AI‐Driven Mapping of Seizure Spread Patterns

Abstract

Objective

The focus of epilepsy research has largely been on seizure onset; however, physicians typically examine the patterns of seizure spread past seizure onset as well. This study aims to align automated seizure analysis with clinical practice, leverage deep learning to standardize seizure annotations that varies among physicians, and understand common seizure spread patterns across patients.

Methods

We developed deep learning algorithms on a small subset of patients to detect seizure activity and deployed these algorithms across 275 seizures in 71 patients to analyze the patterns of seizure spread (extent, timing, surgical outcomes, and common patterns) along with incorporating diffusion‐weighted imaging to understand how these patterns relate to the structural connections of the brain.

Results

Deep learning algorithms outperform single features (line length, absolute slope, and power) in ranking seizure onset contacts using physician annotations as a benchmark. We also find that poor outcome patients have more extensive brain regions involved in their seizures while also having more rapid spread between temporal lobes. Incorporating diffusion‐weighted imaging, we find that an increase in structural connectivity between temporal lobes is associated with quicker seizure spread. Finally, we identify clusters of spread patterns common across patients based on spread timing, location, and extent.

Interpretation

Analyzing seizure spread can reveal new insights into seizure evolution and its relationship with surgical outcomes in patients with epilepsy. The findings also suggest that focusing beyond seizure onset is crucial for understanding and treating epilepsy. ANN NEUROL 2026;100:59–73

Seizure onset, timing, extent of activity, and other patterns of seizure activity captured during a seizure are used in the clinical interpretation of electroencephalogram (EEG) to plan treatment for refractory epilepsy. ¹ , ² , ³ , ⁴ , ⁵ Surgical removal of epileptogenic tissue through resection or ablation may be appropriate given sufficient clinical evidence that the removal of localized brain tissue can cure a patient of epilepsy or improve their quality of life. ⁶ , ⁷ In other cases, patterns captured on EEG may instead indicate other treatment modalities, such as implantable neuromodulatory devices, ⁸ , ⁹ or other palliative options, ¹⁰ , ¹¹ are more favorable.

Correct identification of the seizure onset zone and its surgical removal offer the best chance for complete seizure freedom for patients with refractory epilepsy. ¹² , ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ Accurate localization of the seizure onset has thus been a primary focus in epilepsy research to improve outcomes. Yet, overall seizure freedom rates after surgery have remained relatively stagnant over the last 30 to 40 years and vary greatly across centers and studies quantifying outcomes. ⁶

To improve outcomes of patients with refractory epilepsy, the focus in epilepsy research perhaps should also include efforts in identifying patterns of seizure activity beyond seizure onset – a seizure's location, timing, and extent of spread may be important for localizing epileptogenic circuits or determining the best course of treatment based on specific patterns of seizure spread. Recent studies have reported relationships between the spatial organization of seizure recruitment and surgical outcomes ² , ¹⁸ supporting this notion. However, we lack automated and validated measures to reliably quantify seizure activity spread. Indeed, inter‐rater and intra‐rater reliability of human physicians in the evaluation of intracranial EEG is poor. ¹⁹

Here, we quantify seizure spread with both simple EEG features and deep learning models and identify the best performing deep learning model based on clinician markings. We used the best performing deep learning model to answer three main questions: (1) Is the extent and timing of spread associated with patient outcomes? (2) Is the timing of seizure spread related to the structural connectivity of the brain? (3) Do low‐dimensional embeddings of seizure spread patterns reveal clusters of patients and seizure attributes?

Materials and Methods

Clinical Data and Outcome Scoring

Patients were included if they had a diagnosis of treatment‐resistant epilepsy, underwent intracranial EEG (iEEG) implantation as part of epilepsy surgical planning, and exhibited at least one seizure during iEEG monitoring. Seventy‐one patients (mean age = 33 ± 12, 31 female patients) with seizures captured by electrocorticography (ECoG, n = 23) or stereoelectroencephalography (SEEG, n = 48) met inclusion criteria. Patients provided written informed consent for iEEG data analysis, in line with the University of Pennsylvania's institutional review board‐approved protocol (reference number 821778). Across the 71 patients, 275 seizures were captured (Supplementary Fig S1, mean length = 85 ± 94 seconds; mean number of seizures captured per patient = 3.9 ± 3.8 seizures). Fifty‐eight patients had Engel outcome scores at 2 years after undergoing epilepsy surgery. Engel I outcome scores were classified as good outcomes and Engel II to IV were classified as poor outcomes.

Intracranial EEG Acquisition

ECoG and SEEG electrodes were implanted in patients based on clinical necessity. Continuous iEEG signals were obtained for the duration of each patient's stay in the epilepsy monitoring unit. Intracranial data were recorded at 256, 512, or 1,024 hertz (Hz) for each patient. Seizure onset times were defined by the unequivocal electrographic onset (UEO). ²⁰ Interictal data were taken at least 6 hours before seizure onset and were 180 seconds in length. All annotations were verified by neurologists and consistent with detailed clinical documentation. The spacing between SEEG contacts is 5 mm and the contacts are 2.41 mm in size.

Electrode Localization

In‐house software ²¹ was used to assist in localizing electrodes after registration of pre‐implant and post‐implant images (T1‐weighted [T1W] and computed tomography [CT] images). All electrode coordinates and labels were saved and matched with the electrode names on IEEG.org. All electrode localizations were verified by a board‐certified neuroradiologist (author J.S.).

Pre‐Processing of EEG

Following removal of artifact‐ridden electrodes, iEEG signals were bipolar referenced. Signals were notch‐filtered at 60 hertz (Hz) to remove power line noise and low‐pass and high‐pass filtered at 127 Hz and 1 Hz to account for noise and drift. The iEEG signals were downsampled to 128 Hz because a larger sampling rate would not fit into memory of a GPU for training and testing of deep learning algorithms. Signals were then pre‐whitened using a first‐order autoregressive model to account for slow dynamics. All iEEG signals for each channel were normalized to each respective channel's interictal data. This was done by applying the Python package “sklearn” robust scaler function. This function scales features by removing the median and scales the data according to the interquartile range.

