Development of individualized risk assessment models for predicting post‐traumatic epilepsy 1 and 2 years after moderate‐to‐severe traumatic brain injury: A traumatic brain injury model system study

Abstract

Objective

Although traumatic brain injury (TBI) and post‐traumatic epilepsy (PTE) are common, there are no prospective models quantifying individual epilepsy risk after moderate‐to‐severe TBI (msTBI). We generated parsimonious prediction models to quantify individual epilepsy risk between acute inpatient rehabilitation for individuals 2 years after msTBI.

Methods

We used data from 6089 prospectively enrolled participants (≥16 years) in the TBI Model Systems National Database. Of these, 4126 individuals had complete seizure data collected over a 2‐year period post‐injury. We performed a case‐complete analysis to generate multiple prediction models using least absolute shrinkage and selection operator logistic regression. Baseline predictors were used to assess 2‐year seizure risk (Model 1). Then a 2‐year seizure risk was assessed excluding the acute care variables (Model 2). In addition, we generated prognostic models predicting new/recurrent seizures during Year 2 post‐msTBI (Model 3) and predicting new seizures only during Year 2 (Model 4). We assessed model sensitivity when keeping specificity ≥.60, area under the receiver‐operating characteristic curve (AUROC), and AUROC model performance through 5‐fold cross‐validation (CV).

Results

Model 1 (73.8% men, 44.1 ± 19.7 years, 76.1% moderate TBI) had a model sensitivity = 76.00% and average AUROC = .73 ± .02 in 5‐fold CV. Model 2 had a model sensitivity = 72.16% and average AUROC = .70 ± .02 in 5‐fold CV. Model 3 had a sensitivity = 86.63% and average AUROC = .84 ± .03 in 5‐fold CV. Model 4 had a sensitivity = 73.68% and average AUROC = .67 ± .03 in 5‐fold CV. Cranial surgeries, acute care seizures, intracranial fragments, and traumatic hemorrhages were consistent predictors across all models. Demographic and mental health variables contributed to some models. Simulated, clinical examples model individual PTE predictions.

Significance

Using information available, acute‐care, and year‐1 post‐injury data, parsimonious quantitative epilepsy prediction models following msTBI may facilitate timely evidence‐based PTE prognostication within a 2‐year period. We developed interactive web‐based tools for testing prediction model external validity among independent cohorts. Individualized PTE risk may inform clinical trial development/design and clinical decision support tools for this population.

Key points.

• Information available during admission and early acute care predicts post‐traumatic epilepsy (PTE) occurring between acute inpatient rehabilitation and 2 years after msTBI for adults (16 years and older).

• Cranial surgeries (including craniectomy and craniotomy), acute care seizures, intracranial fragments, and traumatic hemorrhages were leading PTE predictors in this cohort.

• A suite of interactive web‐based calculator tools has been developed for testing prediction model external validity among independent cohorts. Individualized prognosis for PTE risk may inform clinical trial development/design and clinical decision support tools for use with msTBI survivors.

1. INTRODUCTION

Epilepsy is a common complication arising from traumatic brain injury (TBI), and TBI survivors are 50 times more likely to die of a seizure than their non‐injured peers. ¹ Yet, limited tools exist to assess individual post‐traumatic epilepsy (PTE) risk to inform treatments and clinical decision‐making that facilitate prevention and management. Large scale clinical trials of PTE treatment are rare, ² but a systemic review and meta‐analysis suggest that phenytoin and levetiracetam can be administered for 7 days after moderate‐to‐severe (ms)TBI with similar efficacy, although with variable adverse effects. ³ , ⁴ No additional standardized guidelines exist for specific anti‐seizure medication (ASM) prophylactic use beyond the first week post‐TBI, ⁵ , ⁶ and extended ASM use clinically and in experimental models can confer detrimental effects. ⁷ , ⁸

Accurate PTE risk prediction could inform much needed clinical trials as well as clinical management, including ASM use and decisions concerning daily activities. Although previous work has identified prognostic variables, no tool exists to quantify individual PTE risk. Previously we generated a well‐performing multivariable logistic regression model to predict PTE status 1 and 2 years post‐TBI. ⁹ Significant predictors included subdural hematoma (SDH) and presence of intracranial fragments, craniotomy, craniectomy, seizure during acute hospitalization, and pre‐injury incarceration, among others. ⁹ Despite the large sample size, the number of individuals with seizures per variable modeled was relatively low, and overfitting was a relevant concern regarding external generalizability of model parameters to individual PTE risk. ¹⁰

Least absolute shrinkage and selection operator (LASSO) regression for feature selection addresses overfitting with low event rate data sets; thus, we developed prospective predictive models to quantify individual PTE risk occurring from inpatient rehabilitation (IPR) through Year 2 post‐injury, when 80% of PTE diagnoses are made. ¹¹ LASSO was used to “shrink” coefficients of less important predictors to zero, which reduces overfitting and facilitates translating complex prognostic models to simple, clinically viable tools, ¹² particularly for PTE clinical management scenarios where limiting false negatives is important. We developed four models predicting: Model 1, PTE by Year 2 post‐injury using baseline predictors readily available during acute care; Model 2, individual seizure risk rapidly to “enrich” study enrollment for randomized clinical trials (RCTs) in the acute care setting; Model 3, new/recurring seizure at Year 2 with proximal (Year 1) variables; and Model 4, new seizure at Year 2 with proximal variables for the adult TBI population (16 years and older).

