EEG Ictal Power Dynamics, Function-Structure Associations, and Epilepsy Surgical Outcomes

Abstract

Background and Objectives

Postoperative seizure control in drug-resistant temporal lobe epilepsy (TLE) remains variable, and the causes for this variability are not well understood. One contributing factor could be the extensive spread of synchronized ictal activity across networks. Our study used novel quantifiable assessments from intracranial EEG (iEEG) to test this hypothesis and investigated how the spread of seizures is determined by underlying structural network topological properties.

Methods

We evaluated iEEG data from 157 seizures in 27 patients with TLE: 100 seizures from 17 patients with postoperative seizure control (Engel score I) vs 57 seizures from 10 patients with unfavorable surgical outcomes (Engel score II–IV). We introduced a quantifiable method to measure seizure power dynamics within anatomical regions, refining existing seizure imaging frameworks and minimizing reliance on subjective human decision-making. Time-frequency power representations were obtained in 6 frequency bands ranging from theta to gamma. Ictal power spectrums were normalized against a baseline clip taken at least 6 hours away from ictal events. Electrodes' time-frequency power spectrums were then mapped onto individual T1-weighted MRIs and grouped based on a standard brain atlas. We compared spatiotemporal dynamics for seizures between groups with favorable and unfavorable surgical outcomes. This comparison included examining the range of activated brain regions and the spreading rate of ictal activities. We then evaluated whether regional iEEG power values were a function of fractional anisotropy (FA) from diffusion tensor imaging across regions over time.

Results

Seizures from patients with unfavorable outcomes exhibited significantly higher maximum activation sizes in various frequency bands. Notably, we provided quantifiable evidence that in seizures associated with unfavorable surgical outcomes, the spread of beta-band power across brain regions is significantly faster, detectable as early as the first second after seizure onset. There was a significant correlation between beta power during seizures and FA in the corresponding areas, particularly in the unfavorable outcome group. Our findings further suggest that integrating structural and functional features could improve the prediction of epilepsy surgical outcomes.

Discussion

Our findings suggest that ictal iEEG power dynamics and the structural-functional relationship are mechanistic factors associated with surgical outcomes in TLE.

Introduction

Temporal lobe epilepsy (TLE) is recognized as the most common form of focal epilepsy.¹ Resective and ablative surgeries represent the primary therapeutic approach for patients with unilateral drug-resistant TLE with a 50%–65% success rate.^2-4 However, it remains unclear which patients with TLE will respond favorably to surgical intervention. Therefore, elucidating the underlying mechanisms of surgical outcomes in TLE could improve the prediction of postoperative clinical trajectories.

Intracranial EEG (iEEG) is commonly used in clinical epilepsy care to evaluate seizure ictal onset, providing high temporal resolution with comprehensive—albeit variable—spatial resolution based on sampled structures. The evaluation of ictal iEEG can help identify the epileptogenic zone (EZ), which is crucial for the success of epilepsy surgery.⁵ An important focus of iEEG assessment is the identification of seizure onset patterns,⁶ which provides insight into mechanisms supporting seizure generation,^7,8 can reveal the targets of surgery and guide the surgical outcomes.^9,10 For example, early research in iEEG recognized “fast” or “rapid” activity at ictal onset as a key feature of the EZ.¹¹ Other frequency bands, including high-frequency oscillations and slow waves at onset, have also been correlated with the EZ.^12,13

In recent years, methods based on frequency-specific power characteristics at ictal onsets, such as the epileptogenicity index¹⁴ and EZ “fingerprints,”¹⁵ have improved our understanding of EZ localization. However, persistent surgical failure points to inadequate identification of the EZ. Clinical observations have demonstrated that a more rapid and extensive propagation of seizures limits the precise localization of the EZ, which can hinder surgical planning.¹⁶ Therefore, transitioning the focus from solely the identification of the EZ based on which channels are involved during a seizure toward a comprehensive understanding of seizure dynamics in time-space is essential. A critical but unresolved question remains in epilepsy research: Does power activation in different frequency bands during the seizure period relate to surgical outcomes? A comprehensive quantitative analysis of the cumulative distribution of seizure-period activity combining spatial and temporal dimensions is necessary to address this question.

In addition, the ictal spatial propagation of seizures has been acknowledged as associated with the structural organization of brain. Our group previously showed that different ictal propagation axes are associated with distinct patterns of white matter organization,¹⁷ with other studies demonstrating that ictal propagation relies on the underlying brain structural network.¹⁸ Clearly, investigating the relationship between iEEG power propagation and structural features, such as diffusion tensor imaging (DTI) metrics, holds great potential for providing valuable insight into the organization of the epileptogenic network and optimizing surgical planning.

