Research ArticleEpilepsy

Cognitive refractory state caused by spontaneous epileptic high-frequency oscillations in the human brain

See allHide authors and affiliations

Science Translational Medicine  16 Oct 2019:
Vol. 11, Issue 514, eaax7830
DOI: 10.1126/scitranslmed.aax7830

A brief refractory state

Patients with epilepsy present abnormal activity in specific brain areas, resulting in the development of seizures. The removal of these epileptic regions is often the best treatment available. Seizure-originating tissue is thought to be chronically dysfunctional; however, it is unclear how epileptic tissue responds to cognitive stimuli before, during, and after seizures. Now, Liu et al. detected and analyzed brain oscillations in nonlesional epileptic tissue of patients while performing cognitive tasks. The author found that the epileptic tissue generated physiological responses to cognitive stimuli except when the stimulus arrived at the 1-s window preceding seizure. The results suggest that the epileptic tissue might not be as chronically impaired as previously thought.


Epileptic brain tissue is often considered physiologically dysfunctional, and the optimal treatment of many patients with uncontrollable seizures involves surgical removal of the epileptic tissue. However, it is unclear to what extent the epileptic tissue is capable of generating physiological responses to cognitive stimuli and how cognitive deficits ensuing surgical resections can be determined using state-of-the-art computational methods. To address these unknowns, we recruited six patients with nonlesional epilepsies and identified the epileptic focus in each patient with intracranial electrophysiological monitoring. We measured spontaneous epileptic activity in the form of high-frequency oscillations (HFOs), recorded stimulus-locked physiological responses in the form of physiological high-frequency broadband activity, and explored the interaction of the two as well as their behavioral correlates. Across all patients, we found abundant normal physiological responses to relevant cognitive stimuli in the epileptic sites. However, these physiological responses were more likely to be “seized” (delayed or missed) when spontaneous HFOs occurred about 850 to 1050 ms before, until about 150 to 250 ms after, the onset of relevant cognitive stimuli. Furthermore, spontaneous HFOs in medial temporal lobe affected the subjects’ memory performance. Our findings suggest that nonlesional epileptic sites are capable of generating normal physiological responses and highlight a compelling mechanism for cognitive deficits in these patients. The results also offer clinicians a quantitative tool to differentiate pathological and physiological high-frequency activities in epileptic sites and to indirectly assess their possible cognitive reserve function and approximate the risk of resective surgery.


Millions of patients with epilepsy are refractory to medications and may require surgical resection of the epileptic tissue given that a complete removal of the tissue has been associated with favorable outcomes for seizure control (1, 2). Surgical outcomes, however, are less impressive for nonlesional epilepsies (3), and there is indirect evidence that postoperative cognitive deficits ought to be considered as a concern in these patients (47).

To date, we know little about the deleterious effects of surgical resections of nonlesional epilepsies partly because of the heterogeneity of surgical procedures and partly because of the lack of systematic prospective studies probing regionally specific cortical functions in postoperative patients. What complicates the case of nonlesional epilepsies even more is the existence of conflicting evidence from clinical recordings referring to high-frequency oscillations (HFOs) as a biomarker for pathological epileptic tissue (8), whereas there is a large body of studies in the field of cognitive neuroscience referring to high-frequency broadband (HFB) activity or high gamma as a biomarker for normal physiological responses of the functional tissue (9). The distinction between pathological and physiological high-frequency activity has only recently been explored (10, 11), and tools to reliably distinguish the two are largely absent in today’s clinical practice. The absence of clear means by which clinicians can distinguish the two types of high-frequency activities may have practical consequences. For instance, several clinical studies have failed to show improved surgical outcomes when resections were aimed to remove sites with high-frequency activities (1214). One might argue that a possible explanation for the failure of these studies is rooted in the mixing of pathological HFOs with physiological HFBs.

What remains to be achieved in the current clinical practice of epilepsy treatment is a means by which clinicians can distinguish pathological from physiological high-frequency activity and estimate at the individual patient level the extent of reserve physiological function in the pathological epileptic tissue. This information can be used to infer the potential cognitive side effects of removing the epileptic tissue and, as such, guide clinical treatment decisions.

Toward this aim, we designed the current study to explore the presence or absence of stimulus-locked physiological responses in the epileptic tissue, to determine their relationship with ongoing pathological HFOs, and to measure the behavioral effects of interictal HFO activities during experimental cognitive conditions. We also devised a method by which regionally specific physiological high-frequency activities can be reliably differentiated from pathological epileptic HFOs at the individual patient level.


Data collection

We recorded spontaneous (interictal) epileptic activity during rest and induced electrophysiological responses during experimental conditions using direct intracranial electroencephalography (iEEG). Our study subjects consisted of two different groups of participants (fig. S1). Group 1 consisted of three patients (S1 to S3) with neocortical seizures that were confirmed to originate from higher visual association cortices. These patients had implanted subdural grids of electrodes. Group 2 consisted of three other patients (S4 to S6) with nonneocortical seizures that were confirmed to originate from the medial temporal lobe (MTL; hippocampus or amygdala). These patients were monitored with depth electrodes. The detailed description of demographics, seizure types, and electrode implantation is given in Table 1. These two groups of patients also performed different experimental tasks: The experimental condition for group 1 consisted of visual object recognition task, whereas group 2 participants performed a memory task. The goal of the study was not to provide a larger-scale populational study but to prove the consistency and potential generalizability of results across two entirely different groups of patients with different epilepsies, monitoring hardware, and experimental conditions.

Table 1 Subject demographical information.

RNS, responsive neurostimulation.

