Predicting posttraumatic epilepsy using admission electroencephalography after severe traumatic brain injury

Abstract

Objective:

Posttraumatic epilepsy (PTE) develops in as many as one third of severe traumatic brain injury (TBI) patients, often years after injury. Analysis of early electroencephalographic (EEG) features, by both standardized visual interpretation (viEEG) and quantitative EEG (qEEG) analysis, may aid early identification of patients at high risk for PTE.

Methods:

We performed a case–control study using a prospective database of severe TBI patients treated at a single center from 2011 to 2018. We identified patients who survived 2 years postinjury and matched patients with PTE to those without using age and admission Glasgow Coma Scale score. A neuropsychologist recorded outcomes at 1 year using the Expanded Glasgow Outcomes Scale (GOSE). All patients underwent continuous EEG for 3–5 days. A board-certified epileptologist, blinded to outcomes, described viEEG features using standardized descriptions. We extracted 14 qEEG features from an early 5-min epoch, described them using qualitative statistics, then developed two multivariable models to predict long-term risk of PTE (random forest and logistic regression).

Results:

We identified 27 patients with and 35 without PTE. GOSE scores were similar at 1 year (p = .93). The median time to onset of PTE was 7.2 months post-trauma (interquartile range = 2.2–22.2 months). None of the viEEG features was different between the groups. On qEEG, the PTE cohort had higher spectral power in the delta frequencies, more power variance in the delta and theta frequencies, and higher peak envelope (all p < .01). Using random forest, combining qEEG and clinical features produced an area under the curve of .76. Using logistic regression, increases in the delta:theta power ratio (odds ratio [OR] = 1.3, p < .01) and peak envelope (OR = 1.1, p < .01) predicted risk for PTE.

Significance:

In a cohort of severe TBI patients, acute phase EEG features may predict PTE. Predictive models, as applied to this study, may help identify patients at high risk for PTE, assist early clinical management, and guide patient selection for clinical trials.

1 |. INTRODUCTION

Nearly one in three severe traumatic brain injury (TBI) survivors will develop posttraumatic epilepsy (PTE).^1–4 PTE is the leading cause of epilepsy for those with onset between ages 15 and 24 years and accounts for 20% of symptomatic and 5% of all cases of epilepsy.^5–7 PTE has a deleterious effect on recovery from TBI, as epileptogenic pathways induce central nervous system (CNS) inflammation and degeneration, inhibiting regeneration after injury.^8,9 Severe TBI patients with PTE have worse functional outcomes in the short and long term.^4,10–12

Recent advances in data science allow for earlier, more accurate prognostication in medicine, with a potential for earlier disease intervention.¹³ Within TBI, advanced modeling techniques integrating multimodal data have improved early prognostication compared to traditional techniques.^14–16 For patients with epilepsy, early treatment leads to improved cognitive and seizure-free outcomes.^17–19 Despite the morbidity of PTE and potential for improving epilepsy outcomes with early detection and intervention, no clinical tool exists to identify patients at high risk for PTE.

Electroencephalography (EEG) is a widely used clinical tool in TBI populations both to determine the presence of seizures and seizure risk and to prognosticate acute recovery.²⁰ Both standardized visual EEG (viEEG) and quantitative EEG (qEEG) features have been used to predict outcomes after a variety of brain insults.^21–24 Specifically in TBI patients, features such as variability or power in specific frequency bands are associated with mortality.^25,26 Early work predicting PTE using EEGs, however, was largely unsuccessfully.²⁷ Using visual features and classifying EEGs as normal or abnormal, a large, multicenter group from the 1970s did not find any differences at various time points posttrauma.²⁸ Results in more recent cohorts were mixed, with some finding that EEG had potential predictive power.^29,30 These efforts were limited by a reliance on visual, rather than quantitative, EEG features and utilizing late EEGs acquired weeks to months after trauma.