This study examined the spatiotemporal dynamics of seizures in TLE from iEEG analysis and its association with surgical outcomes. Specifically, we hypothesized that higher power spreading rapidly to more brain regions would be associated with unfavorable postoperative outcomes. To test this, we used time-frequency analysis to investigate the characteristics of presurgery iEEG power in patients with unilateral TLE who had favorable or unfavorable surgical outcomes through the following: (1) employing a realistic anatomical mapping technique to accurately localize the power activation, (2) providing quantitative measurements of the extent of seizure propagation within the brain, and (3) exploring the structure-function relationship by integrating fractional anisotropy (FA) values derived from DTI.

Methods

Participants

We studied 27 patients with drug-resistant TLE from 2 Level IV academic centers: the Medical University of South Carolina (MUSC: n = 16, 7 Females) and the Hospital University of Pennsylvania (HUP: n = 11, 6 Females). Table 1 provides the summary of clinical information, and inclusion and exclusion criteria are detailed in eMethods. Based on standard-of-care practices, 10 patients (8 from MUSC, 2 from HUP) underwent anterior temporal lobectomy (ATL) and 17 patients (8 from MUSC, 9 from HUP) underwent laser interstitial thermal therapy (LITT). This research was performed through retrospective review, and it was approved by the institutional review board of the MUSC and HUP with a waiver of informed consent as well as according to the Declaration of Helsinki.

Table 1.

Clinical Information

ID	Age at SZ onset	SZ frequency (per month)	Visible MRI temporal lesion	No. of IEEG SZ clips	IEEG SZ duration (s)	IEEG targets (no. of contacts and sides)	Age at first surgery	Surgery type	Engel score	Follow-up time
1	22	2	No	16	67.88 ± 38.83	70 L, 110 R	50	R ATL	I	1.5 y
2	22	4	No	6	88.17 ± 26.87	30 L, 40 R	28	R ATL	I	1 y
3	25	10	No	4	78.00 ± 5.60	70 L	55	L LITT	I	1 y
4	30	3	No	1	49	70 L, 70 R	39	R ATL	I	5 y
5	21	30	Yes	4	32.75 ± 17.56	70 L	31	L LITT	III	3 y
6	23	1	No	7	81.29 ± 8.73	70 L, 70 R	33	R ATL	I	4 y
7	4	4	Yes	2	157.50 ± 44.55	40 L, 40 R	19	R ATL	I	6 y
8	3	4	No	9	107.56 ± 36.46	80 L, 70 R	29	L LITT	I	6 y
9	35	2	Yes	9	76.67 ± 17.54	70 L, 70 R	41	L LITT	I	4 y
10	40	8	No	3	122.67 ± 78.50	40 L	42	L ATL	I	5 y
11	8	30	No	9	116.44 ± 46.94	76 R	21	R ATL	I	3 y
12	17	4	No	2	97.00 ± 4.24	66 L, 66 R	27	L LITT	I	3 y
13	31	8	No	9	52.67 ± 11.30	70 L, 70 R	33	R ATL	I	7 y
14	19	6	Yes	3	120.67 ± 91.68	20 L, 40 R	44	R LITT	II	7 y
15	5	4	Yes	3	118.33 ± 10.21	60 L, 62 R	30	R LITT	IV	2 y
16	16	1	Yes	4	114.75 ± 20.60	70 L	21	L LITT	III	1 y
17	22	20	No	8	30.75 ± 4.27	114 L, 12 R	27	L LITT	I	2 y
18	29	—	No	5	101.20 ± 34.60	60 L, 52 R	38	L LITT	IV	2.5 y
19	26	—	Yes	5	230.20 ± 173.69	74 L, 20 R	47	L LITT	I	4 y
20	16	—	No	5	77.80 ± 21.14	81 L, 89 R	26	L LITT	III	4 mo, then a second surgery
21	16	0.1	Yes	3	162.67 ± 105.22	83 L, 59 R	31	L LITT	III	1 y
22	18	—	No	18	56.95 ± 19.12	94 R	24	R LITT	II	1 y
23	16	2	No	7	76.86 ± 39.21	154 L	28	L LITT	IV	3 mo, then a second surgery
24	6	3	No	5	75.60 ± 28.99	154 R	47	R ATL	II	1 y
25	20	—	Yes	2	41.50 ± 13.44	126 L	40	L LITT	I	1 y
26	16	3	No	4	102.00 ± 75.60	142 L	37	L LITT	I	1 y
27	24	—	No	4	92.5 ± 11.82	104 L, 12 R	44	L ATL	I	3.5 y

Abbreviations: ATL = anterior temporal lobectomy; LITT = laser interstitial thermal therapy; SZ = seizure.

Surgical outcomes were classified according to the Engel Epilepsy Surgery Outcome Scale¹⁹ at least 1 year after surgery (Table 1). Patients were hence identified as having either favorable post-operative seizure control (seizure free, SF, Engel I, 17 patients/9 Females, 100 seizure clips) or unfavorable surgical outcomes (non-seizure free, NSF, Engel II–IV, 10 patients/3 Females, 57 seizure clips). We found no significant differences in age at surgery between NSF (32.0 ± 8.5 years) and SF groups (36.0 ± 10.1 years; Wilcoxon rank-sum p = 0.327), in presurgical seizure burden (seizure frequency x disease duration/months NSF: 1,206.0 ± 1,273.1 vs SF: 1,044.0 ± 1,377.6; p = 0.881), and in seizure event durations (SF: 88.6 ± 62.7 seconds vs NSF: 81.2 ± 46.7 seconds; p = 0.729).

