Visual Features in Stereo‐Electroencephalography to Predict Surgical Outcome: A Multicenter Study

Abstract

Objective

Epilepsy surgery needs predictive features that are easily implemented in clinical practice. Previous studies are limited by small sample sizes, lack of external validation, and complex computational approaches. We aimed to identify and validate visually stereo‐electroencephalography (SEEG) features with the highest predictive value for surgical outcome, and assess the reliability of their visual extraction.

Methods

We included 177 patients with drug‐resistant epilepsy who underwent SEEG‐guided surgery at 4 epilepsy centers. We assessed the predictive performance of 10 SEEG features from various SEEG periods for surgical outcome, using the area under the receiver operating characteristic curve, and considering resected channels and surgical outcome as the gold standard. Findings were validated externally using balanced accuracy. Six experts, blinded to outcome, evaluated the visual reliability of the optimal feature using interrater reliability, percentage agreement (standard deviation ± SD) and Gwet's kappa (κ ± SD).

Results

The derivation cohort comprised 100 consecutive patients, each with at least 1‐year of postoperative follow up (40% temporal lobe epilepsy; 42% Engel Ia). Spatial co‐occurrence of gamma spikes and preictal spikes emerged as the optimal predictive feature of surgical outcome (area under the receiver operating characteristic curve 0.82). Applying the optimized threshold from the derivation cohort, external validation in 2 datasets showed similar performances (balanced accuracy 69.2% and 73.2%). Expert interrater reliability for gamma spikes (percentage agreement, 96% ± 2%; κ, 0.63 ± 0.16) and preictal spikes (percentage agreement, 92% ± 2%; κ, 0.65 ± 0.18) were substantial.

Interpretation

Spatial co‐occurrence of gamma spikes and preictal spikes predicts surgical outcome. These visually identifiable features may reduce the burden of SEEG analysis by reducing analysis time, and improve outcome by guiding surgical resection margins. ANN NEUROL 2025;98:547–560

Epilepsy is one of the most common neurological conditions that affects individuals across various age groups, social classes, and geographical locations. It presents a significant healthcare challenge, impacting >70 million people worldwide. ¹ Seizures persist in up to 40% of patients, despite the use of optimal antiseizure medications. ² In such cases, epilepsy surgery is the only therapeutic option to achieve seizure freedom and improve overall quality of life, by resecting of the presumed epileptogenic zone (p‐EZ). ³ However, before surgery, an accurate delineation of the p‐EZ, derived from noninvasive and, in many cases, additional invasive investigations, is crucial.

Stereo‐electroencephalography (SEEG) is an invasive diagnostic procedure, increasingly utilized worldwide, and it is considered the gold standard to delineate the p‐EZ, especially in complex cases. ⁴ During SEEG interpretation, epileptologists primarily rely on several visual features, such as spikes, ⁵ ictal fast activity (FA) at seizure onset, ⁶ postictal slowing and/or suppression, ⁵ and cortical stimulation results. ⁷ , ⁸ Additional features have been proposed, based on computational approaches, including high‐frequency oscillations (HFOs), ⁹ the epileptogenicity index, ¹⁰ and, more recently, gamma spikes. ¹¹ , ¹²

Although previous studies showed that these features provided valuable information about the p‐EZ and to help prognosticating the surgical outcome, ¹³ surgical decision remains challenging, as approximately 40% of individuals continue to experience disabling seizures after SEEG‐guided resective surgery. ¹⁴ Hence, there is a need for predictive features of epilepsy surgical outcome, which are accurate and easy to use in clinical practice. Quantitative measures derived from computational methods are not broadly implemented in clinical practice, due to technical difficulties and a lack of demonstrated superiority to the standard visual approach. ¹⁵ Ideally, predictive features of surgical outcome should be easily accessible to every epileptologist worldwide. Furthermore, validating the results across external datasets is essential.

Therefore, given that visual assessment remains the standard approach in clinical practice, with most epileptologists relying primarily on visual identification, we visually extracted various features from interictal, preictal, ictal, postictal periods, and cortical stimulation data, from a large SEEG dataset of patients with drug‐resistant focal epilepsy who underwent surgery, and then validated the results in 2 additional datasets. Our goal was to find a reliable and easily accessible surgical predictive feature. To fulfill this objective, the performance of single features and their spatial co‐occurrence in classifying seizure‐free (Engel Ia) versus nonseizure‐free (Engel Ib–IV) patients were assessed using the resected channels and surgical outcome as the gold‐standard. More specifically we pursued the following 3 aims: (1) identify the optimal visual predictive feature; (2) validate its performance using 2 external datasets; and (3) assess its visual extraction reliability among several epileptologists trained at different epilepsy centers.

Materials and Methods

Phase 1: Identification of the Optimal Visual Predictive Feature

Study Design and Derivation Cohort

Consecutive patients with drug‐resistant focal epilepsy, who underwent SEEG investigation at the Montreal Neurological Institute (MNI [2008–2021]) and the Grenoble Alpes University Hospital Center (CHUGA [2009–2018]), as part of their presurgical work‐up, were selected based on the inclusion and exclusion criteria outlined in Figure 1. Briefly, we included patients with unifocal epilepsy who had at least 1 recorded habitual electroclinical seizure during SEEG investigation, along with available preoperative, post‐implantation and postoperative imaging, and who underwent cortical stimulation as part of the SEEG investigation. Patients presenting nonhabitual seizures were excluded, as these cases were likely multifocal. In such cases, surgical approaches vary between centers, and the primary goal is often palliative rather than curative. All patients underwent SEEG‐guided surgery with a minimum follow‐up period of 1 year. Resection margins were determined through a multidisciplinary consensus and individualized for each patient based on SEEG findings, aiming to remove the p‐EZ, with the seizure onset zone (SOZ) serving as a proxy for the p‐EZ. In cases where an epileptogenic lesion was visible on the magnetic resonance imaging (MRI) and matched SEEG findings, the resection was adjusted accordingly. If eloquent cortex was involved, the surgical strategy was further adjusted based on the identified extent of the eloquent cortex relative to the SOZ, aiming to maximize resection while preserving function. In cases involving multiple SEEG procedures followed by multiple surgeries, only the SEEG prior to the last surgery was considered.

