Effects of stimulation intensity on intracranial cortico-cortical evoked potentials: a titration study

Abstract

Objective

The aim of the present study was to investigate the optimal stimulation parameters for eliciting cortico-cortical evoked potentials (CCEPs) for mapping functional and epileptogenic networks.

Methods

We studied 13 patients with refractory epilepsy undergoing intracranial EEG monitoring. We systematically titrated the intensity of single-pulse electrical stimulation at multiple sites to assess the effect of increasing current on salient features of CCEPs such as N1 potential magnitude, signal to noise ratio, waveform similarity, and spatial distribution of responses. Responses at each incremental stimulation setting were compared to each other and to a final set of responses at the maximum intensity used in each patient (3.5–10 mA, median 6 mA).

Results

We found that with a biphasic 0.15 ms/phase pulse at least 2–4 mA is needed to differentiate between non-responsive and responsive sites, and that stimulation currents of 6–7 mA are needed to maximize amplitude and spatial distribution of N1 responses and stabilize waveform morphology.

Conclusions

We determined a minimum stimulation threshold necessary for eliciting CCEPs, as well as a point at which the current-dependent relationship of several response metrics all saturate.

Significance

This titration study provides practical, immediate guidance on optimal stimulation parameters to study specific features of CCEPs, which have been increasingly used to map both functional and epileptic brain networks in humans.

1. Introduction

Single pulse electrical stimulation (SPES) is increasingly utilized in patients with drug-resistant epilepsy who are undergoing pre-surgical intracranial monitoring to map functional connectivity (Matsumoto et al., 2004; Catenoix et al., 2005; David et al., 2013) and probe epileptogenic networks (Iwasaki et al., 2010; Enatsu et al., 2012; Mouthaan et al., 2016). SPES has been used since the 1960’s (Buser et al., 1968) as a safe, robust experimental method to characterize the integrity of network-based systems and disorders. Compared to structural tractography and functional covariance measures, SPES provides dynamic (activity-dependent) directional information regarding connectivity; specifically, the influence that one neural system can exert over another (Friston, 2011). Despite its widespread utilization, there is no standardized method of conducting SPES, with each study center independently developing their own stimulation parameters (Matsumoto et al., 2017; Prime et al., 2018). Since parameter choice, such as current intensity or pulse duration, is known to influence SPES results (Donos et al., 2016; Prime et al., 2018; Kundu et al., 2020), significant variations in methodology could lead to potentially disparate findings, highlighting the need to more thoroughly understand the effects of these choices.

The overall goal of the present study was to investigate the optimal SPES current intensity parameters for eliciting cortico-cortical evoked potentials (CCEPs). For the purpose of this manuscript, we assume that overall connectivity is isotropic and relatively homogeneous, instead of focusing on the influence of brain location and underlying networks. We conducted incremental stimulation titration trials in thirteen patients undergoing invasive EEG monitoring to assess the effect of increasing stimulation intensity on salient features of CCEPs such as the magnitude of the early evoked response (N1), signal to noise ratio, waveform similarity and spatial distribution of the responses. We demonstrate that there is a minimum stimulation input that is needed to differentiate between non-responsive and responsive sites, and a maximum at which all features reach saturation. Our findings provide additional evidence that CCEPs elucidate direct connections between stimulation pair and response sites (Matsumoto et al., 2004) and offer practical guidance on the range of SPES parameters for eliciting these responses, providing a useful method for further applications of CCEPs in studies investigating both physiological and epileptogenic brain networks.

2. Methods

2.1. Patients 

We studied 13 patients who were undergoing intracranial EEG monitoring prior to surgery for treatment of drug-resistant epilepsy at the Johns Hopkins Epilepsy Center from January 2019 through March 2021. The study was approved by the Johns Hopkins School of Medicine Institutional Review Board (IRB 00247294) and was conducted using guidelines established in accordance with the Code of Ethics of the World Medical Association (1964, Declaration of Helsinki).

2.2. Electrode placement  

Patients were implanted with stereo electroencephalography (S-EEG) electrodes (N = 11) or S-EEG and subdural grid (N = 2). The type, number and location of the electrodes were determined by the suspected location of the epileptogenic zone in each patient according to noninvasive tests including clinical seizure history, neuroimaging, neuropsychology and scalp EEG recordings. Patients underwent S-EEG implantation if: (1) the suspected seizure onset zone (SOZ) was in deep-seated locations such as the mesial structures of the temporal lobe, but imaging (MRI or PET) was non-lesional, (2) failure of previous invasive studies to clearly outline the exact location of SOZ, or (3) there was a need for bilateral exploration for possible multifocal seizure onset.

