Practical measurements distinguishing physiological and pathological stereoelectroencephalography channels based on high‐frequency oscillations in the human brain

Abstract

Objective

The present study aimed to identify various distinguishing features for use in the accurate classification of stereoelectroencephalography (SEEG) channels based on high‐frequency oscillations (HFOs) inside and outside the epileptogenic zone (EZ).

Methods

HFOs were detected in patients with focal epilepsy who underwent SEEG. Subsequently, HFOs within the seizure‐onset and early spread zones were defined as pathological HFOs, whereas others were defined as physiological. Three features of HFOs were identified at the channel level, namely, morphological repetition, rhythmicity, and phase–amplitude coupling (PAC). A machine‐learning (ML) classifier was then built to distinguish two HFO types at the channel level by application of the above‐mentioned features, and the contributions were quantified. Further verification of the characteristics and classifier performance was performed in relation to various conscious states, imaging results, EZ location, and surgical outcomes.

Results

Thirty‐five patients were included in this study, from whom 166  104 pathological HFOs in 255 channels and 53 374 physiological HFOs in 282 channels were entered into the analysis pipeline. The results revealed that the morphological repetitions of pathological HFOs were markedly higher than those of the physiological HFOs; this was also observed for rhythmicity and PAC. The classifier exhibited high accuracy in differentiating between the two forms of HFOs, as indicated by an area under the curve (AUC) of 0.89. Both PAC and rhythmicity contributed significantly to this distinction. The subgroup analyses supported these findings.

Significance

The suggested HFO features can accurately distinguish between pathological and physiological channels substantially improving its usefulness in clinical localization.

Plain Language Summary

In this study, we computed three quantitative features associated with HFOs in each SEEG channel and then constructed a machine learning‐based classifier for the classification of pathological and physiological channels. The classifier performed well in distinguishing the two channel types under different levels of consciousness as well as in terms of imaging results, EZ location, and patient surgical outcomes.

Key points.

• Novel quantitative features, including morphological repetition, rhythmicity, and phase–amplitude coupling, were useful in separating pathological and physiological high‐frequency oscillations (HFOs).

• A machine‐learning model based on these features could distinguish between the two HFO types on channel level with high accuracy.

• The results were rigorously verified and found to be consistent across different states of consciousness, epileptogenic zone locations, imaging findings, and surgical outcomes.

1. INTRODUCTION

Epilepsy is a chronic neurological disease characterized by repeated seizure attacks that can seriously affect brain development and function. Approximately 30% of patients with epilepsy are drug‐resistant ¹ and are potential candidates for further preoperative evaluation and surgical intervention. ² Precise identification of the epileptogenic zone (EZ) is essential for planning surgical strategies. ³ The use of stereoelectroencephalography (SEEG), characterized by its excellent temporal and spatial resolution, is valuable in locating EZs. ⁴ In clinical practice, interictal high‐frequency oscillations (HFOs) from SEEG recordings have emerged as a potentially useful diagnostic for EZ localization. ⁵ , ⁶

An HFO is a transitory paroxysmal (~8–60 ms) electrophysiological activity that was first detected using microelectrodes in the human hippocampal region by Bragin et al. who observed synchronous fast firing within coupled neurons resulting in intense action‐potential firing, which is one of the processes involved in HFO formation. ⁷ In addition, other hypotheses that are gaining popularity suggest that inhibitory postsynaptic potentials and pyramidal action potentials could be sources of ripple processes. On the other hand, an out‐of‐phase fire may cause a quick ripple to originate from a collection of hyperexcitable but disconnected pyramidal cells. ⁸ The latest definition of HFOs suggests the presence of at least four consecutive cycles of oscillatory activity that can stand out from the background after 80 Hz high‐pass filtering. ⁹

Furthermore, epileptologists are paying significantly more attention to HFOs because of their strong correlation with EZ and predictive power over seizure outcomes. ⁵ , ⁶ , ¹⁰ To reduce the reviewer's burden, multiple automated detectors have been utilized. ¹¹ , ¹² In a previous study, a two‐stage HFO analysis pipeline was developed, involving initial detection and a convolutional neural network (CNN)‐based classification system following the framework proposed by Maeike Zijlmans et al. through which artifact‐free HFOs could be reliably extracted. ¹³

HFOs can be categorized into ripple (80–250 Hz) and fast ripple (250–500 Hz) HFOs according to their frequency range. It has been established that physiological HFOs are spontaneously generated in the 70–200 Hz range during physiological conditions. This can also be detected in various brain regions, especially in the visual and motor‐related cortexes and medial temporal lobe structures, ¹⁴ , ¹⁵ and can be considered a sign of focal neuronal activation. ¹⁶ The two HFO categories share frequency bands and thus present difficulties in distinguishing their relative amplitude, duration, and other characteristics. The comprehensive explanation of a pathological HFO remains unknown. ¹⁷ These nonepileptiform high‐frequency events could inevitably affect HFO‐based EZ localization.

