of view (FOV) = 192, matrix 64³, 3-mm cubic voxel, and flip angle = 25°. MREG data were reconstructed by L2-Tikhonov regularization with lambda = 0.1, with the latter regularization parameter determined by the L-curve method (Hugger et al., 2011). An analysis of the point-spread function revealed that the resulting effective spatial resolution was 4.5 mm. This resting state MREG data acquisition lasted 10 min and patients were imaged 2–4 times consecutively with 2–5 min apart, because we intended to capture possible spontaneous epileptiform activity. Because all patients had at least two scans, both of them were further analyzed. HC group were imaged only once and some only for 5 min due to schedule reasons, because epileptiform activity was not to be expected.

Subjects were given ear plugs to reduce noise and soft pads were fitted over the ears to minimize motion. During scanning, all participants received the instructions to simply rest, stay awake, and focus gaze on a cross on a screen, which they saw via a mirror mounted on the head coil. High-resolution T1-weighted MPRAGE (TR = 1,900 ms, TE = 2.49 ms, TI = 900 ms, FA = 9°, FOV = 240, slice thickness 0.9 mm) images were obtained for co-registration of the MREG data to Montreal Neurologic Institute (MNI 152) in 4 mm ISO standard space.

2.4 Cardiorespiratory data and electroencephalography data

Anesthesia monitoring signals (PPG, ETCO₂, ECG, and cuff-based blood pressure) were also registered using 3T MRI-compatible anesthesia monitor (GE Datex-Ohmeda^™; Aestiva/5 MRI). In some cases, the ETCO₂ data measured from by the anesthesia monitor were corrupted. Thus, we used scanners inbuilt PPG and respiratory bellows data for groups comparisons.

EEG was recorded with an MR-compatible BrainAmp system (Brain Products, Gilching, Germany) with 32 Ag/AgCl electrodes (including one ECG electrode) placed according to the international 10–20 system. To get low electrode impedances (<5 kΩ), the skin potential was removed with the stick abrasion technique (Vanhatalo et al., 2003). Data sampling rate was 5 kHz and band pass from DC to 250 Hz. Signal quality was tested outside the scanner room by recording 30-s eyes open and eyes closed. MR-scanner optical timing pulse and BrainAmp SyncBox were used to ensure that the EEG and fMRI data were in synchrony. The amplifier was placed outside the bore, and cables were stabilized to avoid motion artifacts.

2.5 Data preprocessing

2.6 Coefficient of variation (CV_MREG) maps

Figure 1 illustrates the main analysis schema of the study. CV was calculated as follows:

(1)

where X is signal time series, μ is mean, and σ is standard deviation (SD). For the fMRI data, fslmaths was used to calculate voxel-wise time domain signal mean (μ) and SD (σ) values for every subject. Calculation of CV_MREG was executed for both full-band and filtered data in the same way. Whole brain global signal values were calculated using fslmeants for full band MREG data. From this individual mean signal, the signals μ, σ, CV_MREG were calculated. For the physiological verification data (PPG, respiratory waveforms), these variables were calculated in MATLAB.

The normal range of voxel-wise CV_MREG values were calculated within the HC group to produce a voxel-wise threshold map of CV_MREG. The threshold voxel-wise CV map CV_{MREG_thr} was calculated as follows:

(2)

where μ is mean value of CV_MREG in HC group, and σ is its SD (Figure 1). This threshold encompassed the normal range of CV values. Individual CV_MREG maps in the epilepsy group were then thresholded at CV_{MREG_thr} to identify voxels above this range. As patients had two measurements, the thresholded maps from each measurement were binarized and combined, preserving only the overlapping voxels from both measurements that exceeded CV_{MREG_thr}. Thresholding was performed only for full band data, because of the previously mentioned voxel-wise adding of full-band mean to filtered data. We visually inspected the corresponding areas exceeding the threshold and compared them to clinical symptoms.