2. METHODS

We followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) reporting guidelines for prognostic model development (eTable 1—Data S1).

2.1. Participants

Our eligible cohort included 6089 participants prospectively enrolled from 21 centers participating in the National Institute on Disability, Independent Living, and Rehabilitation Research TBI Model Systems (TBIMS) National Database Study. Each TBIMS center gathers baseline data spanning acute hospitalization through IPR. TBIMS centers enroll individuals meeting the following eligibility criteria: (≥16 years) with msTBI (post‐traumatic amnesia [PTA] >24 h, positive computed tomography [CT] scan findings, or Glasgow Coma Scale [GCS] score <13) who present to a TBIMS designated acute care hospital within 72‐h post‐injury and receive IPR from a designated TBIMS rehabilitation hospital. Participants are interviewed at 1, 2, 5 years and every subsequent 5 years post‐TBI. Participants, or legal proxies, provided written informed consent to participate. Each center's institutional review board approved data collection and analyses protocols, with data sharing overseen by the TBIMS National Data and Statistical Center (TBIMS NDSC). Data were gathered using standard TBIMS NDSC procedures. ¹³ Self‐reported measures were collected from participants or designated proxies. Baseline data were collected through medical record review and participant/proxy interview. Follow‐up data were collected via telephone or in‐person interview.

Included participants were injured between 2010 and 2018 with seizure‐event data available at Year 1 or Year 2 post‐TBI. A total of 5685 participants had all baseline predictors and seizure data on either of the first 2 years, and 404 participants were missing data for at least one baseline predictor. The derivation of the analytic sample is provided in Figure 1. ¹³

FIGURE 1.

Visual summary of cohort derivation for “Models 1 and 2”, “Model 3”, and “Model 4”. Individuals in “Models 1 and 2” were excluded if missing outcome or predictor data. Individuals in “Model 3” were excluded if missing Year 1 predictors. Individuals in “Model 4” were excluded if having a seizure in the first‐year post‐injury or if missing Year 1 predictors.

2.2. Predictor variable data collection

Demographic, pre‐injury, and TBI‐related clinical variables tested as predictors for all models are defined in eTable 2—Data S1. Greater injury severity, longer hospital length of stay (LOS), acute care seizure, positive CT findings, concurrent spinal cord injury, cranial surgeries (craniotomy/craniectomy, not including bolt or extraventricular drain placement), and history of any conditions listed in eTable 2—Data S1 were considered PTE “risk categories.” Variable selection was based on author expertise, biological rationale, prior literature, and data availability. Contusion load was categorized into five categories (0, 1, 2, 3, and 4+). Other Year 1 predictors, for Models 3 and 4, included all‐cause rehospitalization (excluding seizures) and reported alcohol/drug use during Year 1 post‐TBI. Model 3 also included Year 1 seizure occurrence as a predictor.

2.3. Post‐traumatic seizure data collection

Data collectors asked participants to self‐report (yes/no) if a physician told them that they have had a seizure since rehabilitation (Year 1) or over the past year (Year 2) as published previously. ⁹ Although the question is anchored temporally in reference to the time of last interview, it is otherwise similar to the epilepsy question in the National Health Information Survey (NHIS). ¹⁴ PTE was operationalized for this cohort as at least one seizure occurring between discharge from acute care and 2 years post‐injury. Acute care seizure was defined using International Classification of Diseases, Ninth and Tenth Revisions (ICD‐9/10 codes)—listed in eTable 2—Data S1.

2.4. Statistical analysis

All statistical data analysis was done using R (version 4.3.2). Descriptive statistics characterizing the study cohort by PTE status are provided overall and for each logistic regression model (eTables 2–4—Data S1).

2.4.1. LASSO regression overview

We employed group LASSO logistic regression ¹⁵ using a complete set of independent variables to build generalizable PTE risk prediction models that reduce overfitting. Over‐fitted models can underperform when applied to new data sets, thereby limiting clinical utility by misclassifying more persons than expected. LASSO reduces overfitting by using a regularization parameter (λ), to shrink the β‐coefficients of certain predictor variables to zero, excluding them from the model. ¹² Group LASSO excludes variables with a priori defined groups, which may include single or multiple predictors. Dummy variables created for different categories of a predictor were considered as the “group” so that all categories of a multi‐level categorical variable were kept if at least one dummy variable had a non‐zero coefficient.

2.4.2. Lambda selection and cross‐validation (CV)

When choosing λ, we performed repeated 10‐fold cross‐validation (CV) using 100 iterations, with equal proportion of PTE incidence across all 10 folds and 100 iterations. In each iteration, we fit group LASSO logistic regressions for a range of λ‐values on the training sets and obtained the area under the receiver‐operating characteristic curves (AUROCs) for the corresponding test sets. Each λ‐value yielded an average AUROC over the 10 test folds. In each iteration, we retained the λ‐value yielding the maximum shrinkage, such that the average AUROC was within 1 standard error (SE) of the maximum average AUROC, maximizing model parsimony within a range of acceptable prediction performance. The most common λ from the 100 repetitions of the 10‐fold CV was chosen as the shrinkage parameter for each group LASSO model using the corresponding full cohorts. These models yielded covariates with non‐zero coefficients. The 10‐fold CV ensured that the tuning parameters were chosen in an approximately unbiased manner. Parameter tuning was done similarly for all models, differing only in the participants and/or predictors included.