View this table:

Automatic mapping of HFO sites

Overall, 139 channel hours of iEEG data were used for the automatic detection of HFOs during resting state in all subjects. HFOs were identified using a previously validated method (see Materials and Methods) (10). The method was established upon the broadly accepted definition of spontaneous interictal HFO, namely, a transient oscillatory activity having a minimum of four oscillatory cycles that are distinct from background activities (15). The identified HFOs exhibited abrupt and transient power increase in the high band (recognized as an isolated light “blob” over 80 Hz in the time-frequency map), known as one of the spectral properties making HFOs distinct from spikes. Moreover, the detected HFOs presented repetitive waveform patterns that formed compact and inhomogeneous clusters in the original high-dimensional space of the raw time series, showing consistency with the stereotyped waveform patterns that characterize pathological HFOs that are specifically indicative of seizure onset zones (fig. S2).

We identified channels with mean HFO rate ≥0.2/min of recording (a minimum of two spontaneous HFOs were recorded) during rest for each subject and constructed spatial maps of HFO distribution (Fig. 1B and data file S1). A total of 3302 HFO events were found in 191 electrodes, accounting for 23% of the total recorded channels. Eighty-six percent of the electrodes with HFO were located inside the clinically defined epileptic zones (sites that were deemed to be epileptic by trained epileptologists visually inspecting the patient’s EEG during their admission).

Fig. 1 Spatial locations of pathological HFOs and functional HFBs.

(A) Data recorded from a subset of channels in two representative subjects, S1 (grid) and S4 (depth), showing early involvement of the HFO site in seizure onset. Epileptiform discharges are marked by red stars. (B) Electrode coverage in all subjects with the load of HFOs in each electrode. Results are derived after coregistration of individual preoperative magnetic resonance imaging and postoperative computed tomography images. The selected ROIs in each subject are circled in red; the contacts with HFB activation within the ROI are pointed at by white arrows.

Evaluation of physiological HFB responses in epileptic sites

To explore the presence of physiological responses in epileptic sites, we defined three inclusion criteria to define an epileptic region of interest (ROI). These criteria were (i) presence of pathological HFOs, (ii) presence of epileptiform interictal discharges, and (iii) involvement in ictal onset. Criterion 1 was determined automatically by our algorithm as described above. Criteria 2 and 3 were determined by the information (EEG notes prepared by clinical epilepsy teams reviewing the interictal and ictal EEGs) available in each patient’s online medical records (Fig. 1A). This led to the inclusion of 15 sites at the lateral occipital gyrus (S1) and the ventral temporal cortex (S2 and S3) and 23 sites at the amygdalohippocampal regions (S4 to S6). The selected ROIs represented brain regions that most clinicians would consider to be epileptic (Fig. 1B).

In the next step of our study, we wanted to determine whether the selected epileptic sites were capable of generating stimulus-locked physiological responses to relevant cognitive stimuli. In each ROI, we explored the presence of stimulus-locked responses in the HFB (70 to 180 Hz) range that are known to be reliable markers for physiological engagement of the brain site during cognitive and behavioral tasks (9). In group 1, with coverage over the visual regions, we measured HFB responses to visual stimuli (task 1), and in group 2, with electrode coverage in the MTL, we analyzed HFB responses during the encoding of memory items (task 2, see Behavioral Task Paradigm Supplementary Materials).

In the selected epileptic sites, we recorded 957 spontaneous HFO events. The average HFO frequency per channel in the selected sites was 1.9 ± 0.5 for group 1 and 4.1 ± 0.5 for group 2 (Fig. 1B and data file S1). The same sites generated a total of 1020 task-induced HFB epochs during the cognitive tasks (281 ± 29 for group 1 and 59 ± 0.6 for group 2; data file S1). We confirmed that the HFB responses from epileptic sites presented comparable activation patterns to those of the nonepileptic functional regions (fig. S3).

Temporal and spectral characteristics of HFOs and HFBs

We compared the spectral and temporal properties of task-induced HFB responses with spontaneous HFOs originating from the same epileptic tissue. Group-level time-frequency analysis demonstrated that the spontaneous epileptic HFOs involved power augmentation in a frequency range spanning from 79 ± 3.5 to 192 ± 21.4 Hz, which was similar to that of the HFBs (67 ± 3.8 to 204 ± 13.3 Hz; P = 0.87; Fig. 2E and data file S1). The spectral frequency showing the power augmentation was centered at 128 Hz for HFOs and 123 Hz for HFBs (Fig. 2, B and E, and fig. S4). However, the averaged power time course in the high band (>80 Hz) for HFO and HFB events revealed a clear difference in duration between these two types of activities (HFO = 77 ± 5.4 ms, HFB = 513 ± 33.6 ms, P = 0.002, η2 = 0.94; Fig. 2E), indicating that HFOs are transient events with a sharp power increase above 80 Hz, whereas HFB activities have sustained high-band power increase (Fig. 2C and fig. S4). After aligning the single spectrum of each event by its peak and computing the average, we measured the full width at half maximum (FWHM) corresponding to this peak, which was the range in frequency where the signal attenuated to half of its maximum power, for both HFO and HFB data. The spectral width of power increase for spontaneous HFO was substantially smaller than task-driven HFBs (HFO = 36 ± 2.6 Hz, HFB = 94 ± 11.9 Hz, P = 0.002, η2 = 0.87; Fig. 2E and fig. S4), implying that pathological HFOs are more narrow banded (Fig. 2, D and E, and fig. S4).

Fig. 2 Temporal and spectral profiles of HFO and HFB signals.