Consequently, methods for the accurate classification of normal and pathological HFOs have been the subject of many investigations. For example, Su Liu and colleagues demonstrated that pathological HFOs show consistent morphological patterns, ¹⁸ while, in contrast, Annika Minthe and colleagues proposed an association between HFOs originating from a quiet background and epilepsy. ¹⁹ Various studies have also shown that pathological HFOs often coincide at the event level with spikes, which may be one of the distinguishing features. ²⁰ Nevertheless, there remains a deficiency of specific measurements that can effectively differentiate the occurrences of different HFOs. ⁷ , ²¹

It is usually recognized that the EZ location is based on the classification of each SEEG channel (pathological or physiological channel). Therefore, the present study designed four channel‐level quantitative metrics for HFOs using three distinct categories, namely, morphological repetition, rhythmicity, and phase–amplitude coupling (PAC). After selecting the manually labeled pathological and physiological channels as the gold standard, machine‐learning methods were developed using the proposed measurements for accurate differentiation between the two channels using these HFO features. The efficiency of the classifier was also examined in relation to different states of consciousness, EZ locations, imaging outcomes, and surgical outcomes to enhance the understanding of the classification in different scenarios. The true extent of the EZ cannot be determined before surgery as it is a hypothetical region, representing the smallest area of the brain that can be safely removed while resulting in freedom from seizures after surgery. The seizure‐onset zone (SOZ) and early spread zone (ESZ) displayed by the ictal SEEG are commonly used in clinical practice to identify the region for resection. Accordingly, in this study, interictal HFOs found in the SOZ and ESZ were classified as abnormal and potentially reflecting the EZ.

2. MATERIALS AND METHODS

2.1. Patient selection and grouping

The study retrospectively enrolled patients with focal epilepsy who had received depth electrode implantation and SEEG monitoring at Beijing Tiantan Hospital between January 2019 and December 2020. The inclusion criteria were (i) patients who had undergone preoperative evaluation and had been diagnosed with focal epilepsy, (ii) aged 7 years and older, ²² (iii) had electrode contacts in the sensorimotor or visual regions, ¹⁴ and (iv) had ictal SEEG data that could define the SOZ and ESZ. The exclusion criteria were (i) a history of previous resection surgery, (ii) age 6 years and younger, and (iii) inability to define SOZ and ESZ by ictal SEEG or the presence of habitual seizures that were not captured during SEEG recording. The Ethics Committee of the Beijing Tiantan Hospital approved the study.

Patients were then divided into subgroups according to various criteria (i) According to SOZ and ESZ location: the temporal lobe epilepsy (TLE) subgroup, characterized by the localization of SOZ and ESZ within the temporal lobe, and the extratemporal lobe epilepsy (eTLE) subgroup, in which SOZ and ESZ were located outside the temporal lobe; patients with SOZ and ESZ covering the temporal lobe and beyond were excluded from the subgroup analysis due to their limited number. (ii) According to imaging findings: the magnetic resonance imaging (MRI)‐positive (MRI+) subgroup, defined as the patients where epilepsy‐related lesions were found on MRI, and the MRI‐negative (MRI−) subgroup, including patients where MRI showed no significant findings. (iii) According to surgical outcome: including the seizure‐free subgroup, defined as the group of patients who reached International League Against Epilepsy (ILAE) 1 after resection surgery, and the nonseizure‐free group, defined as patients who did not reach ILAE 1 after surgery; patients who did not undergo resection surgery were excluded. The procedure described above was implemented on the subgroups and the entire patient population.

2.2. SEEG data acquisition and HFO detection

Intracranial electrodes (Huake‐Hengsheng Medical Technology, Beijing, China; 8–16 contacts) were placed using SinoRobot (Sinovation Medical Technology Company, Beijing, China) to record SEEG. Electrophysiological signals were then recorded on a video‐EEG system (Nihon‐Kohden, Tokyo, Japan) with a sampling rate of 2000 Hz. Long‐term video monitoring was performed simultaneously to capture habitual seizures.