Deep Learning Algorithms

The structure of the deep learning algorithms and their parameters are as follows. Python packages and versions are listed at the end of the Methods section. Python code can be found at https://github.com/andrewyrevell/revellLab/ in the SeizureSpread package. The code for the deep learning algorithms is provided in the supplementary material because these algorithms are central to measuring seizure spread.

Single Features – Absolute Slope, Line Length, and Broadband Power

The single features absolute slope, ²² line length, ²³ , ²⁴ and broadband power were calculated on the preprocessed EEG data. Broadband power was calculated using the Scipy Python package version 1.5, ²⁵ and the function scipy signal.welch (default parameters, with FFT epoch length equal to 1 second).

Line length:

$$\mathrm{abs}\left(\left[x\left(k-1\right)-x\left(k\right)\right]=\frac{L\left(n\right)}{K}\right)$$ (1)

where LL(n) is the normalized line length value at a discrete time index n; L(n) is the sum of distances between successive points within the sliding window of size N sample points; x[k] is the value at the kth sample; and K is the normalization constant. ²³

Absolute slope:

$$S_i\left(t\right)=\left|\frac{\mathrm{\Delta }{\mathit{EEG}}_i\left(t\right)}{\mathrm{\Delta }t}\right|*\frac{1}{{\sigma }_i}$$ (2)

where i runs over all channels, t denotes time, and ‡ _i denotes the standard deviation of S _i (t) during the interictal period of channel i.

Training and Testing, and Measuring Seizure Spread

Training and testing data were broken into 1‐second windows with 0.5 seconds overlap on interictal data (“definitely not seizing”) and sections of ictal data on each channel that an annotator determined to be “definitely seizing.” Performance was quantified with area under the curve (AUC) for differentiating interictal and “definitely seizing” windows. A leave‐one‐out cross validation approach was used in a pilot subsample of 13 patients.

To quantify seizure spread after training and testing, 180 seconds of preictal and 180 seconds of postictal data were collected in addition to the seizure. For each channel, 1‐second windows with 0.5 seconds overlap were used to calculate probability of seizure (for the deep learning algorithms) or the normalized single feature. Probabilities or single features were smoothed over 20 seconds. Onset of activity for each channel was determined by the time window at which the smoothed value crossed a predetermined set threshold after the unequivocal onset.

Validation of Seizure Spread Algorithms: Median Rank Percent of Seizure Onset Contacts

After measuring seizure spread, the performance of each algorithm was assessed by its ability to appropriately rank seizure onset contacts with physician markings. For patients with seizure onset annotations (n = 55), the agreement between physician markings and each algorithm marking is calculated, and the algorithm ranking of the seizure onset contacts were averaged (median). Tied ranks were settled by electrode label. This median ranking is divided by the total number of implanted contacts to normalize for differences in implantations. Note, this calculation penalizes ranking scores for physicians who marked a large number of contacts. However, we focused on comparing algorithms, and this penalty as well as biases introduced by our tie ranking scheme is equal between algorithms. The performance of each algorithm was calculated at varying thresholds (Supplementary Fig S3). The optimal probability thresholds for each algorithm were 0.69 (WaveNet), 0.60 (1D CNN), 0.94 (LSTM), 0.26 (absolute slope), 0.11 (line length), and 0.02 (broadband power).

Atlas Choice

We chose the Automated Anatomical Labeling (AAL) atlas ²⁶ , ²⁷ , ²⁸ in this study because (1) it is a common structural atlas used to create structural connectivity; (2) it contains regions with sufficient depth to include depth electrodes (contacts that fall outside the atlas are excluded from analysis, reducing power); and (3) the AAL atlas provides appropriate power to study the structure–function relationship of the brain ²⁹ (ie, its parcellation scheme is appropriate to use in studies incorporating both structural data from diffusion‐weighted imaging and functional data from iEEG). Adding additional atlases may reduce the power of our study.

Calculating Extent of Spread

The extent of seizure spread can be quantified in two ways: (1) by the number of contacts and (2) by the number of brain regions activated over time. In patients with intervention outcome data (n = 58), each metric was converted into a percentage by (1) dividing the number of active contacts by the total number of contacts implanted (excluding contacts that fell outside the brain or artifact contacts), and (2) dividing the number of active brain regions by the total number brain regions sampled. If multiple contacts fell in a brain region, then that region was still counted only once. The activity (ie, probabilities or single feature metrics) of all the contacts within a single region were averaged together. That region was considered active if the average probability or metric fell above a predetermined threshold.

Calculating Timing of Spread

In patients with intervention outcome data (n = 58), activation times of all contacts in the temporal lobe were averaged together for each of the left and right hemispheres to calculate inter‐lobar spread time. The absolute value difference in the average activation times was recorded in seconds. To account for seizure with no spread, inverse spread times were calculated (1/spread time). For example, spread time could not be calculated or would be considered infinite if the left temporal lobe was active and the right temporal lobe never became active. Therefore, inverse spread time would be adjusted to zero and could be compared to seizures with spread.

Contingency Tables, Sensitivity, and Specificity

Cutoff times at 5, 10, 15, 20, 30, and 60 seconds were used to differentiate good and poor outcomes. Outcomes at 2 years were used. Sensitivity, specificity, positive predictive values (PPVs), negative predictive values (NPVs), chi‐square test, and Cramer's V are reported for each cutoff.

Structural Connectivity

The below subsections detail the methodology for calculating structural connectivity.