The implanted electrodes consisted of arrays of macroelectrodes (2.3 mm exposed diameter, 1 cm spacing, Adtech [Racine, WI] or PMT Corp. [Chanhassen, MN]). Depth electrodes were implanted stereo tactically using ROSA robotic assistant device (Medtech, Montpellier, France) as part of standard patient care. The S-EEG depths (AdTech Medical Instruments Corp., WI, USA) were multi-contact, and consisted of 4–14 cylindrical 2.3 mm long platinum contacts separated by 5 mm between centers of adjacent electrodes of the same bundle.

2.3. Electrode localization

Final electrode locations were obtained by combining the information from post-implantation CT and brain MRI using BioImage Suite (Duncan et al., 2004). Using the FreeSurfer parcellation and visual verification with post-implant MRI, electrodes within cortical and subcortical gray matter were identified (Fischl, 2012).

2.4. Electrophysiology

Intracranial evoked potentials elicited by SPES were recorded using a NeuroPort amplifier (Blackrock Microsystems, Salt Lake City, UT), filtered (analog Butterworth antialiasing filters: first-order high-pass at 0.3 Hz, third-order low-pass at 7500 Hz), digitized at 16-bit resolution, and down-sampled to 1 kHz with a digital antialiasing filter. SPES was delivered using a CereStim R96 (Blackrock Microsystems, Salt Lake City, UT). Stimulation was applied in a bipolar manner to pairs of adjacent electrodes using biphasic pulses (0.15 ms/phase) at 0.4 (P1-P4) or 0.5 (P5-P13) Hz frequency. Stimulation was conducted on all possible pairs in gray matter including sites within the SOZ and sites with considerable inter-ictal discharges. At each stimulation site, the SPES procedure consisted of a series of titration blocks followed by one full block (Figure 1A). Titration blocks consisted of 10 pulses delivered at each current intensity, starting from 0.5 mA (P1) or 1.0 mA (P2-P13) and increasing by 0.5 mA (P1) or 1.0 mA (P2-P13) up to a maximum current intensity (range of 4–10 mA, depending on the patient. The maximum current intensity used varied by patient and was determined by visual confirmation of consistent evoked potentials during real-time visualization by two board-certified neurologists (JYK, NEC). The charge ranged from 0.075–1.5 μC/phase, and with a surface area of 0.081 cm², the charge density range was 0.926–18.519 μC/ cm²/phase. After the titration blocks were completed for an electrode pair, a full block of 50 (P1-P4) or 40 (P5-P13) pulses was delivered at the maximum current intensity at the same electrode pair before proceeding to the next stimulation site.

Figure 1.

Experimental Methods. (A) Single-pulse electrical stimulation (SPES) was applied in a bipolar manner to adjacent electrodes at 0.4 or 0.5 Hz. At each stimulation site, SPES consisted of titration blocks at incrementally increasing current intensities for 10 trials each and a full block at the maximum current intensity for 40 or 50 trials. (B) Example response traces at the same site to SPES of increasing current intensity. Trials at each current intensity were centered using the baseline mean (−500 to −10 ms) and averaged to obtain average responses of each stimulation-response pair at each current intensity. (C) The amplitude, latency, and polarity of the N1 peak within 10 to 50 ms post-stimulus was identified in each average response trace. (D) The amplitude of the N1 peak normalized by the pre-stimulus baseline was used to quantify the Z-score. Responses with Z-score greater than 6 were considered significant evoked potentials, representing an electrophysiological relationship between stimulation and response site. Example significant responses from a full block are mapped here, as circles located at the response site colored according to the magnitude of the Z-score with lines coming from the red stimulation site. Three response electrodes are highlighted to show how the average response waveform varied with current intensity at different locations.

Electrode channels with excessive noise were rejected by visual inspection, and all channels were re-referenced using a bipolar montage. Stimulus artifacts in response channels were removed by replacing artifactual signals from −5 to 10 ms with reversed, tapered copies of the signals immediately before and after this period, to preserve the time-frequency composition of the signal surrounding the stimulus (Crowther et al., 2019). Following artifact removal, all signals were forward and backward low-pass filtered at 50 Hz using a finite impulse response (FIR) filter to isolate the CCEP waveform and remove line noise. Filtered signals were segmented into stimulus-locked analysis windows of −500 ms pre-stimulus to 1500 ms post-stimulus. All pre-processing and analysis were performed using custom scripts in Matlab (R2020a, MathWorks, Natick, MA).

2.5. Evoked potential N1 peak and Z-score

Signals within the analysis window from −500 ms to 1500 ms were baseline-centered and averaged across the 10 trials for each titration block and across the 40 or 50 trials for each full block, to obtain average responses of channels to each stimulation site at each titrating and maximum current intensity (Figure 1B). Evoked responses from SPES can typically be characterized by an early negative deflection occurring 10 to 50 ms post-stimulus, known as the N1 potential, thought to represent direct excitatory connectivity (Matsumoto et al., 2004). The timing, amplitude, and polarity of the N1 peak in each average response were identified using a peak detection algorithm. The absolute value of the N1 potential voltage in the average response was marked to quantify the magnitude of the evoked potential (Figure 1C). The N1 voltages were also normalized by the standard deviation of the pre-stimulus baseline period from −500 to −10 ms. This normalized N1 potential is defined as the Z-score of the channel’s response and quantifies the effective connectivity between stimulation and response sites, with Z-score > 6 considered significant (Keller et al., 2011). We used the Z-score of responses to the full block stimulations at the maximum current intensity to determine the significance of connections.