Subsequently, two 10‐min segments were randomly selected from the awake stage (awake group), and two 10‐min segments were selected from the sleep stage (sleep group) to detect HFOs. To mitigate the potential impact of seizures, portions spaced more than 2 ho from ictal occurrences were selected. Further analysis as described in Sections 2.4 and 2.5 was then performed on the merged data (sleep and awake segments) and separate consciousness data (sleep or awake segment only). The automatic detection and classification of HFOs were then performed using a two‐stage detector we had previously developed and validated. ¹³ In brief, an envelope detector was used to filter out the candidate HFOs (cHFOs) which were subsequently classified as quality HFOs (qHFOs) or artifacts using a CNN‐based classifier. ¹³ The location of each HFO peak and its surrounding ±300 ms data were extracted for further analysis. The HFO detection was performed separately on each 10‐min SEEG segment from individual patients.

2.3. Electrode localization and classification

The electrodes were manually positioned according to the preoperative MRI and co‐registered postoperative computed tomography (CT) images. After re‐referencing the bipolar montage (i.e., A1–A2, A2–A3), white matter channels and channels designated as artifacts were excluded from the analysis. Moreover, as the PAC analysis requires at least 3000 ms data, ²³ channels in which fewer than five HFOs had been detected per 10‐min section in the sleep or awake stage were also excluded. The channels were then divided into pathological groups, including the channels located in SOZ and ESZ, and the physiological group, including the channels located outside SOZ and ESZ and not showing epileptiform discharges. The presence of SOZ was autonomously detected in all included patients by the two expert epileptologists, who examined the initiation of ictal discharge before the manifestation of clinical symptoms during documented seizures. ESZ was defined by channels showing the propagation of ictal activity within 3 s of the start of a seizure. ²⁴ HFOs recorded from the pathological channels were deemed pathological HFOs (pathoHFOs), and similarly, HFOs recorded from physiological channels were defined as physiological HFOs (phyHFOs).

2.4. Feature building

Based on previous studies and clinical observation, we constructed four metrics of three categories, namely, morphological repetition (referred to as “repetition” below), ¹⁸ rhythmicity, and PAC. ²⁵ Figure 1 displays the schematic diagram of the present study. Each feature was calculated for each of the two HFO types.

FIGURE 1.

Workflow diagram of the high‐frequency oscillations analysis. (A) This figure depicts the distribution of the various included channels. The purple area represents the visual site, and the blue area represents the somatomotor site. Red dots indicate the pathological channels, blue dots represent the physiological channels located in the somatomotor site, purple dots represent the physiological channels situated on the visual side, and gray dots represent the physiological channels other than those mentioned above. (B) Stereoelectroencephalography trace and interictal high‐frequency oscillation (HFO) automatic detection; the red trace indicates the detected HFOs. (C) Repetition calculation and the difference between EZ, defined as seizure‐onset zone plus early spread zone, and non‐EZ, defined as region outside the seizure‐onset zone plus the early spread zone. (D) HFO interval distribution in EZ and non‐EZ regions. (E) Phase–amplitude coupling (PAC) calculation and the difference between the epileptogenic zone (EZ) and nonepileptogenic zone (non‐EZ) between the phase of low frequency (3–10 Hz in the red block) and amplitude of high frequency (50–200 Hz in the red block).

2.4.1. Repetition

HFO signals were first subjected to high‐pass filtering at 3 Hz to avoid the influence of the slow wave. The repetition was then measured by the Pearson correlation coefficient r _HFOs between each HFO and the HFO template, defined as the average of all HFOs in the same channel. The mean value of all r _HFOs in the same channel was then calculated, with the result r _HFOs indicating the channel‐level repetition.

2.4.2. Rhythmicity

The skewness and kurtosis of the distribution of HFO intervals were used to quantify rhythmicity. After extracting the time stamp of each HFO peak from the automatic detection step, an interval distribution was constructed for each channel by choosing all of the intervals defined as the interval between each neighboring HFO pair. The skewness and kurtosis values were then used to describe the interval distribution's asymmetry and tailedness to represent the channel‐level rhythmicity.

2.4.3. PAC

The PAC, measured by the modulation index (MI), represents the coupling strength between the low‐frequency component phase and the high‐frequency component amplitude. ²⁵ MI was calculated using Phase–Amplitude Coupling Toolbox (PACT), a plug‐in for EEGLAB. ²⁶ A Hann window was added to each 600‐ms‐length epoch to calculate MI and eliminate the edge discontinuity. After that, all epochs in the same channel were concatenated to form a vector, from which the MI was finally calculated. We first selected the low‐frequency range between 1 and 35 Hz (from the delta band to the gamma band) and the high‐frequency range between 50 and 500 Hz (high gamma band). Subsequently, to represent a channel‐level PAC, the mean value of MI was assessed within the phase range of 3–10 Hz and the amplitude range of 50–200 Hz based on the conspicuous portion of the data (Figure 1E).