Data Acquisition

Intracranial EEG

Patients underwent iEEG implantation following standard in-house neurosurgical protocols. The implantation targets were defined by the clinical care team based on a combination of semiology and history, neuroimaging, and scalp neurophysiologic data. Patients had monitoring with sampling of at least 1 temporal lobe, and the remainder of iEEG targets included both extra-mesial temporal regions (e.g., anterior temporal region for encephalocele) and extratemporal (e.g., medial frontal, insular) according to each case's hypotheses. The details of acquisition parameters of iEEG recordings at the 2 centers are described in the eMethods. The identification of seizure onsets was illustrated in eFigure 1.

Brain Imaging Acquisition

Preimplantation brain imaging, including 3T T1-weighted images and diffusion-weighted images, was obtained through a 3T MRI protocolized for epilepsy. After electrode implantation, spiral CT images and postimplantation T1-weighted images were obtained clinically for electrode localization (acquisition parameters are described in the eMethods).

Data Processing

iEEG Data Processing

To investigate the spatiotemporal characteristics of seizure dynamics, we investigated frequency-specific power dynamics. The significance of power activation during seizures in predicting surgical outcomes is illustrated in Figure 1. The observed patterns of power spreading in 2 example patients indicate a potential association between surgical outcomes and both the timing and spatial distribution of power within specific frequency bands. Specifically, we proposed a priori that more rapid (i.e., faster distribution) and broader (i.e., more spatially distributed) seizure activation would be associated with unfavorable postsurgical outcomes (i.e., NSF status).

Figure 1. Importance of Frequency-Specific Power Spreading in Surgical Outcomes Differentiation.

This figure illustrates the significance of frequency-specific power spreading in distinguishing surgical outcomes. The data from 2 example patients are presented. One patient belongs to the group with good surgical outcomes (Engel score = I, complete seizure-freedom, SF), while the other patient belongs to the group with unfavorable surgical outcomes (Engel score ≥II, nonseizure freedom, NSF). The left schemes show the electrode sampling on 3D brain models for the 2 example patients. The spread of beta power across electrodes over time is shown in the middle. The power activation time maps and the spread patterns are then represented by the 3D brain schemes with activation time in inverted “hot” colormap. The power was calculated as the beta power normalized by an interictal baseline obtained at least 6 hours away from the seizure activity. Darker electrode colors indicate areas where power activation required more time (i.e., was slower). The SF patient exhibits larger areas in darker colors, suggesting a slower activation process. By contrast, the electrodes of the NSF patient predominantly show lighter electrode colors, indicating that most areas were activated before 5 seconds. On the 3D brain model, blue areas highlight regions activated within 5 seconds after seizure onset. The sample NSF patient shows a broader activation area than the sample SF patient. Taken together, this example pair may indicate that the spatiotemporal power dynamics during a seizure can potentially influence surgical outcomes.

To enable an unbiased, automatic, and systematic analysis of seizure dynamics, as well as to capture information within a realistic anatomical space, we have improved on a method based on prior anatomical mapping.⁶ Figure 2 provides a visual summary of the methods to estimate power dynamics. We used custom MATLAB scripts using functions of commonly used EEG analysis toolboxes (i.e., Brainstorm,²⁰ Fieldtrip,²¹ Statistical Parametric Mapping²²). The methodology is detailed in the Data Processing section of the eMethods. Briefly, as shown in Figure 2A, after preprocessing, we first constructed seizure-activated power representations for 6 frequency bands, including theta (4–7 Hz), alpha (8–12 Hz), beta (13–30 Hz), gamma1 (35–55 Hz), gamma2 (65–115 Hz), and gamma3 (125–175 Hz). These representations were created for multiple electrodes in each clip. Next, we mapped the electrode power representation onto the anatomical space to enable the visualization of seizure power propagation within the anatomical framework (Figure 2B). For subsequent statistical analyses, we constructed seizure-activated representation of the brain regions (Figure 2B). Specifically, the individual preimplantation T1 images were segmented and parcellated using the Atlas of Intrinsic Connectivity of Homotopic Areas (AICHA), which provides detailed definitions of brain regions and facilitates the grouping of iEEG electrodes.²³ For each region, we grouped the smoothed power values where electrodes landed.

Figure 2. Schematic Representation of the Process for Power Mapping Analysis.