Imaging Protocol

Prior to electrode implantation, magnetic resonance imaging (MRI) data were collected on a 3T Siemens Magnetom Trio scanner using a 32‐channel phased‐array head coil. High‐resolution anatomic images were acquired using a magnetization prepared rapid gradient echo (MPRAGE) T1‐weighted (T1W) sequence (repetition time = 1810 ms, echo time = 3.51 m, flip angle = 9, field of view = 240 mm, and resolution = 0.94 × 0.94 × 1.0 mm³). High Angular Resolution Diffusion Imaging (HARDI) was acquired with a single‐shot EPI multi‐shell diffusion‐weighted imaging (DWI) sequence (116 diffusion sampling directions, b‐values of 0, 300, 700, and 2000 s/mm², resolution = 2.5 × 2.5 × 2.5 mm³, and field of view = 240 mm). Following electrode implantation, spiral CT images (Siemens) were obtained clinically for the purposes of electrode localization. Both bone and tissue windows were obtained (120 kV, 300 mA, and axial slice thickness = 1.0 mm).

Diffusion Weighted Imaging Preprocessing

HARDI images were subject to the preprocessing pipeline, QSIPrep, to ensure reproducibility and implementation of the best practices for processing of DWIs. ³⁰ Briefly, QSIPrep performs advanced reconstruction and tractography methods in curated workflows using tools from leading software packages, including FSL, ANTs, and DSI Studio with input data specified in the Brain Imaging Data Structure (BIDS) layout.

Structural Network Generation

DSI‐Studio (http://dsi-studio.labsolver.org, version: December 2020) was used to reconstruct the orientation density functions within each voxel using generalized q‐sample imaging with a diffusion sampling length ratio of 1.25. ³¹ Deterministic whole‐brain fiber tracking was performed using an angular threshold of 35 degrees, step size of 1 mm, and quantitative anisotropy threshold based on Otsu's threshold. ³² Tracks with lengths shorter than 10 mm or longer than 800 mm were discarded, and a total of 1,000,000 tracts were generated per brain. Deterministic tractography was chosen based upon prior work indicating that deterministic tractography generates fewer false positive connections than probabilistic approaches, and that network‐based estimations are substantially less accurate when false positives are introduced into the network compared with false negatives. ³³ To calculate structural connectivity, the AAL atlas was used. Structural networks were generated by computing the number of streamlines passing through each pair of atlas regions. Streamline counts were log‐transformed and normalized to the maximum streamline count, as is common in prior studies. ³⁴ , ³⁵ , ³⁶ , ³⁷ For each left and right hemispheres, all the temporal lobe, frontal lobe, and parietal lobe structures were combined and the structural connectivity between each hemisphere of each lobe were totaled together to represent the structural connectivity between the hemispheres of each lobe.

Generalized Linear Models to Quantify the Relationship Between Structural Connectivity and Timing of Spread

The Statsmodel Python package was used to construct a Tweedie regressor with 1.1 power. Structural connectivity was the independent variable and inverse spread time was the dependent variable. The percent deviance explained, D ², was calculated using the Sklearn linear_model Python package and the TweedieRegressor score method. D ² indicates the percentage of deviance explained, a generalization of the coefficient of determination R ².

Python Packages and Versions

The conda environment for the analyses can be found in https://github.com/andyrevell/revellLab/ in the environments folder.

Results

Deep Learning Algorithms Effectively Detect Ictal Versus Interictal States on Single Channels

To investigate the hierarchical organization of seizure spread patterns across seizures and patients, we need robust measures of seizure spread. Currently, a limited number of studies deploy automated algorithms to quantify seizure spread and usually rely on single features, such as line length ³⁸ or power ² ; however, we did not know if such algorithms reliably measure seizure spread. Many studies that do quantify spread are performed with a small number of patients or require manual annotations by an epileptologist. ³⁹ , ⁴⁰ We compared the performance of different large‐scale, and completely automated seizure onset ranking algorithms to capture spread. We used both simple EEG features (absolute slope, line length, and broadband power) and 3 deep learning algorithms with different neural network architectures designed for time‐series data.

We chose the single EEG features because they have been shown to correspond with clinical annotations for seizure onset ¹⁷ , ²² , ²³ , ²⁴ and they are a relatively small number of simple features to compare and preserve power in our study. The deep learning algorithms were chosen because they are effective predictors of time series data. ⁴¹ The deep learning algorithms were (1) WaveNet, ⁴² a one‐dimensional conventional neural network (1D CNN) with a causal and dilated neural network architecture, (2) a 1D CNN, denoted here as a default CNN as opposed to WaveNet with a specific neural network architecture, and (3) a long‐short‐term‐memory (LSTM) neural network originally applied for sequence modeling (Fig 1). ⁴³

FIGURE 1.

Seizure detection on single channels. (A) Seizure spread is quantified by measuring seizure activity across multiple channels. A pattern of spread is characterized by the extent, timing, and locations of seizure activity. (B) Schematic showing how seizure spread can be measured. Machine learning algorithms can be deployed to differentiate 2 classes on single channels: definitely seizing (ictal) states, and definitely not seizing (interictal) states. Once excellent performance has been achieved to differentiate states, fully automated algorithms can be deployed to determine state transitions on peri‐ictal data. Simple features such as absolute slope, line length, and power – all associated with seizure onset – can be used. EEG data can also be used in the case of deep learning algorithms such as (1) WaveNet, a 1D CNN with a causal and dilated neural network architecture, (2) a 1D CNN, denoted here as a default CNN as opposed to WaveNet with a tailored architecture, or (3) an LSTM neural network. (C) Seizure spread can be visualized in 4 different ways, all aided to enhance our understanding of seizure spread patterns. 1D CNN = one‐dimensional conventional neural network; EEG = electroencephalogram; LSTM = long‐short‐term‐memory. [Color figure can be viewed at www.annalsofneurology.org]

To measure spread, we first trained the deep learning algorithms to differentiate between 2 states, ictal and interictal states at the single‐channel level, so that we could eventually quantify when the state transition happens across channels (Fig 2A). In a pilot subsample (n = 13), the area under the curve (AUC) was calculated for differentiating the 2 states with a leave‐one‐patient‐out cross validation across varying learning rates in a pilot subsample (Fig 2B). Similarly, a leave‐one‐out cross validation was performed on the single features for differentiating ictal and interictal states (Fig 2C). The deep learning algorithms at a default learning rate of 0.001 outperform the AUC of the single features (p < 0.001, Wilcoxon Signed‐rank test, 2‐sided, False Discovery Rate (FDR) correction for 15 tests pairwise across the 6 algorithms), except the comparison between LSTM and power (p > 0.05).