Building on these findings and recognizing the importance of early, within days of trauma, risk stratification of PTE, we evaluated early viEEG and qEEG features in an age- and outcome-matched severe TBI cohort. The discovery of novel early electrophysiologic biomarkers of PTE will not only aid current clinical management of TBI patients, but also support evaluation of early antiepileptogenic interventions in TBI populations.

2 |. MATERIALS AND METHODS

We performed a case–control study sampled from a prospective database of severe TBI patients admitted to a level 1 trauma center between March 2011 and December 2018.³¹ This study received approval from the University of Pittsburgh Human Research Protection Office with consent from subjects’ legal representatives. Research procedures followed the ethical standards of the responsible committee on human experimentation and the Helsinki Declaration of 1975. Data are available to qualified researchers upon request. We followed the STARD (Standards for Reporting Diagnostic Accuracy Studies) reporting guidelines.³²

Our prospective database has been previously described^1,14,31 and includes patients 16–80 years old with severe TBI, defined as a postresuscitation Glasgow Coma Scale (GCS) score ≤ 8. Our database excluded patients with imminent brain death (GCS = 3 with fixed and dilated pupils), pregnancy, and/or penetrating trauma. Trained neuropsychologists assessed outcomes through a structured interview at 3, 6, 12, and 24 months posttrauma using the Glasgow Outcome Scale Expanded (GOSE): 1 = death, 2 = vegetative state, 3 = lower severe disability, 4 = upper severe disability, 5 = lower moderate disability, 6 = upper moderate disability, 7 = lower good recovery, 8 = upper good recovery.

We retrospectively identified incidence and timing of PTE by completing a structured review of the electronic health records at our health system and nearby systems. As previously described,^1,33,34 we defined PTE as a single late posttraumatic seizure (i.e., >7 days posttrauma). A seizure included any clinical event deemed to be a seizure by the treating health care team or an electrographic seizure on EEG. All patients typically received a 7-day course of prophylactic phenytoin after their initial trauma.

For our cohort, we identified all patients in the database who survived 2 years after their severe TBI. We excluded patients with pre-existing epilepsy or alcohol withdrawal seizures, or who had an early post-traumatic seizure prior to or during EEG recording. From this cohort, we identified all patients who had PTE and an interpretable, early EEG. Then, we identified one matched control patient with severe TBI without subsequent PTE for each case. To do so, we identified patients within the same decade of life (age) who had a similar GCS score ± 1 point. As our overall cohort was unbalanced, with fewer PTE patients, we increased our non-PTE cohort size through performing a second round of matching 15 non-PTE to 15 PTE randomly selected PTE patients (i.e., repeat sampling). After matching, we excluded any patients without PTE and EEG files that we were unable to analyze.

All patients underwent continuous, video-EEG monitoring for 3–5 days after injury as part of standard of care. A board-certified epileptologist, blinded to PTE outcomes, identified the earliest 5-min clip of artifact-free EEG for qEEG analyses. This EEG included any epileptic or pathological events, if present. The epileptologist described viEEG features over the first 30 min of initial recording using standard American Clinical Neurophysiology Society (ACNS) nomenclature.³⁵ Given the low prevalence of individual categories of epileptiform discharges and their lack of association with PTE (Table S1), we grouped together all epileptiform features according to ACNS guidelines, including lateralized rhythmic delta activity, lateralized periodic discharges, sharp waves, spikes, ictal–interictal continuum, and brief potentially ictal rhythmic discharges.³⁵ We also evaluated presence of background alpha activity, as this is associated with normal brain activity and has been found to be correlated with TBI outcomes.²⁶

For quantitative EEG analyses, we utilized Persyst Version 14, an automated EEG analysis software, to extract the following qEEG features calculated at a frequency of 1 Hz.³⁵ We averaged both hemispheres together for our first analysis to identify features different in patients with PTE compared to those without PTE:

1. Bandpass filtered spectral power from fast Fourier transformation, averaged across all electrodes and summed within frequency bins: delta (.5–4 Hz), theta (4.01–8 Hz), beta (8.01–13 Hz), and alpha (>13 Hz);