(A) Constructing iEEG time-frequency power representation: The example iEEG data are a 40-second raw intracranial EEG recorded during a seizure, displayed using a bipolar montage. Electrode leads were marked by colors and displayed on 2D brain image. Time-frequency analysis was then applied to calculate the power spectrum representation of the sample iEEG data across a frequency range of 1–200 Hz. Subsequently, power is grouped into 6 distinct frequency bands for further analysis. As a result, we obtain the spread of power over time in 6 frequency bands across electrodes, which are represented by 6 time-electrode images in the figure. The power spectrum in 6 frequency bands is normalized to baseline, which is a 3-minute resting recording taken 6 hours away from any seizure. This normalization process results in z-scores. Any z-scores with FDR-corrected p-values larger than 0.05 were set to zero. (B) Transformation: Beta power is used as an example to illustrate the interpolation and smoothing of the electrodes' power spectrum onto the anatomical space by mapping electrodes onto the T1-weighted MRI. The 2D axial slices demonstrate the propagation of beta power over time. For more advanced statistical analysis, the anatomical power representation can be regionally grouped based on the atlas definition, where electrodes are positioned (AICHA atlas used in this study). (C) Six lines (1 for each frequency band) represent the spatial spread of power across regions starting 5 seconds before seizure onset. Owing to variances in the number of regions sampled between individuals, attributable to their respective iEEG electrode implantation plans, we computed the proportions of seizure-activated regions to the total number of regions sampled by iEEG electrodes rather than relying on the absolute count of seizure-activated regions. (D) 3D view of electrode locations of all patients presented in MNI template. AICHA = Atlas of Intrinsic Connectivity of Homotopic Areas; FDR = false discovery rate; iEEG = intracranial EEG; MNI = Montreal Neurological Institute.

Subsequently, the temporal progression of seizure-activated regions was quantified to understand the spatiotemporal dynamics. To accommodate the variation in the number of brain regions across patients, the seizure activation size was computed as the ratio of activated to the total number of regions sampled per patient. The changes in activation size across all 6 frequency bands are depicted in Figure 2C, illustrating the variability of activated regions across different frequency bands. The minimum and maximum proportions of activated regions were then identified in a predefined period, reflecting the range of activation observed during that interval. We determined the spreading rate of seizure propagation by dividing the difference between maximum and minimum activation sizes by the time interval, indicating the extent to which the seizure spread from focal regions to more widespread areas. We initially used a period from 5 seconds before to 10 seconds after the seizure onset for statistical analyses. To demonstrate the stability of results beyond the choice of specific time windows, we investigated alternative estimation periods, as detailed in the Results section.

Figure 2D provides an overview of electrode locations across patients, and the specific implantation strategies can be found in Table 1. Because it has been previously reported that seizure spread can be affected by number of sampled regions,²⁴ we tested this in our sample. On average, electrodes sampled 33.6 ± 13.3 regions per patient (according to the AICHA atlas). We did not find statistically significant difference in the number of regions (rank-sum, p = 0.514) sampled between the NSF group (33.0 ± 13.4) and the SF group (35.1 ± 13.2). In addition, to determine whether the strategy of electrode implantation influenced the propagation of seizure power activities, potentially biasing the comparisons between the SF and NSF groups, we analyzed the differences in electrode distribution. The results shown in eFigure 2 and eTable 1 indicated no significant differences in skewness, kurtosis, or distribution volume of electrode sampling between the 2 groups.

DTI Processing

For DTI processing, we extracted mean FA values corresponding to the regions defined by the AICHA parcellation. The FA value of a brain region represents the general amount of diffusion asymmetry within a region, and this metric has been shown to be among the most sensitive to structural network changes in epilepsy.^25,26 The processing steps followed the standard pipeline implemented in DSI Studio, and the details are described in the section of DTI Processing in eMethods.

Structural-Functional Association

A noteworthy inquiry pertains to the association between alterations in power and the underlying structural organization of brain regions. To investigate this relationship, we examined the correlation between the structural integrity of brain regions, as measured by FA derived from DTI, and the functional aspect represented by power values obtained from iEEG. We computed the Spearman correlation coefficient between FA values and power across sampled brain regions.^18,27 Given that power undergoes temporal changes throughout the course of seizure propagation, we performed the correlation analysis at each time point along the seizure progression timeline. This allowed us to examine how the relationship between FA and power evolves during different stages of the ictal event.

Statistical Analyses

We applied Wilcoxon rank-sum tests to evaluate the differences between SF and NSF groups and Wilcoxon sign-rank tests to compare differences across time points within a seizure. The nonparametric tests were chosen to account for potential imbalances in the cohort sizes of the 2 groups, which is expected given typical epilepsy surgery outcomes. False discovery rate (FDR) correction was applied to control for multiple comparisons. To ensure sufficient statistical power, comparisons were performed on seizure clips, which were divided as derived from SF or NSF participants. Owing to the exclusion of 4 patients without high-quality DTI images, the statistical comparisons of structural-functional association were conducted in a cohort comprising 146 seizure clips from 16 SF patients and 8 NSF patients. To investigate the potential of structural and functional features in predicting surgical outcomes, we implemented binary logistic regression models on 3 primary factors: the average FA across regions, power spreading rate, and structural-functional association. The effect size of these models was evaluated using adjusted R² values, which give a measure of the proportion of sample variance explained by the models. Fisher R² to Z transformation was applied to measure the significance of differences between models. The performances of models were also assessed by comparing evaluation of area under the receiver operating characteristic curves (AUC), as well as the predictive accuracy obtained by leave-one-out cross-validation (LOOCV).