2.4.3. Logistic regression approach overview

Because LASSO model parameter estimates are inherently biased, ¹² final models were generated using logistic regression models containing only the predictors with non‐zero coefficients independently derived for each final group LASSO model. We also developed an R Shiny application to calculate PTE risk scores for each participant using the final logistic regression models. Model formulas are included in the app, which is available upon request for collaborative research. Using these formulae, we estimated PTE risk for every participant in our data set and plotted AUROCs to display sensitivity and specificity at all thresholds for estimated PTE risk and cohort‐specific PTE risk distribution density plots.

2.4.4. Logistic regression prediction model development and PTE case adjudication

We used complete case data from the 6089 cohort to generate four models that included baseline variables. Year 2 models considered baseline variables and additional Year 1 predictor variables (seizure event and rehospitalization in Year 1). Those lost to follow‐up or who died within the 2‐year period without having a seizure were excluded. The missingness definitions and adjudication process for each model are described below.

Model 1 (n = 4126) predicts PTE through Year 2 post‐TBI, using information collected during acute hospitalization. Model 2 (n = 4126) quantifies seizure risk during acute care, mirroring the classification window needed for determining participant eligibility to enroll into RCTs. This model excludes acute care seizure and acute care LOS as covariates (which would require the duration of the acute hospitalization to adjudicate). Participants with a post‐acute care seizure documented in either Year 1 or Year 2 but missing seizure status in the other year were considered PTE cases for both Models 1 and 2. Those without a seizure in Year 1 and missing seizure data in Year 2 were considered missing, as we could not adjudicate if missing data included a seizure event. Model 3 (n = 3788) included participants with complete baseline data, along with Year 1 covariates and Year 1 seizure status, to predict seizure events in Year 2. Seizure status in Year 2 was defined as new PTE or recurrent seizure event in Year 2. Participants missing post‐acute care seizure information in either Year 1 or Year 2 and without seizure event reported in the other year were considered missing. Model 4 (n = 3333) fit a model predicting only new in Year 2 seizure (PTE) using baseline and Year 1 covariates. New seizure was defined as no seizure in Year 1 but with a Year 2 (new) seizure.

2.4.5. Logistic regression prediction model internal validation

We evaluated overall performance, stability, and internal validity for all models through 5‐fold CV. We set a decision threshold with specificity of at least 60% to balance it with a high negative predictive value (NPV), ensuring a high true‐negative rate, which is crucial for many aspects of clinical decision‐making when managing PTE risk. Sensitivity, specificity, and AUROC were assessed for each iteration of the 5‐fold validation procedure. AUROC, sensitivity, specificity, positive predictive value (PPV), and NPV were averaged across the five test folds. Information from each participant with complete data was used once in a test set. β‐coefficients for final models were generated using the full data set and used to generate model equations used in the online app.

3. RESULTS

The baseline cohort included primarily men (73.8%) with an average age of 44.1 years (SD = 19.7). Severe TBI accounted for 76.1% of injuries. Table 1 shows that 916 (22.2%) had a seizure after acute care hospitalization and within 2 years post‐injury.

TABLE 1.

Patient characteristics of primary baseline model cohort.

	No seizures N = 3210	Seizures N = 916	p‐value a
Sex, n (col %)
Female	881 (27.4%)	201 (21.9%)	.001
Male	2329 (72.6%)	715 (78.1%)
Age, mean (SD)	45.0 (20.1)	41.0 (17.9)	<.001
Length of Acute Stay, mean (SD)	19.6 (17.4)	24.7 (23.6)	<.001
Previous Moderate‐to‐Severe TBI, n (col %)
No	3134 (97.6%)	882 (96.3%)	.035
Previous TBI	76 (2.37%)	34 (3.71%)
Acute Seizure Events, n (col %)
No	2985 (93.0%)	729 (79.6%)	<.001
Yes	225 (7.01%)	187 (20.4%)
Alcohol Present at Time of Injury, n (col %)
No	2377 (74.0%)	625 (68.2%)	.001
Yes	833 (26.0%)	291 (31.8%)
Previous Stroke (Bleeding), n (col %)
No	2918 (90.9%)	823 (89.8%)	.365
Yes	292 (9.10%)	93 (10.2%)
Previous Stroke (Ischemic), n (col %)
No	3112 (96.9%)	888 (96.9%)	>.999
Yes	98 (3.05%)	28 (3.06%)
Diagnosed Neurodegenerative Disease, n (col %)
No	3135 (97.7%)	901 (98.4%)	.251
Yes	75 (2.34%)	15 (1.64%)
Pre‐morbid Headaches, n (col %)
No	3173 (98.8%)	907 (99.0%)	.799
Yes	37 (1.15%)	9 (.98%)
SAH
No	933 (29.1%)	198 (21.6%)	<.001
Yes	2277 (70.9%)	718 (78.4%)
Pre‐injury Suicide Attempt, n (col %)
No	3061 (95.4%)	853 (93.1%)	.009
Yes	149 (4.64%)	63 (6.88%)
Pre‐injury Incarceration, n (col %)
No	2946 (91.8%)	798 (87.1%)	<.001
Yes	264 (8.22%)	118 (12.9%)
Previous Military Service, n (col %)
No	2797 (87.1%)	822 (89.7%)	.039
Yes	413 (12.9%)	94 (10.3%)
Fragments Present in CT, n (col %)
No	3042 (94.8%)	804 (87.8%)	<.001
Yes	168 (5.23%)	112 (12.2%)
EDH, n (col %)
No	2867 (89.3%)	772 (84.3%)	<.001
Yes	343 (10.7%)	144 (15.7%)
SDH, n (col %)
No	1576 (49.1%)	300 (32.8%)	<.001
Yes	1634 (50.9%)	616 (67.2%)
TBI Severity, n (col %)
Severe	835 (26.0%)	152 (16.6%)	<.001
Moderate	2375 (74.0%)	764 (83.4%)
Craniotomy, n (col %)
No	2778 (86.5%)	699 (76.3%)	<.001
Yes	432 (13.5%)	217 (23.7%)
Craniectomy, n (col %)
No	2892 (90.1%)	609 (66.5%)	<.001
Yes	318 (9.91%)	307 (33.5%)
Drugs Present, n (col %)
No	2633 (82.0%)	676 (73.8%)	<.001
Yes	577 (18.0%)	240 (26.2%)
Previous Military Combat, n (col %)
No	2490 (77.6%)	690 (75.3%)	.168
Yes	720 (22.4%)	226 (24.7%)
Spinal Cord Injury, n (col %)
No	3014 (93.9%)	880 (96.1%)	.015
Yes	196 (6.11%)	36 (3.93%)
Contusion Load, n (col %)
0	1136 (35.4%)	193 (21.1%)	<.001
1	676 (21.1%)	157 (17.1%)
2	648 (20.2%)	224 (24.5%)
3	384 (12.0%)	166 (18.1%)
4+	366 (11.4%)	176 (19.2%)
History of Mental Health Disorder, n (col %)
No	2490 (77.6%)	690 (75.3%)	.168
Yes	720 (22.4%)	226 (24.7%)
Previous Psychiatric Hospitalization, n (col %)
Yes	3010 (93.8%)	820 (89.5%)	<.001
No	200 (6.23%)	96 (10.5%)
Race, n (col %)
White	2184 (68.0%)	555 (60.6%)	<.001
African American	456 (14.2%)	184 (20.1%)
Other	570 (17.8%)	177 (19.3%)