2. Mean amplitude integrated EEG: description of the overall EEG amplitude that covers a broad range of frequencies, inclusive of .16–1 Hz;

3. Peak envelope: a peak-to-peak measure of EEG amplitude that is inclusive of EEG frequencies from the 0–25-Hz range; and

4. Rhythm spectrogram: quantitative measure of the EEG rhythm similarly stratified by the frequency bands listed above.

Given the large proportion of patients who underwent decompressive hemicraniectomy (DHC), resulting in asymmetric breach rhythms and potential confounding effects on qEEG data, we performed a secondary subanalysis stratifying cranial hemispheres by presence or absence of cranial defects from DHC. We first identified all features significantly different in our initial analysis (all patients with hemispheres averaged together) and subsequently evaluated these qEEG features in hemispheres with cranial defects from DHC stratified by PTE. We performed a similar analysis for the hemispheres without cranial defects from DHC. As an example, in a patient with a right-sided DHC, the right hemisphere with a cranial defect would be analyzed against other hemispheres with cranial defects from DHC, and the left hemisphere without a cranial defect would be analyzed against nondecompressed hemispheres. Patients without DHC (i.e., an intact cranium) would have both hemispheres averaged together and included in the no cranial defect from DHC analysis, whereas patients with bilateral frontotemporal craniectomies (i.e., the Kjellberg craniectomy) were counted as having bilateral cranial defects with both hemispheres averaged together and included in the cranial defect from DHC analysis.³⁶

2.1 |. Statistical methods

We summarized clinical and EEG features using descriptive statistics. Where appropriate, we used two-sided Student t-test for normally distributed variables and χ2 or Fisher exact test for categorical variables. We adjusted p-values for our qEEG and viEEG features’ false discovery rate using the method popularized by Benjamini and Hochberg.³⁷

We performed two multivariable analyses to assess: (1) our ability to predict PTE using early EEG features; and (2) the independent association of viEEG, qEEG, and clinical information with PTE. First, we used random forest modeling to predict long-term risk of PTE. Random forest is a robust, tree-based model that can handle multiple data types and minimizes the potential for overfitting, especially in small samples.³⁸ Our multivariable modeling approach with random forest has been used several times in previous qEEG studies in TBI evaluating functional and mortality outcomes.^22,23 For all of our classification tasks, our random forest models had 500 trees, with a maximum depth of m/2, where m is the number of model features.^22,23

We developed four random forest models with different feature sets. We developed a random forest model using all of the qEEG features we collected (Model 1). Then, we developed subsequent random forest models using all of the viEEG features (Model 3) and clinical features (Model 4). Next, to compare the relative effects of viEEG, qEEG, and clinical data, we developed a random forest model using all features that were significantly different on univariable analysis when stratified by PTE (Model 5). We excluded nonsignificant features to avoid overfitting.

For our second multivariable modeling approach, we developed an adjusted logistic regression model (Model 2) using only qEEG features. We used logistic regression, another machine learning modeling approach, for several reasons. Logistic regression has highly explainable features in the form of odds ratios, a possible benefit over random forest. Additionally, by using a second methodologically distinct approach, we minimize the risk that our results reflect overfitting from one particular modeling structure.

Our initial logistic model suffered from multicollinearity across measures of spectral power.³⁹ This finding is a result of the inherent interdependent nature of the raw features. Using isolated frequency band power as an example, an increase in power of one frequency band (i.e., delta) indicates lower power in the other frequency bands (i.e., theta, alpha, beta). To account for this, we normalized these measures to compare their ratio of each combination of power bands (i.e., delta:theta, delta:beta, theta:alpha, etc.). Power ratios are commonly used in qEEG studies, because these ratios are predictive of outcomes and eliminate the need for the nonintuitive units of spectral power (μV²/Hz).^40–42

When building our logistic regression model, we excluded all qEEG features and ratios that failed the linearity assumption. We then used a probabilistic approach for feature selection using the Akaike information criterion.⁴³ We scaled final variables for ease of interpretation, and checked model assumptions by evaluating for multicollinearity and outliers in the final model. We report odds ratios (ORs) with 95% confidence intervals (CIs).