2.6. Signal-to-noise ratio of early response

We also computed a more general measure of the evoked response by quantifying the signal-to-noise ratio (SNR) of each channel’s early response to stimulation. Not only is this measure more suited to quantifying responses at lower current intensities that may not have clear N1 peaks, but it also enabled us to determine at what current intensity did evoked potentials rise above noise level. We computed the SNR as the ratio of the variance of a channel’s signal from 10 to 100 ms post-stimulus to the variance of the channel’s pre-stimulus baseline. The 10 to 100 ms window was chosen to sufficiently capture the rise and fall of the N1 peak of the evoked potential and the fast positive deflection that immediately follows N1, known as P2 (Araki et al., 2015).

The significance of the observed SNR was determined by comparing it to a permutation distribution, generated by applying random circular shifts in time to each trial and repeating the SNR calculation 500 times for each response to each stimulation (methods similar to Crowther et al., 2019 for evoked spectral responses). Because of the random misalignment of the responses with respect to the stimulus, this distribution is representative of the values expected if the signal is not modulated by the stimulus. The resulting P-values for the probability of greater SNR values than those observed were corrected for multiple comparisons using false-discovery rate, and significance was determined using adjusted P < 0.05 (Benjamini and Hochberg, 1995). We then determined at what current intensity above which all responses had statistically significant SNRs, as this is indicative of a minimum threshold that must be reached to elicit significant evoked responses.

2.7. Waveform similarity

While evoked responses are typically quantified based on the magnitude of specific temporal features, such as the amplitude or root mean square (RMS) of the N1 potential, it is also useful to investigate how the overall shape of the response waveform changes with current intensity. While there can be variability in the exact timing and scale of these features in the waveforms of the same stimulation-response pair at different current intensities, overall shapes can look very similar once enough current is applied. Therefore, elastic measures such as dynamic time warping (DTW) that provide time-invariance can be useful in quantifying the similarity of response waveforms in ways that time locked comparisons, such as correlation, cannot.

Time series data can be compared using DTW, in which the alignment of samples in each time series is stretched or compressed to best align the series with each other (Sakoe and Chiba, 1971). Similarity can then be determined using the Euclidean distance between the realigned time series, which we will call the DTW distance. The smaller the distance, the more similar the signals. For applications such as clustering of time series, averaging methods that incorporate DTW, such as DTW barycenter averaging (DBA), can provide robust averages of times series as cluster centers whose DTW distance to all time series included in the cluster is minimized (Petitjean et al., 2011). Specifically, DBA k-Means differs from k-Means in that cluster centers are determined by DBA of the time series in the cluster, and the cluster assignment of each response is based on the closest cluster using DTW distance. Since this method allows for time series that have distinctive features occurring at slightly different times to be considered similar, this is particularly useful for comparing and clustering the waveforms of evoked potentials.

All average responses were normalized by subtracting the mean and dividing by the standard deviation. DBA k-Means was performed using the tslearn package (Tavenard et al., 2020) to cluster all responses of a given channel from stimulation of the same electrode pair into two groups based on titrating current intensity. This is hypothesized to separate responses that contain an evoked potential from those that do not, and the cluster with higher current intensities may be representative of a waveform characteristic of that channel’s maximal response to that stimulation site. The DTW distance of each response to the cluster center of the group most assigned to the responses at greater current intensities was computed as a measure of how similar each response is to an identified characteristic response, if one exists for that channel. The distances for each stimulation-response pair were normalized by the maximum distance across all current intensities in that stimulation-response pair. We also determined at what current intensity above which responses no longer switched cluster assignment, as this was indicative of a threshold that must be reached for responses to shift to a consistent, characteristic shape, if one existed for that stimulation-response pair. An example process for one stimulation pair is shown in Figure 2.

Figure 2.

Dynamic time warping (DTW) and DTW barycenter averaging (DBA) k-Means Analysis. (A) Average responses of the same stimulation-response pair to each titrating current intensity were normalized and clustered using DBA k-Means. (B) The DTW distance from the normalized average response to the center of the cluster at greater current intensities was computed. Example traces for the responses to 1, 3, and 5 mA are shown.