Data Availability

The data and scripts for this study are available on reasonable request from the corresponding author; however, restrictions may apply to ensure participant consent and anonymity.

Results

Figure 3A illustrates the dynamics in seizure activation size, calculated as the proportion of the seizure-activated region from 5 seconds before seizure onset to 30 seconds postseizure onset across all 6 frequency bands. An upward slope in the percentage of seizure-activated areas around the point of seizure onset was evident in both the SF group and the NSF group, while the differences between groups varied across frequency bands. To further compare the characteristics of power spreading in the early seizure onset stages, a specific time window from 5 seconds before to 10 seconds after the seizure onset was evaluated, as highlighted by the gray square in Figure 3A. Figure 3B presents a comparison of maximum activation size within this time window, demonstrating significant differences between the SF and NSF groups in the theta (FDR p = 0.001), alpha (FDR p = 0.006), beta (FDR p < 0.001), and gamma1 (FDR p = 0.061) bands. The significant differences in maximum seizure activation size between the groups suggest a broader involvement of brain areas in the NSF group. No significant differences were detected in the minimum activation size.

Figure 3. Power Spreading Dynamics at Seizure Onset.

(A) Error-shaded plots depict the escalation in the proportion of seizure-activated regions from the seizure onset in both groups. To statistically compare the power spreading characteristics in the early stages of seizure onset, we derived features within the time window (from 5 seconds before to 10 seconds after onset), denoted by the gray square in the plots. (B) Box plots providing a statistical comparison of spreading metrics, specifically the maximum and minimum proportions of seizure-activated regions, as well as the spreading rate (calculated as the difference between maximum and minimum proportions of activated regions dividing the corresponding time interval), within this specified time window. NSF = non–seizure-free; SF = seizure-free.

It is noteworthy that on comparing the spreading rates between the 2 groups, there was an accelerated seizure spread specifically within the beta bands in the NSF group as compared with the SF group (FDR p = 0.018). Moreover, we used a sequence of sliding windows to evaluate the differences in beta power spread between groups, ensuring that differences in the spread between groups were not an artifact of the specific choice of the estimated period. This approach also allowed us to explore the temporal window within which the spread of beta power can effectively distinguish surgical outcomes. These windows had their end of the estimation period ranging from 1 second to 30 seconds postseizure onset, as illustrated in Figure 4. Statistical analyses revealed that the differences in spreading rates between the NSF and SF groups started as early as 1 second and peaked at 11 seconds postseizure onset This discovery may indicate the significance of early power spread as a predictive factor for surgical outcomes.

Figure 4. Spreading Rate of Beta Power in Different Time Windows.

The top line plot presents the percentage of regions activated by seizures from 5-second before onset to 30-second postonset in beta frequency band. The end time of the estimation window for spreading rate calculation ranged from 1-second to 30-second postseizure onset. An error bar contrasts the spreading rates in each estimation window between groups with good and poor surgical outcomes using the rank-sum test. The error bar interval shown in the figure is 2 seconds to enable a clear representation, whereas in our analysis, the sliding window was processed at an interval of 1 second. The orange line symbolizes the statistic values (z-scores) of between-group statistical tests. The area shadowed by light orange highlight the significant differences post-FDR correction. FDR = false discovery rate; NSF = non–seizure-free; SF = seizure-free.

Figure 5 summarizes the assessment of the structural-functional association, as evaluated by the correlations between beta power and DTI-extracted FA values across regions at each time point during seizure propagation (Figure 5A). As shown by the sample seizure displayed in Figure 5B, we found a trend of increased association between beta power and FA values after seizure onset. The overall dynamics of structural-functional correlation coefficients (R values) in both the SF and NSF groups were initially estimated by correlating regional FA and iEEG beta power in all sampled regions, as shown in eFigure 3. Both SF and NSF groups revealed a significant increase in structural-functional associations immediately after seizure onset, and this association was stronger in the NSF group, beginning at 4 seconds postonset.

Figure 5. Assessment of Structural-Functional Association.