Figure 1.

Flow chart. CHUGA = Grenoble Alpes University Hospital Center; CHUNA = Nancy University Hospital Center; HUP = HUP iEEG Epilepsy Dataset (https://openneuro.org/datasets/ds004100/versions/1.1.3); MNI = Montreal Neurological Institute; SEEG = stereo‐electroencephalography. *Clusters of seizures were excluded due to the potential mixing of postictal and preictal phases from the subsequent seizure after the initial one.

The present retrospective study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. ¹⁶ This study was approved by the respective research ethics boards (MNI REB IRB00010120, Cogepistim MR004 11.05.21 DRCI CHUGA) and written informed consent was obtained from all patients.

Feature Selection and Definitions

The methods of data acquisition were previously described. ¹¹ In brief, the MNI SEEG recordings were recorded with either Harmonie (Stellate, Montreal, Quebec, Canada) or Nihon‐Kohden EEG amplifiers (Tokyo, Japan) with homemade MNI or commercial DIXI Medical (Besançon, France) electrodes. The CHUGA recordings were acquired using Micromed EEG amplifiers (Mogliano Veneto, Italy), with DIXI Medical or ALCIS (Besançon, France) electrodes. Recordings used a reference placed on a SEEG channel without EEG abnormalities, typically located far from the most active SEEG sites. The exact location varied by center and implantation scheme.

All features were visually extracted using a bipolar montage of neighboring contacts of the SEEG electrodes, utilizing Stellate. Notch filters (50 Hz for the CHUGA dataset and 60 Hz for the MNI dataset) and a bandpass filter between 0.3 and 100 Hz were applied. Additionally, a gamma band filter (30–100 Hz) was applied specifically for the visual identification of gamma spikes. A total of 10 features (Fig 2) from 5 different segments (interictal, preictal, ictal, postictal, and cortical stimulation if a typical electroclinical seizure was induced) were visually extracted by an epileptologist (C.A.) with SEEG analysis training, who was blinded to the list of resected channels and the surgical outcome at the moment of the visual analysis. No automatic detector was used.

Figure 2.

Feature selection and definitions. EEG = electro‐encephalography; NREM = nonrapid eye movements; sec = second. [Color figure can be viewed at www.annalsofneurology.org]

These features enabled us to define the regions of interest utilized for analysis (Table S1). Segments were chosen at least 24 hours post‐SEEG implantation to minimize the potential impact of anesthesia and acute implantation. To address the variability of interictal features related to vigilance state and reduce the impacts of reduced or discontinued medication, we primarily selected data from the first days of implantation (>24 hours), focusing on wakefulness and sleep segments. If one state was found to be most effective in predicting surgical outcomes, only that state was included. For ictal data, we also primarily selected data from the first days of implantation. In addition, as the vigilance state does not affect ictal spatiotemporal dynamics, ¹⁷ only one typical electroclinical seizure irrespective of the vigilance state was analyzed. If no segment was meeting the criteria from the first days of implantation, segments meeting the criteria at any time of the SEEG investigation were considered.

From the interictal segments (10‐minute segments during wakefulness and nonrapid eye movement [NREM] sleep), we extracted the primary irritative zone defined by channels showing the most frequent (highest frequency visually identified) spikes during wakefulness and NREM sleep (channels that consistently showed the most frequent spikes across both states of vigilance, including their spatial spread during NREM sleep; Fig S1). Precisely, we visually identified channels consistently showing prominent, repetitive spiking activity. This qualitative approach relied on pattern recognition, where channels with higher spike frequency appeared more active with frequent spikes over time. This region has been demonstrated to contain meaningful information regarding the p‐EZ location. ¹⁸ Additionally, NREM sleep has been identified as optimal for localizing the p‐EZ. ¹⁹ , ²⁰ These segments were required to be free from seizure activity, with a buffer period of 2 hours before and after the selected segments.

The second feature from the interictal data belonged to the gamma spike zone, known for its effectiveness in predicting surgical outcome. ¹¹ We utilized the 10‐minute wakefulness segment, as this state was shown to be most effective in predicting surgical outcome with gamma spikes. ¹¹ This zone was defined as set of channels with at least 1 spike with a minimum of 3 oscillations in the gamma band (30–100 Hz) preceding the spike‐onset (Supplementary Method, Fig S2–S4). First, we identified spike onset at the channel level, applied a gamma bandpass filter, increased SEEG gain, and extracted channels showing gamma activity before spike onset, requiring at least 4 oscillations, including 3 before the spike onset. All channels displaying spikes were analyzed until the criteria for identifying gamma spikes were met.

Only 1 typical and habitual electroclinical seizure was analyzed for each patient. We utilized segments with at least 1 minute prior to the ictal EEG onset and at least 1 minute after the end of the seizure. With the segment containing the ictal activity, we extracted the preictal spike zone, known for its relevance in localizing the p‐EZ. ²¹ This zone included the channels that showed spikes occurring within the 30‐second window preceding the ictal EEG onset (Fig S5), regardless of spike count or morphology, which largely vary depending on the underlying etiology or anatomical location (for example, spikes associated with hippocampal sclerosis differ in shape and frequency from those seen in focal cortical dysplasia type II). ²² Additionally, we extracted the last‐preictal spike zone, defined as the list of channels with the last spikes closest to the ictal EEG onset, regardless of the delay, given the heterogeneity of epilepsy (Fig S6). Subsequently, we extracted ictal features that belonged to the SOZ, ² defined as the list of channels with the first sustained EEG modification (at time zero) that led to a seizure (Fig S6). As the propagation speed of ictal activity varies significantly by anatomical region, with longer latencies in mesial temporal lobe epilepsy compared with neocortical temporal lobe epilepsy and other neocortical areas, such as the frontal lobe, ²³ we defined the SOZ + early spread zone as including both the SOZ and the subsequent first rhythmic ictal activity identified in another set of channels within a distinct anatomical/subanatomical region, irrespective of the time of the spread (Figs S6, S7). Finally, the fast activity (FA) zone, which has been shown to have valuable information regarding the localization of the p‐EZ, ²⁴ was defined as channels visually manifesting ictal activity (≥13 Hz) within the time interval of 0–1 second, where 0 represents the seizure onset (Fig S6).