2.7 Voxel-wise analysis between-group differences in CV maps

The contrasts between the localization of CV_MREG values of two groups were analyzed with FSL's randomize with 10000 random permutations implanting threshold-free cluster enhancement (TFCE) correction in both directions (HC > DRE, HC < DRE) separately (Smith & Nichols, 2009). The t-statistic maps with corrected p-values (p < 0.05) were created to evaluate statistical significance of the CV_MREG maps between the group (Nichols & Holmes, 2002; Winkler, Ridgway, Webster, Smith, & Nichols, 2014). This analysis was performed with full band and filtered bands accordingly.

2.8 Template based analysis and global signal

For the template-based analysis the gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF), masks were created for standard MNI 2 mm space and were used to mask CV maps. The mean values of masked CV_MREG values were calculated afterward to compare groups.

The global signal was calculated from masked, with brain only MNI space voxels mean (FSL command: fslmeants) MNI space voxels mean. This mean was calculated for every subject over time (5 min) to get global signal over time. The mean and SD was calculated from previously acquired global signal for the CV calculation.

2.9 Evaluation of potential sources of CV differences

3.1 Individual mapping of DRE patients

Thresholding of DRE patients’ CV maps with CV_{MREG_thr} enabled the detection the areas of elevated CV_MREG >3SD in each patient separately. We further investigated whether the overlapping areas include only voxels that were above the CV_MREG limit in both first and second measurements (CV_MREG > μ +3σ). There were no significant differences between groups and individuals with decreased values (CV_MREG < μ-2σ).

All of the DRE patients presented a reproducible CV_MREG abnormality (Figure 2) at the individual level. The repeatedly detected CV_MREG abnormality was linked to the anatomical area triggering individual seizure manifestation in eight DRE patients (Figure 2). With four patients, there were also other interesting areas (Figure S1). Importantly, no voxels remained after CV_thr thresholding in the HC group (Figure 6b).

Subjects with DRE tend to have abnormal autonomous nervous system activity. In terms of autonomous nervous system activity, seven of the 10 DRE patients had altered CV_MREG in the brainstem at or near autonomous nervous system nuclei (Figure 3). The individual differences were mostly located in the ventral and dorsal respiratory areas near solitary and ambiguus nuclei, which are part of the reticular activating system responsible for autonomic nervous system control. These areas also showed significant differences in all group analyses (Figures 3,4 and 5).

3.2 Spatial CV map differences between groups

The CV_MREG of DRE patients was significantly increased compared to HC (p < 0.05, Figure 4a). In the full-band fMRI data, differences were predominantly located in the left hemisphere. Symmetrical changes were detected in primary sensorimotor cortices extending to parietal lobule and to visual cortices V1–V3. Major frontal lobe changes were predominantly located in left cerebral cortex. Additional significant changes were located in left callosal body, the cingulum, in the upper and lower brainstem in ventral and dorsal respiratory centers, and in the cerebelli, where the most dominant areas were left crus I–II, left VII–VIII, vermis, right crus II, right VII–VIII and small lobes in left and right I–IV.

The second measurement showed an increase in the CV_MREG difference between patients and controls. At a lowered threshold of p < 0.01, extensive differences between DRE patients and HC were observed (Figure 4b). Compared to first measurement, group differences were both more widespread and symmetrical.

3.3 Differences between brain pulsation frequencies

Group level analysis of different sub-bands revealed significant differences in first measurement only in the respiration band (p < 0.05, Figure 5a). The second measurement showed group differences in cardiovascular at p < 0.05 and in both respiratory and VLF bands at p < 0.01 (Figure 5b). Compared to the first measurement, group differences in CV_MREG maps of the respiratory band were more widespread in the second measurement.

Also, in the CV_MREG maps of second measurement, the respiration band markedly dominated group differences. The VLF changes were localized along the midline and lack the peripheral extent of those changes seen in respiratory band. Interestingly, the differences in the cardiac band included these peripheral cortical structures.

