Leveraging electrophysiologic correlates of word encoding to map seizure onset zone in focal epilepsy: Task-dependent changes in epileptiform activity, spectral features, and functional connectivity

Abstract

Objective

Verbal memory dysfunction is common in focal, drug resistant epilepsy (DRE). Unfortunately, surgical removal of seizure generating brain tissue can be associated with further memory decline. Therefore, localization of both the circuits generating seizures and those underlying cognitive functions is critical in pre-surgical evaluations for patients who may be candidates for resective surgery. We used intracranial electroencephalographic (iEEG) recordings during a verbal memory task to investigate word encoding in focal epilepsy. We hypothesized that engagement in a memory task would exaggerate local iEEG feature differences between the seizure onset zone (SOZ) and neighboring tissue as compared to wakeful rest (“non-task”).

Methods

Ten participants undergoing presurgical iEEG evaluation for DRE performed a free recall verbal memory task. We evaluated three iEEG features in SOZ and non-SOZ electrodes during successful word encoding and compared them with non-task recordings; inter-ictal epileptiform spike (IES) rates, power in band (PIB), and relative entropy (REN; a functional connectivity measure).

Results

We found a complex pattern of PIB and REN changes in SOZ and non-SOZ electrodes during successful word encoding compared to non-task. Successful word encoding was associated with a reduction in local electrographic functional connectivity (increased REN) which was most exaggerated in temporal lobe SOZ. The IES rates were reduced during task, but only in the non-SOZ electrodes. Compared with non-task, REN features during task yielded marginal improvements in SOZ classification.

Introduction

An important step in the surgical evaluation and treatment of drug resistant focal epilepsy is mapping the brain networks underlying normal physiological functions, like memory, and the pathological network(s) generating spontaneous, unprovoked seizures^1–3. This is particularly relevant for verbal memory, where surgical resection and stimulation therapies can cause speech and memory deficits. A fundamental clinical challenge is to improve prediction of pre-operative seizure and memory outcomes. The need for better tools to define the seizure onset zone (SOZ), map verbal memory function, and guide resection margins is well recognized⁴. Currently, SOZ localization requires capturing spontaneous, unprovoked seizures during intracranial electroencephalographic (iEEG) monitoring⁵. Yet, when considering all patients undergoing epilepsy surgery, only approximately 50 – 60 % of patients achieve long term seizure freedom⁶. Considerable effort has been applied to improving epileptogenic brain localization using inter-ictal iEEG recordings: the long periods wherein patients are not having seizures. With the prevalence of cognitive difficulties in people with epilepsy (PWE), inter-ictal cognitive tasks may constitute an additional tool for epileptogenic brain localization and ultimately support the assessment of post-operative risk of memory decline^7–10.

To develop a better understanding of the inter-ictal electrophysiologic correlates of word encoding and the inter-ictal biomarkers of epileptic brain, we investigate iEEG, which has been widely used to characterize electrographic features of both the SOZ and memory impairments in PWE^11–13. Features commonly studied and incorporated into multi-feature approaches include inter-ictal epileptiform discharges, power in band, and multivariate measures^14–19. Inter-ictal epileptiform activity, including interictal epileptiform spikes (IES) and high frequency oscillations (HFO), during verbal encoding has been associated with memory disruptions^{12, 20}. Reductions in theta band power, an important feature of memory processing in rodent navigation, are also related to memory disturbances in PWE^{21, 22}. Local network alterations may also help explain the observed links between spectral changes and memory performance in PWE. Functional dissociation of SOZ from the surrounding tissue has been previously identified in inter-ictal iEEG, warranting further exploration of bivariate features in different behavioral states^{19, 23, 24}. Continued evaluation of these features during cognitive tasks may further support cognitive state-specific approaches to SOZ localization and to map the networks underlying normal verbal memory^{9, 11, 16, 25}.

The utility of many of these features for inter-ictal SOZ characterization is well established, but they remain imperfect tools and their performance in different brain states remains a limited area of research. Previous investigations of SOZ biomarkers in task have been in highly focused, comparing high frequency oscillations in epileptic and non-epileptic hippocampus^{9, 10}. Here, we take a more general approach, using a heterogeneous set of recordings, primarily frontal and temporal subdural, electrocorticographic (ECoG) grids and strips, to determine if engagement in a cognitive task known to activate these brain regions would improve the SOZ localization performance of commonly evaluated iEEG features. Such insights into how SOZ and the surrounding brain respond to cognitive stimuli are necessary to motivate further investigation into the use of cognitive tasks for inter-ictal SOZ localization.