Next, using the postictal segment, we delineated the postictal zone (Fig S8), comprising channels showing suppression of the SEEG background compared with the preictal SEEG or slow‐wave activity (typically in the delta band) subsequent to the seizure. ²⁵ These zones may offer valuable insights into the localization of the p‐EZ. ²⁶ Moreover, we identified the first postictal spike zone, comprising channels showing the initial spikes following the seizure. Finally, considering the association between inducing a typical electroclinical seizure during cortical stimulation and favorable surgical outcomes, ⁷ , ⁸ we extracted the induced SOZ, obtained either at a low frequency (1 Hz) or a high frequency (50–55 Hz). This zone comprised the list of channels with sustained modification that led to a seizure similar to the typical electroclinical seizure. If multiple induced seizures met these criteria, all were analyzed. The intersection of all the corresponding induced SOZ was considered to maintain the specificity of this region.

Of note, the selection of those features combined both exploratory and hypothesis‐driven approaches. Exploratory features included first postictal spikes, last preictal spikes, and postictal suppression/slow‐wave, as their value in predicting surgical outcomes is less established. ²⁷ , ²⁸ In contrast, features, such as the SOZ, preictal spikes, and gamma spikes, were selected based on prior evidence supporting their predictive value in surgical outcomes. ¹¹ , ²¹

Optimal Predictive Features of Surgical Outcome

Each single feature and all possible spatial co‐occurrences among the 10 features were considered for analysis. Once the performance of the co‐occurrence reached saturation, indicating no further improvement in terms of area under the receiver operating characteristic curve (AUC) value with additional features, we concluded the analysis. Each spatial co‐occurrence resulted in a new feature. The gold standard utilized was the list of resected channels, visually identified, from the postoperative MRI (co‐registered with the preoperative and post‐implantation SEEG data ⁷ ) and the surgical outcome. A bipolar channel was considered resected if both contacts were resected on the co‐registered postoperative image. To account for sagging, co‐registration error, and partial contact resection, contacts located within or near (<5 mm) the cavity were also classified as resected. ⁷

To identify the optimal predictive feature of surgical outcome, the trade‐off between classification performance in seizure‐free (Engel Ia) versus nonseizure‐free (Engel Ib–IV) patients and the number of features required to achieve such performance was considered. The optimal predictive feature was anticipated to demonstrate high performance (see Statistical Analysis) in classifying seizure‐free and nonseizure‐free patients while using a minimal number of features. This strategy allows to streamline SEEG analysis, and reduce the associated burden in clinical practice by minimizing the time and complexity of analysis.

Phase 2: External Validation

After identifying the optimal visual surgical predictive feature, we computed its performance in 2 external cohorts. Visual identification of this feature was performed by C.A. The first cohort included 50 patients with drug‐resistant focal epilepsy who underwent SEEG followed by resective brain surgery at the Nancy University Hospital Center (CHUNA). This study was approved by the review board of the CHUNA, and all included patients gave written informed consent. The second cohort included 27 patients from the open HUP iEEG Epilepsy Dataset who had undergone SEEG‐guided surgery (ablation or open brain surgery; https://openneuro.org/datasets/ds004100/versions/1.1.3). ¹³ Patients were selected based on the inclusion and exclusion criteria outlined in Figure 1.

Phase 3: Interrater Reliability Analysis

We asked 6 experts (O.A., I.D., D.M., P.N., P.P., J.S.), trained in different centers worldwide, blinded to the surgical outcome, to independently extract the optimal visually predictive feature identified in phase 1. Each expert reviewed a consecutive and anonymized dataset of 20 patients. Additionally, we requested the extraction of the SOZ, considered as the primary proxy for the p‐EZ, ²⁹ and the 3 optimal single predictive features. The number of experts was determined based on a previous study suggesting that the optimal number of experts for scoring EEGs is likely within the range of 6–10. ³⁰ For the analysis, experts were asked to use the Stellate Reviewer, and a bipolar montage was created for each patient. Each expert received a document detailing the process of extracting the features of interest (Supplementary eDocument).

Statistical Analysis

The normality of feature distributions was first assessed using the Kolmogorov–Smirnov test. ³¹ As the data did not follow a normal distribution, we applied nonparametric Wilcoxon rank‐sum tests ³² to compute the p values and Cliff's Delta ³³ to calculate effect sizes. Cliff's Delta ranged between −1 and 1, with extreme values indicating a better separation. In addition, we applied the Bonferroni–Holm correction ³⁴ for p values when performing multiple comparisons across the 10 features. The overlap ratio between each single/spatial co‐occurrence feature (features occurring within the same channels) and the resected channels was computed as ¹⁵ :

$\text{Overlap ratio}=\left(\mathrm{A}\bigcap \mathrm{B}\right)/\left(\mathrm{A}\bigcup \mathrm{B}\right)=\mathrm{TP}/\left(\mathrm{TP}+\mathrm{FP}+\mathrm{FN}\right).$

Where A represents the list of channels belonging to the feature of interest, B represents the list of resected channels, TP (true positive) is the number of feature channels that were resected, FP (false positive) is the number of feature channels that were not resected, and FN (false negative) is the number of resected channels that were not present in the feature. By incorporating both sensitivity (TP / [TP + FN]) and precision (TP / [TP + FP]) in the context of surgery, the overlap ratio provides a more comprehensive assessment compared with the simple overlap, which only considers sensitivity. As the goal of epilepsy surgery is to accurately identify and resect the p‐EZ while minimizing both FP and FN, the overlap ratio serves as a relevant and informative metric.