2.5. Channel‐level HFO classification

A feature vector was created for each channel using the four metrics described above. First, the features within the subject were standardized by computing z‐scores. The feature vectors were partitioned into the training and testing sets at a 3:1 ratio. The training set was fed into an extreme gradient boosting (XgBoost) classifier, ²⁷ and the testing set was used to evaluate the classifier's performance. The training set was then subjected to a 5‐fold cross‐validation (CV) based on the data for each patient. Additionally, the feature‐based classifier's maximum tree depth, minimum child weight, gamma, subsample ratio, and column subsample ratio were tuned during CV. Furthermore, each feature's contribution was assessed as the feature significance. Classifier testing procedures were performed repeatedly on the whole dataset and the awake/sleep, TLE/eTLE, MRI+/MRI‐ and seizure‐free/not seizure‐free subgroups.

2.6. Statistical analysis

Differences in channel‐level characteristics were compared between pathoHFOs and phyHFOs using either Student's t tests or Wilcoxon signed‐rank tests, depending on the normality of the data distribution. The same statistical analysis was applied to the whole data set and the subgroups. For the demographic data, Student's t‐test was used for the continuous variables.

The potential performance of the classifier was assessed by calculating various metrics, including accuracy, sensitivity, specificity, F1‐score, the receiver‐operating characteristic (ROC) curve, and the area under the curve (AUC). We defined the pathoHFOs, which were and were not classified as pathoHFOs as true‐positive (TP) and false‐negative (FN), respectively, and phyHFOs, which were and were not classified as phyHFOs as true‐negative (TN) and false‐positive (FP), respectively. Thereafter, the measures were computed as follows: accuracy = (TP + TN)/(TP + FN + TN + FP), sensitivity = TP/(TP + FN), specificity = TN/(TN + FP), F1‐score = 2TP/(2TP + FP + FN). The binomial test was used to determine whether the accuracy exceeded the chance levels. ²⁸ Chi‐square tests were then used to compare the accuracy of each subgroup to see if there were any significant differences. The statistical significance level was set at p < 0.05. Feature extraction, Student's t tests, chi‐square tests, and Wilcoxon signed‐rank tests were performed using MATLAB 2019a version (R2019a, The Mathworks, Natick, MA, USA). The construction of the feature‐based classifier and testing of its performance testing were done using Python 3.8.8.

3. RESULTS

3.1. Patient characteristics

Thirty‐five patients (21 females and 14 males) met the inclusion criteria after retrospective screening. The mean age at the time of SEEG was 24 years (range 7–49). The mean epilepsy duration before surgery was 12.3 years (range 2–48), and an average of 9.6 electrodes (range 5–13) were implanted for each patient. For TLE/eTLE analysis, 18 and 14 patients were grouped into TLE and eTLE subgroups, respectively, and three patients were excluded from the subgroup analysis. Twenty‐one patients were assigned to the MRI− subgroup, whereas 14 were assigned to the MRI+ subgroup for MRI analysis. For outcome analysis, 18 patients were allocated to the seizure‐free group, whereas 11 patients were included in the nonseizure‐free group. Additionally, six individuals were removed from the study as they did not undergo resection surgery. The epilepsy duration of patients was longer in the TLE group compared with the eTLE group, and there were no significant differences between age and number of electrodes among all subgroups (Table 1).

TABLE 1.

Characteristics of the included patients.

	All patients	TLE group	eTLE group	p‐value	MRI+ group	MRI‐ group	p‐value	Seizure‐free group	Not seizure‐free group	p‐value
Patients numbers	35	18	14		14	21		18	11
Age (years)	24.3 ± 8.7	25.9 ± 7.4	21.5 ± 9.3	0.14	22.3 ± 3.2	25.7 ± 10.9	0.24	20.67 ± 7.1	25.82 ± 6.3	0.16
Epilepsy duration (years)	12.3 ± 8.9	15.8 ± 10.3	7.9 ± 4.8	0.01*	12.1 ± 7.0	12.4 ± 10.1	0.88	9.72 ± 6.9	12.27 ± 4.8	0.18
Electrode numbers	9.6 ± 1.9	9.8 ± 2.0	9.4 ± 1.9	0.55	9.1 ± 1.4	10.0 ± 2.1	0.12	9.17 ± 2.1	10.18 ± 1.6	0.11

Note: p‐value indicates the difference between the characteristics of two subgroups.