We hypothesized that engagement in a memory task would exaggerate local differences in electrographic features as compared to wakeful rest, and better delineate the SOZ. We evaluated two univariate features and one bivariate feature during task and non-task states in the same ten participants undergoing iEEG monitoring. For task data, we used inter-ictal iEEG recordings while the subjects performed a delayed, free-recall memory task that activates the temporal lobe and frontal cortex (Brodmann areas 9 and 10)²⁶. To lessen the potential impact of epileptiform activity and keep our analysis within a more constrained, “physiologic” brain state, we only used encoding period data from successfully recalled trials. For non-task data, we used inter-ictal iEEG recordings while the subjects were awake and not performing any formal cognitive assessments or tasks. We explored (1) power in band (PIB), (2) inter-ictal epileptiform spikes (IES), and (3) relative entropy (REN), a measure of functional connectivity between electrode pairs. We found that: 1) IES were reduced during word encoding in non-SOZ, but not SOZ; 2) differences in the bivariate feature REN between SOZ and non-SOZ (NSOZ) electrodes were exaggerated during task; and 3) SOZ classification using REN improved in task compared to non-task.

Methods

Study participants

Ten subjects participated in this study while undergoing iEEG monitoring for epilepsy surgery evaluation at the Mayo Clinic, Rochester MN (Table 1). The Mayo Clinic Institutional Research Board approved the research protocol and informed consent was obtained from each participant. The number and placement of intracranial electrodes and duration of iEEG recording were determined by the clinical team with the goal of localizing epileptogenic brain tissue for resective surgery. The research studies were performed during the clinical monitoring period at times without electrographic or clinical seizure activity.

Table 1.

Participants. The gender, age, and handedness of each participant are reported here. Seizure onset zone locations and clinically relevant imaging findings are also highlighted. Histopathological findings from resected tissue are reported if available. Outcome scores refer to the Engel Surgical Outcome Scale where “IA” denotes complete seizure freedom since surgery and “IIIB” denotes worthwhile seizure reduction. Subjects where the surgical outcome was “NA” were not deemed surgical candidates and received neuromodulatory treatment.

Subject	Gender, Age, Handedness	Seizure onset zone	MRI / Histopathology	Surgical Outcome
S1	M, 40, R	R mesial temporal, L mesial temporal	Lesional: suspected MTS	N/A
S2	F, 28, L	R posterior temporal neocortex	Lesional: suspected focal cortical dysplasia	N/A
S3	M, 26, R	L anterior lateral frontal, L parietal neocortex	Lesional: suspected focal cortical dysplasia	N/A
S4	F, 29, R	R mesial temporal	Unclear significance: Mild increase T2/FLAIR signal right hippocampus. Hypoplasia of the posterior aspect of the corpus callosum Path: Subpial (Chaslin) gliosis	Engel IA
S5	M, 34, R	R anterior frontal, R middle temporal	MRI-Negative Path: Focal Cortical Dysplasia Type IIb	Engel IA
S6	M, 57, R	L mesial and lateral superior frontal	Lesional: Focal encephalomalacia in the anterior and inferior right frontal and temporal lobes, small foci in the inferior left frontal and temporal lobes	N/A
S7	M, 33, R	L anterior temporal, L posterior frontal	MRI-Negative Path: Focal Cortical Dysplasia Type IIa	Engel IA
S8	F, 40, R	R insula, R superior and inferior temporal, L amygdala and hippocampus	Lesional: Prior right anterior temporal lobectomy Path: Subpial (Chaslin) gliosis	Engel IA
S9	F, 22, R	L posterior temporal	Lesional: Suspected tuberous sclerosis	N/A
S10	M, 42, R	R temporal and mesial temporal	Lesional: Polymicrogyria in the right posterior perisylvian region, prior anterior right temporal lobectomy Path: Mild, nonspecific gliosis	Engel IIIB

Abbreviations: MTS, mesial temporal sclerosis; R, right; L, left.

Anatomical localization

Pre-implant, volumetric T1-weighted MRI sequences were used to generate cortical surface parcellations²⁷ and to determine electrode coordinates in MNI space via co-registration with post-implant CT. Supplementary figure 1 depicts electrode recording sites across all participants.

Electrophysiological recording

iEEG data were recorded with the Natus XLTek EMU 128 clinical acquisition system (Natus Medical Inc.) and sampled at 500 Hz in a referential montage. Recordings included ECoG and stereoelectroencephalographic (SEEG) electrodes. A full list of electrode types by patient is available in Table S3. A bipolar montage of spatially adjacent pairs of contacts was calculated post-hoc^{28, 29}.