FIGURE 2.

Training and testing of binary seizure states to measure state transitions on single channels. (A) Examples of 1‐second windows of non‐seizing (interictal >6 hours before seizure) and seizing states of the same channel in each row at the same gain. These 2 states were used for training and testing of deep learning algorithms to be deployed on peri‐ictal data nearing the transition to a seizure state. Peri‐ictal data show windows 1 to 20 seconds before seizure states of the same channel. (B) Using a pilot subsample (n = 13 patients), leave‐one‐out cross‐validation and AUC as a function of learning rate is shown for the three deep learning algorithms. At a default learning rate of 0.001, the AUC was compared with each of the single feature AUC (***p < 0.001, Wilcoxon Signed‐rank test, 2‐sided, FDR correction for 15 tests pairwise across the 6 algorithms). Shaded areas represent 95% CIs. (C) Using a pilot subsample (n = 13 patients), leave‐one‐out cross validation and AUC for each of the 3 single feature algorithms in detecting seizing versus non‐seizing binary states on single channels. (***p < 0.001, Wilcoxon Signed‐rank test, 2‐sided, FDR correction for 15 tests pairwise across the 6 algorithms). (D) Schematic showing that these algorithms were deployed on peri‐ictal data to measure state transitions. Heatmap colors indicate seizure probabilities (for the deep learning algorithm) or normalized feature values (for the single features) to measure seizure spread across time (x‐axis) and across channels (y‐axis, order is the same across heat maps) of the example seizure shown at the bottom left EEG. Yellow arrows point to seizure onset channels. AUC = area under the curve; CIs = confidence intervals; FDR = false discovery rate. [Color figure can be viewed at www.annalsofneurology.org]

Once the algorithms are developed to differentiate ictal and interictal states for each channel, they can be deployed on peri‐ictal data to measure the time that the state transition occurs. The timing of state transitions across multiple channels effectively measures the spread of seizure activity across the brain (Fig 2D). In the case of the deep learning algorithms, the transition from interictal to ictal states occurs when the probability of an ictal state surpasses a set threshold. In the case of single features, the transition from interictal to ictal states occurs when each respective normalized feature (rather than probabilities) surpasses a set threshold. The deep learning algorithms have larger contrasts between inter‐ictal and ictal states than the single features (see Fig 2D).

Deep Learning Algorithms Outperform Single Features in Ranking Seizure Onset Contacts

To validate the algorithms measuring seizure spread, we first examine their performances on ranking the initial spread points – the seizure onset contacts (Fig 3A).

FIGURE 3.

Validation of seizure spread algorithms (ranking single channels). (A) A patient example with physician markings is shown. A corresponding thresholded activation map (a thresholded heatmap from Fig 2D) of the WaveNet algorithm is to the right. (B) This schematic shows how the performance of an algorithm is validated with human clinician markings. The algorithm ranking of seizure onset contacts (red) were averaged (median) and divided by the total number of implanted contacts to normalize for differences in implantations. Note, this calculation penalizes ranking scores for physicians who marked large number of contacts. (C) Box plots showing the median rank percent for all patients with seizure onset contact annotations (n = 55). A lower percentage means that the algorithm closely matches a clinician's marking. All 6 algorithms perform better than chance (one‐sample Wilcoxon signed rank test, 2‐sided, FDR correction for 6 tests, null hypothesis is 50% median rank – if a seizure onset contact is randomly assigned a rank, it would be 50% of all implanted contacts). WaveNet, 1D CNN, LSTM all perform better than the single feature algorithms (p < 0.01, Wilcoxon signed rank test, 2‐sided, FDR correction for 15 tests pairwise across the 6 algorithms), but none of the deep learning algorithms outperform each other. 1D CNN = one‐dimensional conventional neural network; FDR = false discovery rate; LSTM = long‐short‐term‐memory. [Color figure can be viewed at www.annalsofneurology.org]

The agreement between physician markings and each algorithm marking is calculated. For example, a clinician may mark LA1 and LA2 as the seizure onset contacts (Fig 3B). The seizure spread algorithm also independently determines the rank order of seizure activity for all contacts. The rank order from the algorithm is averaged (by computing the median) for just the clinician onset contacts. In other words, if the algorithm determines that contact LA1 started to seize first and contact LA2 started to seize second, the median rank of these contacts is 1.5. This median ranking is divided by the total number of implanted contacts to normalize for differences in number of contacts across patients. In contrast, if an algorithm determines that contact LA1 was the 100th seizing contact, then it does not agree with the clinician marking and would have poor performance.

All 6 algorithms perform better than chance in ranking seizure onset contacts (Fig 3C), p < 0.05, one‐sample Wilcoxon signed rank test, 2‐sided, FDR correction for 6 tests, null hypothesis is 50% median rank – if a seizure onset contact is randomly assigned a rank, it would be 50% of all implanted contacts.

All 3 deep learning algorithms perform better than the single feature algorithms in differentiating states (p < 0.01, Wilcoxon signed rank test, 2‐sided, FDR correction for 15 tests pairwise across the 6 algorithms), but none of the deep learning algorithms outperform each other. For the rest of the manuscript, we rely on WaveNet to quantify seizure spread as it had the lowest median rank percent and has been shown to outperform other algorithms across different scenarios. ⁴²

The Extent of Seizure Spread – Poor Outcome Patients Have More Distributed Regions Involved in Seizures

The extent of seizure spread over its evolution can be quantified in 2 ways: (i) by the number of contacts activated over time and (ii) by the number of brain regions activated over time (Fig 4). Note that in the following spread dynamics analyses, we quantify seizure spread by the timing and extent of channel/region activation derived from algorithm‐defined state transitions, rather than estimating the physical propagation speed of electrical activity through brain tissue itself (ie, we aim not just to quantify how quickly any electrical activity propagates across in the brain with respect to time, but rather, quantify when and where a state transition happens).