(A) Illustration of the method used to estimate the structural-functional association by correlating beta power and FA values across regions over time. Mosaic axial slides show the activation of beta power at 5 seconds postonset and the DTI-FA map of a patient example. The red square highlights an enlarged area on the right side. Power values and FA values were grouped in predefined ROIs corresponding to the electrode locations. The correlation between beta power and FA values across regions (S-F correlation) was then computed. Each dot in the scatter plot represents the position of an ROI (red area as an example in the figure) in the power-FA coordinate system. (B) The sequence of axial slices demonstrates the spatial distribution of activated power values at different time points. We calculated the power-FA correlations over time. This example illustrates an increase in the power-FA correlation after seizure onset. (C) Line plots show the dynamics of the correlation coefficient (R values) over time. In the left 2 plots, individual clips are represented by gray lines, while the average for the SF group and NSF group is indicated by blue and red lines, respectively. The areas shadowed by light orange in the 2 plots represented significant increase of structure-function association compared with preonset periods after FDR correction. The shaded error plot on the right indicates that the group with poor surgical outcomes exhibits a higher DTI-power association than the group with good surgical outcomes during ictal activities. The light-orange shadowed regions in this plot emphasize significant group differences post-FDR correction. The orange bar on line plot represents the effect sizes of group comparison along with time. The effect size for the group comparison was determined by dividing the statistical z score by the square root of the total number of samples. DTI = diffusion tensor imaging; FA = fractional anisotropy; FDR = false discovery rate; NSF = non–seizure-free; ROI = region of interest; SF = seizure-free.

However, given the smaller activation size in the SF group, the lower functional-structural correlation might be related to regions not activated by seizures. To ensure the validity of our analysis, we repeated the analysis focusing exclusively on correlations within activated regions and only when the count of seizure-activated regions exceeded approximately 25% of the overall average number of sampled regions across patients (i.e., at least 8 regions for correlation analysis). The results shown in Figure 5C demonstrated that our initial observation remains unchanged, even when excluding nonactivated regions. In addition, the differences between the 2 groups became apparent immediately after the seizure began, suggesting potential differences in the underlying network dynamics or pathophysiology associated with surgical outcomes.

Considering the significant relationship found between regional FA and beta power during seizures and its pronounced effect in the NSF group, we hypothesized that calculating the grand average FA value across all brain regions sampled by the electrodes could also reveal significant differences between groups. When comparing the average FA values between groups, we found a higher mean FA in the NSF group compared with the SF group (p < 0.001), which may suggest the potential use of this factor to predict surgical outcomes. We then examined how the structural and functional features aid in predicting surgical outcomes. Therefore, we constructed binary logistic regression models aimed at distinguishing surgical outcomes based on 3 factors: average FA values, beta power spreading rate, and the associations between regional FA and beta power, as illustrated in Figure 6. In Figure 6A, the model incorporating structural-functional association exhibited the largest effect size, as assessed by the adjusted R² values. In addition, when integrating all structural and functional features into a binary logistic model (Figure 6B), performance in distinguishing surgical outcomes significantly improved relative to models solely considering averaged FA values (p = 0.056) or beta power spreading values (p < 0.001). As shown in Figure 6C, the AUC values demonstrated that the model integrating structure and function features outperformed models only considering single factor. Furthermore, we used LOOCV to assess the predictive accuracy of the 4 models (Figure 6D). The outcomes by LOOCV demonstrated that models account for the factor of functional-structural associations (model 3: 73.4%, model 4: 73.1%) outperformed those using average DTI-FA values (model 1: 58.2%) or iEEG spreading rates (model 2: 62.1%) independently. The models in Figure 6 were constructed using a postonset time of 10 seconds. Comparable findings were obtained when considering the estimation time points around 10 seconds postonset (eFigure 4). Therefore, this finding may indicate the potential benefits of integrating both structural and functional characteristics in predictive models for predicting surgical outcomes.

Figure 6. Binary Logistic Regression Models for Predicting Surgical Outcomes.

(A) Three separate logistic regression lines are displayed, representing the predictions based on regional FA values, beta power spreading rate, and the associations between fractional anisotropy and beta power. (B) A binary regression model incorporating all 3 factors is presented. R² values are reported as an evaluation of the model's ability to explain the variance in the samples. (C) ROC curves for the 4 logistic regression models. AUC values are presented to show the performance of models. (D) A bar plot shows the accuracy of these models through the LOOCV approach. [t] represents the time point after seizure onset used for estimating the spreading rate of beta power activities and the FA-power correlation. For the construction of models in this figure, [t] was set to 10 seconds postonset. The effects of different [t] values on the evaluation of models were illustrated in eFigure 4. AUC = area under the ROC curve; FA = fractional anisotropy; LOOCV = leave-one-out cross-validation; ROC = receiver operating characteristic.

Discussion

This study improves our understanding of the complex relationship between seizure dynamics and surgical outcomes, introducing a systematic approach to quantify ictal power spreading to understand seizure propagation and its association with structural measures. We demonstrated that seizures in patients with unfavorable surgical outcomes exhibit anatomically broader spread in temporally faster fashion across frequency bands. We also showed that tissue diffusion properties support beta power dynamics among patients with unfavorable outcomes.