For all models, we reported the area under the curve (AUC) using leave-one out cross-validation (LOOCV). We compared each model’s AUC to the base qEEG model to assess for significant improvements in performance.⁴⁴ Additionally, we tested whether models had discriminatory ability by comparing each model to chance (i.e., a model with a 50% chance of correctly identifying the correct choice).

To understand the relative importance of the features in our final, all-inclusive random forest Model 5, we implemented Shapley Additive Explanations (SHAP), a common post hoc analysis approach used in machine learning. Briefly, SHAP is an additive feature attribution technique that explains a tree ensemble’s overall performance by analyzing the relative importance of each model feature.^45,46 This relative importance is visually displayed in a bee swarm plot. We used R and Python (version 3.8.3) for all analyses. Statistical significance was defined as p < .05.

3 |. RESULTS

Overall, 124 patients with severe TBI from March 2011 through December 2018 were eligible for our study, excluding the 78 who died within 2 years posttrauma, eight with pre-existing epilepsy or alcohol withdrawal seizures, and five with seizures prior to or during EEG recording (Figure 1). Of the 29 patients with PTE, 27 had usable EEG data. The median time to onset of PTE was 7.2 months posttrauma (interquartile range = 2.2–22.2 months). After matching 42 patients using age and admission GCS, seven patients had corrupted or missing EEG files, leaving 35 of 82 patients without PTE for our final cohort. Only one patient had an early seizure (within 7 days of trauma) on posttrauma day 7, and this was after discontinuation of the EEG used for this analysis that was obtained on posttrauma day 2. This patient went on to have a late (>7 days) posttraumatic seizure and thus was included in the PTE group. Table S2, which compares clinical and outcome information for those included and not included in our study, does not show any significant differences.

FIGURE 1.

Consort diagram. For our case–control study, we included all patients from a prospective database of severe traumatic brain injury (TBI) patients who survived 2 years posttrauma and did not have pre-existing epilepsy, alcoholic (EtOH) withdrawal seizures, or seizures within 72 h of trauma, when our electroencephalograms (EEGs) were collected. We subsequently matched all patients with posttraumatic epilepsy (PTE) and available EEG to those without using age and admission Glasgow Coma Scale. We matched an additional 15 patients without PTE to 15 randomly selected patients from our PTE cohort, as our dataset was unbalanced, with more patients without PTE than with PTE. After matching, we excluded all patients without PTE who had missing EEG.

Table 1 displays clinical characteristics of the patient cohort. Consistent with our matching, patients were similar in age (p = .11) and GCS (p = .63). PTE patients had significantly higher rates of DHC than those without PTE (63% [n = 17] versus 26% [n = 9], p < .01). Whereas outcomes for PTE patients were worse at 3 months (p = .02), by 6, 12, and 24 months posttrauma, the GOSE scores were similar (p > .15; Table 2). Our follow-up rates at 3 (88%), 6 (82%), 12 (78%), and 24 months (45%) posttrauma are consistent with historical norms for clinical trials in TBI, although our follow-up rates are artificially depressed by excluding patients who died within 2 years.⁴⁷

TABLE 1.

Patient characteristics.

Characteristic	PTE, n = 27	No PTE, n = 35	p
Age (years)	31 ± 14	37 ± 16	.11
Sex, male	81% (22)	71% (25)	.36
DHC	63% (17)	26% (9)	<.01a
GCS
3	15% (4)	22% (8)	.63
4	0% (0)	6% (2)
5	4% (1)	9% (3)
6	15% (4)	17% (6)
7	59% (16)	43% (15)
8	7% (2)	3% (1)

Note: All patients survived to 2 years and were matched based on age and admission GCS.

Abbreviations: DHC, decompressive hemicraniectomy; GCS, Glasgow Coma Scale; PTE, posttraumatic epilepsy.

^aStatistically significant.

TABLE 2.