Abbreviations: eTLE: extratemporal lobe epilepsy, MRI, magnetic resonance imaging, TLE: temporal lobe epilepsy.

^* p < 0.05.

3.2. HFO distribution

In total, 1 015 604 HFOs in 4066 channels were detected within the whole data set. After the exclusion of artifact channels, white matter channels, and channels with ≤5 HFOs observed per 10‐min segment, 219 478 HFOs in 537 channels were finally included for the feature extraction, among which 166 104 HFOs in 255 channels were labeled pathological and 53 374 HFOs in 282 channels were labeled physiological. The details of the subgroups are shown in Table S1. The entire data set was replicated in that the HFO rate, computed as the HFO numbers per minute, within the SOZ and ESZ, was significantly higher than that outside. The phyHFO rate in the visual and somatomotor areas was considerably higher than that observed in other nonepileptogenic regions (Figure 2). In addition, there was no significant difference in the pathoHFO rate between all subgroups (Figure S1). Similar results were found in the awake and sleep stages (Figures 2 and S1).

FIGURE 2.

Comparison of the high‐frequency oscillation rate between epileptogenic and nonepileptogenic zones. In the entire data set, it was found that the HFO rate in the epileptogenic zone (EZ), defined as the seizure‐onset zone plus early spread zone, was higher than that outside the EZ. In the normal brain region, the HFO rate in the functional area (defined as the visual site and the somatomotor site) was higher than in the outside functional area. Similar results were also found in the awake and sleep stages. EZ, epileptogenic zone. **p < 0.01; ***p < 0.01.

3.3. Differences in features between the two types of HFOs

Repetition: The correlation coefficients of pathoHFOs were significantly higher than phyHFOs on the channel level (p < 0.01), indicating that the morphology of pathoHFOs was more stereotyped (Figure 3A). Rhythmicity: In terms of the HFO interval distribution skewness and kurtosis, pathoHFOs were found to be significantly higher than phyHFOs from both metrics on the channel level (p < 0.01), thereby suggesting that the distribution of pathoHFO intervals was more skewed and sharp (Figure 3B,C). PAC: A significant increase in MI was observed between the amplitude of the high‐frequency band (50–200 Hz) amplitude and that of the low‐frequency band (3–10 Hz) phase in pathoHFO signals (p < 0.01), while a similar pattern was not seen in phyHFOs (Figure 3D). Besides, the results revealed that the phases of slow waves coupled with HFOs were different between pathoHFOs (mainly 90°–180°, Figure 3E) and phyHFOs (mainly −90° to 0°, Figure 3F). These comparisons were the same between pathoHFOs and phyHFOs in all subgroups (Figures S2 and S3).

FIGURE 3.

Differences between pathological high‐frequency oscillations (HFOs) and physiological HFOs. The results showed a significant difference in the repetition (A), skewness (B), and kurtosis (C) of the HFO interval distribution and phase‐amplitude coupling (PAC, D). (E, F) The PAC of pathological HFOs and physiological HFOs, the numbers on the radius mean the frequency of HFOs and the angles around the figure indicate the phase of slow waves, with color representing the modulation index. (G) The receiver‐operating characteristic (ROC) curve of the classification of the testing set, with the binomial test showing that the performance exceeded the level of chance. (H) Kurtosis and PAC were found to play critical roles in the classification. (I) An example of HFO classification in an individual patient; the red region refers to the surgical resection area, and the color of contacts indicates the normalized HFO rate, with a poor localization performance before HFO classification. (J) After the classification, the pathological HFO contribution was more consistent with the resection area. This patient has been seizure‐free for >1 year. EZ, epileptogenic zone; MI, modulation index; and PAC, phase–amplitude coupling.

3.4. Feature‐based classifier performances

After dividing the channels according to the ratio described in the methods section, the whole dataset (537 channels) was divided into the training data (402 including 196 pathological channels) and the testing data (135 including 59 pathological channels). The classifier correctly classified 51 of 59 pathological channels (TP = 51, FN = 8) and 62 of 76 physiological channels (TN = 62, FP = 14) in the testing data set. Therefore, the accuracy was 0.837, with a sensitivity of 0.864, a specificity of 0.816, and an F1 score of 0.823. The ROC analysis showed an AUC of 0.89 (Figure 3G), and the binomial test confirmed that the accuracy markedly exceeded the chance levels (p < 0.01, Figure 3G). Both PAC and rhythmicity appeared to have more prominent roles in the classifier (Figure 3H). Figure 3I,J shows a single patient who received resection surgery at our center and has been seizure‐free for >1 year, in which the red region refers to the surgical resection area; the localization performance was poor before the HFO classification (Figure 3I), while after the classification, the pathoHFO contribution was more consistent with the resection area (Figure 3J).