Memory task

Each subject performed the delayed, free-recall, verbal memory task (Figure 1). Subjects were asked to memorize lists of 12 common nouns presented on a laptop screen. PET studies have shown increased activation during free recall in the temporal lobe and in Brodmann areas 9 and 10 in the frontal lobe for word lists of this length²⁶. Words from the list were shown sequentially; each word remained on the screen for 1600 ms, followed by a randomly jittered 750 – 1000 ms blank inter-stimulus interval. Immediately following the final word in each list, participants performed a math distractor task (> 20 s). Following the distractor task, participants were given 30 seconds to verbally recall words from the list in any order. Forgotten and recalled words were noted. Each session consisted of 25 different lists. Each subject completed at least 18 lists. From the iEEG recording at each electrode, a 3000 ms signal segment was chosen for each encoding trial (word) which consisted of 1600 ms of stimulus, and 700 ms each of pre-stimulus and post-stimulus signals. The encoding period of each recalled trial was retained for further analysis under the assumption that recalled trials represent periods where the subject was fully engaged in the task and epileptiform activity did not interfere with task performance¹¹. Each subject, on average, completed 390 trials and recalled 30.1% of the trials, resulting in 54 minutes (n = 1074 trials) of recordings of recalled trials.

Figure 1.

Task Description and Data Processing. (A) depicts the free recall verbal memory task. During encoding, participants are shown a list of twelve consecutive words. Participants then complete a series of math problems during the distractor phase before freely recalling as many words from the list as they can remember. (B) shows the spectrogram of the encoding period from a single electrode, averaged over all words that were successfully recalled. The associated time series average is overlaid in white. Dashed lines delineate the onset and offset of word (stimulus) presentation, which is also highlighted by the orange bar at the bottom. (C) An average spectrogram and time series for the non-task period at the same electrode. The dashed lines describe the periods defined as “pre-stim” and “stimulus”, although no cognitive task was being performed at this time. Part D shows data processing block diagram. Classifier performance in identifying electrodes as SOZ or non-SOZ was evaluated for each feature (power in band, inter-ictal spikes, and relative entropy) from each brain state (task and non-task). Abbreviations: iEEG, intracranial electroencephalography; PIB, power in band; IES, interictal epileptiform spikes; REN, relative entropy; SOZ, seizure onset zone; AUC, area under the receiver operating characteristics curve.

Non-task data

Inter-ictal, non-task data were defined as awake periods occurring outside of task and at least 40 minutes preceding or following seizures. Non-task data of approximately the same duration and day as the task were used in all but one subject for whom increased seizure frequency and decreased awake time precluded the use of non-task data recorded on task-day. Periods of awake iEEG were identified using a previously published approach for behavioral state classification ^{30, 31} (using opensource published code https://github.com/vkremen/Semi_Automated_Sleep_Classifier_iEEG). Awake iEEG was then manually reviewed and channels with major artifacts and discontinuities were excluded from further analysis. Fifteen-minute periods of awake iEEG were then segmented into 3000 ms segments to reflect the total number and length of encoding period trials during one session of the task. Three such 15-minute periods were segmented for each subject. We will refer to each 3 second segment of non-task iEEG as a “trial”. Because non-task data were not time locked to stimulus presentation, “pre-stimulus” and “stimulus” segments in each trial were based purely on the durations of the pre-stimulus and stimulus periods in the task data.

Seizure onset zone

SOZ channels were identified by visual review of the ictal iEEG recordings by a board certified epileptologist as part of standard clinical practice. Bipolar contact pairs were classified as: (1) SOZ – if both channels in the bipolar pair were SOZ, and (2) non-SOZ - both channels were NSOZ. We excluded “irritative zone” (IZ) electrodes: bipolar pairs where one channel was in the SOZ and the other was not¹⁷. In the rest of the manuscript, we refer to bipolar contact pairs as “electrodes”. Of the 10 subjects, SOZ electrodes were located in the temporal lobe (n = 34), frontal lobe (n = 24), mesial temporal structures (n = 8), parietal lobe (n = 6), and subcortical regions (n = 8). The total number of electrodes used for the task-based analysis (n = 80 for SOZ, n = 769 for NSOZ) and non-task-based analysis (n = 80 for SOZ, n = 706 for NSOZ) differed slightly due to the presence of recording discontinuities in non-task data.