Comparing different sub-bands of patients’ fMRI CV_MREG maps also to secondary verification group (with better physiological data quality with similar results in cardiac PPG and respiratory ETCO₂ rates during the scanning in both groups, Figure S2) revealed that full-band CV_MREG was even more elevated in DRE patients versus HC (Figure S3).

3.4 Template-based CV-analysis

Data from both groups indicated that GM has significantly higher CV_MREG-values than WM and CSF (Figure 6a). Compared to HC, DRE patients had significantly elevated CV-values for GM, WM, and CSF in both measurements, while each GM had highest values. Importantly, CV_MREG in GM was elevated in the second measurement for every patient.

3.5 Thresholded data

Thresholding of the data was performed based on the HC data statistics of the whole brain data. While the HC dataset showed voxels exceeding the >2 σ threshold, but not >3σ (Figure 6b), maps of DRE patients showed voxels at >3σ threshold. The second measurement showed more voxels than the first (Figure 6b). Based on this, a voxel-wise threshold was applied to the CV_MREG-maps.

3.6 Global fMRI signal motion and physiological parameters

3.6.2 Motion

There were no differences in motion parameters between the HC group and the DRE patients (Table 2, Figure 7a) Furthermore, neither CV_Motion of absolute nor relative movement of time domain motion signal were significantly different between groups (Table 2, Figure 7a). The only significant difference was a reduction of the absolute movement in the DRE patients second measurement compared to the first (p = 0.00044).

4.1 Individual thresholded diagnostic mapping of abnormal physiological signal in epilepsy

To the best of our knowledge, this is the first time that an fMRI signal variation has been used to show significant individual brain signal changes in DRE patients in the absence of epileptic activity. DRE patients showed repeating increases in physiological variation of the fMRI signal measured as CV_MREG and band passed SD. On the group level, changes in CV_MREG showed widespread increases in fMRI signal variation against two control groups (one being age and sex matched and the other with verifiably more similar physiological pulsation data).

On the individual level, these changes were confined to large anatomical areas in majority of cases and in three cases to a sub-centimeter frontal foci. This suggests that CV_MREG may enable subject-specific mapping of case-specific changes. Previously, due to limited statistical power of slowly sampled fMRI data, it was not feasible to threshold DRE patients’ CV_MREG maps at individual level with the healthy control population threshold of 3 SD. Since we do not find CV_MREG changes >3 SD in the control groups, we feel that CV_MREG may be a marker at individual level.

The reason for the increased sensitivity stems from the ultra-fast fMRI image sampling as it removes several major confounds that have previously prevented comprehensive analysis of human brain physiology in fMRI. First of all, the critical signal sampling of physiological brain pulsations prevents aliasing of these pulsations on top of each other, which then enable clear distinction between the physiological phenomena after band pass. Secondly, the prompt 80-ms acquisition of the k-space in MREG captures the spreading pulse effects accurately, while commonly applied interleaved slice sampling mixes the propagation due to 1,3,5…2,4,6 timing of slices. Thirdly, the statistical power of fast-fMRI is elevated due to 20-fold increase in samples. Previously, these advances enabled separation of single epileptic spikes in an EEG-MREG study (Jacobs et al., 2014). The CV analysis of signal variation benefits from more accurate and statistically powerful data and thus enables the formation of individual CV_MREG maps, when thresholded against values 3σ above normal level.

Combined EEG-fMRI usually markedly improves the fMRI mapping of the epileptic activity (Moeller et al., 2009), with the requirement of the activity to be present during the multimodal scanning. However, our patients did not exhibit epileptic activity. Fast fMRI CV_MREG-maps showed novel information at individual level regarding the brain pathophysiology in DRE that matched overall clinical assessment in majority of cases.