Various techniques^14,28 have been identified to extract seizure characteristics with clinical translation. Time-frequency power spectrum transformations have been extensively used^6,15,29 to highlight frequency-specific power dynamics along with the temporal progression of seizures. Compared with commonly used approaches, such as the epileptogenicity index,¹⁴ the method implemented in this study addresses certain limitations by reducing the reliance on subjective human decision-making, controlling for baseline activity during interictal states, and providing insights into seizure power dynamics across various frequency bands in anatomical space. Although overlapping with other available frameworks that “imaging” seizure activities,^6,28 our method greatly improves the computational efficiency by projecting the normalized time-frequency power activities from channels to anatomical spaces, markedly reducing the computational time required for data processing.

Previously, ictal onset patterns, locations, and spread features have been related to surgical outcomes mainly through visual exploration of clinical data.¹⁶ Objective quantification of seizure spread across brain regions can be challenging given patient-specific implantation for iEEG recordings based on personalized epilepsy evaluations.²⁴ To circumvent this complexity, our study introduced a metric for power dynamics, designated as seizure activation size, which considers the individualized nature of iEEG recording placements across patients. This metric simplifies the propagation of seizures across brain regions by the increase in the proportion of seizure-activated areas over time. Our findings demonstrate an escalation in the proportion of seizure-activated regions from the onset of the seizure, indicating widespread involvement of brain regions during seizure activity. In addition, we proposed the quantification of spreading rate, a metric that synthesizes information about the spatial and temporal spread of seizures, providing valuable use in distinguishing surgical outcomes. Notably, we found that seizures associated with unfavorable surgical outcomes reached their maximum activation size more rapidly in the beta frequency band, particularly during the early seconds of seizure onset. Beta band activities have been acknowledged to play an important role in ictal onsets,³⁰ as well as in functional connectivity during seizures.^31,32 Our observation aligns with prior research that identified correlations between widespread seizure onset³³ and early seizure spread in the initial seconds^34,35 in patients with unfavorable surgical outcomes.

It was also reported that analyzing iEEG data from clinical care poses limitations because brain sampling is influenced by case-specific clinical hypotheses, resulting in variable data sources per participant. For instance, the granularity of brain area coverage can vary depending on the hypotheses formed by the clinical team prior to implantation. However, our systematic evaluation demonstrated no differences between groups in terms of the number of sampled regions and electrode distributions, suggesting that electrode sampling is unlikely to confound the comparative analysis of seizure-related activities. Future studies could consider alternative metrics or additional neuroimaging techniques^36,37 to test any other potential confounding factors related to sampling.

This study also investigated how structural properties support neurophysiologic activities by correlating ictal beta power with DTI-derived FA values in the iEEG electrode sampled regions. Prior research noted heightened structural-functional coupling between DTI connectomes and iEEG coherence networks as seizures progress from preictal to ictal states, suggesting seizure function depends on brain structure.^17,18,27,38 While FA is typically applied to the study of white matter bundles, recent investigations have also identified associations between FA values computed in brain regions containing gray matter and deficits in neurologic diseases, indicating disease-related tissue damage at the microstructural level.^39,40 Studies have demonstrated that atrophy in cortical gray matter is predominantly attributed to neuronal shrinkage and axonal degeneration and is considered clinically significant because it is associated with the disease development and progression.⁴¹

In this study, we built on previous work by demonstrating a pronounced increase in structural-functional association, specifically between iEEG beta power dynamics and DTI-derived FA maps during seizures. Notably, we observed that this structural-functional association was significantly stronger within the NSF group, especially during the initial stages of seizure spread. Such a differential relationship points to a potential variation in how functional activities and structural attributes interact at the beginning of ictal propagation between groups, suggesting the existence of a more complex and widespread network for seizure spread. That is, the accelerated spreading rate observed in the NSF group may be due to tissue damage at multiple locations, enabling the concurrent activation of various loci and, thus, faster seizure propagation. This network likely covers an expansive range of brain regions, indicating extensive microstructural reorganization and a more diffuse epileptogenic network.^28,42-44 This form of aberrant neural plasticity presents additional challenges for surgical intervention, potentially rendering seizures more resistant to treatment in TLE.

These findings underscore that both the degree and the timing of structure-function integration play a key role in seizure propagation and outcomes. We note, however, that this approach relies heavily on the brain region–based integrity, which involves both gray and white matter. Given the proliferation of multiple techniques to assess white matter integrity and connectivity, future studies should further test the structural-functional association by using alternative approaches, such as region-to-region tractography, to highlight the structural integrity of white matter tracts connecting such regions.

Finally, our investigation sheds light on the potential capacity of integrated structural and functional features in predicting surgical outcomes in patients with TLE. Our findings demonstrate that the combination of structural and functional characteristics, including averaged FA across regions, beta power spreading rate, and FA-beta power associations greatly improves prediction of surgical outcomes. In addition, to demonstrate the translational significance of our proposed models, we compared its LOOCV accuracy with that of a conventional clinical model based on well-established predictors of seizure outcome after epilepsy surgery.^45,46 The conventional model, which predicted surgical outcome (dependent variable) based on clinical predictors (sex + type of surgery + preoperative seizure frequency + MRI findings + seizure duration), had a LOOCV accuracy of 61.9%. The value is lower than our proposed models that incorporate structural-functional associations or the integration of iEEG power and DTI-FA features, which have LOOCV accuracies of 73.4%, and 73.1%, respectively. Thus, our study provides evidence that considering structural-functional association may further improve the prediction of surgical outcomes in epilepsy.