(A) Data plots for exemplar HFO and HFB in the same sites in two representative cases for groups 1 and 2 subjects (S1 and S4). For each sample event, raw data (top) and high-pass–filtered data above 80 Hz (bottom) are shown. For an HFO event, time = 0 indicates the time point corresponding to the peak amplitude; for an HFB event, time = 0 indicates the onset of stimulus onset. (B) Time-frequency maps averaged across all HFO and HFB events in S1 and S4. (C) Averaged power time course in the high-frequency band above 80 Hz for HFO and task-induced HFB in S1 and S4. Task-induced low-band deactivation is also presented. (D) Averaged signal spectra for HFO and HFB in S1 and S4. The spectral width is given by its FWHM (dashed vertical lines). (E) Averaged time-frequency maps, signal duration, and spectral width for HFO and HFB activities in all subjects. **P < 0.01.

We sought to answer how the spectral power changes as a function of frequency for both HFBs and HFOs and whether HFOs present similar broadband power increases as task-induced HFB activities. This was done by assessing the power activation through the change in magnitude and slope of signal power spectral density (PSD). In Fig. 3A, we give examples of PSD plots for single HFO and HFB signal in S1. The PSD analysis for HFO showed its peak at about 120 Hz, and the power increase was not constant over the entire HFB. By contrast, the PSD for task-induced HFB response showed a parallel shift in the magnitude compared with its baseline, indicating its broadband nature, which is in line with a series of past studies (16). This difference was consistently observed in all subjects (Fig. 3B). For each subject, we fit the PSD of HFB data and their baselines to the power law function within the frequency range of 40 to 200 Hz to avoid the spectral flattening in the lower band and the noise floor at higher frequencies. In addition, it has been suggested that a frequency at 40 to 50 Hz is a point where a reliable power increase could be measured across subjects (17). In all subjects, HFO events presented substantial changes in exponent x (Ps1 = 0.0005, Ps2 = 0.0004, Ps3 < 10−4, Ps4 = 0.0027, Ps5 = 0.005, Ps6 = 0.0002, η2 = 0.05 to 0.41). By contrast, no significant change in x was observed for HFB in all subjects (P > 0.1, η2 = 10−4 to 0.01), which was in concordance with earlier studies (16, 17), suggesting a parallel change in the HFB during behavioral tasks (Fig. 3C). The change in the slope/shape of PSD (Δx) for HFOs may reveal different dynamics within the neuronal population (16, 17), which was substantially different from HFBs (P = 0.004; Fig. 3D).

Fig. 3 Characterization of signal PSD.

(A) Signal PSD for a single HFO event and an HFB event in S1. Compared with its corresponding baseline (black dashed line), the sample HFO shows a change in the shape of PSD, with its peak localized at about 120 Hz; HFB shows an overall shift in the amplitude, but not the shape of PSD. (B) Average PSD for HFOs and HFBs in all subjects. Plots are shown in scaled windows for visualization purposes. (C) Histogram of exponent coefficient x derived from power-law model fitting for HFOs and HFBs. Results show a shift in the distribution of exponent values only for HFOs. Different window scaling was used for visualization purposes. (D) Comparison between the change in exponent coefficient x for HFO and HFB events in all subjects. **P < 0.01.

Differentiating HFOs and HFBs using the machine learning method

Next, we aimed to explore the features in the time and frequency domains that could differentiate HFOs from HFBs. Three classic features of peak amplitude (maximum value of the signal envelope), spectral centroid, and high-band power were used. In addition, on the basis of our inspection of the data, we also computed four other features optimized for the discrimination between HFO and HFB. These included sub-band power ratio, spectral entropy, short-time line length, and relative peak amplitude. Sub-band power ratio and spectral entropy were computed to assess the power distribution of signals in the high and full spectrum. Short-time line length and relative peak amplitude were used as measures of signal complexity: A large line length value would have indicated greater variation within the time series, whereas a large relative peak amplitude (>80 Hz) would have suggested an abrupt jump in the signal relative to its local background. A detailed description of each feature is reported in Materials and Methods.

Using the first approach considering the classical features, HFO signals exhibited significant difference (P < 0.05) compared with HFB in two of the three features but without cross-subject reliability. By contrast, consistent trends could be observed across all subjects for the four optimized features (Fig. 4A and fig. S5). Overall, HFOs comprised a smaller sub-band (high-to-low band) power ratio compared with HFBs (P < 10−4), suggesting stronger low-frequency components in the pathological HFOs or weaker low-frequency power in the HFBs. In addition, HFBs had higher entropy (P < 10−4), suggesting a more uniform distribution of energy above 80 Hz compared with HFO events, in which the energy was less distributed. Moreover, as expected, HFBs showed two additional distinct features, namely, larger line length (P < 10−4) and smaller relative peak amplitude values (P < 10−4). These could be related to longer temporal span and greater signal complexity of HFBs compared with HFOs.

Fig. 4 Feature distribution of HFBs and HFOs.

(A) Box plots showing the normalized values for three classical features and four optimized features in S1 and S4 (the first two rows) and in all subjects (the third row). (B) Three-dimensional scatterplots showing the distribution of the three most distinctive features (sub-band power ratio, line length, and relative peak amplitude) in all subjects. ****P < 10−4, ***P < 10−3, and **P < 0.01.

Importance estimation and commonality analysis of the features showed that all metrics captured some unique properties of neural activity relevant for HFO and HFB discrimination, among which relative peak amplitude contributed most as a unique feature accounting for 29% of the total variations explained by predictors. Sub-band power ratio, entropy, line length, and relative peak amplitude jointly explained 18% of the total variance, with 0.9% of the variance being common to all six features (Table 2). Figure 4B demonstrates the distribution of the three features with larger total effect coefficient values, as presented in Table 2. Consistently, in all three subjects, the feature distributions showed clear separation with minor overlap. We performed supervised support vector machine (SVM) classification with listed features and calculated the receiver operating characteristic (ROC) curves for the prediction. The classifier was able to separate HFOs from HFB response with the area under the ROC curve (AUC) value reaching 0.96 (fig. S6) with leave-one-subject-out cross-validation (see Materials and Methods).