4.2 Group results

Both mesial temporal lobe and idiopathic generalized epilepsy have shown increased amplitude of low-frequency BOLD signal fluctuations in thalami, temporal cortices, and default mode network (DMN) (Wang et al., 2014; Zhang et al., 2010). It is also shown that there might be relation between VLF differences on white matter and BOLD signal detection (Gawryluk, Mazerolle, & D'Arcy, 2014). The epilepsy patients’ functional connectivity of resting state BOLD signal also reduced within DMN, and between epileptic areas and DMN (Constable et al., 2013; J. Gotman et al., 2005; Lee, Smyser, et al., 2013; Robinson et al., 2017). Zhang et al. (2015), have found that a combined measure of BOLD signal amplitude and functional connectivity density is abnormal in epilepsy. However, recent studies showed both increased and decreased changes in BOLD amplitude and connectivity measures (Centeno & Carmichael, 2014; Robinson et al., 2017). Our previous evidence in resting state fMRI also showed both increases and decreases in regional and long-range connectivity measures in epilepsy (Mankinen et al., 2011, 2012).

The group level results indicated only increased variation in the fMRI signal in DRE patients. The group level changes were widespread, especially in the respiratory and VLF bands. As the underlying signal characteristics were altered in DRE, they were bound to have an influence on interleaved BOLD data with 2- to 3-s sampling rate. The increased physiological noise can further alias over low frequency in interleaved echo-planar imaging BOLD signal of 2–3 s in spurious ways, and some of the previous mixed results can be explained by differential aliasing of physiological pulsations in the brain (Kiviniemi, Kantola, Jauhiainen, & Tervonen, 2004).

4.3 Pathophysiological mechanism behind pulsation abnormality

The recently discovered glymphatic brain clearance system uses physiological pulsations to clear the brain tissue by convecting water through the brain tissue via aquaporin AQP4 water channels especially during sleep (Iliff & Nedergaard, 2013; Jessen, Munk, Lundgaard, & Nedergaard, 2015; Nedergaard, 2013). In human brain, glymphatic system may be driven partially by respiratory and vasomotor control waves in addition to cardiac pulses (Kiviniemi et al., 2016).

The most common cause in DRE is MTS, which causes 40%–50% of all DRE cases (Jutila et al., 2002; Lapalme-remis & Cascino, 2016). Tissue samples of MTS have been shown to present a striking lack of peri-vascular AQP4 water channels that mediates the glymphatic water clearance (Eid et al., 2005). The absence of AQP4 channels in MTS areas could lead to altered extracellular electrolyte concentrations that sensitizes neuronal tissue to seizures (Eid et al., 2005; Lundgaard et al., 2017; Marchi et al., 2007).

The results of our study suggest that the physiological pulsations, that drive glymphatic clearance, may be abnormal in drug-resistant epileptic brain (Kiviniemi et al., 2016; Nedergaard, 2013). As none of our DRE patients had MTS lesions, we suggest that seemingly normal DRE patients’ brain tissue may suffer from failed physiological drive of the glymphatic clearance, rather than direct AQP4 channel loss as in MTS. The pulsation abnormality becomes evident as abnormal signal variation in the specific physiological frequency bands, especially respiratory and very low frequencies. These frequencies further overlap with autonomic disturbances known to be present in epilepsy (Ansakorpi et al., 2002, 2011; Liu et al., 2017; Suorsa et al., 2011; van der Kruijs et al., 2016).

Furthermore, the pulsation abnormalities could be linked to the detected abnormalities in neurovascular coupling in epilepsy, where an abnormal and prolonged dilation in the blood vessels are detected prior to epileptiform activity (Jacobs et al., 2009; Mäkiranta et al., 2005; Moeller et al., 2009; Osharina, Aarabi, Manoochehri, Mahmoudzadeh, & Wallois, 2017). Prolonged vasodilatations may minimize glymphatic pulsation of the perivascular CSF reservoir and lead to altered tissue homeostasis and further to epileptic activity (Jessen et al., 2015).