Next, we used the overlap ratios as predictions, with seizure‐free patients considered as the positive class and nonseizure‐free patients as the negative class. Classification results were presented using receiver operating characteristic curves and the AUC was considered as a measure of detection accuracy by thresholding the values of the overlap ratios from 0 to the maximum value. A random AUC value corresponds to 0.5, and a higher AUC value indicates better performance in classifying patients from both groups. We utilized the DeLong ³⁵ test to assess whether the AUC values were statistically different across features (single features and their spatial co‐occurrences). For the AUC confidence intervals (95% CI), we computed the values 100 times, and each time, a random 20% of the patients were removed from the dataset.

From the derivation cohort, the optimized threshold was identified based on the highest balanced accuracy defined as: BA = (sensitivity + specificity) / 2, where sensitivity is TP / TP + FN and specificity is TN / TN + FP; TP (true positive) is the number of seizure‐free patients correctly predicted, FN (false negative) is the number seizure‐free patients incorrectly predicted to be nonseizure‐free, TN (true negative) is the number of nonseizure‐free patients correctly predicted, FP (false positive) is the number of nonseizure‐free patients incorrectly predicted to be seizure‐free. Using this threshold, we evaluated the performance of the optimal predictive feature in 2 external datasets, and computed the sensitivity, specificity, and balanced accuracy for each external dataset. Moreover, the performance of the optimal predictive feature was calculated in each dataset after applying specific dataset optimized threshold.

We also examined the pairwise interrater reliability (IRR) among the 6 experts using the mean percentage agreement (PA) and chance‐adjusted Gwet's kappa. ³⁶ Gwet's coefficient (κ) was calculated as follows ³⁶ :

$$\mathrm{\kappa }=\frac{p-e\left(\gamma \right)}{1-e\left(\gamma \right).}$$

Where e(γ) represents the chance agreement probability and p is the overall PA. We utilized standard conventions for interpreting κ; 0–0.20 (slight), 0.21–0.40 (fair), 0.41–0.60 (moderate), 0.61–0.80 (substantial), and 0.81–1.00 (almost perfect). ³⁷ IRR was assessed both with and without accounting for the class imbalanced problem inherent in SEEG data, where channels without the feature of interest are typically much more frequent than those with it. To address this problem, we used the following procedure: for each expert pair, we randomly selected channels identified as showing the feature of interest and matched them with an equal number of channels without the feature, as identified by the experts. The number of randomly selected channels was half of the minimum number of channels identified with the feature of interest. This ensured an optimal balance for randomizations. If the number of channels was <2, it was adjusted to 2. This process was repeated 100 times per patient, and the resulting κ values were averaged per patient and then across patients for each feature. Only the IRR results after addressing the class imbalance are reported in the manuscript, since this approach is more appropriate. Results without addressing the imbalance problem are provided in the Supplementary Material.

Finally, to ensure that the performance of the optimal predictive feature was not driven by a specific epilepsy type, MRI findings, spatial extent of the SOZ, or pathology, we evaluated the performance of the identified optimal feature in classifying seizure‐free versus nonseizure‐free patients according to epilepsy types (temporal lobe epilepsy [TLE] and extra TLE), MRI findings (lesional and nonlesional), spatial extent of the SOZ (focal and nonfocal [regional to widespread]) following the classification used in Abdallah et al., ³⁸ and pathologies (focal cortical dysplasia [FCD] type II, hippocampal sclerosis, gliosis/scar, other malformations of cortical development, and vascular malformation/low‐grade epilepsy‐associated neuroepithelial tumor). For this analysis, available data from all 4 centers were pooled for each category.

Results

We included 100 patients (46 from the MNI, 54 from the CHUGA) in the derivation cohort (53 women; median age 30.8 years, interquartile range 22.49–38 years). A total of 60 patients had extra TLE. A total of 44 had a nonlesional MRI. A total of 42 patients remained seizure‐free at their last post‐surgical follow up (median 48 months, interquartile range 26.25–79.75 months). Clinical characteristics are shown in Table 1. No clinical differences were observed between seizure‐free and nonseizure‐free groups, except for MRI findings. Seizure‐free patients showed fewer other lesions, such as atrophy, cavernoma, polymicrogyria, and so on compared with nonseizure‐free patients.

Table 1.

Demographic and Clinical Characteristics in the Derivation Cohort

	Engel Ia group (n = 42)	Engel Ib–IV (n = 58)	p value (2‐sided test)
Sex, n (%)
Female	24 (57.1)	29 (50)	0.54
Male	18 (42.9)	29 (50)
Median age, yr (IQR)
At seizure onset	9.5 (4–16)	12 (7–19)	0.15
At SEEG investigation	29 (16.1–38)	31 (24–38)	0.23
Median duration of epilepsy at SEEG, yr (IQR)	14.25 (8–23)	15 (9–28)	0.45
MRI findings (lesional vs nonlesional), n (%)
Lesional	21 (50)	35 (60.3)	0.31
Nonlesional	21 (50)b	23 (39.7)
MRI findings, n (%)
FCD II	14 (33.3)	10 (17.2)	0.03
HS	2 (4.8)	2 (3.4)
Normal	20 (47.6)b	23 (39.7)
Others	6 (14.3)	23 (39.7)
No. electrodes, median (IQR)	14 (10–15)	12 (9–15)	0.3
Type of epilepsy (TLE vs ETLE), n (%)
TLE	18 (42.9)	22 (37.9)	0.68
ETLE	24 (57.1)	36 (62.1)
Type of epilepsy, n (%)
Frontal	16 (38.1)	20 (34.5)	0.65
Temporal	20 (47.6)	26 (44.9)
Posterior	4 (9.5)	10 (17.2)
Insulo‐opercular	2 (4.8)	2 (3.4)
Side of surgery, n (%)
Right	24 (57.1)	24 (41.4)	0.15
Left	18 (42.9)	34 (58.6)
Median follow up, months (IQR)	38 (23–72)	60 (32–80)	0.11
Pathology, n (%)a
FCD II	16 (44.4)	13 (25.5)	0.19
HS	4 (11.1)	3 (5.9)
Gliosis, scar	7 (19.4)	14 (27.5)
Cavernoma, tumor	2 (5.6)	3 (5.9)
Other MCD (FCD I, FCD III, FCD unspecified, PNH, ectopic neurons)	7 (19.5)	18 (35.2)