Table 2 Unique and common effects for each independent variable.

View this table:

“Seizing” effect of HFOs on HFB responses

Once we established that the epileptic tissue was capable of generating normal physiological responses to relevant cognitive stimuli, we asked whether the cognitive function of a cortical tissue is compromised by spontaneous pathological activity and, if so, for how long? For this, we first visually reviewed the iEEG data in all subjects to confirm that no clinical or subclinical seizures occurred during behavioral tasks. We then investigated the temporal relationship between task-induced HFBs and HFOs during the two experimental tasks. We identified spontaneous (not stimulus-locked) HFOs during the experimental task. A total of 464 HFOs were captured from the sites in which HFB responses had been recorded during behavioral tasks (average HFO rate per channel: group 1 = 2.5, group 2 = 3.3).

We tested whether the presence of HFOs during a cognitive task might be undesirably affected by the presence of stimulus-locked HFBs. For this, we evaluated the probability of an HFO being detected during a specific period of time (before or after a stimulus) and compared the value with randomized surrogates (see Materials and Methods). Results showed that, in all subjects, the detected HFOs were present at near-random time points throughout the recording (P = 0.23 to 0.66), suggesting that the HFO detection during experimental tasks was not influenced by the presence of HFB in the background.

Subsequently, we obtained the concatenated power time series for stimulus-locked HFBs and spontaneous HFOs during cognitive experiments (Fig. 5, A to C, and fig. S7) and computed the “discordance value” as well as the Pearson’s correlation coefficient between the HFB and HFO traces using 1-min windows. The discordance values between HFOs and HFBs, defined as the ratio of the number of HFO samples occurring outside the task-induced HFBs to the total number of HFO samples, were calculated for each 1-min window bin. In all subjects, the discordance values are significantly larger compared with the chance value (group 1: 0.81 ± 0.04 versus 0.46 ± 0.03; group 2: 0.92 ± 0.01 versus 0.78 ± 0.02; P = 0.03, η2 = 0.37), indicating that HFOs are spontaneous neural activities that do not coincide in time with stimulus-locked HFB responses (data file S1).

Fig. 5 Temporal correlations between HFOs and HFBs during cognitive tasks.

(A) An illustration of a single-trial HFB response in S1. The power increase was evaluated by calculating the area (green) defined by the power time course of HFB and its corresponding local threshold derived from the prestimulus baseline (red dashed line). (B) Concatenated power time series (100 s) showing spontaneous HFOs and stimulus-locked HFBs during cognitive task in S1. Onsets of the visual stimuli are shown in gray vertical lines. (C) Ten-second scaled window showing the same data as in (B). Five behavioral trials are included. Note the absence of HFB activation during the second and the third trials where two incidents of HFOs had been detected. (D) Relationship between the magnitude of HFO versus HFB per unit of time. Each data point represents the sum of HFB or HFO power within each 1-min window bin. The results are normalized between 0 and 1 and ranked according to the HFO power (from low to high). Data points and error bars represent means and SEM, respectively (darker shades: mean values; lighter shades: individual data). *P < 0.05.

HFO and HFB activity in each 1-min window bin was measured from the concatenated traces; the bins were then ranked according to the HFO power (from low to high), and the HFB activity in each corresponding bin was calculated (Fig. 5D). The Pearson’s correlation was derived from the values from all time bins computed over the entire task (for instance, 10 values in each patient who performed the visual task). We observed that HFO and HFB power showed negative correlation in five of six subjects (group 1: r¯ = 0.46, PS1 = 0.04, PS2 = 0.04, PS3 = 0.09; group 2: r¯ = 0.81, PS4 = 0.05, PS5 = 0.02, PS6 = 0.03).

We noticed that, in each subject, there were experimental trials in which no HFB activation was observed—even though there were no spontaneous HFOs present at the time of stimulus onset or shortly afterward. This could mean that ongoing HFOs before the onset of stimulus could have directly led to the absence of stimulus-induced HFBs. To test this hypothesis, we examined the co-occurrence of HFOs and HFBs within a wide temporal window. We first established that the missed HFB responses were not due to the characteristics of cognitive stimuli being presented or methodological limitations in detecting HFBs mixed with ongoing HFOs (see Materials and Methods). We examined the occurrence of HFOs within the full-length window of –400 ms to +1000 ms after the stimulus onset and obtained the number of trials with HFOs, but with no HFBs. The number was divided by the total number of trials wherein HFOs appeared. This ratio was labeled as the discordance value at the single-trial level. In all subjects, these discordance values were substantially larger than the values derived from shuffled data, indicating that the probability of a trial missing HFB was significantly higher than chance when HFO co-occurred within the −400- to +1000-ms window around the trial onset (0.62 versus 0.54; P = 0.01, η2 = 0.35); same results (with stronger difference) were also observed when HFOs occurred in the −400- to 0-ms prestimulus period (0.85 versus 0.55; P = 0.01, η2 = 0.85). These results suggested that the presence of HFB responses correlated inversely with the presence of HFOs before or during stimulus presentation (data file S1).

To clarify whether the likelihood of a trial with missed HFB response depends on the time of HFO occurrence, we computed the temporal discordance value using different window sizes from −2 to +2 s relative to the onset of stimulus, with a step size of 50 ms (Fig. 6A). We noted a strong association between the proportion of missed HFB events and the timing of HFO occurrence for task 1: If HFOs appeared shortly before (and slightly after) the onset of visual stimulus, then HFBs were more likely to be missed; the likelihood is strongly correlated with the relative time of HFO occurrence to the stimulus onset (r¯ = 0.82, P < 10−4). This effect arose as early as 1050 ms before the stimulus (P < 0.05) and increased most drastically at −200 ms before the onset of visual stimuli (P < 10−4), reaching its peak of 77% at +50 ms (50 ms after the onset) and immediately dropping to chance likelihood after +100 ms (around the time that HFBs are expected to begin to occur). That is, the occurrence of HFO 100 ms after the onset of cognitive stimulus does not cause a missed cognitive response, whereas that of HFOs shortly before the onset of cognitive stimulus will have the highest likelihood to lead to a missed HFB response. These robust results were found during a visual task and in a neocortical epileptic site within the higher visual association cortices.