FIGURE 4.

Extent and timing of seizure spread across multiple channels. The extent of seizure spread over time is quantified by (i) the number of channels and (ii) the number of regions active. (A) The percentage of channels and regions active over time for a patient with 14 seizures is shown. Darker lines indicate seizures captured earlier during their hospital stay. (B) Extent of seizure spread can be biased by the number of contacts or regions implanted. Poor outcome patients (n = 30) have higher number of contacts implanted than good outcome patients (n = 28, **p < 0.01, Mann Whitney U test, 2‐sided, null hypothesis: number of contacts implanted is equal between outcomes). However, poor outcome patients (n = 30) did not have significantly higher number of regions sampled than good outcome patients (n = 28, x denotes trending p < 0.10, Mann Whitney U test, 2‐sided, null hypothesis: number of sampled regions is equal between outcomes). (C) Even with larger sampling (more contacts), poor outcome patients (n = 30) have higher percentage of contacts active than good outcome patients (n = 28) at the 30 second mark (dashed line, *p < 0.05, Mann Whitney U test, 2‐sided, null hypotheses: percentage of active contacts at 30 seconds is equal between outcomes). Similarly, poor outcome patients have higher percentages of sampled regions active than good outcome patients (n = 28) at 30 seconds (p < 0.05, Mann Whitney U test, 2‐sided, null hypothesis: percentage of active implanted regions at 30 seconds is equal between good and poor outcome patients). The shaded areas represent 68% CIs. (D) Effect sizes of the differences between outcomes are shown over time. Dashed line is at 30 seconds. (E) Contingency tables comparing the time of spread between temporal lobes in good and poor outcomes. Sensitivity, specificity, PPVs, NPVs, chi‐square test, and Cramer's V are reported for each cutoff. *p < 0.05, FDR correction for 6 tests for the 6 cutoffs (null hypothesis: no association between timing of spread at a specific cutoff and outcome. Red highlights indicate cutoffs with significant associations or high specificity. CI = confidence interval; FDR = false discovery rate; NPVs = negative predictive values; PPVs = positive predictive values. [Color figure can be viewed at www.annalsofneurology.org]

The percent of contacts and regions activated over time in an example patient with 14 seizures is shown in Figure 4A using the WaveNet algorithm at its optimal threshold. Darker lines represent seizures captured earlier during their hospital stay and lighter lines represent seizures captured near the end of their hospital stay. Earlier seizures (darker lines) have more rapid activation (larger slopes) of contacts and regions. The velocity of activation has a noticeable shift with smaller slopes and longer seizures at approximately the sixth to eighth seizure. Here, we see evidence that the evolution of seizures across time and across seizures themselves can change – the quality of the seizures within a patient can change during their hospital course and may be due to a variety of factors (for example, medication changes). In other words, the changes in seizure patterns within a patient can be captured with algorithms designed to quantify seizure spread (Are their seizures stereotypical? Is their pattern changing? Should we use this seizure to localize seizure onset for surgery given its stereotypical nature?). This is purely automated and did not require an epileptologist's manual labor for the 14 seizures.

We also hypothesized that the pattern of seizure spread between good and poor outcome patients is different. We reasoned that poor outcome patients may have more distributed (extensive) regions involved during the seizure. ⁴⁴ Before testing this hypothesis, however, we reasoned the extent of seizure spread can be biased by a different number of contacts and regions targeted for implantation between the 2 groups (see Fig 4B). This reasoning is based on prior work that has found that more expansive coverage is associated with non‐focal epilepsy networks, ⁴⁵ less confident preimplantation judgment of the epileptogenic network, ⁴⁶ and poor surgical outcome. ⁴⁷ We found that poor outcome patients (n = 30) typically had a higher number of contacts implanted than good outcome patients (n = 28, *p < 0.01, Mann Whitney U test, null hypothesis: the number of implanted contacts is the same between good and poor outcome patients). In contrast, we found that poor outcome patients (n = 30) did not have significantly higher number of brain regions sampled than good outcome patients (n = 28, x indicates trending p < 0.10, Mann Whitney U test, null hypotheses: the number of sampled regions is the same between good and poor outcome patients).

Despite having more contacts and a similar number of sampled regions, poor outcome patients demonstrated a higher percentage of contacts and regions active over time (see Fig 4C). In other words, poor outcome patients have more distributed (extensive) regions involved during their seizures despite a bias that would be expected to decrease the percentage of contacts or regions activated over time (poor outcome patients have more contacts implanted, and if the extent of seizure spread was equal between good and poor outcome patients, then the percentage of contacts activated would be less, not more).

We also calculated effect size differences across time between good and poor outcome patients (see Fig 4D). Effect sizes are largest between 20 and 50 seconds into a seizure indicating that the prediction of outcome using the extent of seizure activity may be best in this time window. However, effect sizes are low‐to‐moderate (<0.8) and the percentages of active contacts or regions may not be solely sufficient for clinical use. These results provide evidence that the pattern of spread may be different between good and poor outcome of patients with epilepsy, validating past work showing that highly localized seizures are related to improved intervention outcomes. ⁴⁸ The observed pattern of spread can then be used to build predictive models to help direct treatment and indicate if surgical intervention may result in persistent seizure freedom.

The Timing of Seizure Spread – Poor Outcome Patients Have More Rapid Spread Between Temporal Lobe Regions

To expand on the relationship between seizure spread patterns and patient outcomes, we evaluated whether surgical success can be predicted based on the timing of bilateral seizure spread. Only patients with bilateral temporal lobe sampling and patients with outcome scores at 2 years are used. Figure 4E shows contingency tables comparing the timing of spread between temporal lobes in good and poor outcome surgical epilepsy patients. The activation time of all contacts in the temporal lobe structures are averaged together for each of the left and right hemispheres. The difference between these average activation times between the left and right lobes are recorded. Cutoff spread times of 5, 10, 15, 20, 30, and 60 seconds are used to differentiate good and poor outcome patients. Sensitivity, specificity, positive predictive values (PPVs), negative predictive values (NPVs), chi‐square test, and Cramer's V are reported for each cutoff.