^a p‐values for continuous variables were based on two‐sample t‐tests and p‐values for categorical variables were based on chi‐square tests.

3.1. Model 1

A λ = .01769 was selected as the tuning parameter for LASSO‐derived variable selection. For Model 1, regression β‐coefficients for Model 1 baseline PTE predictors selected at this λ are presented in Table 2A and include craniectomy, craniotomy, intracranial fragments, contusions, SDH, SAH, acute hospitalization seizure, pre‐injury psychiatric hospitalization history, prior incarceration, pre‐injury drug use, alcohol at hospitalization, and acute care LOS.

TABLE 2.

Logistic regression coefficients for LASSO‐selected variables.

(A) 2‐year cumulative seizure model		(B) RCT model (excludes acute seizure and LOS as predictors)		(C) Year 2 seizure model		(D) Year 2 new seizure incidence model
Predictor	Beta coefficient	Predictor	Beta coefficient	Predictor	Beta coefficient	Predictor	Beta coefficient
Craniectomy	1.192	Craniectomy	1.232	Yr‐1 Seizure Incidence	3.067	Craniectomy	1.047
Acute Seizure	1.172	Intracranial Fragments	.321	Craniectomy	.981	Acute Seizure	.892
Intracranial Fragments	.364	Craniotomy	.544	Acute Seizure	.684	Preinjury Drug Use	.653
Craniotomy	.496	SDH	.362	Intracranial Fragments	.434	Craniotomy	.563
SDH	.359	Pre‐injury Psychiatric Hospitalization	.530	Preinjury Drug Use	.629	Contusion Load (1) a	<.001
SAH	.114	Contusion Load (1) a	.230	Craniotomy	.414	Contusion Load (2) a	.603
Contusion Load (1) a	.212	Contusion Load (2) a	.491	Acute care LOS	.007	Contusion Load (3) a	.846
Contusion Load (2) a	.452	Contusion Load (3) a	.638	(Intercept)	−3.371	Contusion Load (4+) a	.866
Contusion Load (3) a	.556	Contusion Load (4+) a	.668			Acute‐Care LOS	.005
Contusion Load (4+) a	.574	Prior Incarceration	.327			(Intercept)	−3.801
Pre‐injury Psychiatric Hospitalization	.482	Preinjury Drug Use	.231
Prior Incarceration	.427	Race (African American)	.381
Preinjury Drug Use	.263	Race (Other)	<.001
Injury severity	.171	Alcohol at Hospitalization	.209
Acute care Length of Stay	.003	Injury severity	.206
Alcohol at Hospitalization	.188	(Intercept)	−2.607
(Intercept)	−2.773

Note: Description: Table of logistic regression model coefficients for LASSO‐selected variables.

Abbreviations: CT, computed tomography; CV, cross‐validation; EDH, epidural hematoma; LASSO, Least Absolute Shrinkage and Selection Operator; LOS, length of stay; RCT, randomized controlled trial; SCI, spinal cord injury; SDH, subdural hematoma; SAH, subarachnoid hemorrhage; TBI, traumatic brain injury.

^a Contusion load (number) represents the number of contusions noted on CT scan (0 = no contusion, reference, 1 = 1, 2 = 2, 3 = 3, 4 = 4+ contusions).

The AUROC for Model 1 (Figure 2; left column) was .74. The 5‐fold CV demonstrated an average AUROC of .73 ± .02 across the 5 test folds—a comparatively more precise performance metric than the full model's AUROC value, demonstrating acceptable predictive model accuracy to categorize new cases. CV performance metrics are available in Table 3A.

