Low impedance electrodes improve detection of high frequency oscillations in the intracranial EEG

(cid:1) The low impedance electrode reduced noise in the intracranial EEG. (cid:1) The reduced noise facilitated detection of evoked fast ripple oscillations. (cid:1) Low impedance electrodes may facilitate detection of fast ripple oscillations that are biomarkers for epileptogenic tissue.


Introduction
In patients suffering from pharmaco-resistant focal epilepsy, pathological high-frequency oscillations (pHFO) are considered as promising interictal biomarkers of the epileptogenic zone. It is being discussed whether pHFO in the intraoperative electrocorticography (ECoG) may delineate epileptogenic tissue better than interictal epileptic spikes and thereby may aid in tailoring epilepsy surgery and consequently improve postoperative seizure freedom of the patients (Zweiphenning et al., 2022). In particular, fast ripple oscillations (FR, in the ECoG are thought to indicate epileptogenic tissue with high specificity (van 't Klooster et al., 2017). However, FR have not reached clinical practice yet.
From a measurement perspective, FR have rare appearance (a few per minute), short duration (some 10 0 s of milliseconds), low amplitude (up to a few lV), and the generators of FR are confined to a small area of brain tissue (up to a few mm). All these characteristics make FR demanding to detect against the noise level. One needs to optimize the signal-to-noise ratio (SNR) for FR detection in the measurement chain. As one link in the chain, a low noise amplifier has shown improvement in FR detectability (Fedele et al., 2017a;Fedele et al., 2017b). As another link in the chain, a higher contact density on the electrode improves the detection of FR (Boran et al., 2019;Zweiphenning et al., 2020) and epileptic spikes (Sindhu et al., 2023). Unfortunately, a higher contact density requires a smaller contact area, which reduces contact impedance and thereby increases thermal noise in the recording, thus reducing SNR. As a possible remedy, an ECoG electrode with low impedance contacts (LoZ) has been certified for clinical use (Sarnthein et al., 2022). It remains an open question whether the LoZ indeed improves FR detection over a standard electrode with high impedance contacts (HiZ).
As a first step to answer this question, we study here the detectability of physiological activity in the FR frequency band, in particular the high frequency components associated to the somatosensory evoked potential (SEP) elicited in the ECoG by median nerve stimulation. The SEP is a standard tool for intraoperative neurophysiological monitoring. The N20 response of the SEP is generated from area 3b in the primary somatosensory area. In the spatial distribution of the N20, a phase reversal occurs over the central sulcus, i.e. recordings from contacts in proximity of the central sulcus show opposite N20 polarity over sensory and motor cortex (Antonakakis et al., 2019). For this reason, N20 phase reversal analysis is used to localize the central sulcus when lesions are operated on in the sensorimotor region (Gregorie and Goldring, 1984). The evoked N20 response concurs with an evoked FR event and an evoked high frequency oscillation (HFO, > 500 Hz) in the ECoG Maegaki et al., 2000;Urasaki et al., 2002) that is associated with cortical neuron firing (Baker et al., 2003). In this physiological model system, the timing and the total number of events in the FR band is well defined. This allows us to quantify how the FR detection rate is improved for LoZ electrodes.

Patients
We included 10 patients (median age 40 y, range 19-56 y, 6 female) who underwent brain tumor resections in the perirolandic region at our institution. We recorded the phase reversal SEP to localize the healthy tissue of the somatosensory cortex in the central sulcus. Knowing the location of the healthy tissue helped the surgeon (MCN) to spare it from neurosurgical manipulation and thereby to avoid postoperative adverse events for the patient. The collection of personal patient data and their analysis was approved and performed in accordance with the guidelines and regulations of the local research ethics committee (Kantonale Ethikkommission Zürich 2018-02171). Patients gave informed consent to the scientific analysis of their data.

Anesthesia management
According to our standard protocol for neurosurgical interventions, anesthesia was induced with intravenous application of Propofol (1.5-2 mg/kg) and Fentanyl (2-3 lg/kg). Intratracheal intubation was facilitated by Atracurium (0.5 mg/kg), which was stopped afterwards to avoid muscle relaxation. Anesthesia was maintained with Propofol (5-10 mg/kg/h) and Remifentanil (0.1-2 lg/kg/min). The depth of anesthesia was monitored by the bispectral index (BIS) typically in the range 35-40.