Patient outcomes.

GOSE	PTE, n = 27	No PTE, n = 23	p
3 months
2	30% (8/26)	4% (1/27)	.02a
3	42% (11/26)	48% (13/27)
4	12% (3/26)	11% (3/27)
5	12% (3/26)	11% (3/27)
6	0% (0/26)	15% (4/27)
7	0% (0/26)	11% (3/27)
8	4% (1/26)	0% (0/27)
6 months
2	8% (2/25)	0% (0/24)	.46
3	52% (13/25)	54% (13/24)
4	12% (3/25)	0% (0/24)
5	4% (1/25	4% (1/24)
6	8% (2/25)	17% (4/24)
7	8% (2/25)	17% (4/24)
8	8% (2/25)	8% (2/24)
12 months
2	4% (1/23)	0% (0/26)	.93
3	52% (12/23)	45% (12/26)
4	13% (3/23)	12% (3/26)
5	18% (4/23)	15% (4/26)
6	0% (0/23)	4% (1/26)
7	9% (2/23)	12% (3/26)
8	4% (1/23)	12% (3/26)
24 months
2	0% (0/12)	0% (0/12)	.15
3	41% (5/12)	7% (1/16)
4	25% (3/12)	25% (4/16)
5	0% (0/12)	25% (4/16)
6	17% (2/12)	19% (3/16)
7	17% (2/12)	12% (2/16)
8	0% (0/12)	12% (2/16)

Note: GOSE was recorded by a trained neuropsychologist at 3, 6, and 12 months postinjury.

Abbreviations: GOSE, Glasgow Outcome Scale Expanded; PTE, posttraumatic epilepsy.

^aStatistically significant.

Table 3 displays the viEEG features. After correcting for false detection rate, none of the viEEG features was significantly different when stratifying by PTE. Table 4 compares qEEG features for those with and without PTE. Patients with PTE have significantly higher power in the delta spectrogram (Figure 2A), as well as variance in power in the delta and theta frequencies, and peak envelope. To assess whether our findings were related to cranial defects from DHC, we separately compared hemispheres with and without cranial defects. We found similar effects (Table 4, Figure 2B) as when we did not control for DHC.

TABLE 3.

Visual EEG features.

Characteristic	PTE, n = 27	No PTE, n = 35	p
Continuous	93% (25/27)	74% (26/35)	.16
Disorganized	96% (26/27)	91% (32/35)	.70
Voltage attenuation present	44% (12/27)	40% (14/35)	.78
Focal slowing present	70% (16/23)	51% (18/35)	.25
Background alpha activity present	30% (8/27)	49% (17/35)	.21
Epileptiform activity	11% (3/27)	11% (4/35)	.99

Note: A board-certified epileptologist, blinded to PTE outcomes, scored 30-min clips of EEG using standard reporting from the ACNS. We did not report breach or symmetry, as these were strongly influenced by decompressive hemicraniectomy. We grouped together all epileptiform features according to ACNS guidelines, which include lateralized rhythmic delta activity, lateralized periodic discharges, sharps, spikes, ictal–interictal continuum, and brief ictal rhythmic discharges. Individual epileptiform discharges had no association with PTE (Table S1). Hemispheres with breach rhythm were not analyzed independently for the visual EEG features. We only evaluated background alpha activity, as this is associated with normal brain activity and has been found to be correlated with traumatic brain injury severity.²⁶

Abbreviations: ACNS, American Clinical Neurophysiology Society; EEG, electroencephalographic; PTE, posttraumatic epilepsy.

TABLE 4.

Quantitative EEG features.