2.8. Spatial variation of responsive channels

We were also interested in how the set of responsive channels at different current intensities compared to those seen in the full block. Since the responses to full blocks were generated using the greatest number of trials and maximum current intensity, this comparison allowed us to quantify how closely the spatial distribution of responses from lower current intensities matched the best observed representation of the spatial distribution of responses. First, the Pearson’s correlation of the vector of N1 responses in titration blocks with the vector of N1 responses from the stimulation site in the full blocks was computed, since the correlation will be greater when the set of the most responsive channels in a titration block is more similar to that in the full block. Additionally, for each stimulation site, we used each channel’s response Z-score from each titration block as a predictor of whether that channel had a significant response in the full block (Z-score > 6). Using simple thresholding of the Z-scores of all stimulation-response pairs for a given current intensity in each patient and varying a discriminative threshold value over the range of Z-scores, we generated receiver operating characteristic (ROC) curves for each titration level. The area under the curve (AUC) was calculated to quantify how well the responses at each titrating current predicted significant responses at the maximum current intensity. This can characterize the similarity of the sets of responsive channels because similarly responsive channels would be better predictors.

2.9. Statistical Analysis

Stimulation-response pairs at each titration current intensity were separated into significantly responsive and non-responsive groups based on significance of Z-score of the same stimulation-response pair in the full block, as described above. To determine how N1 voltage, SNR, and DTW distance varied at different current intensities, Kruskal-Wallis tests were used to calculate whether the current intensity had a significant effect on each of these response variables within channel group, followed by post-hoc Dunn’s tests were used to compute pairwise differences in metrics between different current intensities. Wilcoxon rank sum tests were used to calculate differences between responsive and non-responsive groups at each current intensity. All P-values were Bonferroni corrected for multiple comparisons, and adjusted P < 0.05 was considered significant. Effect sizes of significant results were calculated using eta-squared on the Kruskal-Wallis H statistic or Pearson’s correlation r on the Dunn’s test and Wilcoxon rank sum test z statistic (Tomczak and Tomczak, 2014). Detailed information on resulting test statistics and effect sizes are shown in Supplementary Material. Statistical calculations were computed in R (R Core Team 2019) with additional package rstatix (Kassambara, 2021).

2.10. Data availability statement

The data that support the findings of this study are available from the corresponding upon reasonable request.

3. Results

Thirteen patients (6 males, 7 females, median age 42, range 20–54) were included in this study (Table 1). A median of 102 electrodes were implanted (range 52–214) and a median of 38 electrodes stimulated (range 16–90) across all patients. Across all patients we observed 155,712 total average responses (per patient: median 7,639, range 3,486–30,070), with 48,852 classified as significant evoked potentials (per patient: median 2,294, range 437–13,950). Titrating current intensities varied from 0.5 to 10 mA depending on the patient, with the range kept around 5 mA for P1-P7 but extended up to 10 mA for P8-P13. While stimulation location varied across patients, each stimulation-response pair was treated the same, regardless of brain region, under the simplified assumption that overall connectivity was homogeneous and isotropic.

Table 1:

Patient characteristics.

Patient No.	Sex	Age	Implant Type	Number of Electrodes Implanted (Stimulated)	Titration Block Current Intensity Range (Increments) [mA]	Full Block Current Intensity [mA]	Range of Number of Significant Responses Observed for Each Full Block Stimulation (Median)
P1	F	51	S-EEG	88 (16)	0.5–4.0 (0.5)	4.0	3–25 (10.5)
P2	F	42	S-EEG, Grid	214 (18)	0.5–4.5 (1.0)	4.5	4–47 (20)
P3	M	48	S-EEG	103 (34)	0.5–5.5 (1.0)	5.5	1–39 (11)
P4	F	23	S-EEG	108 (35)	1.0–6.0 (1.0)	6.0	7–39 (22.5)
P5	F	23	S-EEG	168 (20)	1.0–5.0 (1.0)	5.0	2–48 (7.5)
P6	M	51	S-EEG	114 (18)	1.0–4.0 (1.0)	5.0	3–24 (14)
P7	M	45	S-EEG	70 (47)	1.0–5.0 (1.0)	5.0	1–26 (9)
P8	F	39	S-EEG	112 (68)	1.0–10.0 (1.0)	5.0	6–63 (41)
P9	M	20	S-EEG, Grid	98 (90)	1.0–10.0 (1.0)	10.0	0–64 (13)
P10	F	48	S-EEG	54 (34)	1.0–10.0 (1.0)	10.0	6–34 (16)
P11	M	33	S-EEG	52 (42)	1.0–10.0 (1.0)	10.0	5–33 (15)
P12	M	24	S-EEG	102 (62)	1.0–10.0 (1.0)	10.0	0–59 (14)
P13	F	54	S-EEG	92 (50)	1.0–10.0 (1.0)	10.0	3–54 (34)