In the awake state, the classifier showed accuracy/sensitivity/specificity/F1‐score values of 0.723/0.712/0.750/0.765, respectively, and an AUC of 0.80. In the sleep state, the accuracy/sensitivity/specificity/F1‐score values were 0.780/0.833/0.736/0.775, respectively, and the AUC was 0.85 (Figure 4A). In the TLE subgroup, the classifier showed accuracy/sensitivity/specificity/F1‐score values of 0.831/0.833/0. 829/0.820, respectively, and an AUC of 0.89, whereas in the eTLE group, the accuracy/sensitivity/specificity/F1‐score values were 0.851/0.708/1.000/0.829, respectively, and AUC was 0.95. The classifier achieved an accuracy, sensitivity, specificity, and F1‐score of 0.841, 0.869, 0.803, and 0.864 in the MRI+ subgroup and 0.865, 0.865, 0.864, and 0.848 in the MRI subgroup, respectively, with AUCs of 0.90 and 0.92, respectively. The classifier accuracy, sensitivity, specificity, and F1 score values for the seizure‐free subgroup were 0.832, 0.831, 0.832, and 0.821, respectively, with an AUC of 0.91, while for the nonseizure group, these values were 0.888, 0.892, 0.883, and 0.902, respectively, with an AUC of 0.94 (Figure 4).

FIGURE 4.

The performances of classifiers in the subgroup analysis. The receiver‐operating characteristic (ROC) curve of the classification of the testing set in (A) awake/sleep subgroup, in which the areas under the curve (AUCs) were 0.89 in the whole data set, 0.80 in the awake group, and 0.85 in the sleep group. (B) The temporal lobe/extratemporal lobe epilepsy (TLE/eTLE) subgroup, in which the areas under the curve (AUCs) were 0.89 in the TLE group and 0.95 in the eTLE group. (C) The magnetic resonance imaging positive/negative (MRI+/MRI−) subgroup, in which the areas under the curve (AUCs) were 0.90 in the MRI+ group and 0.92 in the MRI‐ group. (D) In the seizure‐free/nonseizure‐free subgroup, the areas under the curve (AUCs) were 0.91 in the seizure‐free group and 0.94 in the nonseizure‐free group. The chi‐square test showed no significant difference in classification accuracy in all four pairs of subgroups.

The classification accuracy was found to be comparable between all pairs of subgroups according to the results of the chi‐square test (Figure 4), and the binomial test also confirmed that the accuracy exceeded the chance levels in the whole group as well as in all the subgroups (p < 0.01, Figure S4). The PAC and rhythmicity also played more critical roles in the subgroup classifiers (Figure S4). These measurements exhibited varying efficacies in different forms of focal epilepsy depending on the patient's level of consciousness.

4. DISCUSSION

Addressing physiological and pathological HFOs is essential to achieving efficient EZ localization. However, the use of these traits for the successful differentiation of HFOs remains in an exploratory stage. The results of this investigation confirmed that SOZ and ESZ exhibited substantially higher HFO rates, thus closely resembling EZs seen in clinical practice. The putative phyHFO rates were also higher in the eloquent cortex according to functional network and epileptic conditions, echoing previous studies. ¹⁴ , ²⁹ Four metrics from three distinct categories that differed significantly between the HFO types were proposed. These features were further incorporated into a machine‐learning classification algorithm to classify pathoHFOs and phyHFOs at the channel level.

Given the significant amount of human labor required for HFO analysis, their automatic detection has become necessary and common. However, currently available detectors focus on the accurate identification of HFOs, and the classification procedure is insufficiently incorporated, especially for the classification of pathoHFOs and phyHFOs. ¹² , ³⁰ Although the previous detectors assisted with labeling the various HFO components, such as ripple, fast ripple, and spike, ¹³ , ³¹ the frequency‐band overlap between the two HFO types may still jeopardize the localization accuracy. ³² Consequently, this study developed a systematic and effective framework based on multiple features and machine‐learning algorithms that could be successfully used in an automatic pipeline for HFO analysis to prevent the interference of phyHFOs. Sufficient data from the two unique categories are required to determine the stable distinguishing metric of the two different types of HFOs. To increase the probability of successful phyHFO recordings, this study specifically recruited patients with depth electrodes placed in the sensorimotor or visual regions, which are known to produce phyHFOs more frequently. ¹⁴ The findings indicated that these two expressive areas displayed greater spontaneous HFOs compared to the other regions.