GOSE	PTE, n = 27	No PTE, n = 23	p
FFT spectrogram
Delta mean	1.28 ± .36	.94 ± .23	.001a
Delta variance	.53 ± .29	.24 ± .14	<.001a
Theta mean	.70 ± .22	.60 ± .16	.14
Theta variance	.12 ± .07	.07 ± .04	<.01a
Alpha mean	.57 ± .19	.52 ± .14	.37
Alpha variance	.08 ± .06	.05 ± .03	.10
Beta mean	.44 ± .16	.41 ± .11	.58
Beta variance	.05 ± .04	.05 ± .05	.99
Mean EEG amplitude	10 ± 6	8 ± 4	.15
Peak envelope	26 ± 13	16 ± 7	<.01a
Rhythm spectrogram
Delta	.51 ± .39	.32 ± .22	.07
Theta	.82 ± .60	.65 ± .38	.28
Alpha	.56 ± .77	.38 ± 23	.33
Beta	.33 ± .24	.28 ± .18	.37
No DHC	PTE, n = 25	No PTE, n = 33	p
FFT spectrogram
Delta mean	1.19 ± .33	.94 ± .24	<.01a
Delta variance	.41 ± .23	.23 ± .15	<.01a
Theta variance	.09 ± .05	.07 ± .04	.18
Peak envelope	22 ± 11	17 ± 7	.06
DHC	PTE, n = 17	No PTE, n = 9	P
FFT spectrogram
Delta mean	1.48 ± .48	.96 ± .21	<.01a
Delta variance	.68 ± .39	.22 ± .09	<.01a
Theta variance	.16 ± .09	.07 ± .04	<.01a
Peak envelope	36 ± 19	16 ± 7	.02a

Note: We extracted 14 quantitative EEG features per patient, which are described in more detail in the Materials and Methods section. To control for the differences in rates of DHC between those with PTE and those without, we evaluated decompressed hemispheres stratified by PTE, and nondecompressed hemispheres, as a separate subanalysis. This allows us to control for changes in power and amplitude from a cranial defect. For example, if a patient had a right-sided DHC, the right side with the cranial defect would be included in the “DHC” subanalysis, whereas the left side without a cranial defect would be included in the “No DHC” analysis. We adjusted p-values for false discovery rate using the method popularized by Benjamini and Hochberg.³⁷

Abbreviations: DHC, decompressive hemicraniectomy; EEG, electroencephalographic; FFT, fast Fourier transform; PTE, posttraumatic epilepsy.

^aStatistically significant.

FIGURE 2.

(A) Box plot of spectral power stratified by frequency band. The mean power was significantly different in the delta frequency, but similar for all other frequencies. (B) Delta power stratified by decompressive hemicraniectomy (DHC). To control for differing rates of DHC between patients with and without posttraumatic epilepsy, we compared the spectral power in the delta frequency in hemispheres with and without DHC. For example, for a patient who had a right-sided DHC, the right, decompressed hemisphere would be included in the “DHC” analysis, whereas the left, nondecompressed hemisphere would be included in the “No DHC” analysis. For patients who have not had a DHC, the hemispheres were average together and included in the “No DHC” analysis.

To assess our ability to predict the long-term risk of PTE and assess the relative importance of viEEG, qEEG, and clinical features, we developed a series of machine learning models using LOOCV (Table 5). Our first model (Model 1) using all 14 qEEG features had an AUC of .70 (95% CI = .55–.84). Model 2, our logistic regression model with the same qEEG feature set, had a similar AUC (.70; 95% CI = .57–.84, p = .95). The logistic regression model, after feature selection, included delta:theta power ratio (OR = 1.3 for a .1 increase in the ratio, 95% CI = 1.1–1.6, p < .01) and peak envelope (OR = 1.1, 95% CI = 1.0–1.2, p < .01). Both delta:theta ratio (1.9 for PTE [95% CI = 1.8–2.0] vs. 1.6 for no PTE [95% CI = 1.5–1.7], p < .001) and peak envelope were significant using descriptive statistics. Models 3 (AUC = .57, 95% CI = .42–.73) and 4 (AUC = .55, 95% CI = .40–.70) included all of the viEEG and clinical features, respectively. Neither model was able to discriminate risk of PTE better than chance (p > .36 for both).

TABLE 5.

Models.