Recording setup
We used the ISIS system (inomed Medizintechnik GmbH) for IONM. We placed 20 mm needle electrodes with impedance < 1 kO for electrical stimulation of the median nerve at the wrist. We stimulated with a square-wave constant current pulse with dura-tion 200 ls. Stimulation intensity was in the range of 10-40 mA and evoked a twitch response of the thumb. The stimulus repetition rate was 4.7 Hz.
We used two different ECoG strip electrodes. The low impedance electrode (LoZ, WISE Cortical Strip, WISE Co., Milan, Italy) was a 4-contact cortical electrode strip (strip thickness: a b c V. Dimakopoulos, M.C. Neidert and J. Sarnthein Clinical Neurophysiology 153 (2023) 133-140 0.25 mm; material of electrodes: pure platinum; exposed electrode diameter 2.3 mm; inter-electrode distance 10 mm center-tocenter, Fig. 1a right). The electrode was manufactured using cluster beam implantation to embed a thin, conducting platinum layer on a silicone substrate to realize a flexible, soft and thin cortical strip electrode. The implanted platinum nanoparticles range between 10 nm and 100 nm in size so that the nanostructure has roughness in the same order of magnitude (Gnatkovsky et al., 2019). Due to the roughness, the surface area of each contact is around 70 times larger than a flat platinum surface. The large surface area allows a highly efficient exchange of charge at the interface between the surface and the surrounding fluids. These microscopic properties become evident during application of the electrode as they result in low impedance of the electrode. The standard electrode (HiZ, Ad-Tech Medical Instruments Corporation, Racine, Wisconsin, USA) was a 4-contact cortical electrode strip (strip thickness 0.8 mm; platinum electrode contact diameter 2.3 mm; inter-electrode distance 10 mm center-to-center, Fig. 1a left). It consists of platinum platelets with a flat surface that are mechanically fixed to a silicone strip and is used by many neurosurgery centers worldwide. After dura opening, the LoZ electrode was placed on the brain surface tangentially over the central sulcus so that it covered both primary motor and primary somatosensory cortex (Fig. 1b). The position of the LoZ electrode was optimized for SEP recording at the hand area of somatosensory cortex, the hand omega (X). In all patients, the HiZ electrode was placed next to the LoZ for simultaneous recording. We used these 10 recordings for further analysis unless stated otherwise. A dura needle served as electrical reference. During the measurements, the surgical field was covered by a moist gauze to ensure optimal impedance. The neuromonitoring device measured the impedance of the electrode contact at 140 Hz and confirmed the lower impedance for the LoZ electrode (median HiZ 6.9 kO, median LoZ 3.4 kO, Fig. 1c).
The averaged SEP and the continuous ECoG were recorded at sampling rate 20 kHz in the amplitude range 10 mV peak-topeak and16 bit analog-to-digital conversion, resulting in a quantization step of 0.153 lV/Bit. For intraoperative viewing during IONM, we set the recording sweep length to 50 ms for the SEP and filtered the recordings for display (30-300 Hz). Concurrently, we recorded the continuous ECoG for offline processing. The ECoG was filtered in the pass band 0.5-5000 Hz. All recordings had at least 25 s duration, corresponding to >200 sweeps.