We show that there is an association between timing of spread and surgical outcome if seizures spread between temporal lobes within 5, 10, 15, and 20 seconds (p < 0.05, FDR correction for 6 tests for the 6 cutoffs, chi‐squared test, null hypothesis: there is no association between the timing of spread at a specific cutoff and surgical outcome). Spread within 5 seconds has the highest specificity (94%), thus patients with seizures that spread quickly have a high likelihood of a poor outcome. Timing of spread at any evaluated cutoff did not provide good sensitivity (40–60%, ie, many patients with slow or no spread still have poor outcomes).

Cramer's V are reported to show that effect sizes are low to moderate (<0.6), and timing of spread may not be a singularly sufficient metric for clinical use, however, these results provide evidence that the pattern of spread may be different between good and poor outcome patients with epilepsy.

Structural Connectivity Between Temporal Lobes Is Associated With the Timing of Spread between Regions

We hypothesized that the timing of spread between temporal lobes is associated with the structural connectivity between these lobes ⁴⁰ – greater connectivity between temporal lobes would entail a quicker spread. Of the 71 patients, 22 acquired High Angular Resolution Diffusion Imaging (HARDI). We separated these patients with bilateral temporal lobe sampling (n = 15) and unilateral implantation (n = 7). We totaled the strength of structural connectivity between all temporal lobe regions of the left and right hemisphere (Fig 5A).

FIGURE 5.

Structural connectivity between temporal lobes is associated with the timing of spread between regions. (A) Schematic showing how structural connectivity was measured in a subset of patients (n = 22) with HARDI. The heatmap represents the structural connectivity matrix of an example patient and color bar represents the strength of structural connectivity between all regions. The streamline counts between the regions of left and right lobes of each hemisphere was summed (ie, the total streamline counts was computed between all the regions in the left and right temporal lobes). (B) Scatter plots and a generalized linear model showing the relationship between the strength of connectivity between lobes and the timing of seizure spread between temporal lobes (n = 15 patients with bilateral sampling). X‐axis indicates the total normalized streamline counts between the respective left and right lobes of each patient. Y‐axis indicates inverse spread times (1/seconds). Lower numbers indicate slower spread and higher numbers indicates quicker spread. An inverse spread time of zero indicates no spread was observed between the temporal lobes of the bilaterally sampled patient. (*p < 0.005; **p < 0.001, FDR correction with 4 tests, null hypothesis: no association between structural connectivity and spread time). (C) Boxplots showing the structural connectivity strengths of patients (n = 22) divided into 3 cohorts: bilaterally sampled patients with any spread between temporal lobe structures (spreaders), bilaterally sampled patients with no spread (non‐spreaders), and patients who had only unilateral sampling (unilateral implant, this cohort was excluded in panel b). Mann Whitney U test was performed between the spreaders (n = 11) and non‐spreader groups (n = 4). *p < 0.01 (FDR correction with 4 tests. null hypothesis: no differences in structural connectivity between spreaders and non‐spreaders). Patients with unilateral implantation were plotted ad hoc and initially excluded, but not tested. FDR = false discovery rate; HARDI = High Angular Resolution Diffusion Imaging. [Color figure can be viewed at www.annalsofneurology.org]

We found a relationship in the strength of connectivity between temporal lobes and the timing of seizure spread between temporal lobes using a generalized linear model (Fig 5B, p < 0.005, n = 15 patients, percent deviance explained, D ² = 0.43, FDR corrected for 4 tests, null hypothesis: there is no relationship between the strength of structural connectivity and the timing of spread). D ² indicates the percentage of deviance explained, a generalization of the coefficient of determination R ². Here, higher connectivity strength between temporal lobes is associated with quicker spread between the temporal lobes. We found no relationship between the strength of connectivity between other lobes – frontal and parietal – and the timing of seizure spread between temporal lobes (p > 0.05, n = 15 patients, percent deviance explained, D ² = 0.02 and 0.21 for the frontal and parietal lobes, respectively, FDR corrected for 4 tests, null hypothesis: there is no relationship between the strength of structural connectivity and the timing of spread).

Patients With Spread Btween Temporal Lobes Have Higher Structural Connectivity than Patients Wth No Spread

We divided the 22 patients with structural connectivity into 3cohorts: patients with any spread between the bilaterally sampled temporal lobes (n = 11), patients with no spread (n = 4), and patients with unilateral sampling who were excluded from the previous section’ analysis (n = 7). Patients who had unilateral implantation already had sufficient clinical suspicion that seizure semiology was unilateral and may not have spread to the contralateral hemisphere – they can be considered similar to the cohort with no spread (but we still kept them separate for comparison). We tested the hypothesis that patients with no spread between temporal lobes have lower structural connectivity between temporal lobes than patients with spread (Fig 5C, left box plot, Mann Whitney U test, p < 0.01 with FDR correction for 4 tests). Null hypothesis: no differences in structural connectivity between spreaders and non‐spreaders. Similar to the previous section's analysis, we found no difference in the structural connectivity between the frontal and parietal lobes in spreaders versus non‐spreader (p > 0.05, middle two box plots), and found a difference at a smaller parcellation scale by only considering the structural connectivity between the lateral temporal gyri (p < 0.01, right box plot). Patients with unilateral implantation were plotted for comparison and show similar trends in structural connectivity to patients with no spread.

Cluster of Seizure Spread Patterns

We performed principal component analysis (PCA) to assess the relationship of seizure spread clusters in a low‐dimensional embedding with clinically relevant variables. Four scatter plots show different attributes of the seizure clusters in Figure 6A. Five example patients and their seizures are highlighted, revealing that seizure spread patterns from the same patient are often proximal in this low‐dimensional representation, although this latent space representation reveals some intrasubject variability across seizures. We observed that first principal component corresponds closely with epilepsy laterality, an effect potentially driven by differences in electrode sampling strategies between these patient subgroups.