Data preprocessing

We employed separate iEEG pre-processing methods for the time-domain features (IES and REN), and frequency-domain features (PIB) prior to feature extraction.

To process the time domain features for each trial, we (i) removed 60 Hz line noise, (ii) bandpass filtered between 0.1 Hz - 115 Hz, (iii) clipped 200 ms of signal from both ends, and (iv) linearly detrended the signal (see SI). Further, we z-score normalized each trial using the mean and standard deviation of that trial (Figure 1D). The resulting 2600 ms signal was used for all further analyses.

Spectral features for each trial were processed as previously published ²⁹. We (i) removed 60 Hz line noise; (ii) clipped 100 ms of signal from both ends; (iii) filtered the signal into different frequency bands: low theta (2–5 Hz), high theta (6–9 Hz), alpha (10–15 Hz), beta (16–25 Hz), low gamma (36–55 Hz), and high gamma (65–115 Hz); and (iv) created a spectrogram with a 2-Hz frequency resolution and 500 ms sliding window with 50 ms slide length (see SI). Finally, we log-normalized and z-scored the spectrogram within each frequency bin using the mean and standard deviation for the given bin and clipped two time points from both ends.

Inter-ictal iEEG Biomarkers

We extracted two univariate (PIB and IES) and one bivariate feature (REN) for further analysis.

Inter-Ictal Epileptiform Spikes

We used a previously validated automated spike detector to determine the number of spikes at each electrode for every task and non-task trial^32,17. Minor modifications were made to the detector’s signal scaling steps and thresholds after visual review of the IES detections (see SI). We calculated the spike rate as the number of spikes per minute of iEEG. Electrodes with spike rates higher than 60 spikes/min were excluded to reduce the likelihood of artificially high spike rates from artifactual activity. This excluded <1.9% electrodes from the analysis.

Power in Band

Power was extracted from low theta, high theta, alpha, beta, low gamma, and high gamma bands for task and non-task trials. Power was computed separately for the ‘pre-stimulus’ period (−500 ms to 0 ms), and ‘stimulus’ period (0 ms to 1600 ms) because stimulus is known to cause major PIB changes associated with memory encoding^{28, 33–35}. PIB was computed by averaging the power values over all the time bins and frequency bins from the spectrogram corresponding to the given frequency band and period across trials. This resulted in 12 PIB features (6 bands x 2 durations) each for task and non-task data for each electrode.

Relative Entropy

Relative entropy between a pair of electrodes describes the dissimilarity between the signals at those electrodes. The dissimilarity is measured using Kullback-Leibler divergence (KLD) between the signals’ amplitude distributions. The REN between a pair of electrodes was calculated as follows. (i) We computed the KLD for each trial by comparing the signal amplitude distributions between the two electrodes while using one electrode as reference and averaged the value across trials. (ii) We repeated step (i) with the other electrode as reference. (iii) We chose the larger of the two average KLD values as the REN value¹⁹ (see SI). Since REN is a bivariate measure, for a participant implanted with N electrodes, there will be N(N – 1)/2 total REN values, i.e., one corresponding to each pair of electrodes. Due to its bivariate nature, REN is dependent on spatial relations between electrodes. Since the exclusion of IZ electrodes would have artificially exaggerated separation between existing contacts inside and outside the SOZ, IZ electrodes were included in REN calculations.

Statistical analysis and classification

For group-based comparisons of task and non-task features, we used the Wilcoxon Rank-Sum test. Since recordings included ECoG (grid/strip) and SEEG (depth) electrodes (Table S3), we performed the analysis of the ECoG and SEEG electrodes pooled together as well as separately (S5, S7, S8). Unless stated otherwise, results are presented from the pooled analysis. Further, we assessed the SOZ localizing performance of each feature separately for task and non-task using a kernel-SVM classifier (see SI). Average area under the curve of receiver operating characteristics (AUC-ROC) over leave-one-subject-out cross-validation was used as the evaluation metric of the classifier. Classification was performed using ECoG and SEEG electrodes pooled together.

Results

Power in band during task in SOZ and Non-SOZ brain regions

In keeping with the literature on task-induced power changes, we observed a complex pattern of low and high frequency changes induced by the memory task with a highly significant increase (p < 0.001) in high gamma power with word presentation in task-activated electrodes (Figure S2 and SI). There were complex spectral power changes in each frequency band in the task and non-task state for SOZ and NSOZ electrodes across all brain regions (Figure S3) and in the temporal lobe (Figure S4).