Offline data analysis
The offline data processing used the continuous ECoG that was recorded in parallel to the SEP recordings. Data analysis was performed with custom scripts in Matlab (https://www.mathworks.com) and is very similar to our publication on the intraoperative SEP in the scalp EEG (Dimakopoulos et al., 2023). To detect the SEP stimulation artefact, we first filtered the ECoG (high pass cutoff = 200 Hz) and performed local peak detection (minimum peak prominence between peaks = 30 ms, minimum peak width = 4 ms, samples = 0.2 ms). We used the times of the detected stimulus artifact as triggers to define sweeps with poststimulus recording sweep length 50 ms. We classified sweeps with amplitude ± 100 lV as artefact-ridden and excluded them from further analysis.
We averaged 100 sweeps and filtered the averaged trace (bandpass [30 300] Hz, IIR filter, response roll-off À12 db per octave, forward and reverse filtering to avoid phase distortion). We visually inspected the data and selected one optimal channel with high N20 amplitude (positive or negative) for further analysis. From the averaged N20 trace, we determined the N20 peak latency. To obtain the N20 peak amplitude and the SNR, we inspected the latency of the N20 peak. If the N20 latency was >20 ms, we selected a signal window [20 25] ms. If the N20 latency was 20 ms, we selected a signal window [17 22] ms. In the same way, we filtered the averaged trace in the [250 500] Hz band to obtain the evoked FR and in the [500 1000] Hz band to obtain the evoked HFO. We doubled the largest deflection in the signal window of the N20 frequency band to define the N20 signal amplitude. In the FR and HFO bands we used the peak-to-peak amplitude.
To estimate the noise level, we performed ± averaging, a procedure in which measurements from every other sweep are inverted prior to creating the averaged result (van Drongelen, 2018). In particular, we computed the difference between Sweep(2n) and Sweep(2n + 1) to estimate the noise distribution. We then calculated the root-mean-square (rms) of the noise trace in the signal window to quantify the noise level. We prefer this noise estimate over choosing a pre-stimulus epoch that may include effects of late SEP components; similarly, an early post-stimulus epoch may include effects of subcortical potentials. We divided the signal amplitude by the noise level to obtain the SNR.

Statistics
We based our statistical analysis on the evoked signal amplitude and the noise level. We randomly selected a subset of 100 sweeps to compute the median amplitude in the [30 300] Hz N20 band. We repeated this process 1000 times to obtain a median amplitude for each recording sessions. The distribution of the amplitude values with its median and interquartile interval was parametrized in a boxplot. In the boxplots we do not show the outliers outside of the whiskers (>1.5 times interquartile range). Similarly, we obtained one boxplot for each frequency band.
For the noise decay we assumed the power function Amplitude = a * N k and fitted the line log(Amplitude) vs. log(N) to obtain the exponent k and the intercept = a for N = 1 sweep. For the line fit, we used the function fitlm.m of Matlab. We evaluated the significance of the fit with the F-test that indicates if the regression model provides better fit than a model that contains no independent variables. To compare the noise amplitude between the two electrodes, we created a linear mixed effects model (Bates et al., 2015). For comparisons between two groups, we used the paired Wilcoxon test. Statistical significance was established at p < 0.05.

Results
3.1. Response to median nerve stimulation in N20, FR and HFO bands Fig. 2 shows an example of the average trace recorded with the LoZ electrode for the medianus SEP from one patient. We averaged 100 sweeps and then filtered the average trace in the N20 frequency band (Fig. 2a), in the FR band (Fig. 2b), and in the HFO band (Fig. 2c). The SEP trace in the N20 frequency band had a local negative peak that defines the recording from the somatosensory cortex.
3 Fig. 1. Electrodes, recording site and impedance. (a) Electrodes in the surgical field. The low impedance electrode (LoZ, right) has dark platinum electrode contacts embedded in the carrier material. The high impedance electrode (HiZ) has platinum platelets as electrode contacts. (b) Schematic representation of the site. The strip electrode was placed perpendicular to the central sulcus over the hand area (hand omega X). Only the LoZ was placed in an optimal position to record the somatosensory evoked potential (SEP) after stimulation of the median nerve. The HiZ was placed in parallel in the surgical field. (c) For the recordings included in this study, the median impedance was 6.9 kX for HiZ and 3.4 kX for LoZ.

Noise characteristics
We first analyzed the noise characteristics of the continuous ECoG in the frequency domain. To calculate the linear spectral density (LSD, Fig. 3) for the two electrodes, we resampled the data to 4 kHz. The LSD of the LoZ electrode (red line) was lower than the LSD of the HiZ electrode (blue line) across the whole spectral frequency range.
We calculated the noise ratio ¼ median noiseðLoZÞ median noiseðHiZÞ in each frequency band and obtained 75% for the N20 band, 59% for the FR band, and 72% for the HFO band. The distribution of the noise resembles Gaussian white noise in all three bands where recording artefacts add a high amplitude tail (Fig. 4d, e, f). We next averaged the ± noise level over N randomly selected sweep pairs and repeated the process 1000 times to obtain a median noise level. The median noise decayed with the number of averaged sweeps N with the power law noise = a*N k where k = 0.500 in all three frequency bands (linear fit of log(Amplitude) vs. log(N), Fig. 4g, h, i). The value of k = À0.500 indicates that the averaging of N sweeps reduces the noise level by 1/ p N, as expected for Gaussian white noise. The intercept a equals the median of the noise distribution ( Fig. 4a, b, c) for each frequency band.

