Fast oxygen dynamics as a potential biomarker for epilepsy

Changes in brain activity can entrain cerebrovascular dynamics, though this has not been extensively investigated in pathophysiology. We assessed whether pathological network activation (i.e. seizures) in the Genetic Absence Epilepsy Rat from Strasbourg (GAERS) could alter dynamic fluctuations in local oxygenation. Spontaneous absence seizures in an epileptic rat model robustly resulted in brief dips in cortical oxygenation and increased spectral oxygen power at frequencies greater than 0.08 Hz. Filtering oxygen data for these fast dynamics was sufficient to distinguish epileptic vs. non-epileptic rats. Furthermore, this approach distinguished brain regions with seizures from seizure-free brain regions in the epileptic rat strain. We suggest that fast oxygen dynamics may be a useful biomarker for seizure network identification and could be translated to commonly used clinical tools that measure cerebral hemodynamics.


Results
To address whether hemodynamic changes could be a useful epilepsy biomarker, we selected a rodent model of epilepsy with a high rate of spontaneous seizures. The Genetic Absence Epilepsy Rat from Strasbourg (GAERS) displays frequent corticothalamic seizures consistent with childhood absence epilepsy 7,8 . During 30-minute sessions, we recorded 198 seizures from 5 rats (Fig. 1a) that were never associated with severe postictal hypoxia (Fig. 1b). Since absence seizures were free of severe and long-lasting postictal hypoxia, which could be elicited by induced focal seizures (Fig. 1c) 9,10 , this is an ideal model to study oxygen dynamics associated with brief, hypersynchronous network events.
We categorized seizures as discrete, double, or stringed events depending on whether a seizure was associated with another seizure within 20 seconds of seizure termination. Following discrete seizures (Fig. 2a), we observed brief dips in neocortical oxygen of 1.6 mmHg which peaked at 10.6 s following seizure onset (Fig. 2c). As expected, no seizures were observed in the non-epileptic control (NEC) strain (Fig. 2b) and no oxygen changes were observed at randomly chosen times (sham seizure onsets) (Fig. 2e). While longer seizures did not significantly alter the magnitude of this dip ( Supplementary Fig. 1), they significantly delayed the peak oxygen dip (Fig. 2d vs. Fig. 2f, Supplementary Fig. 1). Notably, our results are consistent with previous clinical characterization of hemodynamics during absence seizures 11 , suggesting our recordings are clinically valid.
We then assessed whether double (Fig. 2g) or stringed seizures (Fig. 2h) would summate and drive oxygen to lower levels. In either case, repeated seizures did not increase the magnitude of the oxygen dip and revealed an oxygen overshoot following the final seizure ( Fig. 2i-m), highlighting the brain's ability to regulate oxygen. Importantly, oxygen inflections occurred ~10 s following each seizure onset and gave rise to fast oscillatory activity, clearly evident following a string of seizures (Fig. 2h). Bolstered by this observation, we postulated that the presence of absence seizures could be determined solely based on the spectral characteristics of oxygen data. Prior research determined that cerebrovascular dynamics generally fluctuate at frequencies less than 0.1 Hz 6 . Given the fast kinetics of oxygen dips, especially during stringed events, seizures could increase the amplitude of high frequency oxygen changes. We performed power spectral density analysis on 30-minute recordings from GAERS and NEC rats (Fig. 3a) and, indeed, observed significantly increased power between 0.08-0.1 Hz (Fig. 3b). Since these changes were not observed at a control site without seizures (hippocampus) (Fig. 3c) and are correlated with the occurrence of seizures (Fig. 4a,b), fast oxygen dynamics reflect pathological network activity.
To further isolate the increase in fast oxygen dynamics, we applied a digital high-pass filter to the data to observe only the oscillatory components at these high frequencies (Fig. 3d). By simply measuring the standard deviation of the filtered data, we could distinguish between epileptic vs. non-epileptic rat strains (Fig. 3e). Moreover, in a brain region without seizures (hippocampus) in the epileptic rat strain, none of the data points crossed the threshold to be classified as epileptic (Fig. 3f). While this measurement was associated with the occurrence of seizures (Fig. 4a,b), these oxygen characteristics were not strictly unique to the epileptic brain since the control strain also had oscillations at this frequency, albeit markedly reduced (Fig. 3b). Rather, seizures shifted the distribution of high-frequency oxygen oscillations to higher amplitudes in the GAERS rats (Fig. 4c). Therefore, this analysis is suitable for identifying active epileptic networks, rather than the occurrence of individual seizures.
Lastly, we assessed the ability of this analysis to separate data into epileptic vs. control during smaller sampling intervals. Recordings of 7 minutes or less were associated with reduced accuracy, but separation was maintained  (Fig. 5a-c). Thus, we have identified fast oxygen dynamics as a potential biomarker of absence epilepsy and is sufficient to distinguish active epileptic brain networks from non-epileptic within 14 minutes of recording.