For NSOZ electrodes, PIB was significantly different between task and non-task during pre-stimulus in the high theta, alpha, beta, and high gamma bands; and during stimulus for the high theta, alpha, beta, and high and low gamma bands. For electrodes in the SOZ during stimulus PIB was not significantly different between task and non-task in all but the low theta band (Figure S3). In task-activated SOZ electrodes, PIB was significantly different between task and non-task in the gamma bands (Figure S2). Those differences were absent in SOZ electrodes in the temporal lobe (Figure S4). For NSOZ electrodes, differences were observed across several bands for various subsets of electrodes (Figures S2, S3, S4). These data further suggest that NSOZ responds to word presentation “physiologically” with anticipated spectral changes associated with word encoding, whereas the less densely sampled SOZ, which can respond “physiologically”, may do so less consistently, even for words that are successfully remembered¹¹. PIB changes observed for ECoG electrodes alone were very similar to changes observed when considering all the electrodes (Figure S5).

The PIB changes for task and non-task data in various bands and in SOZ and NSOZ electrodes are modest (Figure S3), and suggest that PIB features from task are unlikely to provide a significant advantage over non-task PIB features for SOZ localization as confirmed by the classification analysis (Table S1).

Inter-ictal spike rates decreased in non-SOZ during task but did not improve SOZ localization

IES rates and the effect of task were consistent with previous work that showed IES rates are reduced with physiological activation with a task^{36, 37} (mean for task = 9.3 spikes/min; for non-task = 10.2 spikes/min; p = 0.013). We found that the spike rate was reduced during task compared to non-task only for NSOZ (mean for task = 9.6 spikes/min, for non-task = 10.6 spikes/min, Wilcoxon rank sum test, p = 0.013) (Figure 2). Generally, there was a substantial overlap in the range of IES rates between SOZ and NSOZ electrodes for task as well as non-task, suggesting limited utility of IES for SOZ localization. The results remained the same when including electrodes with spikes rates higher than 60 spikes/min (Figure S6) or while considering ECoG electrodes only, but not for the pool of SEEG electrodes, which did not show significant differences between task and non-task (Figure S7). SOZ localization classifier performance was close to chance for task as well as non-task iEEG using IES as the feature (Table S1).

Figure 2.

Inter-ictal epileptiform spike (IES) rates during verbal memory task and non-task. In the box plots, the central bar denotes the median, the box edges denote the 25^th and 75^th percentiles and the whiskers stop at the most extreme data points not considered to be outliers. Spike rates in SOZ and non-SOZ electrodes are shown during task (hatched-bar) and non-task (light shaded box). (A) Across all electrodes and all subjects, spike rate was decreased during task in non-SOZ electrodes (Wilcoxon rank-sum test, p < 0.05=*) (Task: SOZ n = 80, NSOZ n = 754; Non-Task: SOZ n = 78, NSOZ n = 644). (B) In the temporal lobe specifically, the trend of decreased spike rate during task remains significant (Wilcoxon rank-sum test, p<0.05=*) (Task: SOZ n = 34, NSOZ n = 249; Non-task: SOZ n = 34, NSOZ n = 193).

Relative entropy during task improved SOZ localization

SOZ has been shown to be functionally disconnected from the surrounding brain^{23, 24} and to drive synchronizing and desynchronizing interactions between brain regions in seizure spread³⁸. We explored the effect of a task on functional connectivity in SOZ and NSOZ using REN for all pairs of electrodes within a subject during task and non-task (Figure 3). The pairs are categorized based on the type of electrode contacts involved in computation of REN. During non-task, REN was higher in SOZ-SOZ pairs as compared to NSOZ-NSOZ pairs^{23, 24}. This difference was further exaggerated during task where median REN increased in SOZ-SOZ pairs and decreased in NSOZ-NSOZ pairs. A larger increase in REN for SOZ-SOZ pairs indicates that during task, the distributions of signal amplitudes at SOZ electrodes becomes more dissimilar than the distributions of amplitudes of signal at NSOZ electrodes. Those differences were mainly driven by ECoG electrodes (Figure S8).

Figure 3.