No., Number; M, Male; F, Female; S-EEG, Stereo electroencephalography

3.1. N1 peak voltage increases and plateaus with greater current intensity

In each patient, the current intensity had a significant effect on the amplitude of the N1 voltages in significantly responsive stimulation-response pairs (Kruskal-Wallis tests, P < 0.05) and the effect size was large (eta-squared > 0.14) or moderate (eta-squared > 0.06) (Figure 3, Supplementary Table 1). While the effect was also significant for non-responsive stimulation-response pairs, effect size was small in every patient (eta-squared < 0.06). Post-hoc pairwise comparisons showed significant differences in median N1 voltages between lower current intensities (Dunn’s tests, P < 0.05), while differences between N1 voltages at greater currents did not show significance above a certain current intensity in each patient (Figure 3, Supplementary Figure 1A). The current intensity above which all pairwise comparisons were no longer significant was indicated by the bar marked n.s. on each plot in Figure 3 and varied between 2 and 7 mA, depending on the patient. The maximum in this range, 7 mA, may be considered a conservative threshold needed to be reached for the N1 voltage of responses to reach a plateau (Figure 3). These differences are also visually apparent from the plots in Figure 3, particularly in patients with the 10 mA range, where median N1 voltages were more variable at lower currents, then increase and level off at greater currents.

Figure 3.

N1 voltage comparisons across current intensity. The N1 peak voltages of stimulation-response pairs at each titration current intensity were compared using Kruskal-Wallis tests separately for stimulation-response pairs classified as significantly responsive or non-responsive in the full block. There was a significant effect of current intensity in each patient, and post-hoc Dunn’s test pairwise comparisons showed significant differences between median N1 voltages at low current intensities. At higher current intensities pairwise differences between N1 voltages of significantly responsive stimulation-response pairs were no longer significant, indicated by the bar labeled n.s. (non-significant) on each plot. Error bars represent 95% confidence interval of the median.

Comparisons between responsive and non-responsive stimulation-response pairs at each current intensity showed a significant difference in median N1 voltage at almost every current intensity for each patient (Wilcoxon rank sum tests, P < 0.05). The effect size increased with current intensity and reached a moderate effect size from 3 to 8 mA (r > 0.3) in all patients except P5 and P6 (Figure 3, Supplementary Figure 1B). While it is expected that the difference between responsive and non-responsive stimulation-response pairs increases with current, knowing what current intensity is necessary for sufficient separation is informative since a common goal of SPES is to distinguish significantly responsive channels from non-responsive channels.

3.2. SNR increases and plateaus with greater current intensity

Observed SNR values are shown in Figure 4A. These generally followed the same trend as the N1 peak voltages, which was expected since the signal time range for calculating SNR included the N1 peak. We saw low SNR at low current intensities, and as current increased, the observed SNR increased and began to plateau in some patients. Current intensity had a significant effect on the SNR of the responsive stimulation-response pairs (Kruskal-Wallis tests, P < 0.05) with a moderate effect in P3 and P6 (eta-squared > 0.06) and large effect among the rest (eta-squared > 0.14) (Figure 4A, Supplementary Table 1). While non-responsive stimulation-response pairs had a significant effect, effect size was very small in each patient (eta-squared < 0.06). Post-hoc pairwise comparisons showed similar results to N1 voltage, that there were significant differences between median SNRs at lower current intensities, but there was a certain current above which there were no longer any significant differences, as indicated on the plots by the bar marked n.s., ranging between 2.5 to 7 mA (Figure 4A, Supplementary Figure 2A). Like the N1 voltage results, the maximum of 7 mA may represent a conservative threshold for SNR of responses to reach a plateau (Figure 4A). Using Wilcoxon rank sum tests for pairwise comparisons between SNRs of responsive and non-responsive stimulation-response pairs at each current intensity, there was a significant difference at every current intensity, with a moderate or large effect size (r > 0.3) reached by 2 to 4 mA and above, across all patients (Figure 4A, Supplementary Figure 2B).

Figure 4.

Early response signal to noise ratio (SNR) comparisons across current intensity. (A) The observed SNR at each titration current intensity was compared separately for stimulation-response pairs classified as significantly responsive or non-responsive in the full block. Post-hoc Dunn’s test pairwise comparisons show significant differences between observed SNR at low current intensities. At higher current intensities pairwise differences between median observed SNR of significantly responsive stimulation-response pairs were no longer significant, indicated by the bar labeled n.s. on each plot. Error bars represent 95% confidence interval of the median. (B) Box plots indicating the distribution of the current intensities above which the observed SNR stayed significant, for each patient.

We also computed the significance of the observed SNR values based on a generated permutation distribution of randomly shifted trials for each stimulation-response pair. The current intensity above which the SNR stayed significant was determined. The distribution of this threshold for every stimulation-response pair classified as responsive in the full blocks is shown in the boxplots for each patient in Figure 4B. Notably, the median value was relatively consistent across patients, between 2 and 4 mA. This is the same range shown in Figure 4A in which the difference between SNR of significantly responsive and non-responsive channels reached a moderate effect size and is therefore indicative of a minimum current intensity that must be reached to elicit evoked potentials from SPES that rise above noise.