FIGURE 2.

Left column: Receiver‐operating characteristic (ROC) curves derived from logistic regression model fits with variables selected via the Least Absolute Shrinkage and Selection Operator (LASSO) regression model. Right column: Density plots of predicted probabilities of seizure events. The dashed line represents the distribution of probabilities among individuals who experienced a seizure event (PTE: post‐traumatic epilepsy), whereas the solid line represents the distribution of probabilities for those who did not (no PTE). The dark vertical dashed line represents the classification threshold. (A) Model 1, (B) Model 2, (C) Model 3, and (D) Model 4.

TABLE 3.

Internal validation (5‐fold cross‐validation) results for logistic regression.

Fold number	Validation specificity	Validation sensitivity	Validation NPV	Validation PPV	Validation AUC
(a) Baseline model
1	.6106	.7609	.8991	.359	.7382
2	.6075	.7049	.8784	.3386	.7311
3	.6137	.765	.9016	.3608	.7539
4	.6153	.7322	.8896	.3517	.7237
5	.5483	.765	.8911	.3256	.711
Mean	.5991	.7456	.8920	.3471	.7316
SD	.0273	.0246	.0084	.0142	.0163
(b) RCT model excluding acute seizure and LOS as predictor
1	.5966	.6902	.8705	.3290	.6866
2	.6558	.6393	.8645	.3462	.7135
3	.5981	.7268	.8848	.3402	.7207
4	.6262	.6721	.8701	.3388	.7032
5	.5685	.6940	.8670	.3144	.6772
Mean	.6090	.6845	.8714	.3337	.7002
SD	.0326	.0315	.0073	.0120	.0166
(c) Second year model (new and recurring seizure model)
1	.6292	.8614	.9673	.2628	.8648
2	.6261	.8100	.9559	.2477	.8227
3	.6484	.7900	.953	.2548	.8001
4	.6849	.8200	.9615	.2837	.8679
5	.6819	.8400	.9655	.2867	.8393
Mean	.6541	.8243	.9606	.2671	.8390
SD	.0265	.0273	.0064	.0158	.0257
(d) Second year new seizure incidence model
1	.6116	.7949	.9775	.123	.7705
2	.7047	.4737	.9525	.0968	.6052
3	.6169	.7895	.9777	.1210	.7459
4	.6250	.7368	.9726	.1162	.7247
5	.7007	.5263	.9567	.1053	.6642
Mean	.6518	.6642	.9674	.1125	.7021
SD	.0421	.1345	.0102	.0094	.0642

Note: Description: Summary of test‐data performance in 5‐fold cross‐validation of logistic regression models based on LASSO‐selected variables. Values reported are averages of each performance metric across the 5 test folds.

Abbreviations: AUC, area under the curve; NPV, negative predictive value; PPV, positive predictive value.

A PTE prediction probability threshold was set for Model 1 at .1702 to maximize sensitivity while fixing specificity at ≥.6, accurately predicting 697 of 916 participants (76.1%) with PTE (219 false negatives). Of 1980 participants predicted to have seizures, 697 were true positives (PPV = 35.2%), leaving 1283 false positives, an expected finding given the low seizure incidence. Of the 2146 participants predicted to not develop seizures, 1927 were true negatives (NPV = 89.9%). PTE risk by PTE status is described in eTable 5a—Data S1 and Figure 2A. PTE participants demonstrated a tri‐modal distribution of predicted PTE risk using the LASSO‐selected predictors including cranial surgeries, acute care seizure, intracranial fragments, subdural hematoma, subarachnoid hemorrhage, contusions, prior incarceration, preinjury drug use, injury severity, acute care LOS, and alcohol use at hospitalization. The initial peak for each group is separated at a PTE probability of ~.17; a secondary peak occurs at an estimated probability near .40.

3.2. Model 2

Model 2 considered all baseline variables except acute care seizure and LOS for use in RCTs. We used λ = .01769 as the tuning parameter and a classification threshold of .1824. Coefficients resulting from the fitted logistic regression model using this λ are presented in Table 2B. The AUROC for Model 2 (Figure 2 left column) was .72. In 5‐fold CV, the model demonstrated an average AUROC = .70 ± .02 (Table 3B). Predicted vs observed PTE status is presented in eTable 5B—Data S1. The density curves of estimated PTE risk for participants with/without PTE are illustrated in Figure 2 right column. LASSO‐selected predictors include cranial surgeries, intracranial fragments, subdural hematoma, hemorrhage, contusions, prior incarceration, preinjury drug use, race, injury severity, and alcohol use at hospitalization.

3.3. Model 3

Of the 4126 participants in the first two models, 3788 individuals with both Year 1 and Year 2 seizure status and Year 1 variables (rehospitalization, substance abuse) were included in Model 3 to predict Year 2 new and recurring seizure prevalence. Here, 209 (5.5%) had a first‐time seizure in Year 2, and 292 (7.71% among all and 64.18% among participants with Year 1 seizure) had recurring seizure (eTable 3—Data S1).