Note: Table summarizes the demographic and clinical characteristics of the derivation cohort. Categorical variables are displayed as number (percentage), and were analyzed with the Fisher's exact test and Pearson's χ²‐test. Ordinal variables are shown as the median (interquartile range [IQR] 25–75) and were analyzed with the Mann–Whitney test. ^aSome missing data (13/100 including 6 in the Engel Ia group and 7 in the Engel Ib–IV group); ^b1 patient with an agenesis of the corpus callosum on the magnetic resonance imaging (MRI) was classified as a nonlesional case, but grouped under “others” when considering the details MRI due to abnormal MRI findings.

Abbreviations: ETLE, extratemporal lobe epilepsy; FCD, focal cortical dysplasia; HS, hippocampal sclerosis; MCD, malformation of cortical development; MRI, magnetic resonance imaging; PNH, periventricular nodular heterotopia; SEEG, stereo‐electroencephalography; TLE, temporal lobe epilepsy.

Single Features

As shown in Fig S9, the spatial distribution of each feature was normalized by both the number of resected channels and the number of implanted SEEG electrodes. The distribution patterns remained consistent across both approaches. We observed that gamma spikes showed the highest AUC, achieving a value of 0.79 (95% CI 0.79–0.80), followed by preictal spikes with an AUC of 0.71 (95% CI 0.70–0.73), first‐postictal spikes with an AUC of 0.70 (0.69–0.70), and seizure‐onset + early‐spread with an AUC of 0.65 (95% CI 0.64–0.66). The SOZ, usually considered as the primary proxy of the p‐EZ, achieved an AUC of 0.64 (95% CI 0.63–0.65), followed by the induced SOZ (AUC 0.62; 95% CI 0.61–0.63), and the postictal zone (AUC 0.61; 95% CI0.60–0.62). Additionally, except for the FA zone, primary irritative zone, and the last‐preictal spike zone, which achieved AUC values <0.60, seizure‐free patients had a higher overlap ratio between each single feature and the resected channels compared with nonseizure‐free patients (Fig 3).

Figure 3.

Performance of single/top 5 spatial co‐occurrence features. Boxplot distributions showing the median and interquartile range of the overlap ratio between each single feature (A), and top 5 spatial co‐occurrence features and the resected channels (B). The distribution of each feature is also illustrated. The red color represents seizure‐free patients (Engel Ia), whereas the blue color represents nonseizure‐free patients (Engel Ib–IV). Each dot represents the overlap ratio between the feature of interest and the resected channels for each patient. The effect size is indicated (Cliff's d), and the statistical significance (Wilcoxon rank‐sum test) is reported with a p value <0.05 after Bonferroni–Holm correction. ∩ = spatial co‐occurrence. [Color figure can be viewed at www.annalsofneurology.org]

Spatial Co‐Occurrence

We investigated all possible spatial co‐occurrences among the 10 features, and observed that the performance of co‐occurrence plateaued with 5 features, with no further improvement observed with additional features. This represents an analysis of 637 possibilities. The highest performance was obtained with the co‐occurrence of gamma spikes, preictal spikes, seizure onset + early spread, and first postictal spikes, as well as the co‐occurrence of gamma spikes, preictal spikes, seizure onset + early spread with an AUC of 0.83 (95% CI 0.83–0.84). This was closely followed by the co‐occurrence of: gamma spikes, seizure onset + early spread and first postictal spikes; gamma spikes, preictal spikes, and first postictal spikes; and gamma spikes and preictal spikes with an AUC of 0.82 (95% CI 0.82–0.84; Table S2). Among these top 5 features, which demonstrated almost identical classification performance, no statistical difference was found (p > 0.05). Considering the number of features required to achieve a high performance (AUC >0.80) within this top 5 features, the co‐occurrence of gamma spikes and preictal spikes emerged as the optimal predictive feature (Fig 4), as only 2 features are needed, whereas more features are required in the remaining top 5 features. Finally, this optimal predictive feature performed statistically significantly better than the SOZ as proxy of the p‐EZ (p = 0.002).

Figure 4.