Model	Features	Model	AUC	95% CI	p
1	qEEG	RF	.70	.55–.84
2	qEEG	Logistic	.70	.57–.84	.95
3	viEEG	RF	.57	.42–.73	.23
4	Clinical	RF	.55	.40–.70	.14
5	All	RF	.76	.63–.89	.06

Note: We developed a series of models to predict long-term risk of posttraumatic epilepsy. Models 1, 3, 4, and 5 were built using RF. Models 1, 3, and 4 used all of the qEEG, viEEG, and clinical features, respectively, in the dataset. The visual and clinical models lacked significant discrimination compared to chance (p > .36). Model 5 included all clinical and qEEG features significant using descriptive statistics (i.e., decompressive hemicraniectomy, delta mean power, delta variance, theta variance, and peak envelope). Model 2 was a logistic regression model, described further in the Materials and Methods section. To assess model performance, we compared all model performances using the AUC to our base, qEEG model. Model 5, including qEEG and clinical features, led to an improvement in performance that did not reach the preset threshold for statistical significance (p = .06). Of note, we were unable to score focal slowing for four patients due to technical/artifactual limitations, and Model 3 is build using 46 patients.

Abbreviations: AUC, area under the curve; CI, confidence interval; qEEG, quantitative electroencephalography; RF, random forest; viEEG, visual electroencephalography.

Our last model (Model 5) combined all clinical, visual, and quantitative EEG features significant using descriptive analysis. From Tables 3 and 4, this included no viEEG features, one clinical feature (DHC), and four qEEG features (delta power, delta variance, theta variance, and peak envelope). The AUC for Model 5 was .76 (95% CI = .63–.89), an improvement on Model 1 (qEEG features; AUC = .70), which did not reach the prespecified level for statistical significance (p = .06). We assessed the relative importance of various features in Model 5 using SHAP (Figure 3). Delta variance, delta mean power, and DHC were the three most important features for model performance.

FIGURE 3.

Shapley Additive Explanations (SHAP) summary plot and feature ranking. This plot demonstrates the relative importance of the various features in our Model 5, including significant quantitative electroencephalography and clinical features (i.e., decompressive hemicraniectomy [DHC], mean delta power, delta variance, theta variance, and peak envelope). The higher the SHAP value, the higher the risk of developing posttraumatic epilepsy. Model features are listed according to their relative importance as assessed by SHAP values, with delta variance as the most important feature. A dot is created for each feature attribution value for the model for each patient and smoothed to a violin shape. The violin is colored according to the values of the features, with red representing higher feature values and blue representing lower feature values.

4 |. DISCUSSION

In a cohort of long-term severe TBI survivors, we demonstrated that early EEG biomarkers, acquired within a few days of trauma, can accurately predict the long-term risk for developing PTE. Our results build upon prior literature utilizing early EEG features to prognose long-term mortality and functional outcomes in a variety of neurological diseases.^22–25,48 For TBI specifically, a variety of early qEEG features, including absolute power and variability, accurately predict mortality, although often at an early time point (i.e., discharge from hospital).^22,25,26,48 We extended these findings to predicting the long-term risk of PTE, a potentially treatable condition that lowers quality of life and increases health care burden.^49,50 Unlike early mortality prediction, our models identify risk of disease occurring months to years after initial trauma.² Although patients at high risk for PTE frequently have more severe injuries, our EEG biomarkers are not simply a reflection of injury severity, as our patients were clinically matched, survived 2 years posttrauma, and had similar functional outcomes at 1 and 2 years. Taken together, our modeling results show that machine learning approaches using early EEG biomarkers can discriminate the long-term risk of epilepsy. This finding has important clinical and translational implications. Early prediction may not only guide prognosis but also direct more prompt treatment. Ample evidence suggests that inherent CNS plasticity mechanisms are most adaptable during the early postinjury period, and therefore the very early identification of epilepsy risk creates an optimal window for antiepileptogenic therapeutics.⁵¹

We have also demonstrated the relative strength of early quantitative EEG analysis over the clinical gold standard of viEEG analysis in epilepsy outcome prediction. As evidenced by our findings, early qualitative viEEG analysis lacks significant discriminatory ability to predict long-term outcomes. This is likely the result of the lack of salient visual features on these early EEGs, which are heavily impacted by both anesthetic effects and artifacts such as breach rhythm caused by craniotomy. On the other hand, qEEG, and specifically spectral power analysis, can elucidate findings not readily apparent on standard visual analysis. Importantly, classic viEEG markers of epilepsy such as presence of epileptiform activity did not discriminate risk of PTE in our study. This result, albeit surprising, is likely related to the relatively low incidence of these features in early EEG analysis.