We selected a λ = .01676 and a classification threshold of .0551. LASSO‐selected predictors after β‐coefficient shrinkage include Year 1 seizure incidence, acute care seizure, cranial surgeries, intracranial fragments, preinjury drug use, race, injury severity, and acute care LOS (Table 2C). Model 3 had an AUROC = .86, sensitivity = .87, and specificity = .60 (Figure 2 left column). Five‐fold CV test sets had an average AUC = .84 ± .03, sensitivity = .82 ± .03, and specificity = .65 ± .03 (Table 3C). Of 1739 individuals predicted to have seizures, 434 were true positives (PPV = 25.0%). Of 1930 predicted to be seizure‐free, 1875 were true negatives (NPV = 96.7%). eTable 5C—Data S1 shows PTE prediction vs observed PTE status for this model, and Figure 2 right column shows a density curve for estimated Year 2 PTE risk.

3.4. Model 4

After excluding 455 individuals with Year 1 seizures from Model 3, a total of 3333 were included in Model 4 (eTable 4—Data S1). We selected a λ = .0115 and a classification threshold of .05038. Acute care seizure, cranial surgeries, pre‐injury drug use, contusion load, and acute care LOS were retained after LASSO β‐coefficient shrinkage (Table 2D). Model 4 had an AUROC = .73, sensitivity = .74, and specificity = .60. The AUROCs for Model 4 are presented in Figure 2 (left column). Test data sets in 5‐fold CV had a mean AUROC = .70 ± .07, sensitivity = .66 ± .15, and specificity = .65 ± .05 (Table 3D). Model 4 PTE prediction vs observed PTE status and corresponding PPVs and NPVs are presented in eTable 5D—Data S1. Finally, the density curves for estimated Year 2 PTE risk are depicted in Figure 2 (right column).

3.5. Sensitivity analysis

We evaluated selection characteristics associated with each sample, and there were few clinically meaningful differences between the analytic samples and those included/excluded (eTables 6A–D—Data S1). Individuals included for Models 1 and 2 were slightly younger, with higher SAH, SDH, contusion burden, mental health history, craniectomy rates, and White race representation. Although statistically significant, differences across these variables were modest, showing only a >5% difference in the frequency of individuals having no cerebral contusions and a difference in age that was <10% of the SD for both groups. No variables were meaningfully different in Model 3. Acute care seizure and craniectomy frequencies were >5% in Model 4 but <10%. Model covariate missingness was not associated with seizure rate. In addition, we assessed data missingness by center and did not observe any major differences (data not shown).

4. DISCUSSION

To date, there are few published examples of prognostic model development for PTE, ⁹ , ¹⁶ and clinicians cannot easily translate population‐based risk estimates identified from contemporary epidemiological studies, ¹¹ , ¹⁷ or apply identified clinical features from smaller clinical studies to individual patient care, leaving a major gap in both research and clinical care in supporting clinical decision‐making and facilitating the clinical trial pipeline for this population. As a result, systematic reviews and treatment guidelines for the PTE field are few and represent limited advancements in clinical care. ¹⁸ , ¹⁹ Our individualized, quantitative PTE risk prognostication tool begin to address this important research gap, with the goal of advancing treatment trials and personalizing clinical decision support for this population. One small study reports a risk calculator tool for early post‐traumatic seizures (PTS), ¹⁶ yet no work has generated a predictive model to calculate individual PTE risk post‐hospitalization. PTE risk prediction for TBI survivors is complicated by variable injury types, severity, and other factors. A quantitative prognostic model can provide clinical outcome probability estimates that inform and personalize clinical decision‐making. PTE risk calculation may inform seizure‐threshold lowering medication use and dosage, work/driving restrictions, and lifestyle education. Future external validation may support model use in clinical decision support.

The primary PTE risk factors identified here are congruent with our previous prognostic models. There is a large association across all models between neurosurgical procedures and PTE, ⁹ , ¹¹ , ²⁰ , ²¹ which may contribute to epileptogenesis via infection and/or glial scarring around the surgical site. Craniectomy may also be a surrogate marker for persistent focal structural or electrophysiological brain injury and a proxy for overall injury severity. Acute care seizures were another influential covariate predictor in three of four models and the top predictor in Model 3. The association between acute care seizures and PTE is complex, reflecting underlying acute pathophysiology cascades (e.g., excitotoxicity, neuroinhibition, neuroinflammation) that affect both TBI and epileptogenesis. These relationships may also be genetics dependent. ²² , ²³ , ²⁴ , ²⁵ , ²⁶ , ²⁷ Although we do not have the data granularity available to differentiate immediate/early/late seizures during acute care, our prior TBIMS work suggests that immediate seizures may increase PTE risk after acute care discharge. ¹¹ , ²⁸ Like previous work, ⁹ SDH and contusions carried higher PTE risk, likely due to damage from intraparenchymal blood and tissue compression. Intracranial fragments may exacerbate PTE risk by compromising dura and contributing to intracerebral blood. ²⁰ , ²⁹