FIGURE 6.

Clusters of seizure spread patterns. (A) Four scatter plots are shown highlighting different attributes of the patients. (Left) Five subjects with multiple seizures are highlighted. Subjects 2, 3, and 4 have seizures predominantly in one cluster, but have seizures spanning multiple clusters. (Left center) Seizures are highlighted by laterality of seizure onset determined though clinical chart review. (Right center) Seizures are highlighted by good and poor surgical outcomes. (Right) Seizures are highlighted by length. (B) The k‐means clustering was performed at k = 5 based on the optimal number of principal components. (Left) Explained variance ratio as a function of number of principal components to select for an optimal number of clusters. (Right) Clusters of seizures in principal component space based on spread patterns. (C) (Top) Seizure pattern (location and timing) for each cluster is shown through a coronal slice of the brain. Colors indicate the percentage of time in a seizure when a region was active. Darker regions indicate earlier activation. Only regions with at least 2 patients with seizures showing activity are colored, all the else regions are gray. (Bottom) The number of regions activated at any point during a seizure for each cluster. [Color figure can be viewed at www.annalsofneurology.org]

We then clustered seizure spread patterns using k‐means clustering, optimizing k based on explained variance as a function of the number of principal components (Fig 6B). The optimal number of principal components was determined to be 5 based on the elbow method. Additional components explain variance ratios less than 5%, and we find that greater cluster numbers produced clusters comprised of seizures from just one patient. We plot the number of active regions at any time in a seizure across all the seizures in each cluster (Fig 6C). These clusters and their positions in principal component space reveal differences in both implant strategy and network engagement during seizure spread across our patient cohort. Seizure spread patterns associate considerably with epilepsy laterality and, relatedly, implantation strategies along the PC1 axis. For example, seizure spread clusters 4 and 5 are made up of only negative PC1 loadings and exhibit predominantly right sided sampling and epilepsy lateralization, whereas cluster 3 is made up of only positive PC1 loadings and exhibits predominantly left‐sided sampling and epilepsy lateralization. PC2 loadings appear to relate considerably to the number of implanted regions, as exemplified by differences between clusters 1 and 2 (Fig 7).

FIGURE 7.

Taxonomy of Seizure Spread Patterns. (A) Principle components analysis (PCA) showing explained variance ratio as a function of number of components to select for an optimal number of clusters. A vertical dash is at n = 5 components (B) K‐means clustering sum of squared errors (SSE) is plotted as a function of number of k clusters. (C) The taxonomy of seizure spread patterns is shown using hierarchical/agglomerative clustering. Colors of branches correspond to clusters from Fig 6. X‐axis shows the euclidean distance between clusters with a “complete” linkage function, also known as the Farthest Point Algorithm or Voorhees 32 Algorithm. For example, the left mesial cluster is the first branch point indicating that this spread pattern is a distinct evolution of seizure spread and is most dierent from the other clusters. [Color figure can be viewed at www.annalsofneurology.org]

Discussion

In this study, we develop, validate, and compare different algorithms to measure seizure spread with the goal to characterize patterns of seizure spread. We find that deep learning algorithms are highly effective in differentiating ictal and interictal states over single features (see Fig 2) which can be used to rank seizure onset channels and measure spread (see Fig 3). We discover that poor outcome patients have more distributed regions involved in seizures (see Fig 4A–D) and seizure spread within 5 seconds between the average activation time of the left and right temporal lobes yields a specificity of 94% in differentiating good and poor outcome surgical patients at 2 years (42% sensitivity; see Fig 4E). This speed of seizure spread between temporal lobe regions is associated with the strength of structural connectivity between temporal lobes but not between frontal or parietal lobes (see Fig 5). Last, we cluster seizure spread patterns and find that epilepsy lateralization and implantation strategies explain much of the spatiotemporal diversity of seizure spread patterns in our cohort (see Fig 6).

The Focus of Seizure Activity Past Seizure Onset

In our study, we show the pattern of seizure activity may be an important marker that can predict response to epilepsy surgery. Whereas correct identification of seizure onset contacts is essential for the success of surgery and has been a major focus in computational studies attempting to identify location of seizure onset, it perhaps should not be the only focus of EEG interpretation nor the focus to identify ideal surgical candidates and brain regions targeted for surgery. Here, we show that the pattern of spread, whether through the extent of spread (see Fig 4A–D) or time of spread (see Fig 4E) is associated with surgical outcomes at 2 years.

Furthermore, the ability to quantify patterns of seizure spread – whether through complicated deep learning algorithms or though simple features – opens new avenues to study epilepsy pathophysiology and seizure evolution. Although we present evidence that deep learning algorithms may be superior in capturing spread patterns over simple features, such as line length, these singular features still capture onset and spread better than chance (see Fig 3), and many labs or clinical software may be suited for reporting spread patterns. For example, the patient in Figure 4A shows 14 seizures captured during their hospital stay and we can observe how the pattern of spread changes during the days to weeks a patient may stay in the epilepsy monitoring unit. We can quantify whether a seizure is a stereotypical pattern (eg, mesial temporal vs. temporal neocortical spread) and is representative of their semiology to be used for interpretation and localization of seizure onset. We can also observe how certain external factors, such as medications, sleep deprivation, and other habits, may change the patterns of seizure activity over time. Further, these algorithms may also have utility in scalp EEG, which benefits from standard sampling across patients (although more coarse in localization than intracranial implantations).