Fig. 6 Relationship between HFOs and HFBs during cognitive tasks.

(A) Likelihood of missed physiological responses (or absence of stimulus-locked HFBs) compared with the time of HFO occurrence. Cortical physiological HFB responses are more likely to be missed if there is an ongoing spontaneous HFO activity within −200 to +100 ms around the onset of stimuli (time 0) for the visual task and −200 to +200 ms for the memory task (P < 10−4). Note that HFB responses are expected to occur ~100 to 200 ms after the onset of visual or memory stimuli (yellow arrows). Solid line: averaged results; dashed line: data in each individual subject. (B) Pearson’s correlation between the time of onset of HFB activation and HFO occurrence relative to the stimulus onset. Results are derived from trials with both HFO and HFB responses. (C) HFB responses in three time windows grouped by the relative time of HFO occurrence (−1500 to −1000 ms, −1000 to −500 ms, −500 to 0 ms to stimulus onset) in S1 and S4. HFBs are weakened as HFOs are approaching the stimulus onset. (D) Relationship between HFO synchronization/power and physiological brain activity. Disruptive HFOs have more electrodes with synchronous activity, as well as higher power, compared with nondisruptive HFOs. Data points and error bars represent means and SEM, respectively. ****P < 10−4, **P < 0.01, and *P < 0.05.

To examine the generalizability of our findings to other regions of the brain, we examined the same discordance relationship in patients with MTL coverage performing a memory encoding and recognition memory task. The results in task 2 were strongly affirmative (r¯ = 0.92, P < 10−4), with consistent result being observed in every subject. A noteworthy difference was that the maximum likelihood of missed HFBs was seen around +150 ms after the onset of the memory trial compared with +50 ms after the onset of visual stimulus observed in task 1. This is expected because responses in the visual cortices during processing of visual items are earlier than the memory-related responses in the MTL structures. Moreover, we found that in trials when the occurrence of HFOs did not lead to the complete extinction of HFB responses, the response onset latencies for HFB events correlated significantly with the timing of HFO occurrence (task 1: r¯ = 0.39, P = 0.04; task 2: r¯ = 0.73, P = 0.02). Our results suggest that HFO occurrence before a stimulus leads to delayed and weakened HFB response when and if the HFB response is not missed (Fig. 6, B and C).

Of note, HFOs preceding task trials during experimental settings were not different from spontaneous HFOs during resting state in signal peak amplitude, duration, mean frequency, or any of the most distinctive features (P > 0.5; fig. S8). In addition, the duration of prestimulus HFOs (81 ± 2.2 ms) presented no correlation to the relative time of its occurrence in all subjects (rs1 = −0.14, Ps1 = 0.34, rs2 = −0.04, Ps2 = 0.64, rs3 = −0.14, Ps3 = 0.39, rs4 = 0.24, Ps4 = 0.12, rs1 = 0.07, Ps5 = 0.40, rs5 = −0.06, Ps6 = 0.62), indicating that HFOs occurring closer or further away from to the time of stimulus onset were similar in their duration. These results imply that the absence of functional HFB response in a trial is associated with the presence of pathological HFOs during the stimulus presentation and that when HFOs occur right before the stimulus, there is higher possibility that the cortical structure may encounter failure or delay in eliciting a valid HFB response.

Considering the fact that pathological HFOs can be generated from multiple clusters of neurons that fire simultaneously within the same epileptogenic network, we investigated the disruptive effect of HFO synchronization across multiple electrodes on the physiological brain activity. For each prestimulus HFO event, we examined data from epileptic sites outside the ROI to look for synchronous HFO activity across multiple channels and correlated the results with the presence or absence of physiological HFB responses. In two subjects (S2 and S3), we could not identify synchronized HFO activity across multiple electrodes because the electrode coverage over the epileptic tissue was relatively sparse in both of these subjects (Fig. 1B). In four remaining subjects with larger electrode coverage, we could identify synchronized HFO activity across multiple electrodes. In all these subjects, HFOs disrupting behavior had a larger number of electrodes with synchronized activity than HFOs without a disruptive effect (2.9 versus 1.9; P = 0.04; Fig. 6D and fig. S9A). In addition, we looked for a possible relationship between the power of HFO activity in the ROIs and the likelihood of missing physiological responses and observed that in all six subjects, HFOs that were more disruptive had higher z score power (3.2 versus 1.6; P < 10−4; Fig. 6D and fig. S9B).

Behavioral relevance of spontaneous HFOs

We tested the behavioral relevance of pathological HFOs in the three subjects who performed the memory encoding and recognition memory task. Of note, the visual localizer task had no behavioral component except pressing a button when rare trials with red hashtag signs appeared. These trials were too rare for any analysis of behavioral relationship with the occurrence of HFOs. However, task 2 was appropriate for measuring a possible relationship between the occurrence of spontaneous HFOs and task performance. For this reason, we conducted the behavioral correlation analysis in group 2 subjects only. In task 2, subjects were first engaged in incidental encoding task where they saw single images or words and responded if a given item was an indoor or outdoor object. During this condition, we expected that they could incidentally encode to their memory many of the presented items. After a distractor period, they were then presented with an old or new stimulus and were asked to respond if the object had been presented (old) or not (new) during the incidental encoding phase of the experiment. Once they responded old or new, their confidence was tested.