3.3. Response waveforms are more similar at greater current intensities

The DTW distances between the waveforms of stimulation-response pairs at each titration current intensity and the waveform of the DBA cluster center are shown in Figure 5A. As mentioned above, this distance provided a metric of how similar a response’s waveform shape is to an average waveform reached at higher current intensities, with smaller distance meaning more similar. As current intensity increased, the DTW distance of significantly responsive stimulation-response pairs decreased and eventually leveled off at higher current intensities. This was consistent with our hypothesis that the waveforms of responsive channels are more similar at higher current intensities and eventually reach a point where the shape of responses are more consistent. Current intensity had a significant effect on DTW distance in both responsive and non-responsive stimulation-response pairs (Kruskal-Wallis tests, P < 0.05), and the effect size was greater for responsive channels in every patient (eta-squared > 0.14) (Figure 5A, Supplementary Table 1). Post-hoc comparisons between current intensities showed significant differences between median DTW distance at lower currents, but in each patient there was a current intensity above which differences were no longer significant, ranging from 2.5 to 6 mA, indicated by the bar labeled n.s. on each plot (Figure 5A, Supplementary Figure 3A). This could indicate a point where the waveforms become relatively consistent, as evidenced by their DTW distance from the characteristic response reaching a minimum, with the maximum of 6 mA considered a conservative threshold to reach this consistency (Figure 5A). Comparisons of median DTW distance in responsive to non-responsive stimulation-response pairs at each current intensity showed significant differences (Wilcoxon rank sum tests, P < 0.05) that generally increased in effect size as current intensity increased (Figure 5A, Supplementary Figure 3B).

Figure 5.

Comparisons of evoked potential waveforms across current intensity using dynamic time warping (DTW) distance. (A) The DTW distance of stimulation-response pairs at each titration current intensity were compared using Kruskal-Wallis tests separately for stimulation-response pairs classified as significantly responsive or non-responsive in the full block. There was a significant effect of current intensity in each patient, and post-hoc Dunn’s test pairwise comparisons show significant differences between median DTW distance at low current intensities. At higher current intensities pairwise differences between DTW distance of significantly responsive stimulation-response pairs were no longer significant, indicated by the bar labeled n.s. on each plot. Error bars represent 95% confidence interval of the median. (B) Box plots indicating the distribution of the current intensities above which the cluster assignment of each stimulation-response pair remained constant, for each patient.

For each stimulation-response pair, the maximum current above which the cluster assignment of responsive channels did not change was determined and the distributions of these values for each patient are indicated in Figure 5B. Since this is the current after which the waveform of the responsive channel was similar enough to be classified in the same cluster, we might consider this the threshold for having a characteristic shape.

3.4. Locations of significant responses are more similar at greater current intensities

The Pearson’s correlation coefficient between the Z-scores from each titration block and the Z-scores from the full block at maximal intensity of the same stimulated pair is indicated for each patient in Figure 6A. As the current increased, the correlation between these Z-scores increased and started to plateau in some patients. A high correlation (r > 0.80) was achieved between 3 to 7 mA across all patients, indicating that the set of responsive channels at the lower current was similar to the set of the most responsive channels in the full block stimulation of the same pair. Additionally, the accuracy of using the Z-scores from each titration block as a predictor of whether channels had a significant response in the corresponding full block provided another metric of the similarity of responding channels. As current increased, the Z-scores of titration blocks became better predictors of the significant responses in the full block, as indicated by the increasing AUC in Figure 6B. Across all patients, an AUC of at least 0.8 was achieved between 2.5–5 mA across all patients and continued to increase with greater current. This indicates that once a current is reached to sufficiently evoke responses, the locations of these responding channels is relatively similar to final locations of responses, and these locations will become more similar and eventually plateau as current intensity is increased.

Figure 6.

Comparisons of spatial distributions across current intensity. (A) Correlation of N1 Z-scores of responses at maximum current intensity in full block with N1 Z-scores of responses at each titration block shows increasing similarity of response distribution with current intensity. Error bars represent 95% confidence interval of the median. (B) Receiver operating characteristic (ROC) curves were generated by varying the threshold of titration block Z-scores to predict significance of responses in the full block (Z-score > 6). High accuracy (area under the curve [AUC] > 0.80) was achieved from responses to stimulations at 2.5–5 mA, depending on the patient, and accuracy generally increased and plateaued at greater current intensities.