Amplitude and signal-to-noise ratio
Because only the LoZ electrode was placed for optimal SEP recording in 10 patients, we first present the amplitudes for the LoZ electrode. We averaged the SEP amplitude over 100 randomly selected sweeps, repeated the process 1000 times, and obtained a median peak-to-peak amplitude of 19.3 ± 6.4 lV for the N20, 1.1 5 ± 0.70 lV for the FR and 0.52 ± 0.45 lV for the HFO band ( Fig. 5a, Table 1). Dividing the signal amplitude by the noise level for N = 100 sweeps yielded the SNR (Fig. 5b). We then computed the SEP amplitude of LoZ for each single sweep, divided by the ± averaged sweep pair to obtain the single-sweep SNR, and computed the median SNR (Fig. 5c). The single sweep SNR was 10 times lower than for N = 100 sweeps, as expected from the 1/ p N law.
To estimate the SNR for HiZ, we used the LoZ signal amplitude and the median HiZ noise amplitude to extrapolate to the SNR of the HiZ electrode (blue bars in Fig. 5b). By this calculation, the median SNR was reduced by the noise ratio in each frequency band (N20 75%, FR 59%, HFO 72%). Similarly, we extrapolated to the SNR of the HiZ electrode for the single sweep SNR (blue bars in Fig. 5c).

Improved detection of evoked FR
We finally quantified the rate of events that we could detect. We defined the rate of detected events as the rate of sweeps with SNR > 1. We computed the rate for each patient recording in the three frequency bands. Across all 10 patients, median detection rates for LoZ were slightly higher in the N20 band (HiZ: 97%, LoZ: 99%) and in the HFO band (HiZ: 90%, LoZ: 95%). Median detection rate improved most in the FR band (HiZ: 91%, LoZ: 100%). Fig. 6a shows an example where an evoked FR exceeds the noise level of the LoZ electrode (SNR > 1) but the FR goes undetected in the HiZ electrode (SNR < 1). Fig. 6b shows the amplitude distribution of all FR against the noise level of HiZ and LoZ, respectively. The number of FR detected with LoZ slightly exceeds the number of FR detected with HiZ, thus quantifying the advantage of the LoZ electrode for FR detection.

Discussion
We investigated the detectability of events in the FR band in a recording setup where the timing of the FR events is precisely defined. As our main result, the noise level and the SNR in the FR band improved from HiZ (165 nV, SNR = 7.5) to LoZ (90 nV, SNR = 12.8). This improvement facilitated the detection of FR events with a small amplitude.