Like prior work, our findings demonstrate the importance of premorbid mental health (MH) when assessing PTE risk. ⁹ , ³⁰ MH conditions, which affect over half of individuals with msTBI, ³¹ and may be linked to post‐TBI inflammatory burden, ³² should be considered when weighing the risks and benefits of psychotropic medications. However, premorbid and post‐injury psychiatric disease can lead to pharmacotherapeutic treatments that lower seizure threshold. ³³ Some selective serotonin reuptake inhibitors, vs other antidepressant classes, exhibit anticonvulsant properties. ³⁴ Although medications are often clinically necessary, individual PTE risk may guide drug class selection. Because mood disturbances and behavioral side effects can occur with ASM use, ³⁵ limiting ongoing prophylactic use for individuals with low PTE risk is important. Psychogenic non‐epileptic seizures associated with MH are more common in trauma populations vs the general population, ³⁶ and they are challenging to discern from organic seizures via self‐report, which may confound the true role of MH history in PTE risk. However, pre‐injury drug and active alcohol use at the time of hospitalization are significant factors relevant to MH and secondary injury for which epidemiological studies ³⁷ , ³⁸ , ³⁹ , ⁴⁰ support a contributing role. Similar to our previous work involving an earlier cohort of individuals with msTBI, ⁹ non‐White race was a PTE risk factor in some models. Although complex, this finding highlights the need to examine multiple social determinants of health (SDOH) affecting risk for PTE. ⁴¹ , ⁴² Although factors impacting access to acute care (e.g., neurosurgical procedures, advanced intensive care unit [ICU] care, and electroencephalography [EEG] monitoring) are available within TBIMS systems of care, access barriers could be relevant to long‐term follow‐up with TBI specialty care, similar to what is observed in the general population with epilepsy. ⁴¹ , ⁴²

Model 2 required only information available during acute hospitalization for rapid seizure risk adjudication, even prior to acute care seizure onset. Although this model is not as high‐performing as Model 1, it may identify individuals with “high” and “low” PTE risk for RCT enrollment and/or stratification in observational studies. Models 3 and 4 use baseline data, rehospitalization during Year 1, and substance use to predict first‐time and new/recurrent Year 2 seizures with moderate success despite low event rates. Of interest, preinjury alcohol use becomes a significant variable in Year 2 vs Year 1 models, whereas the effect of pre‐injury drug use increases in effect size for Year 2 vs Year 1 models, indicating possible re‐engagement in drug use post‐injury as a potentially modifiable risk factor. Once validated, these models could be useful for clinical surveillance, ongoing PTE prevention efforts, research advancing patient selection and treatment guidelines regarding PTE prophylaxis, ²⁹ , ⁴³ and clinical decision‐making when considering the use of medication that lowers seizure thresholds ⁴⁴ and safe return to driving. ⁴⁵

Our group LASSO penalty models for variable selection support individual‐level risk calculations and have pragmatic application potential by retaining only the strongest, most relevant predictors. We favored a maximally sparse model at the expense of small model performance reductions to avoid overfitting. Our model includes easily obtainable variables that do not require advanced imaging, which may be unavailable in rural and under‐resourced facilities or within far‐forward military hospitals.

We fixed specificity at 60% to obtain a moderate specificity and sufficient sensitivity to generate high NPV. We did not want to miss possible PTE cases when considering questions like who should refrain from driving or who might require additional ASM prophylaxis. With low PTE rates, we hypothesized that setting the specificity at <60% would not be informative, as many more true negatives would accumulate compared to true positives.

Our models correctly identified ~70%–85% of individuals with PTE, yet also had false positives (due to low event rates). Thus clinical decision rules focused on risk mitigation may be somewhat conservative when using these models to inform decisions like when to wean ASMs or prescribe medications that lower seizure threshold to prevent treatment‐associated cognitive deficits when using phenytoin post‐injury. ⁷ , ⁸ Recent clinical trials, practice patterns, and meta‐analyses favor levetiracetam over phenytoin, ⁴⁶ , ⁴⁷ , ⁴⁸ with experimental studies also suggesting that levetiracetam has neuroprotective and anti‐epileptic benefits. ⁴⁹ , ⁵⁰ Yet accurate quantitative approaches for estimating individual PTE risk are critical to facilitating seminal studies that further refine prophylaxis selection and treatment guidelines, allowing practitioners to limit ASMs to only persons at high PTE risk, a practice that could benefit long‐term recovery.

Future directions include external validation and model refinement to include recent cases with new variables capturing pre‐morbid epilepsy and more granular acute care seizure variables operationalized to meet International League Against Epilepsy (ILAE) definitions. We created a suite of PTE risk/classification calculators to support external validation; other users can access the app by request to collaborate and contribute data for external validation purposes. We will develop and test imputation strategies for accommodating missing data in external data sets using this approach. The individual calculator panels demonstrate how the app inherently uses the probability distributions shown in Figure 3 to generate individualized risk assessments. Users can set probability thresholds for categorizing PTE status and calculating sensitivity and specificity. Furthermore, the calculator can export individual or group data in a de‐identified format that is compatible with most database platforms.

FIGURE 3.

R Shiny App Layout of Variable Input Dashboard and Primary Model Post‐Traumatic Epilepsy (PTE) Risk/Classification Calculator with Example “Patient” Use Cases in PTE Risk/Classification Calculator. The input dashboard (left) allows users to select the specific risk factors relevant to any patient with moderate to severe traumatic brain injury (TBI). Users can manually set a threshold value for PTE classification (center). Here the threshold probability for prospective classification of seizure status is set at a .17 to correspond to a specificity threshold of .60 for the primary model. A visual summary is provided displaying each individual's PTE risk relative to the threshold value and with respect to the population distribution for those with and without PTE. (A) Tabular and visual summary (from online application) of predicted post‐traumatic seizures (PTS) probability in a White patient with a present SDH, intracranial fragments, 5 days of acute length of stay (LOS), and a history of pre‐injury drug use. Here the predicted probability of PTE narrowly exceeds the threshold probability of .170 (where specificity is set at .60 and indicated by the dashed vertical line), and the patient is classified prospectively as “at risk” for seizure. When the predicted probability marginally exceeds the threshold, there is a greater chance of misclassification. (B) For the same patient with a single contusion identified in computed tomograph (CT) imaging, the predicted probability of PTE easily exceeds the threshold probability of .17, and the patient would be classified prospectively as “at risk” for seizure. (C) If the same patient had a craniectomy procedure, the predicted probability of PTE far exceeds the threshold probability of .17. At this probability, the likelihood of this classification being a “false positive” is much lower.