From a methodology perspective, we developed an innovative, yet simple, approach for analyzing qEEG features in severe TBI patients by controlling for cranial defects after neurosurgical craniectomies. DHC is a common treatment for elevations of intracranial pressure in severe TBI, occurring in 42% of our cohort, and strongly associated with clinical outcomes.^36,52,53 The resulting skull defect leads to the loss of dispersive filtering, changing the EEG signature in regions underlying the defect.⁵⁴ Previous qEEG prognostic efforts have largely ignored this important technical limitation, confounding prediction.^22,26 To account for the effects of skull defects, we performed a subanalysis of cerebral hemisphere-based skull defects from DHC, only comparing hemispheres with cranial defects to others with cranial defects and vice versa (Table 4, Figure 2B). This analysis yielded similar results after controlling for DHC cranial defects, highlighting the robust nature of our EEG biomarkers.

Building on our prior work, these results suggest that early EEG analysis may augment the predictive capabilities of early clinical data on the outcome of PTE.¹ The high incidence of PTE in severe TBI patients provides a unique opportunity to identify the risk of epilepsy, sometimes years in advance, in a cohort of patients who have yet to have an epileptic seizure. PTE risk modeling should not be limited to EEG analysis; we previously demonstrated that, similar to EEG, deep learning of cranial imaging can predict long-term outcomes in severe TBI.¹⁴ Future analyses will be aimed at incorporating multimodel clinical and radiologic data with early EEG biomarkers to optimize PTE risk prediction.

There are several potential limitations of the current study. We performed a retrospective case–control study, which can introduce bias. Despite this format, the larger TBI sample population from which this cohort was obtained was followed prospectively with standardized clinical assessments, thereby enriching the integrity of the data. Another possible limitation is the unequal distribution of DHC within the two patient populations, which we attempted to control for through our analytic approach. Lastly, our modeling efforts are limited by a low patient volume (n = 62) and high number of features, potentially leading to overfitting. Although 62 is a small number of patients compared to large-scale TBI clinical trials, this is the norm for studies of qEEG due to the data-rich nature of qEEG.^22,25 Our 1-Hz frequency band allowed us to collect each feature 300 times per patient over the 5-min EEG. Similarly, random forest modeling approaches are a commonly used machine learning tool in medicine specifically for their tendency to avoid overfitting.^55,56 Any model using EEG features, including this one, should be externally validated prior to clinical use.

5 |. CONCLUSIONS

In a cohort of severe TBI patients, quantitative early EEG features can distinguish the long-term risk of PTE. We illustrate the limitations of traditional viEEG analysis and provide a methodological innovation for craniotomy-related artifact. Our findings indicate early prognostication of PTE and present an opportunity for future antiepileptogenic interventions.

Key Points.

• Despite one third of severe TBI survivors developing posttraumatic epilepsy, no tools exist to identify patients at high risk for epilepsy

• We performed a case–control study to identify quantitative and qualitative electroencephalographic features that differed in patients with and without posttraumatic epilepsy

• In multivariable modeling using clinical and electroencephalographic features, a random forest model could predict risk of epilepsy with an area under the curve of .76, and logistic regression showed independent associations of delta:theta power ratio and peak envelope with posttraumatic epilepsy

• Very early electroencephalographic features can discriminate patients based on long-term risk of posttraumatic epilepsy

DATA AVAILABILITY STATEMENT

Data will be available to qualified researchers upon request.