Validity of the SNR extrapolation from LoZ to HiZ
The time constraints of surgery did not permit to place the LoZ and the HiZ electrode consecutively on the optimal anatomical location at the hand X. In our analysis design we therefore assumed that the signal amplitude for HiZ would be in the same range as for LoZ if it were placed on the same position. To base our analyses on a large number of patients and for consistency across patients, we chose to use the median signal amplitude of the 10 patients (Fig. 5a) for further analysis. We first filtered the raw signal in the three frequency ranges (a, d, g, N20 30-300 Hz), (b,d,, (c, f, I, HFO 500-1000 Hz). We then segmented into sweeps and calculated the noise amplitude as Sweep(2n + 1) -Sweep(2n) where Sweep is the trace of an individual sweep. (a, b, c) Median and interquartile range (IQR) of the noise for LoZ and HiZ. (d, e, f) Distribution of noise for the low impedance electrode (LoZ, red) and the high impedance electrode (HiZ, blue) and their medians (dashed lines). The distributions have a high amplitude tail. (g, h, i) Averaging over sweeps decreases the noise amplitude. The noise reduction follows the power law a*N k with a the intercept for 1 sweep, N the number of sweeps that were averaged, and k = À0.500 the slope of the decrease in the log-log plot. FR, fast ripple; HFO, high frequency oscillations.
V. Dimakopoulos, M.C. Neidert and J. Sarnthein Clinical Neurophysiology 153 (2023) 133-140 To estimate the noise level, we used recordings where we recorded from LoZ and HiZ simultaneously. Because of its suboptimal anatomical placement, the signal amplitude for the HiZ was obviously lower than for the LoZ electrode (Table 1). The noise estimate was independent of anatomical placement because we used the ± averaging (van Drongelen, 2018), which efficiently eliminates evoked components in the noise. The validity of the noise estimate is supported by the exponent k = À0.500 (Fig. 4) to match the theoretically expected 1/ p N law. Interestingly, the LoZ had lower noise across the whole spectral frequency range (Fig. 3), whereas the comparison between a Low Noise Amplifier and a standard amplifier showed noise reduction mainly above 100 Hz (Fedele et al., 2017a;Fedele et al., 2017b). After having thus validated the evoked signal amplitude and the noise level, we calculated the single-sweep SNR for the LoZ electrode (Fig. 5c, red bars) and then extrapolated to the single sweep SNR for the HiZ electrode (Fig. 5c, blue bars).

Clinical relevance
How can the findings on the detectability of the evoked FR be translated to the detectability of the pathologic FR event that is a biomarker of epileptogenic tissue? The two types of events have been reported with very different amplitude. The median amplitude of the evoked FR in this study was 1.15 lV (Fig. 5a, Table 1). For an epileptic FR, the amplitude was reported with 10 lV for a HiZ electrode of the same contact size and impedance ((Boran et al., 2019), Table 1). Thus, the epileptic FR has much larger amplitude than the evoked FR. On the one hand, this may reflect the pathophysiology, where epileptic interictal discharges are usually seen with much larger amplitude than events related to normal brain function. On the other hand, the observed amplitude difference may be a product of the measurement chain, where all elements including the automated detection algorithms (Boran et al., 2019;Fedele et al., 2017a;Fedele et al., 2016) are optimized for large epileptic FR. The large amplitude FR may thus reflect a measurement bias. Epileptic FR may obviously appear with smaller amplitude when recorded from a larger distance. In this case, events may arise like in Fig. 6a, where the LoZ electrode improves FR detection.
Since the generators of epileptic FR have spatial extent in the mm range, electrodes with denser contact spacing have improved their detection (Boran et al., 2019;Zweiphenning et al., 2020). Denser contact spacing entrains smaller contact diameter and thereby higher impedance, if the contact material remains unchanged. Higher impedance increases thermal noise ( Fig. 3 and Fig. 4), which reduces SNR (Fig. 5) and thereby reduces FR detectability (Fig. 6). If a contact material achieves lower impedance, like the one tested here (Fig. 1c, (Sarnthein et al., 2022)), the implementation of this material in electrodes with high contact density might further improve detection of epileptic FR. To estimate the SNR of the high impedance electrode (HiZ), we divided the median signal amplitude by the median noise amplitude after averaging 100 sweeps (blue bars). (c) SNR for single sweeps with median SNR for LoZ. To estimate the SNR of the high impedance electrode (HiZ, blue bars), we divided the median signal amplitude by the median noise amplitude for single sweeps. The percentage indicates the rate of median SNR reduction. FR, fast ripple; HFO, high frequency oscillations.

Conclusions
Low impedance electrodes for intracranial EEG reduce noise in the FR frequency range and may thereby improve FR detection. Improving the measurement chain may spread the diagnostic value of FR as biomarkers for epileptogenic tissue.  , resulting in (SNR = 0.9 < 1) and the FR went undetected. With the reduced noise level of LoZ (89 nVpp, red lines), the SNR increased to SNR = 1.6 > 1 and the evoked FR was detected. (b) Amplitude distribution of all FR from each patient (colored histograms) and also the noise level of HiZ (blue line) and LoZ (red line). The number of FR detected with LoZ slightly exceeds the number of FR detected with HiZ, thus quantifying the advantage of the LoZ electrode for FR detection. FR, fast ripple.