4. Discussion

In this study, we conducted a rigorous incremental titration of SPES stimulation intensity in thirteen patients undergoing invasive EEG monitoring. We assessed the effect of increasing stimulation intensity on evoked responses as quantified by the following metrics: early response amplitude (N1), signal to noise ratio, clustering of response waveforms using DTW and spatial variation of responses conducted at maximal intensity. We demonstrated that for SPES delivered in a bipolar manner to pairs of adjacent electrodes with biphasic pulses (0.15 ms/phase) at 0.4–0.5 Hz, a minimum current range of 2–4 mA is generally required to elicit evoked potentials, with all response metrics reaching plateau at 6–7 mA. As the current intensity increased within this range (2–7 mA), the amplitude of the evoked response and set of responsive channels increased, while the overall wave shapes became increasingly similar and distinct from background. Our results provide a detailed reference guide for selecting optimal stimulation parameters that are tailored to specific feature(s) of evoked potentials.

A significant limitation in current CCEP research is that there is much variation in the manner at which SPES is conducted between sites. This is a critical methodological inconsistency that may account for discordant findings that potentially limit the clinical utility of CCEPs. Several studies have compared the effect of intensity, duration, and polarity on the magnitude of evoked potentials (Donos et al., 2016; Kundu et al., 2020). Donos et al. reported that underlying applied charge per phase determined the magnitude of intracranial EEG responses to SPES, irrespective of the stimulation current and pulse width combination used. Similarly, Iwasaki et al. 2010 and Enatsu et al. 2012 demonstrated that increasing stimulation intensity (charge/per phase) had a significant effect on N1 amplitude response. Kundu et al. found that this current-dependent relationship was nonlinear across a range of current intensities, and that 5.5 mA may be an asymptotic threshold to consider for monopolar SPES analysis. We opted for bipolar instead of monopolar stimulation for SPES, as it reduced the size of the stimulation artifact (Frysinger et al., 2006) and computational models have suggested that the volume of tissue activated is not dramatically different from monopolar stimulation (McIntyre et al., 2004). Keeping these parameters in mind, we gradually varied stimulation intensity while delivering consistent biphasic bipolar stimulation with constant pulse width at 0.15 ms/phase.

We found that the amplitude of evoked potentials, as quantified by the N1 peak voltage and signal to noise ratio (SNR) of the early response, initially increased with current intensity. At higher current intensities, as Kundu et al. reported, this “dose-response” becomes an asymptotic nonlinear curve. For both N1 and SNR, this asymptote is reached at a stimulation intensity between 2.5–7 mA. This plateau is consistent with our fundamental understanding that evoked potentials represent a direct excitatory relationship between stimulation pair and response sites (Matsumoto et al., 2004); the local and distant response of the network is finite for a given region of stimulation. We also found that the SNR, compared to absolute N1 amplitude peak, is a more sensitive metric to differentiate between significantly responsive (Z-score > 6 in the full block) versus non-responsive sites; the range above which these groups could be differentiated was 2–4 mA for SNR and 3–8 mA for N1 amplitude. This may be because, compared to the N1 amplitude, the SNR is a more general measure of early response magnitude that depends on the variance of the signal within a larger time window, which may better quantify the responses to subthreshold currents that do not have large, distinct N1 peaks. Additionally, the calculation of SNR, in relying on the variance of the signal, is similar to the root mean square (RMS) metric that is commonly used to quantify evoked responses to SPES (Enatsu et al., 2012, 2013, 2015; Lega et al., 2015; Parker et al., 2018) and has been shown to be a robust, sensitive measure of evoked potentials (Prime et al., 2020).

Previous studies have reported qualitative differences in evoked potential morphologies with changes in current intensity (Enatsu et al., 2012; Keller et al., 2014; Kundu et al., 2020). Through correlation of time series at each stimulation intensity, Kundu et al. observed that the evoked response waveform shape shows variability at low current intensities but settles into a characteristic CCEP at higher intensities. We utilized DTW and a DTW-based clustering technique (DBA k-means) (Petitjean et al., 2011) to compare the similarity of waveforms produced per each stimulation set. This not only allowed for slight time-invariance of features when computing waveform similarity, but also provided a means by which to generate average time series as characteristic responses through cluster centers at higher current intensities. Particularly, our hypothesis was that once a threshold current is reached, responses may settle into a characteristic waveform that remains relatively stable as the current increases further.

Therefore, by applying DBA k-Means to evoked potentials in the same stimulation-response pair, we would see the waveforms separate into two clusters based on current intensity, separating current intensities that were able to elicit an evoked potential from those that were not. In this way, we were able to pin-point the stimulation current at which the DBA k-Means cluster shifts to quantify the minimal threshold current that must be exceeded to elicit an evoked response. We found that as current intensity increases, the DTW distance at response channels decreases and eventually levels off at higher current intensities (2.5–6 mA). Since the analysis window captures additional components of the CCEP response, such as the N2 potential, that may reflect recruitment of additional neural ensembles, the stabilization of the response waveform at greater current intensities may reflect maximal recruitment of these ensembles (Kundu et al. 2020). Over a similar range (3–7 mA), the spatial distribution of responses reached a high similarity to that seen at maximal intensity (full block), as quantified by a high correlation (r > 0.80) of response magnitude and predictive value (AUC > 0.80) of response significance. Combined, these findings suggest that once a current is reached to sufficiently activate all regions functionally and/or anatomically connected to the stimulated region, there is also corresponding stabilization of the waveform morphology.