Selection of the optimal predictive feature. The performance of single and all possible spatial co‐occurrences of 5 features from the 10 features (637 possibilities) in surgical outcome prediction is illustrated. Of note, the performance reaches saturation when the area under the receiver operating characteristic curve (AUC) value plateaus. This indicates that adding more features beyond a certain combination does not improve performance. Upon evaluating features with an AUC value >0.80, the spatial co‐occurrence of gamma spikes and preictal spikes (C) was the first feature with only 2 features achieving such a high performance, establishing it as the optimal visually predictive feature. Its performance was almost identical to the spatial co‐occurrence of 4 features (E), which had the highest AUC value. No statistical difference (DeLong test) was observed between A, B, C, D, E. A represents the best single feature: gamma spikes; B denotes the best spatial co‐occurrence of 5 features: gamma spikes ∩, preictal spikes ∩, seizure onset + early spread ∩, first postictal spikes ∩, and postictal suppression‐slow wave activity. C signifies the best spatial co‐occurrence of two features: gamma spikes ∩ preictal spikes; D denotes the best spatial co‐occurrence of 3 features: gamma spikes ∩, preictal spikes ∩, and seizure onset + early spread; and E represents the best spatial co‐occurrence of 4 features: gamma spikes ∩, preictal spikes ∩, seizure onset + early spread ∩, and first postictal spikes. It is important to highlight that the spatial co‐occurrence of 5 features (B) did not increase the performance. Performance of B was lower than the performance of C (2 features), D (3 features), and E (4 features). ∩ = spatial co‐occurrence. [Color figure can be viewed at www.annalsofneurology.org]

External Validation

To assess the generalizability of our results, we computed the performance of the identified optimal surgical predictive feature in 2 external cohorts: CHUNA (50 patients; 27 women [54%]; 13 [26%] nonlesional MRI, 31 [62%] seizure‐free) and HUP (27 patients; 11 [41%] women, 8 sex missing data; 12 [44%] nonlesional MRI, 9 missing MRI results; 7 [26%] seizure‐free).

Similar to the derivation cohort, the spatial co‐occurrence of gamma spikes and preictal spikes correlated with postsurgical outcomes in both external cohorts. The performance of this co‐occurrence achieved an AUC of 0.76 in the CHUNA and an AUC of 0.78 in the HUP datasets. When we applied the optimized threshold of the optimal feature derived from the derivation cohort, corresponding to the highest balanced accuracy (80.8%), to the external cohorts, we found that sensitivity, specificity, and BA were 53.3%, 85%, and 69.2% in the CHUNA, and 71.4%, 75%, and 73.2% in the HUP datasets. The balanced accuracies in the external datasets were similar to the one in the derivation cohort after applying specific dataset‐optimized threshold (Fig 5). Finally, to ensure the reproducibility of our findings, we extracted the 4 features with the best performance (AUC >0.80) in the derivation cohort (Table S2). In the CHUNA dataset, the highest performance was achieved with the co‐occurrence of gamma spikes, preictal spikes, and SOZ + early spread (AUC 0.77), followed by the co‐occurrence of gamma spikes and preictal spikes (AUC 0.76). In the HUP dataset, the highest performance was with the co‐occurrence of gamma spikes and preictal spikes (AUC 0.78; Table S3).

Figure 5.

Performance results of the optimal predictive feature in the derivation and validation cohorts. (A) Boxplot distributions showing the median and interquartile range of the overlap ratio between the optimal predictive feature of surgical outcome (spatial co‐occurrence of gamma spikes and preictal spikes), and the resected channels in the derivation cohort (MNI + CHUGA) and in the two external validation cohorts (CHUNA, HUP). Single subject data corresponds to points on the left of the boxplot, and the distribution of the data on the right side. The red color represents seizure‐free patients (Engel Ia), whereas the blue color represents non‐seizure‐free patients (Engel Ib–IV). The effect size is indicated (Cliff's d), and statistical significance is denoted by a star for a p value less than (red star) or equal to (green star) 0.05 after Bonferroni–Holm correction. Before the Bonferroni–Holm correction, all the p values were <0.05. (B) The performance of the spatial co‐occurrence of gamma spikes and preictal spikes in CHUNA and HUP datasets after applying the optimized threshold derived from the derivation cohort (MNI + CHUGA) and specific dataset‐optimized threshold. Even with the optimized threshold from the derivation cohort, high balanced accuracy values were obtained in the 2 external datasets, demonstrating the generalizability of our results. BA = balanced accuracy; CHUGA = Grenoble Alpes University Hospital Center; CHUNA = Nancy University Hospital Center; HUP = HUP iEEG Epilepsy Dataset (open dataset); MNI = Montreal Neurological Institute. [Color figure can be viewed at www.annalsofneurology.org]

Interrater Reliability

We calculated the IRR agreement across 6 experts in assessing visually the optimal predictive feature from the derivation cohort. We observed an IRR of 0.63 (SD 0.16) for gamma spikes and 0.65 (SD 0.18) for preictal spikes, with a mean PA of 96% (SD 2%) for gamma spikes and 92% (SD 2%) for preictal spikes. The PA for the SOZ was 95% (SD 2%), and the IRR was 0.61 (SD 0.15). To confirm that the sample of 20 patients was sufficient for assessing IRR, we performed a cumulative analysis with increasing sample sizes (5, 10, 15, 20). IRR values were comparable across all sizes, indicating that adding more patients had minimal impact (Fig S10).

For the subgroup analysis, the pooled cohort included 150 patients for epilepsy type (62 TLE), 176 for MRI findings (88 lesional), 176 for the spatial extent of the SOZ (110 focal as assessed in ³⁸ ), and 138 for pathology (39 FCD). The performance of the spatial co‐occurrence of gamma spikes and preictal spikes achieved an AUC of 0.75 in the TLE group (29 seizure‐free patients) and 0.76 in the extra‐TLE group (Fig S11). For MRI findings, this feature achieved an AUC of 0.75 in the lesional group (36 seizure‐free) and 0.76 in the nonlesional group (88 patients with 43 seizure‐free; Fig S12). Regarding the postoperative pathology, the feature achieved an AUC of 0.67 in the FCD type II group (39 patients with 20 seizure‐free), 0.75 in the hippocampal sclerosis group (14 patients with 7 seizure‐free), 0.94 in the gliosis/scar group (31 patients with 16 seizure‐free), 0.76 in the other malformations of cortical development group (45 patients with 20 seizure‐free), and 0.75 in the low‐grade epilepsy‐associated neuroepithelial tumor + vascular malformation group (9 patients with 5 seizure‐free; Fig S13). For the spatial extent of the SOZ, this feature achieved an AUC of 0.75 in the focal group (110 patients with 60 seizure‐free) and 0.76 in the nonfocal group (66 patients with 19 seizure‐free; Fig S14). Finally, to ensure the reproducibility of our findings, we extracted the 4 features with the best performance (AUC >0.80) in the derivation cohort (Table S2), and we assessed their performance in predicting the surgical outcome according to the different clinical subgroups. We found that the co‐occurrence of gamma spikes and preictal spikes emerged as the optimal feature in most of them (Table S4).