There have been many investigations into HFO classifications based on multiple features such as HFO duration and amplitude. ³³ Previous studies have also applied a more sophisticated algorithm to achieve higher classification accuracy. For example, Su Liu et al. defined HFOs recorded within the EZ as pathological, thus suggesting that pathoHFOs were morphologically stereotyped. ¹⁸ In contrast, Jacobs et al. reported that the backgrounds of pathoHFOs were relatively quiet. ¹⁹ Additional research has consistently demonstrated that pathoHFOs frequently co‐occur with spikes. ²⁰ , ³⁴ The presence of morphological repetitions and HFOs riding on the spikes were observed. This study utilized a template‐matching approach to quantify the morphological characteristics.

The similarity between single HFOs and an averaged template was quantified through amplitude correlation. The results indicated the high distinguishing value of such a feature. The rhythmicity of the HFO temporal distribution is frequently described, especially in patients with focal cortical dysplasia. The present study measured the distribution of HFOs across the time on each channel, and the results suggested that the intervals of pathoHFOs appeared to be more consistent than phyHFOs. In terms of the PAC feature, which reflected the potential interaction of HFOs with low‐frequency oscillations (3–10 Hz), the results indicated the MI of pathoHFOs was significantly higher than that of phyHFOs, and prior study has suggested that pathoHFOs may be preferentially coupled with slow waves between 3 and 4 Hz, whereas phyHFOs may be coupled with those between 0.5 and 1 Hz ³⁵ ; thus, the MI of pathoHFOs, which was calculated between HFOs and slow waves between 3 and 10 Hz in the current study, was higher than those of phyHFOs.

Moreover, several studies have also revealed a relationship between HFOs and sleep slow waves in which pathoHFOs were coupled with the “down” state of sleep slow waves. At the same time, phyHFOs were more likely to be associated with the “up” state of slow waves. ³⁶ , ³⁷ These studies may explain the different phases of the slow wave coupling between pathoHFOs and phyHFOs, which was observed here. This study observed that although the difference was more significant during sleep, the MI of pathoHFOs was also significantly higher than that of phyHFOs during the awake state. In addition to the above‐mentioned studies that have detailed the relationship between sleep slow waves and the two types of HFOs, another reason may be that the pathoHFOs ride on the spike events more often than do phyHFOs. ³⁴ , ³⁸ Several studies have also implied that HFOs associated with mesial temporal and neocortical epilepsy might differ in morphological features. ³⁹ , ⁴⁰ The present study did a similar analysis but separated awareness states and grouped patients according to EZ localization and lesion type to address the concern that these results might not be replicated in the varied disease and patient states. The subgroup analysis showed that the results remained stable across different pathologies and states of consciousness.

The classification of pathoHFO‐ and phyHFO‐generating tissues is required for the clinical translation of automatic HFO analysis. Thus, using individual features such as the pathoHFO rate might result in false‐positive localization due to the normal brain‐related phyHFOs. However, when combining multiple pathoHFO features from the morphology and rhythmicity aspect, the XgBoost‐based classifier reached a relatively high distinguishing accuracy well above that of chance. This suggests the potential for future algorithm development. In addition, the present study has also verified the classification using the subgroup analysis, thus enhancing the generalizability of the algorithm to different pathologies. The AUC dropped when dividing the data into the awake and sleep states, possibly explained by the shortened data segments included. It has already been established that longer data segments may yield more stable results considering the time and spatial variation of HFO distribution. ³⁹ , ⁴¹ Following similar speculation, the AUC in the sleep state was significantly higher than that in the awake state, which may partially be due to the higher HFO rate during sleep. ⁴² Consistent with previous research, the results indicated no substantial difference in classification performance across different MRI representations. ⁴⁰ , ⁴³

The current analysis showed that the classifier performed well in terms of surgical outcomes in both the seizure‐free and nonseizure‐free groups. In general, poor surgical results are most likely caused by the use of an incorrect EZ hypothesis before SEEG, implying that the pathHFOs detected by SEEG may have originated from the true EZ. Therefore, it is possible to differentiate the current findings from phyHFOs by observing that these traits continued to exist in propagating pathoHFOs.