Relative entropy (REN) during verbal memory task and non-task. The REN was increased during task performance in SOZ, NSOZ, and SOZ-NSOZ electrode pairs. The largest change, however, was for temporal lobe electrodes. Significant differences between task and non-task REN values for the given electrode pair type and anatomical region are marked (Wilcoxon rank-sum test, p < 0.05 = *, p < 0.001 = **). For SOZ-SOZ pairs, n1 = 414, n2 = 414. For NSOZ-NSOZ pair, n1 = 35,509, n2 = 32,302. Each subject contributes N*(N-1)/2 REN samples where N is the number of electrodes. (A) Across all electrodes, REN was significantly higher during task than non-task for electrode pairs within the SOZ, electrode pairs with only one electrode in the SOZ, and for non-SOZ electrode pairs. (B) REN increased significantly during task for pairs with at least one SOZ electrode within the frontal lobe. (C) For pairs where both electrodes were in the temporal lobe, REN increased significantly across all pair types. (D) For pairs where only one electrode was in the temporal lobe, REN increased significantly across all pair types. (E) Receiver operating characteristics curves for classification with feature REN (k =5) for each subject during the leave-one-subject-out-cross-validation. Each grey dotted line represents the ROC curve for an individual subject. The average ROC curve across all subjects is shown in red with an area under the curve (AUC-ROC) of 0.67.

To further explore the task-based differences in SOZ and NSOZ REN values, we looked at the influence of anatomical location on REN (Figure 3 and Figure S9). REN increased during task for all pair-types in the frontal and temporal lobes compared to non-task (Figure 3). The greatest task-associated increase in REN was seen when both electrodes were in the temporal lobe (Wilcoxon rank sum test, n1 (task) = 178, n2 (non-task) = 178, p < 0.001). Given the importance of the temporal lobe in verbal free recall^{26, 39}, it is reasonable to conclude that this regional effect of REN is attributable to task engagement.

When both electrodes in the pair were in the temporal lobe, the difference between the median REN values for task and non-task was 0.04 for SOZ-SOZ pairs and 0.002 for NSOZ-NSOZ pairs. Interestingly, when only one electrode in the pair was in the temporal lobe, the difference between task and non-task REN decreased and flipped, with relatively lower REN during task: the difference in median REN values for task and non-task was −0.004 for SOZ-SOZ pairs and 0.001 for NSOZ-NSOZ pairs. Overall, the larger change in REN during task for SOZ-SOZ pairs than in NSOZ-NSOZ pairs suggests that task exaggerates the differentiability of SOZ and NSOZ electrodes in the temporal lobe.

We used REN values to classify SOZ and NSOZ electrodes. The best AUC-ROC was achieved when an electrode was represented by the REN value calculated from its 5 closest electrodes (k = 5) as an input to the classifier (AUC-ROC = 0.67 for task, AUC-ROC = 0.6 for non-task; see Table S1). The ROC curves for this classifier during task show large inter-patient variability (Figure 3). The classifier had better than chance level performance for eight out of the ten participants (Table S2). Patients with AUC-ROC < 0.5 did not have any SOZ electrodes in the temporal lobe (Table S2).

Discussion

The localization of the networks underlying verbal memory encoding and their relationship with SOZ is critical for surgical and neuromodulatory management of drug resistant epilepsy. Given the importance of verbal memory function and the potential to improve SOZ localization strategies using inter-ictal iEEG, we investigated the electrophysiologic characteristics of a verbal memory task in PWE. Given the prevalence of cognitive difficulties in PWE, we hypothesized that iEEG recorded during a delayed free-recall verbal memory task might enhance inter-ictal SOZ localization. We directly compared the efficacy of 3 features of task and non-task iEEG activity for SOZ localization: power in band (PIB), inter-ictal spikes (IES), and relative entropy (REN). We show that IES rates are decreased during verbal memory task in non-SOZ, but not in SOZ brain regions. Electrophysiological differences between SOZ and NSOZ were exaggerated during task compared to non-task for REN but not for PIB or IES. Our findings demonstrate that cognitive tasks performed during iEEG monitoring show different electrophysiologic responses in SOZ versus NSOZ and may have localizing potential.

Relative Entropy

REN in the temporal lobe showed the most pronounced differences between task and non-task conditions, indicating that cognitive tasks exaggerate local differences in functional connectivity, especially in the SOZ. We found that REN was higher between SOZ-SOZ electrode pairs than between NSOZ-NSOZ electrode pairs in both task and non-task¹⁹. Moreover, the REN was higher for SOZ-SOZ pairs for task as compared to non-task, showing improved distinguishability for the SOZ during task using REN. It is unclear why using REN with/from its 5 closest neighbors (k = 5) yielded the best performance. The average distance between electrodes at a k=5 is 14 mm (median = 11.0 mm, standard deviation = 7.4 mm). This may constitute a sweet spot that captures local, pathological, network heterogeneity without encompassing too much inter-regional heterogeneity as would be captured with increasingly distant electrodes. Since the temporal lobe is known to be involved in the free recall task²⁸, we hypothesized that the exaggerated REN values during task are attributable to the temporal lobe’s engagement in the task. These findings further suggest that matching a suspected SOZ to cognitive tasks that recruit that region may further improve task-based epileptogenic brain localization.