Our analysis included measures of behavioral performance during the encoding and recall conditions: (i) goal-directed task performance during incidental encoding task (where subjects viewed images or words and decided if the object was an indoor or outdoor item) and (ii) recognition memory condition [where subjects viewed an item and decided if the item was presented during an encoding task (old) or not (new)]. Because our analysis did not show any modality-specific effect (regarding recalling images versus words), we combined the trials for these two types of stimuli, resulting in 60 stimulus trials per condition. We first identified trials with the presence of HFO in the 1-s prestimulus window and then compared the results to trials without HFO preceding the stimulus onset by conducting a paired Friedman test (results for both conditions are combined to increase statistical power). Our analysis revealed that the occurrence of spontaneous HFOs affected the subject’s task performance and memory recall (Fig. 7). Results for each condition were as follows: Subjects showed significantly lower response accuracy during both conditions I and II (condition I: 85% versus 91%; condition II: 54% versus 64%; P = 0.03, η2 = 0.40) for trials co-occurring with the presence of HFOs compared with trials without HFOs. They also showed significantly longer reaction time (condition I: 0.57 s versus 0.48 s; condition II: 1.37 s versus 0.72 s; P = 0.01, η2 = 0.44). Furthermore, they exhibited a slight but significant decrease in recall confidence for trials that were presented with HFOs compared with trials without HFOs before the stimulus onset (70 versus 83; P = 0.01, η2 = 0.01). As seen in Table 3, these effects were seen in all three subjects, and the individual performance seemed to correlate with the degree of baseline pathology (frequency of spontaneous HFO occurrence).

Fig. 7 Behavioral effects of pathological HFO.

(A) Normalized group-level results showing the behavioral effects of prestimulus pathological HFOs during task 2. Square and triangle symbols with gray dash lines depict results for conditions I and II in each individual subject, respectively. Consistently, in all subjects, the reaction accuracy and confidence score decreased, whereas the reaction time increased significantly (P < 0.05, nonparametric Friedman test). (B) Subject with the highest HFO load (S2) had the lowest behavioral performance during memory task, whereas subject with a relatively low baseline HFO rate (S1) had the best performance. These results indicate that subjects’ performance during memory task tends to negatively correlate with the load of HFO during the task and the frequency of HFO during resting state.

Table 3 Behavioral effect of pathological HFOs during task 2.

View this table:

Last, to answer whether the disruptive effect was caused by other epileptic activities, we first confirmed that there was no subclinical seizure during the behavioral tasks; to answer whether the effect was introduced solely by the slow component of interictal activities, we replicated the “refractory window” analysis and examined the behavioral relevance of prestimulus slow spikes (without an isolated high-frequency component above 80 Hz) in three subjects who accomplished the memory task. These spikes were identified using our spike/HFO detection method and were visually validated by reviewing the raw data together with the time-frequency maps. In addition, we looked at the behavioral results in trials with prestimulus spikes during memory task, condition II (old/new recognition). As a result, we did not observe any correlation between the occurrence of slow spike and subjects’ functional response or behavioral performance (fig. S10). The response accuracy and reaction time in trials with spikes were comparable to those without spontaneous HFO [accuracy: S4, 79% versus 81%; S5, 39% versus 34%; S6, 80% versus 77%; P > 0.5; reaction time (s): S4, 0.58 versus 0.64; S5, 1.03 versus 1.10; S6, 1.08 versus 0.83; P > 0.5; see Table 4].

Table 4 Behavioral effect of slow spikes during task2, condition II (old/new recognition).

View this table:


Our findings can be summarized as follows: First, we confirmed that nonlesional epileptic tissue still produces functional responses to cognitive stimuli. Second, we found that intrinsic HFOs generated by an epileptic tissue are different from HFB activities induced by cognitive stimuli and offered distinguishing features of the two signals that can be used to differentiate the two automatically. Moreover, our findings revealed a functional inverse relationship between pathological and physiological high-frequency activities within the neocortical and nonneocortical epileptic tissue: Cognitively relevant stimuli fail to activate the epileptic tissue if they arrive at the brain region within the time window of HFO occurrence. That is, HFO activity seems to have a lasting negative effect on the physiological response profile of the brain tissue for about 1100 ms. Even if the HFO activity has disappeared for hundreds of milliseconds, the tissue seems to be in a “cognitive refractory state.” Furthermore, we learned from the behavioral results in the MTL subjects that if the cognitively relevant stimulus fails to activate the epileptic tissue (in this case, the hippocampus), the subject’s recognition memory performance will be impaired.

We revealed a clear distinction in the spatiotemporal profiles of pathological and cognitively induced physiological high-frequency signals using a quantitative comparison of temporal and spectral properties of epileptic HFOs and functional HFBs. Several research groups have proposed computational methods toward the quantitative identification of HFOs in patients with epilepsy (1820), but the clinical use of the existing approaches is hindered by the presence of physiological high-frequency activities in functional brain regions. Recent studies have documented that pathological HFOs have larger peak amplitudes, short durations, and relatively “clean” waveform patterns that are different from physiological high-frequency activities recorded from nonepileptic structures of motor, visual, and language areas (10, 2123). Our present study lends support to the past studies by documenting that pathological HFOs are spontaneous burst-like events with short durations and relatively homogenous waveforms. We show that pathological HFOs are oscillatory neuronal activities exhibiting increased high-band power within a limited bandwidth and carry large spectral variability. In this regard, it is possible the “physiological HFOs” described in some previous studies (2426) could be related to, or in fact a fragment of, functional HFBs.