Interestingly, despite all the extracted features differing significantly between the TLE and eTLE groups, the classifier tended to be more accurate in the eTLE group (AUC 0.95 compared to 0.89). Therefore, the above‐mentioned characteristics might be more advantageous for eTLE patients; nonetheless, the TLE group's general categorization performance was still deemed excellent. The PACs were all placed first in the feature contribution analysis throughout the entire data set and subgroups, indicating that a major feature of the pathoHFOs was the high‐frequency amplitude interaction with the phase of low frequencies. However, only a small portion of the bias was caused by other hypothesized characteristics.

This study also has several limitations. First, to rigorously differentiate the pathological HFOs from physiological HFOs, we defined the pathological channels as channels in SOZ and ESZ according to ictal SEEG results. However, on the one hand, like an interictal spike, pathoHFOs can also spread from the EZ to normal brain regions, so that pathoHFOs could be detected in some physiological channels, which might partly interfere with the classification. In this study, we found that the pathoHFO rate was significantly higher than that of physiological HFOs under different levels of consciousness, which when combined with the SOZ criteria provide a gold standard for determining an epileptogenic area before surgery. The use of the ESZ criteria avoids the influence of propagated pathological HFOs on physiological HFOs, and we excluded the physiological HFOs from the pathological HFOs to a large extent, with the remaining overlap probably not affecting the channel‐level features and hampering the classification due to the imbalance in the number of the two types. On the other hand, the classifier could thus only differentiate SOZ and ESZ from nonepileptogenic regions, and the ESZ could not be separated from the SOZ, posing a significant difficulty in clinical practice. It was possible to differentiate between the ESZ and SOZ by looking at the relative density of spikes and sleep slow waves, which Chen et al. found to be substantially higher in the irritative zone than in SOZ. ³⁷ Furthermore, they also found that HFOs that did not coincide with spikes in the SOZ were mostly pathological, whereas these types of HFOs were a mixture of pathoHFOs and phyHFOs in the ESZ. ³⁷ Thus, distinguishing the ESZ from the SOZ by computing the ratio of pathoHFOs/phyHFOs, classified by the methods described above, could be a potential solution, given that pathoHFOs and phyHFOs can be differentiated based on deep learning methods ⁴⁴ or other features. ⁷ , ⁴⁵ Second, only brief interictal segments from individual subjects were included, rather than long‐term continuous monitoring. Furthermore, the current classifier could only perform a channel‐level classification due to the limitations of the study's methodology. Therefore, future studies will need to develop an event‐level classifier capable of distinguishing between the pathological and physiological aspects of a single HFO activity.

5. CONCLUSIONS

This study proposed a practical solution to the critical problem that significantly impedes the identification of HFOs in clinical practice, namely, the differentiation between physiological and pathological channels based on HFO features. The current study proposed four distinct quantitative metrics derived from three features, namely, morphological repetition, rhythmicity, and PAC, which could be used in an XgBoost classifier to effectively discriminate the two types of HFOs at the channel level. The localization value of HFOs could thus be markedly improved without interference from physiological HFOs.

AUTHOR CONTRIBUTIONS

Zilin Li: Writing—original draft, writing—review and editing, Methodology, Software, Validation, Formal analysis, Visualization. Baotian Zhao: Writing—review and editing, Methodology, Software. Wenhan Hu: Writing—review and editing, Data curation, Funding acquisition. Chao Zhang: Writing—review and editing, Data curation. Xiu Wang: Writing—review and editing, Data curation. Jianguo Zhang: Writing—review and editing, Supervision, Investigation. Kai Zhang: Conceptualization, Resources, Writing—original draft, Writing—review and editing, Funding acquisition, Supervision, Data curation, Investigation.

FUNDING INFORMATION

This work was sponsored by the National Natural Science Foundation of China (82201600, 82071457, 82201603), National Key R&D Program of China (2021YFC2401201) and Capital's Funds for Health Improvement and Research (2022‐1‐1071, 2020‐2‐1076).

CONFLICT OF INTEREST STATEMENT

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

ETHICS APPROVAL AND CONSENT TO PARTICIPATE

The study was approved by the Ethics Committee of the Beijing Tiantan Hospital with an approval number of KY 2018‐076‐02. The study was conducted per the Declaration of Helsinki, and all patients provided informed consent for using their medical records.

Supporting information

Data S1.

Li Z, Zhao B, Hu W, Zhang C, Wang X, Liu C, et al. Practical measurements distinguishing physiological and pathological stereoelectroencephalography channels based on high‐frequency oscillations in the human brain. Epilepsia Open. 2024;9:1287–1299. 10.1002/epi4.12950

DATA AVAILABILITY STATEMENT

The data sets generated and analyzed during the current study are available from the corresponding author upon reasonable request.