The choice of electrodes (surface ECoG or SEEG depths) and their placement has a nonnegligible influence on iEEG network properties and presumably functional connectivity measures such as REN⁴⁰. ECoG grids have regular spacing between electrodes and record from neighboring cortical sources whereas SEEG can encompass a range of targets with varying density and cover anatomical structures including cortex, white matter, and subcortical areas⁴¹. Given the enrichment of ECoG electrodes in this study (> 85%), it is possible that our findings are specific to ECoG. Although our study was not fully powered to evaluate SEEG electrodes separately in the primary analysis, further exploration of recording technique, electrode placement, and density for task-based SOZ localization is warranted.

Our findings reflect a growing literature describing focal epileptogenic brain and the fact that the same circuits generating seizures often sub serve normal functions¹¹. Episodic memory tasks activate a broad network based on theta synchrony in which the SOZ has been shown not to participate²². There is a complex interplay between the physiological and pathological brain activity^{36, 37}, evident in the functional isolation of the SOZ^{23, 42, 43} under task and non-task conditions. Functional connectivity measures extracted from non-task iEEG have been used to describe the state dependence of SOZ connectivity²³, with maximal isolation at seizure onset²⁴. Given that the SOZ is part of dynamic networks³⁸, future studies of SOZ localization using inter-ictal iEEG could capitalize on the state-dependence of functional connectivity measures.

Despite the significant difference between task and non-task REN values, classifier performance using REN, although higher for task data, was not significantly better than for non-task data (AUC = 0.67 for task vs AUC = 0.60 for non-task for k = 5, Wilcoxon signed-rank test, n1 = 10, n2 = 10, p-value = 0.80; Table S1). REN is a highly patient-specific feature as evidenced by the variable ROC curves for each patient in Figure 3 and Table S2. It is impacted by the number of electrodes, the extent of SOZ and NSOZ coverage by the electrodes, the location and type of electrode (Figures S7, S8), and likely by the relative differences in normal baseline activity of the surrounding brain regions. SOZ classification performance using REN was highest in the subjects with ECoG coverage of temporal SOZ (Table S2). Although further improvement in classification performance is crucial for translation of these methods into practice, understanding the factors that influence performance would help in identifying patients who would benefit from the application of the above approach. It is possible that classifier performance using REN would improve through the use of patient-specific algorithms more suited toward multivariate features⁴⁴.

Power in Band

PIB features extracted from task data did not provide an additional advantage in SOZ localization. Given the extensive literature on the use of spectral features to understand memory processing during task^{28, 45} and PIB features in seizure prediction^{46, 47}, we hypothesized that PIB features during task would help discriminate SOZs. SOZ has been shown to have increased relative PIB values and broader PIB distributions during sleep^{16, 44}. It may be difficult to extrapolate observations from sleep-specific power changes to awake EEG, and the transient power fluctuations observed in different frequency bands in different brain regions during the presentation of cognitive stimuli³³. Our use of short 2600 ms intervals of awake iEEG is not directly comparable to the previously published use of extended periods of non-REM sleep iEEG for SOZ localization.

One limitation of our approach was that the electrodes represented a heterogeneous pool of cortical and subcortical brain regions. Spectral features and high frequency oscillations have proven useful in SOZ localization based on task and non-task states when limited to one brain region, such as the epileptogenic or non-epileptogenic hippocampus⁹. The utility of task-based PIB in SOZ localization might depend on deliberate selection of tasks that recruit a specific region of the brain suspected to hold a SOZ.

Inter-Ictal Epileptiform Spikes

Previous demonstrations of decreased spike rates during cognitive tasks^{12, 36, 37} guided our attempt to use IES during task as an SOZ localization feature. Decreased IES together with reduced functional connectivity could be a manifestation of local neuronal assemblies limiting the pathological hypersynchrony that generate IES. NSOZ spike rates did decrease during task in our study, supporting the concept that non-epileptogenic brain tissue has a stronger modulatory impact on spike rates upon network engagement. Although IES are a hallmark of the epileptic brain, their frequency and location are highly variable^{48, 49}. Previous attempts have been made at SOZ localization using IESs, but with only modest success^{1, 25, 49}. The area with the highest spike rate does not always correlate with the SOZ⁵⁰. In one study, the electrodes with the highest inter-ictal spike rate were located in the SOZ for only 11 out of 19 subjects⁴⁹. Our observations of 1) a greater average spike rate among NSOZ electrodes and 2) near chance classifier performance during non-task iEEG are consistent with these studies.