Our study focused on iEEG ictal power dynamics because they relate to surgical outcomes in TLE, and future work should examine the generalizability of our findings in other epilepsy syndromes.⁴⁷ The current study does not intend to yield a single biomarker to predict surgical outcomes per se. Rather, we introduce a framework that may explain discrepant surgical outcomes in patients with otherwise similar characteristics by quantifying qualitative/visual observations regularly made by clinical neurophysiologists.

In addition, although we propose normalizing ictal data to interictal baseline resting state iEEG activity for analysis, we acknowledge that potential power variability across different frequencies depends on the type of resting state assessed. Future studies should probe the stability of findings when controlling for restful wakefulness vs sleep and across different sleep stages.

Another consideration is the variation in surgical procedures among drug-resistant patients with epilepsy in our study. Our cohort featured a disproportionate distribution of surgical procedure types between NSF and SF groups: While only 1 of 10 patients in the NSF group underwent ATL, 9 of 17 SF patients underwent this procedure. This may be explained by an overall higher proportion of seizure freedom in ATL.⁴⁸ However, to confirm that these results were not confounded by the outcome disparities between LITT and ATL, we replicated the primary comparisons among patients who only underwent the LITT procedure, as shown in eFigure 5. We do, however, suggest that future studies should expand the cohort size to account for such procedural imbalance when testing power dynamics as a predictor of outcomes.

In conclusion, our study contributes to the understanding of seizure spread dynamics and provides a mechanistic explanation for divergent surgical outcomes in TLE. Specifically, our study details the dynamics of seizure power, yielding quantitative evidence that seizures associated with unfavorable outcomes exhibit broader activation and a more rapid spreading rate across brain regions. In addition, we demonstrated a strong iEEG-DTI association, shedding light on the structural underpinnings supporting seizure propagation. We propose that this framework could facilitate multimodal integration to predict epilepsy surgical outcomes. Taken together, our findings highlight the importance of a comprehensive, multidimensional approach to understanding and managing epilepsy, paving the way for future studies to optimize predictive models and interventions aimed at improving patient treatment planning and quality of life.

Glossary

AICHA

Atlas of Intrinsic Connectivity of Homotopic Areas

ATL

anterior temporal lobectomy

AUC

area under the receiver operating characteristic curve

DTI

diffusion tensor imaging

EZ

epileptogenic zone

FA

fractional anisotropy

FDR

false discovery rate

HUP

Hospital University of Pennsylvania

iEEG

intracranial EEG

LITT

laser interstitial thermal therapy

LOOCV

leave-one-out cross-validation

MUSC

Medical University of South Carolina

NSF

non-seizure free

TLE

temporal lobe epilepsy

SF

seizure free

Appendix. Authors

Name	Location	Contribution
Ruxue Gong, PhD	Department of Neurology, School of Medicine, Emory University, Atlanta, GA	Drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data
Rebecca W. Roth, BA	Department of Neurology, School of Medicine, Emory University, Atlanta, GA	Drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data
Allen J. Chang, PhD	Department of Neurology, Medical University of South Carolina, Charleston	Major role in the acquisition of data
Nishant Sinha, PhD	Department of Neurology, University of Pennsylvania, Philadelphia	Drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data
Alexandra Parashos, MD	Department of Neurology, Medical University of South Carolina, Charleston	Major role in the acquisition of data
Kathryn A. Davis, MD	Department of Neurology, University of Pennsylvania, Philadelphia	Drafting/revision of the manuscript for content, including medical writing for content
Ruben Kuzniecky, MD	Department of Neurology, Donald and Barbara Zucker School of Medicine at Hofstra/Northwell, Hempstead, NY	Drafting/revision of the manuscript for content, including medical writing for content
Leonardo Bonilha, MD, PhD	Department of Neurology, School of Medicine, University of South Carolina, Columbia	Drafting/revision of the manuscript for content, including medical writing for content; study concept or design
Ezequiel Gleichgerrcht, MD, PhD	Department of Neurology, School of Medicine, Emory University, Atlanta, GA	Drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data

Study Funding

This study was supported by a National Institute of Neurological Disorders and Stroke (NINDS) award (R01-NS-110347).

Disclosure

N. Sinha received support from American Epilepsy Society (953257), NINDS (R01-NS-125137), and NINDS (R01-NS-116504). K.A. Davis and L. Bonilha received support from NINDS (R01-NS-116504). E. Gleichgerrcht received support from the National Center for Advancing Translational Sciences (NCATS) Awards (UL1-TR-002378, KL2-TR-002381) and NINDS (R01-NS-116504). The other authors report no relevant disclosures. Go to Neurology.org/N for full disclosures.