Discussion
We first showed that absence seizures were associated with brief, mild dips in cortical oxygenation within the normoxic range, but importantly, we never observed severe postictal hypoxia (<10 mmHg). However, severe postictal hypoxia lasting approximately an hour was induced by modelling focal seizures with kindling stimulation. This provides evidence that GAERS rats possess the neurovascular mechanisms that orchestrate this pathological event, but that spontaneous corticothalamic seizures are fundamentally different. Since absence seizures are widely considered to not be associated with a true postictal state and are also free of postictal stroke-like events, severe postictal hypoperfusion/hypoxia could be the physiological driver underlying the postictal state. Moreover, given the potential negative impact of repeated postictal stroke-like events to the pathophysiology of epilepsy 12 , this may also explain the relative severity of different clinical seizure types and why absence seizures are relatively mild. We then demonstrated that isolation of fast oxygen dynamics was sufficient to identify epileptic networks even though the dips in cortical oxygenation were relatively small. Therefore, this simple analytic approach may be a highly-sensitive biomarker for epilepsy. Moreover, the simplicity of this observation and availability of tools to study cerebrovascular dynamics makes this analysis adaptable to other datasets. Clinically, techniques such as functional magnetic resonance imaging (fMRI) or near infrared spectroscopy (NIRS) are suitable candidates. The (e,f) Standard deviation (SD) of filtered oxygen data for neocortical and hippocampal recordings, respectively. GAERS rats have significantly higher neocortical SD than NEC controls (t(6) = 2.55, *p < 0.05). A line drawn at an SD of 0.095 mmHg is able to separate epileptic rats from controls and distinguish non-epileptic brain regions. only requirement for these measurements is a sufficiently high sampling rate (minimum of 0.2 Hz) and appropriate controls. As we have shown, brain regions without seizures can serve as a potential within-subject control. It is also possible that pathological network activity in other epilepsies, independent of overt seizures, could also result in similar deviations in oxygen spectral characteristics. Interictal spikes (IIS) and pathological high-frequency oscillations (HFOs or fast ripples) have been identified as biomarkers of epileptogenic tissue [13][14][15] , but often require implantation of depth electrodes, which is both invasive and costly. IISs, for example, display BOLD-related changes 16 and may be sufficient to perturb oxygen dynamics. Thus, this analytic approach could extend to other types of pathological epileptic activity and be captured by non-invasive imaging techniques. Thus, we have identified fast oxygen dynamics as a potential diagnostic biomarker for epileptiform events and this promising discovery opens the door to further characterize pathological cerebrovascular dynamics.

Methods
Rats. All recordings we performed on four-month-old GAERS and NEC rats (205-265 g) under awake, freely-moving conditions. Rats were bred and housed at the University of Saskatchewan as previously described 17   Oxygen and LFP recordings. Local tissue oxygenation was measured through chronically-implanted optrodes (Oxford Optronix) as previously described 9 . Bipolar electrodes for differential LFP recordings were constructed from insulated nichrome wire with a tip separation of ~1 mm (inVivo1). The following coordinates relative to bregma were used: Neocortex optrode (3.0 mm lateral, 2.0 mm ventral), neocortex electrode (1.0 mm anterior, 1.5 mm lateral, 3.0 mm ventral), hippocampus optrode (3.5 mm posterior, 3.5 mm lateral, 4.0 mm ventral) hippocampus electrode (3.0 mm posterior, 0.5 mm lateral, 4.0 mm ventral). Implants were secured with 3.2 mm stainless steel screws (one served as ground) and dental cement. Wideband LFP signals were amplified 1000X-2000X and digitized at 100 Hz (Grass Technologies). Oxygen data was sampled and digitized at 1 Hz. Prior to recording sessions, rats were extensively handled and then habituated to the recording chamber for 1-2 sessions. All recordings were done during the light cycle. Control and epileptic recordings were alternated to control for potential circadian effects.
Analysis. Python 3.6 and GraphPad Prism 6 were used for analysis and figure generation. Figures were compiled in Adobe Photoshop. Seizures were categorized as outlined in the main text and oxygen data was aligned to seizure onset to observe seizure-induced oxygen changes. To isolate the relative changes and remove the absolute component of the data, a sliding average of 60 s was subtracted from the data. 1 minute of data was removed on both ends of the data. This manipulation enabled us to do power spectral density analysis (using Matplotlib in Python 3.6) and made the data more comparable to techniques only capable of measuring relative changes, which are most commonly used. To filter the data for high frequency components, a 5 th order high-pass Butterworth filter was constructed in Python 3.6 using Signal in SciPy with a 0.08 Hz cut-off. All statistical analyses were performed in Prism 6 and reported in figure captions.

Data Availability
The datasets and code generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Figure 5.
Minimum recording duration to categorize data as epileptic vs. control is between 7 and 14 minutes or about 10 seizures. Neocortical oxygen data was binned as displayed and filtered for fast oxygen dynamics (>0.08 Hz). We computed the differences in group means (a) or between the lowest point in the GAERS group from the highest point in the NEC group (b), such that a positive value in (b) would indicate that there is no overlap between groups. (a) Binning data into shorter epochs did not dramatically alter the differences in group means. (b) Binning data into shorter epochs resulted in the appearance to negative values, indicating that the group distributions started to overlap (i.e. highest value in the NEC group was greater than the lowest of the GAERS group). (c) Data from (b) plotted against the corresponding number of seizures (mean across rats) for that bin and fitted to a sigmoidal curve. The curve crossed from negative to positive values at 9.6 seizures per rat.