Not only do the two kinds of high-frequency activities have different signal properties and features, but they also exhibit clear temporal discordance—the presence of one anticorrelates with the presence of the other. Taking these two findings together, one can conclude that a pathological nonlesional structure responds normally to a relevant cognitive stimulus but that its normal physiological response to these cognitive stimuli is impaired when epileptic HFOs are present. This is important because it redefines nonlesional epileptogenic structures from an entirely sick tissue to a transiently impaired tissue. That is, the epileptic tissue is only deficient, from a cognitive perspective, during a short period of time (about 1 s) when transient epileptic HFO activity is occurring.

We introduced a method that proposes reliable discrimination between HFOs and HFBs from the same brain site using a combination of computational features rather than sole visual inspection—which can be prone to human error. The distribution of our optimized features provided clear separation and showed consistency in all subjects despite large interindividual variability. Our method of differentiating the two types of high-frequency activities may facilitate automatic (and less time-consuming) identification of HFOs indicative of epileptic tissue and, as such, facilitate in the mapping of seizure onset zones. Furthermore, the proposed method will enable one to calculate the “load” of pathological activities in a given period of time and, as such, provide indirect evidence for the reserve functional capacity of an epileptic region or the lack thereof.

In this study, we determined that the occurrence of spontaneous interictal HFOs correlates with poor recognition memory performance. We demonstrated a clear relationship between the presence of HFOs, lack of stimulus-locked HFBs, and patient’s recognition and confidence impairment. By showing that HFOs preceding the presentation of a cognitive stimulus are associated with either absence or delays in physiological responses, we hypothesize that there may exist a causal relationship between the pathological HFOs and cognitive deficits. Hence, treatment paradigms that effectively suppress interictal epileptic HFOs (without altering the baseline physiological function of the epileptic tissue) may enhance the cognitive reserve functions associated with the structure of interest and thus improve the patients’ quality of life. Future prospective studies are needed to determine the functional and causal effects of modulating spontaneous HFOs (for example, with electrical neuromodulation) and measuring their behavioral impact.

In line with our view, there is evidence that the neuromodulation of lesional and nonlesional hippocampal epilepsies provides subtle improvement in memory functions (27, 28), and past studies have reported positive cognitive effects due to disruption of seizures by the responsive neurostimulation system (29, 30). Moreover, observations since the early days of clinical neurophysiology have confirmed that epileptic spikes and seizures cause transitory cognitive impairment whose severity correlates with the frequency of interictal epileptiform discharges (3143). On the basis of these observations, neuropsychological tests have been designed to reveal impairments in specific domains of cognition, which, in turn, are used as indirect information about the possible anatomical origin of seizures in a given individual (44, 45). Despite the earlier studies trying to build a link between gross epileptic abnormalities and behavioral performance, our present work demonstrated a clear relationship between the HFO burden and the electrophysiological response in a pathological brain tissue across neocortical and nonneocortical regions. We defined a temporal window with millisecond resolution for the detrimental effects of pathological activity on cognition, which also provides a possible alternative mechanistic explanation for past clinical observations, namely, that disrupting spontaneous HFOs might be crucially important for enhancing the cognitive reserve function of a given epileptic tissue.

In the current study, we provide important information about the link between pathological epileptic activity, physiological response in a brain region, and subject’s performance in a behavioral task. Our findings will have implications at several levels: At the conceptual level, it changes our view of epilepsy and epileptic brains and suggests that (at least nonlesional) epileptic tissue ought not to be regarded as a functionless sick tissue. The epileptic tissue only “seizes” to function transiently when there are spontaneous HFOs—and also as previous literature has suggested during spikes and seizures. At the scientific level, our findings are important in at least two domains: First, they highlight the signal properties that differentiate normal physiological responses in HFB from those in HFOs. Second, our results suggest that recordings from epileptic tissue (outside the windows of HFOs) will provide close to normal physiological healthy data, and as long as the confounding effects of HFOs are accounted for, intracranial electrophysiological findings in epileptic brains may well be regarded as proxies to findings in normal human brains. Our findings, however, also suggest that intracranial electrophysiology scientists should be careful about their findings if their experimental trials are administered during the cognitive refractory state. Intracranial EEG researchers currently do not differentiate or exclude trials with concurrent HFOs—partly because automatic detection of HFOs has not been widely available. Our methods now offer them a tool to do so. At the clinical level, our findings have several implications: First, they highlight the potential deleterious effects of surgical resection of nonlesional epileptic tissue without careful presurgical assessment of the tissue’s cognitive reserve function. Second, we provide a method by which the cognitive reserve function of an epileptic structure can be approximated by the load of its spontaneous pathological HFOs. Third, our findings raise the possibility of achieving better cognitive outcome by silencing spontaneous HFOs through optimizing medical therapies or using electrical neuromodulatory therapies (such as responsive neurostimulation). In patients with nonlesional epilepsies and with evidence for substantial physiological responses in the epileptic tissue, optimal treatment strategies ought to aim for the minimization of epileptic activities while preserving the integrity of the cortical tissue.

Although our results addressed the electrophysiological and behavioral effects of transient epileptic HFOs on patients’ cognitive functions, the power of the study was limited by the relatively small number of subjects. The study’s design put substantial restrictions in patient inclusion: Only patients in whom electrodes were implanted in the same functional locations and who consented to participate in research activities and perform relevant cognitive and behavioral tasks were recruited. Thus, future studies are needed to quantify the precise relationship between pathological HFOs and task-induced functional HFB responses and establish clear causal evidence that silencing HFOs will have beneficial cognitive effects at the individual patient level.