The VLF (0.01–0.1 Hz) and LF (i.e., respiratory 0.3 Hz) pulsations bands have previously revealed autonomous nervous system changes in heart rate variability (Ansakorpi et al., 2002, 2011; Liu et al., 2017; Lotufo, Valiengo, Benseñor, & Brunoni, 2012; Suorsa et al., 2011; van der Kruijs et al., 2016). The abnormal, noncoupled vasodilation may be linked to the detected VLF vasomotor pulsation increase and to the detected noise changes in the brainstem autonomous nervous system centers (Figure 3). The previous studies have also shown that connectivity of the brainstem and cortical/subcortical structures is altered in temporal lobe epilepsy (Englot et al., 2017). The LF power changes were mapped to brainstem areas in seven of 10 patients in our study indicating high probability of dysregulatory changes in these areas. Our results indicated that fast fMRI offers new diagnostic information on the neurovascular control also at brainstem level.

4.4 Sources of the signal features

In this study, we showed for the first time abnormally increased fMRI signal variation in DRE linked to the intrinsic physiological pulsatility of the brain. While the external markers of cardiovascular physiology and rigid body motion showed no differences compared to two control groups, the brain pulsations in these frequencies still are highly abnormal in DRE. Coefficient of variation reflects stability of the measured signal and increased motion of the subject increases CV (Hajnal et al., 1994; Hao, Khoo, von Ellenrieder, & Gotman, 2017). Also, in this study, the first measurement of the DRE patients had more absolute motion compared to HC. And indeed, in the first measurement, the CV_MREG was found to be increased in the fMRI signal of the DRE patients. In the second measurement, however, where patients’ absolute movement was reduced, CV_MREG was even more increased compared to HC. Therefore, motion does not explain our results. We interpret that in the case of CV_MREG mapping, which is sensitive to any noise source, increased motion may mask the pathological brain tissue CV_MREG changes, since in the absence of motion, the differences become highly significant. We assume that this reduction in absolute movement could be linked to patient habituation to the scanning and verified that it did not affect the CV_MREG results by comparing the control group data.

A recent finding regarding the global fMRI signal SD indicates that indeed the arousal of the subjects may have a widespread effect on the fMRI results (Chang et al., 2016; Liu et al., 2018). Subjects with epilepsy have altered sleep homeostasis and present widespread and complex neural and hemodynamic signal changes that are different in deep brain structures compared to cortex (Boly et al., 2017; Peter-Derex, Magnin, & Bastuji, 2015; Salek-Haddadi et al., 2003). Thus, DRE changes induced by altered arousal mechanisms might explain the differences in 1st vs. 2nd scans. As many subjects may fall asleep after only a few minutes, the second scan may have reduced vigilance that increase sensitivity to abnormal CV values (Tagliazucchi & Laufs, 2014). Furthermore, drowsiness could explain the detected reduction in the motion of the DRE, which was the only difference in motion in the groups. An interesting future project would be investigation of sleep and vigilance fluctuations in combined EEG-fMRI data.

None of the physiological measures showed differences between the groups, which minimizes the possibility of a direct physiological noise source difference. The datasets were rigorously corrected with conventional FSL preprocessing and with advanced ICA-based FIX that further de-noises the data for motion and other noise sources (Griffanti et al., 2014; Salimi-Khorshidi et al., 2014). Without FIX, there were no differences between the HC and first measurement of epilepsy. Again, in the second measurement with less motion artifact, even the raw data had increased CV_MREG in DRE patients (Figure S2).

4.5 Limitations and future prospects

Medication may cause some of the detected signal changes, but preliminary data from drug-naïve epilepsy patients showed converging results (Figure S4), so the found effect is not only seen in DRE, but also with newly diagnosed drug-naïve epilepsy patients. The analysis would naturally benefit from a higher sample size and second measurement verification data also for controls, which are under way. We plan to generate a new CV_MREG threshold image with more controls subjects, and for this aim, we are creating an international imaging cluster for speeding up data acquisition of healthy control subjects.

As CV has been successfully used in various blood flow-related signal measurement techniques in different scales from mouse microscopy to macroscopic human data (Jahanian et al., 2014; Kalchenko, Kuznetsov, Meglinski, & Harmelin, 2012; Makedonov et al., 2013, 2016), it would be interesting if one could use it in multimodal imaging as a way to produce scale invariant measures that could combine animal model data to clinical human brain data.