What is the optimal stimulation parameter to elicit CCEPs? The answer depends on the focus of the study and the specific metric utilized to quantify the CCEPs. We conclude that at least 2–4 mA is needed to differentiate between non-responsive and responsive sites. While this stimulation intensity is great enough to elicit evoked responses, the current-dependent relationship of all observed response metrics suggests that studies should use a consistent current intensity for responses to be fairly compared across conditions. For a study examining the maximum amplitude and spatial distribution of N1 responses and stabilization of the waveform morphology, our findings suggest utilizing higher stimulation closer to 6–7mA. There is no significant quantifiable advantage to using stimulation higher than 7 mA for all response metrics that we studied. The stimulation range is also clinically useful, as not all sites may be able to be stimulated at higher intensities due to clinical symptoms (i.e. motor cortex), after discharges and/or even seizures (Arya et al., 2020).

Our study has several limitations. In our institution, SPES is routinely conducted after patients are restarted on home doses of antiseizure medication and after cortical stimulation for functional mapping is conducted. The effect of antiseizure medication on evoked potentials is not known. The maximum titration currents for patients P1-P7 were lower than for P8-P13, which may account for the less robust plateau effect in these patients at maximum current intensity and may contribute to the somewhat large range of thresholds across patients for observing maximal responses. However, these patients (P1-P7) still robustly demonstrated the minimum current required to elicit evoked potentials that rise above noise (2–4 mA) and showed a clear current-dependent relationship. For this study, we focused on stimulation of all radiographically verified electrodes in gray matter and did not segregate responses based on tissue location (mesial temporal structures versus neocortical, seizure onset zone versus non-seizure onset zone) or control for the distance of stimulation response pairs. Our analysis was based on a simplified assumption that neuronal conductivity was homogenous and isotropic, which may be useful for didactic purposes but may not take into account the contributions of underlying neuronal pattern of activation (Chaturvedi et al., 2010). Responses to SPES have been shown to correlate with structural, functional, and/or effective connectivity (Trebaul et al., 2018; Hebbink et al., 2019; Crocker et al., 2021) and can be influenced by additional factors such as underlying excitability of epileptogenic regions (Valentín et al., 2005; Iwasaki et al., 2010; Enatsu et al., 2012). As a result, variability in the stimulated locations and resulting network activation across patients may have contributed to the somewhat large range of thresholds for observing maximal or stable CCEP responses across patients, in addition to any inherent patient variability. To partially compensate for this unknown, our analysis was conducted on a large database with 155,712 responses amongst 13 patients with heterogenous spatial sampling. The relatively consistent results that emerged with this isotropic assumption points to the general applicability of our findings. While investigating these responses in different brain regions and epileptogenic networks remains an important next step, this study provides a foundation on which future work can build. Studies are currently underway to address whether epileptogenicity and/or location may affect the minimum threshold to elicit responses and/or make responses plateau.

5. Conclusions

This work serves to better understand how the choice of stimulation intensity can affect evoked potentials from SPES. We found that there is generally (1) a minimum current that must be used to elicit any evoked responses, (2) a range of currents over which several metrics used to quantify evoked response magnitude, waveform shape, and distribution are dependent on current intensity, and (3) a threshold current above which all response metrics begin to level off with increasing current intensity. We show that with a biphasic 0.15 ms/phase pulse, 2 to 4 mA is generally required to elicit most responses and all response metrics tend to level off around 6 to 7 mA. As current intensity increased within that range, the amplitude of the evoked responses increased, the overall waveform shapes varied, and the set of responsive channels expanded.

Highlights.

• A minimum stimulation current of 2–4 mA is needed to elicit cortico-cortical evoked potentials (CCEPs) in responsive sites.

• Several metrics quantifying CCEP amplitude, distribution, and morphology all show stimulation current-dependent relationships.

• 6–7 mA is an asymptotic threshold for maximizing CCEP amplitude and spatial distribution and stabilizing waveform morphology.

Abbreviations:

AUC

area under the curve

CCEP

cortico-cortical evoked potential

DBA

dynamic time warping barycenter average

DTW

dynamic time warping

EEG

electroencephalography

RMS

root mean square

ROC

receiver operating characteristic

S-EEG

stereo electroencephalography

SNR

signal to noise ratio

SOZ

seizure onset zone

SPES

single-pulse electrical stimulation