Study design

The primary objective of our current study was to address the question of whether nonlesional epileptic tissue is capable of generating normal physiological responses to relevant cognitive stimuli and how pathological HFOs and induced HFB signals differ from each other and how one affects the other. All hypotheses were related to this objective. We included patients with nonlesional neocortical and MTL epilepsy with near-identical electrode coverage who performed the same behavioral tasks during the EMU (epilepsy monitoring unit) monitoring. Only subjects who met the study inclusion criteria were included. No data were excluded. Subjects were grouped on the basis of their disease types, electrode coverage, and the corresponding behavioral tasks that they participated in. The investigators were blinded to group allocation during data collection. Subsequent data analysis was performed using automatic and semiautomatic techniques of detection and feature extraction. Same data analysis approaches were applied to all subjects’ data.

Temporal correlation analysis

We separately performed HFO detection using data recorded during cognitive tasks to assess the temporal correlation between HFOs and task-induced HFBs. Being cautious that the detectability of HFOs might be hindered by the HFB activity at the background, we calculated the number of HFOs detected in the −400- to 0-ms prestimulus window (400 ms before the presence of cognitive stimuli) as well as the 0- to 1000-ms poststimulus window (the time period starting from the onset to 1 s after). The ratio of these two numbers was compared with the chance value obtained from 10,000 surrogate traces where the time stamps of HFOs were replaced by randomized series.

For each HFO and HFB event, the power time course of the signal was computed and normalized to a threshold derived from the local baseline averaged across all electrodes within a same group. This approach was taken to ascertain that the detection of HFB response was not adversely influenced by the presence of HFO signals in the same site. We then concatenated the epochs according to their time of occurrence. This yielded the digitized version of iEEG data reflecting the magnitude and timing of HFO activity and stimulus-locked HFB activation.

Next, we calculated the discordance value between HFO and HFB, defined as the ratio of HFO samples occurring outside the task-induced HFB response to the total number of HFO samples, and compared the result with chance estimated from 10,000 surrogate traces. The surrogate traces were generated by shuffling the onset of all behavioral trials. The P value of the comparison between real and surrogate traces describes the probability of HFO and HFB events staggered by chance. In addition, we evaluated the temporal relationship between HFOs and HFBs by calculating the Pearson’s correlation between the concatenated traces in 1-min window bins.

To assess the relationship between HFB and HFO, we measured the discordance value at the single-trial level. We counted the number of trials where spontaneous HFOs appeared within the −400- to +1000-ms window around onset of stimulus, but the HFB activation was absent. Similarly, the number of trials where HFOs occurred only during the prestimulus period (400 ms before the stimulus onset) was also identified. These numbers were then divided by the total number of trials encountered with HFOs within the corresponding time window. To clarify whether the likelihood of a trial with missed HFB is correlated with the time of HFO occurrence, we computed the average temporal discordance value using varying window sizes from −2 to +2 s relative to the onset of stimulus, with a step size of 50 ms. Results were compared with 10,000 permutations.

Last, to evaluate the correlation between synchronous HFO activity and its disruptive effect on the physiological response, we identified synchronous HFOs generated from other epileptic sites out of the ROI (by identifying HFOs that shared strong correlation in signal waveforms that occurred less than 50 ms apart from each other). We grouped HFOs according to their time of occurrence (from 1200 ms before to 200 ms after the onset of cognitive stimulus, in 200-ms bins) and labeled the HFOs in each bin with two values: (i) the presence or absence of HFB response (in another word, whether the HFO is disruptive) and (ii) the number of electrode sites with documented synchronized HFO activity. We then calculated the average number of electrodes with synchronous HFO in each bin and performed a pairwise comparison between disruptive versus nondisruptive HFOs.

Statistical analysis

Results and graphs are shown as means ± SEM. Statistical analysis was performed using MATLAB built-in functions. Significant changes in HFO and HFB power were calculated and validated using permutation test with false discovery rate (FDR) correction. Nonparametric Friedman test or two-sided Kruskal-Wallis test was used for group comparison of the high-band power change between HFO and HFB across subjects (46); Friedman test was used for the comparison between signal spectra and its corresponding baseline in each subject. The Kruskal-Wallis test was used for the comparison of features. In the analysis of temporal correlation, we computed the chance value using the permutation method in each subject, and then the Friedman test was used to compare the results across subjects. A randomization test (10,000 permutations, FDR correction) was used for the “seizing effect” window analysis. Last, the behavioral performance was compared using the Friedman test. Differences were considered significant for *P < 0.05. Effect sizes were computed using η2 in the case of group comparison across subjects (the MES toolbox; (47).


Materials and Methods

Fig. S1. Study design.

Fig. S2. Pathological HFOs present isolated high-frequency component and generate compact subclusters in the time domain.

Fig. S3. Comparison between HFB activations in epileptic and in nonepileptic sites.

Fig. S4. Temporal and spectral profiles of HFO and HFB signals in all subjects.

Fig. S5. Feature distribution of HFBs and HFOs in all subjects.

Fig. S6. ROC curves for the SVM classifier.

Fig. S7. Concatenated power time series showing spontaneous HFOs and stimulus-locked HFBs during cognitive task.

Fig. S8. Signal properties of HFOs during behavioral task.

Fig. S9. Comparison between disruptive and nondisruptive HFOs in all subjects.

Fig. S10. Electrophysiological effect of slow spikes during task 2.

Data file S1. Raw data.

References (4852)


Acknowledgments: We gratefully acknowledge all patients involved in the study and thank C. Sava-Segal and other members of the Laboratory of Behavioral and Cognitive Neuroscience for their help during data collection, task execution, and preparation of the earlier drafts of the manuscript. Funding: The study is supported by the Sence Foundation, the Gwen and Gordon Bell family, and the Armatis family. Author contributions: Both authors participated in the data collection, analysis, and interpretation, as well as preparation of the manuscript. Competing interests: The authors declare that they have no competing interests. Data and materials availability: All data associated with this study are present in the paper or the Supplementary Materials.

Stay Connected to Science Translational Medicine

Navigate This Article