Comparing Deep Learning and Non‐Deep Learning Seizure Spread Metrics

We compared a variety of signal‐based features and trained deep learning models for quantifying seizure spread. The choice of features and deep learning models highlight some trade‐offs of both approaches: signal features are validated through prior studies and model neurophysiological phenomena, ²⁰ whereas deep learning models require extensive ground truth labels and more computationally intensive training. A strength of our study is that it contrasts both types of methods while also validating them against clinicians’ markings of seizure onset channels. Of course, other measures have been previously used to characterize the abnormality of an EEG channel based on ictal data, including the Epileptogenicity Index. ¹⁷ This method was developed on mesial temporal lobe epilepsy data and requires parameter tuning for each subject. By contrast, our deep learning methods were trained on a heterogeneous drug‐resistant epilepsy cohort and can be applied across subjects. Whereas the primary objective of this study was to use the seizure spread methods to answer questions about surgical outcomes, structural connectivity, and clustering, the head‐to‐head comparison of spread metrics provides insights into how other researchers should approach deep learning for generalizable and clinically translatable biomarkers of seizure onset and spread.

Limitations

One major limitation to this study is that the clustering of seizure spread patterns may be driven by the implantation and sampling bias of our patients. ⁴⁹ Despite this limitation, however, we find that many patients with unilateral sampling, for example, have their respective implantation schemes for a clinical reason – there is evidence that seizure activity may be limited to the regions targeted for implantation a priori. The patterns of spread classified by physicians are typically implanted in a stereotyped fashion (ie, patients with suspected left mesial temporal lobe epilepsy largely have similar structures targeted for implantation with modifications based on clinical history and other findings). For example, many patients with limited sampling and spread have good outcomes not because their sampling was limited per se, but rather because their spread pattern was predicted to be well‐localized by their physicians, they were implanted to confirm seizure onset, and subsequently had a good outcome after surgery because seizure onset was already biased in its focal localization. Future studies may therefore benefit from investigating seizure spread patterns in a more homogeneous sample with more comparable implantation strategies across patients. Differences in seizure spread patterns may also be explained in part by differences in onset patterns (eg, low‐voltage fast activity vs. rhythmic slow spikes), ⁵⁰ representing an important knowledge gap for further investigation.

Seizure onset annotations included both ECoG and SEEG electrodes in our cohort. We combined these contact types in our analyses and found that seizure spread algorithms can be optimized to characterize spread patterns measured by both ECoG and SEEG in the same patient. However, we recognize that ECoG and SEEG have distinct functional connectivity patterns. ⁴⁹ We therefore urge caution when applying spread algorithms that rely on functional connectivity information rather than single channel features in these cases.

Conclusions

The pattern of seizure activity past seizure onset may help direct treatment of patients with refractory epilepsy and can indicate if surgical intervention or other treatment options may have the best chance to improve a patient's quality of life. We propose a shift in epilepsy research from a primary focus in identification of seizure onset to quantifying the patterns of seizure activity past onset.

Author Contributions

A.Y.R., J.B., J.M.S., S.R.D., B.L., and K.A.D. contributed to the conception and design of the study. A.Y.R., A.R.P., E.C., N.S., B.H.S., A.L., J.M.B., J.S., S.R.D., B.L., and K.A.D. contributed to the acquisition and analysis of data. A.Y.R., B.H.S., J.M.B., B.L., M.J., W.K.S.O., and K.A.D. contributed to drafting a significant portion of the manuscript or figures.

Potential Conflicts of Interest

Nothing to report.

Supporting information

Supplementary Figure S1: Distribution of seizure lengths and number per patient. Left: The distribution of seizure lengths across all 275 seizures in this study. Mean = 85 seconds, median = 68 seconds, and SD = 94 seconds. Right: The distribution of the number of seizures per patient. Mean = 3.9, median = 3.0, and SD = 3.8.

Supplementary Figure S2: Seizures colored by other attributes. Seizures from Figure 6 are colored by other attributes such as sex (top), lesional status (middle), and lobar localization (bottom).

Supplementary Figure S3: Optimal thresholds for seizure spread algorithms. The optimal threshold for each algorithm in agreement with physician markings of seizure onset. Shaded areas represent 95% confidence intervals (CIs). We observed that deep learning algorithms capture relevant seizure onset contacts across more patients and across wider ranges of thresholds than the single features (ie, the deep learning algorithms may more likely capture relevant seizure spread patterns without having to tune specific threshold parameters).

Supplementary Figure S4: Effect size comparisons between seizure spread algorithms in extent and timing. (A) Effect sizes across all thresholds and times for comparing the extent of seizure spread in good and poor outcome patients. Heatmaps and color bars represent Cohen's D. (B) Effect sizes across all thresholds and times for comparing the association between the timing of seizure spread between temporal lobe regions and surgical outcomes. The extent of seizure spread in good and poor outcome patients. Heatmaps and color bars represent Cramer's V.

Supplementary Table S1. Patient‐level clinical summary and analysis‐specific cohort inclusion. The full study cohort included 71 patients (275 seizures). Specific analyses used partially overlapping subsets based on data availability and inclusion criteria. Single‐channel ictal versus interictal state classification model development/testing was performed in a pilot subset (n = 13). Validation of seizure onset contact ranking used the subset with physician seizure onset annotations (n = 55). Outcome‐related spread extent analyses used patients with 2‐year intervention outcome data (n = 58). Structural connectivity analyses used the subset with High Angular Resolution Diffusion Imaging (HARDI) data (n = 22), including a bilateral temporal sampling subgroup (n = 15) for connectivity‐spread timing analyses and a unilateral implantation subgroup (n = 7) shown for comparison in the structural connectivity subgroup analyses.

Data Availability

All analysis code used in this study is publicly available at the repository cited in the manuscript. The underlying intracranial EEG and structural/diffusion MRI data have been curated in Brain Imaging Data Structure (BIDS) format. Because these raw data contain potentially identifiable patient information (including facial features on MRI), they cannot be deposited in an open public repository under the terms of our institutional review board approval. De‐identified derivative data supporting the findings of this study (eg, seizure‐spread metrics, regional activation times, and connectivity matrices) are available from the corresponding author. BIDS‐formatted raw data can be shared with qualified investigators for non‐commercial research purposes upon reasonable request and completion of a data‐use agreement approved by the University of Pennsylvania Institutional Review Board.