Limitation

One important difference between our study and the SOZ localization literature is our use of a bipolar iEEG montage instead of a referential montage. We used a bipolar montage because of its prevalence in cognitive neuroscience research and the decreased risk of introducing false increases in signal correlation due to the common reference of a referential montage⁵¹. Using the same classifier, we observed a lower AUC for SOZ localization using IES than has previously been published²⁵. Further investigation is warranted to determine if the bipolar montage removes signal that may be useful for task-based localization.

Task-based SOZ localization may only be effective when SOZ electrodes are located in anatomical regions recruited by the task. This possibility is supported by our observation that task REN values are highest when at least one electrode involved in the REN calculation is in the temporal lobe. Task engagement varies by region and may change on different regional time scales. Task-based changes in brain activity associated with verbal memory encoding have been mapped out and shown to be highly varied, even between electrodes in the same brain regions³³. Whether or not our findings could be reproduced or improved by alternative cognitive tasks remains unclear.

Given the heterogeneity of epileptic activity and SOZ locations across the general population of patients with epilepsy, it is possible that testing more features in a larger patient cohort would produce different results. Based on previous research¹⁹, we focused our study on PIB, IES, and REN. Other iEEG features for SOZ localization from non-task data such as high frequency oscillations and phase amplitude coupling may also show task-based differences.

Our results from this individual verbal memory task justify further exploration of SOZ localization using an array of cognitive tasks. Cognitive stimuli might be leveraged as naturalistic probes (or stimuli) of local brain function and aid existing approaches for stimulation-based SOZ mapping. Our observation that task engagement influences SOZ iEEG features supports efforts to conduct certain neuropsychological tests that are typically conducted during pre-surgical evaluations or intra-operative neuropsychological mapping during the extended invasive monitoring period^{52, 53}. For a suspected left temporal SOZ, a battery of tests might include reading, object naming, word repetition, phonological discrimination, and picture description⁵⁴. For a suspected parietal SOZ, tests might include attentional orienting tasks, writing, and calculations^{55, 56}. Pairing cognitive tasks to the putative SOZ may aid SOZ localization and better characterize epileptic brain tissue capable of engaging in “physiologic” task responses, aiding in the delineation of resection margins and anticipation of post-surgical deficits.

Supplementary Material

supinfo

Figure S1. Anatomical representation of iEEG electrode placement in aggregate

Figure S2. Power changes in active electrodes during task

Figure S3. Power in band during task and non-task for all electrodes

Figure S4. Power in band during task and non-task for electrodes in the temporal lobe

Figure S5. Power in band during task and non-task for ECoG electrodes

Figure S6. Inter-ictal epileptiform spike rates during task and non-task states including electrodes with spike rates >60 spikes/min

Figure S7. Inter-ictal epileptiform spike rates during task and non-task states for electrodes stratified by electrode type

Figure S8. Relative entropy (REN) for task and non-task stratified by the three possible combinations of electrode type

Figure S9. REN for task and non-task stratified by anatomical location of pairs of electrodes

Table S1. Seizure onset zone classification results

Table S2. Seizure onset zone classification results for REN (k = 5) per patient

Significance.

Previous studies have supported REN as a biomarker for epileptic brain. We show that REN differences between SOZ and non-SOZ are enhanced during a verbal memory task. We also show that IES are reduced during task in non-SOZ, but not in SOZ. These findings support the hypothesis that SOZ and non-SOZ respond differently to task and warrant further exploration into the use of cognitive tasks to identify functioning memory circuits and localize SOZ.

Key Points.

• Seizure onset zone (SOZ) and non-SOZ show similar, but distinct spectral changes during word encoding.

• Functional connectivity, measured by relative entropy (REN), is reduced during word encoding and this reduction is most pronounced in SOZ.

• The inter-ictal epileptiform spike (IES) rate is reduced during word encoding in non-SOZ, but not in SOZ.

• Compared to power in band and IES, REN during verbal memory task yields superior classifier performance for inter-ictal SOZ classification.

• SOZ classification performance with REN during verbal memory task was highest in subjects with temporal neocortical SOZ and subdural electrodes.