Discussion

In this multicentric international cohort study, we report findings from 1 of the largest SEEG series to date, comparing the performance of various visual electrophysiological features in predicting surgical outcomes for patients with drug‐resistant focal epilepsy. First, we found that the spatial co‐occurrence of gamma spikes and preictal spikes was in the top 5 predictive features. When balancing the trade‐off between achieving high performance and minimizing the number of required features, the co‐occurrence of gamma spikes and preictal spikes was the first predictive feature, attaining an AUC >0.80 with just 2 features, thus establishing it as the optimal predictive feature. Second, this optimal feature showed high performances in classifying patients in external datasets, underscoring the robust value of this feature. Third, epileptologists can reliably visually identify these features, providing a practical advantage over challenging computational approaches. This insight holds significant implications for improving clinical efficiency, particularly given the time‐consuming nature of SEEG, and could guide neurosurgeons in determining the optimal surgical resection.

Gamma Spikes Emerges as the Optimal Single Predictive Feature

At the single‐feature level, our results are consistent with previous literature showing that there was a relationship between the gamma spike zone, SOZ, SOZ‐induced zone, and the surgical outcome. ⁹ , ¹³ Interestingly, the visual analysis of gamma spikes led to the superiority of this feature in predicting surgical outcome; a superiority that had been found with computational approaches. ¹¹ However, when considering FA, we found no association with surgical outcome, although a trend was observed. This may be related to our restrictive feature definition. Notably, when using a longer time window (5 seconds vs 1 second), as in computational approaches, seizure‐free patients showed a higher FA overlap ratio than nonseizure‐free patients (p = 0.03 after correction for multiple comparisons, Cliff's d = 0.25, AUC 0.62; Figure S15 in the Supplementary Material). This aligns with results that have demonstrated the weakness of this association, ⁶ particularly when FA was used alone, due to its limited specificity to the p‐EZ. ⁶ , ²¹ , ²⁴ Computational approaches focused on the FA, such as the fingerprint ²¹ or epileptogenicity index, ¹⁰ have shown additional value to conventional analysis. However, those approaches are not adopted in most epilepsy centers, as they require technical implementation and multicenter validation. Even with these methods, FA alone overestimated the p‐EZ, with the best performance achieved when combined with other features, including preictal spikes. ²¹

Spatial Co‐Occurrence of Gamma Spikes and Preictal Spikes Optimally Predicts Surgical Outcome

When examining the spatial co‐occurrence of features, the identified optimal predictive feature included preictal spikes. This aligns with prior research, where the authors emphasized the role of preictal spiking in the typical pattern of the p‐EZ, ²¹ characterized by preictal spikes, narrow‐band FA, and concurrent low‐frequency suppression. However, the authors focused on a subset of patients with FA and included only ictal features. The present study incorporates a comprehensive assessment involving interictal, preictal, ictal, postictal, and cortical stimulation SEEG data, enabling a direct comparison of clinically relevant features. When only patients with FA were considered (n = 79), the co‐occurrence of gamma spikes and preictal spikes also emerged as the optimal predictive feature (AUC 0.82), achieving equal performance to the co‐occurrence of FA (SOZ + early spread) combined with preictal and gamma spikes. This finding highlights the necessity of incorporating interictal feature (gamma spikes) to improve the surgical predictive performance of preictal/ictal features. Further studies incorporating the background suppression would be interesting to compare the performance of our optimal feature with the p‐EZ fingerprint.

Our approach, which involves considering the spatial co‐occurrence of features to preserve their specificity, aligns with the goal of identifying the minimum brain area requiring resection for achieving seizure freedom. ³⁹ Our overlap ratio metric offers the advantage of considering both sensitivity and precision, ¹⁵ informing us on which areas to resect while preserving brain regions where resection is unnecessary for achieving seizure freedom. ⁴⁰ This region likely corresponds to the spatial co‐occurrence of gamma spikes and preictal spikes.

Another important aspect of our results is that demonstrating the superiority of the spatial co‐occurrence between gamma spikes and preictal spikes over the SOZ—principal proxy of the p‐EZ—in predicting surgical outcome could present an alternative target in procedures such as laser ablation or SEEG‐guided radiofrequency thermocoagulation. ⁴¹ This potential improvement could enhance response rates. However, further studies are needed to confirm this possibility.

Generalizability of Our Results

Another point to consider is the performance of the optimal predictive feature in the external datasets. Although slight performance variations were expected across external datasets due to differences in epilepsy types, etiologies, surgical approaches, and SEEG implantations, the optimal predictive feature maintained high balanced accuracy in both external datasets when using a fixed threshold from the derivation cohort. This shows that our results could be generalized across different sites. Additionally, our analysis based on epilepsy type, MRI findings, spatial extent of the SOZ, and pathology showed consistent performance of this optimal predictive feature across these subgroups, except in the FCD type II group, where the performance was slightly lower. This may be due to the limited number of patients in this group (39/138), missing pathology data (12/150 from MNI, CHUGA, and CHUNA) potentially corresponding to this subtype, and incomplete resection of the optimal feature. When pathology was grouped into 2 categories (malformations of cortical development and others), gamma spikes and preictal spikes effectively discriminated Engel Ia from Engel Ib–IV patients. Altogether, our results suggest that this optimal feature may reflect optimal SEEG sampling rather than a “confined” p‐EZ.

Finally, our study demonstrated that gamma spikes and preictal spikes can be easily assessed through conventional analysis, unlike computational approaches, which are challenging to implement. This visual analysis showed a substantial agreement among experts.

Potential Pathophysiological Mechanisms

The key role of the preictal activity in ictogenesis is well established. ²¹ During this period, several studies in human epilepsy have observed a gradual increase in fast‐discharging interneuronal activity, whereas pyramidal activity persists. ⁴² , ⁴³ , ⁴⁴ This progressive increase in interneuronal activity marks the transition from the interictal to the ictal state and may represent the “true” onset of seizures. ⁴⁵ This buildup of activity or isolated synchronized spikes, along with intensified interneuronal firing, leads to a progressive increase in gamma activity with longer duration, amplitude, and power. ⁴² , ⁴⁵ These events are like microseizures, but are not detected on EEG and require local field potential recordings. Sustained low‐voltage fast activity on SEEG may then already indicate seizure spread, emphasizing the limitations of using low‐voltage fast activity alone as a p‐EZ marker. ⁶ , ²¹ Although preictal activity was evaluated in this study, we did not focus on the transition period, as it may be challenging to identify through standard visual analysis. As in previous studies, ²¹ our preictal definition may extend beyond the duration of this transition period. The spatial co‐occurrence of preictal and gamma spikes may correspond to this transition zone, explaining the reproducibility of our results whatever the seizure onset pattern, as attested by our results for epilepsy types, MRI findings, spatial organization of the SOZ, and etiology.

Like interictal HFO (80–500 Hz) spikes, gamma spike activity (30–100 Hz) may reflect increased interneuron firing and a possible “reversal” of inhibitory control during ictogenesis, facilitating seizure initiation. ⁴⁴ , ⁴⁵ Different interneuron populations may generate distinct oscillations, with gamma frequencies playing a crucial role in ictogenesis. ⁴⁶

Strengths and Limitations

This study had several strengths. We comprehensively evaluated the clinical utility of various electrophysiological features, both individually and in their spatial co‐occurrence, using a sizable derivation cohort. We rigorously validated our findings using external datasets, thereby demonstrating the robustness and generalizability of our results. Moreover, by engaging multiple epileptologists trained across diverse epilepsy centers worldwide in feature extraction, we underscored the reliability and reproducibility of this approach across different clinical settings.

However, this study had certain limitations beyond its retrospective design. Only a minority of the patients in the derivation cohort (n = 17) were aged <18 years at the time of the SEEG investigation. Therefore, future studies focusing on children are needed to assess the performance of the identified predictive feature in this population. Although a visual analysis approach was used, as it remains the gold standard in SEEG analysis, quantitative methods could help reduce interexpert variabilities. Another point is the variability in thresholding required to achieve optimal performance across centers, which highlights the inherent challenge of setting universal thresholds for clinical practice. However, our optimal feature performed well, even with a fixed threshold, indicating the generalizability of our findings. Our findings are from a group‐level analysis. Individualized strategies are challenging due to variability in SEEG implantation, epilepsy types, etiologies, and surgical approaches. Larger cohort, data‐driven, and multimodal ⁴⁷ approaches are needed for patient‐specific strategies. Another point is that some resected channels were outside the p‐EZ. However, this is common in epilepsy surgery, as resections often extend beyond the p‐EZ. Finally, we did not include other markers, such as HFOs, spike‐ripples, or DC‐shift, for 2 main reasons. First, there are challenges associated with visual HFO identification and interpretation, along with time constraints. ⁹ , ⁴⁸ Second, assessing DC‐shift and HFOs requires specific sampling frequencies. ⁴⁹ We aimed to prioritize features universally recorded across all centers and that are easily identifiable visually.

Conclusion

The spatial co‐occurrence of gamma spikes and preictal spikes, assessed visually, is a robust predictive feature of surgical outcome. Such knowledge not only contributes to a deeper understanding of the underlying epileptic networks, but also holds significant implications for enhancing clinical efficiency. Prioritizing features can help streamline SEEG analysis by reducing its time‐consuming nature. Additionally, our findings could offer valuable insights for guiding neurosurgeons in determining the optimal extent of resection.

Author Contributions

Chifaou Abdallah: Conceptualization; data curation; formal analysis; funding acquisition; investigation; methodology; project administration; resources; software; validation; visualization; writing – original draft; writing – review and editing. John Thomas: Conceptualization; formal analysis; methodology; visualization; writing – review and editing. Olivier Aron: Data curation; writing – review and editing. Tamir Avigdor: Formal analysis; writing – review and editing. Kassem Jaber: Visualization; writing – review and editing. Irena Doležalová: Formal analysis; writing – review and editing. Daniel Mansilla: Formal analysis; writing – review and editing. Päivi Nevalainen: Formal analysis; writing – review and editing. Prachi Parikh: Formal analysis; writing – review and editing. Jaysingh Singh: Formal analysis; writing – review and editing. Sandor Beniczky: Formal analysis; writing – review and editing. Philippe Kahane: Data curation; writing – review and editing. Lorella Minotti: Data curation; writing – review and editing. Stephan Chabardes: Data curation; writing – review and editing. Sophie Colnat‐Coulbois: Data curation; writing – review and editing. Louis Maillard: Data curation; writing – review and editing. Jeff Hall: Data curation; writing – review and editing. Francois Dubeau: Data curation; writing – review and editing. Jean Gotman: Conceptualization; data curation; writing – review and editing. Christophe Grova: Conceptualization; funding acquisition; investigation; methodology; supervision; writing – review and editing. Birgit Frauscher: Conceptualization; data curation; funding acquisition; investigation; methodology; resources; supervision; writing – review and editing.

Potential Conflicts of Interest

Nothing to report.

Supporting information

DATA S1. Supporting Information.

Data Availability

The data that support the findings of this study are available upon reasonable request and if in accordance with the respective research ethics boards policies.