Once externally validated, companion clinical decision rules can be developed and tested that incorporate the PTE calculator to avoid misclassifying individuals who do go on to have a seizure. In addition, inaccurately classifying someone as at risk could affect prescribing and needlessly restrict daily activities post‐injury. Thus implementation research will be needed to assess calculator utility with varied case‐use scenarios, including return to driving where safety and legal requirements may affect clinical management and involve patient/caregiver education. ⁵¹

5. LIMITATIONS AND CONCLUSIONS

Study limitations include a lack of potentially important PTE risk variables in the TBIMS database, including EEG, pre‐injury epilepsy history, ASMs, and MRI data. Our post‐acute seizure data are self‐reported but do incorporate physician affirmation of seizure within the question format. The study does not address pseudo‐seizures, epilepsy severity, exacerbating factors, or treatment response. Our findings could be subject to recall bias, although interviews do occur with the help of a caregiver whenever possible. We could not differentiate acute care seizure timing with the data available. Sensitivity analyses for those included and not included showed only a few variables with modest clinically meaningful differences. However, the impact of these differences on seizure rates is uncertain given our inability to ascertain PTE status. Although not available here, our previous work shows the potential benefit of considering personal biology with risk prediction. ²² , ²³ , ²⁴ , ²⁵ , ²⁶ , ²⁷ Other limitations include delineating hospital conditions via ICD codes alone. Further work will involve SDOH and other social/structural limitations and characteristics affecting seizure risk for individuals with msTBI. In addition, future work should evaluate characteristics that impact access to IPR and how these differences may impact PTE risk. Although less clinically interpretable, machine learning techniques like tree‐based models or neural networks may improve overall classification rates.

The TBIMS population is considered representative of the national IPR sample. ⁵² , ⁵³ However, the findings may not generalize as well to msTBI survivors who do not receive IPR. This population is necessarily restricted to individuals surviving to acute care discharge, a time frame in which the most mortality occurs after msTBI. ⁵⁴ In addition, we restricted our sample to participants who survived and were followed to Year 2 for seizure adjudication, a timeframe that our prior work suggests >80% of initial late PTS occurs. ¹⁹ Our results may not fully generalize to those lost to follow‐up. However, our sensitivity analysis showed that those excluded from model development (~10% of the case complete cohort) and the case complete cohort, were largely similar in demographic and clinical characteristics, apart from acute care seizure in the Year 2 models. Once externally validated, the prognostic models developed may be generalizable to the broader population of those surviving msTBI and requiring IPR services.

Despite these limitations, we developed satisfactory, parsimonious models to predict individualized PTE risk and introduce a free prognostic web application available to collaborating researchers interested in supporting external validation testing. Over the long term, we hope the study findings will pioneer innovative research on PTE prevention and management.

AUTHOR CONTRIBUTIONS

Data integrity and access: Nabil Awan and Amy Wagner had full access to the study data, and they take responsibility for the integrity of the data and the accuracy of the data analysis. Concept and design: Amy Wagner and Nabil Awan. Acquisition, analysis, or interpretation of data: Amy Wagner, Nabil Awan, Robert Krafty, Raj Kumar, and Shannon Juengst. Drafting of the manuscript: Amy Wagner, Nabil Awan, and Dominic DiSanto. Critical revision of the manuscript for important intellectual content: Amy Wagner, Nabil Awan, Robert Krafty, Raj Kumar, Shannon Juengst, Jerzy Szaflarski, William Walker, Kristen Dams‐O'Connor, Ross Zafonte, Mary Jo Pugh, and Cynthia Harrison‐Felix. Statistical analysis: Nabil Awan, Robert Krafty, and Raj Kumar.

FUNDING INFORMATION

This manuscript was developed with support from the U.S. Department of Defense (DoD) W81XWH1810736 (A.K.W., M.J.P., R.T.K., N.A., D.D.), VA Health Services Research and Development Service Grant Number IK6HX002608 (M.J.P.), and the National Institute on Disability, Independent Living, and Rehabilitation Research [NIDILRR grants 90DP0041 (A.K.W.); 90DPTB0009 (K.D.O'C., R.G.K.); 90DP0033 (W.C.W.); 90DP0084 (C.H.F.); 90DPTB0011 (R.Z.) 90DTB0013, 90DTB0025 (S.B.J.)]; NIDILRR is a Center within the Administration for Community Living (ACL), U.S. Department of Health and Human Services (HHS). The manuscript contents do not necessarily represent the policy of NIDILRR, ACL, HHS, or the Department of Veterans Affairs, and should not be taken as an endorsement by the Federal Government. The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and the decision to submit the manuscript for publication.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

ETHICS STATEMENT

We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

Supporting information

Data S1.

Awan N, Kumar RG, Juengst SB, DiSanto D, Harrison‐Felix C, Dams‐O’Connor K, et al. Development of individualized risk assessment models for predicting post‐traumatic epilepsy 1 and 2 years after moderate‐to‐severe traumatic brain injury: A traumatic brain injury model system study. Epilepsia. 2025;66:482–498. 10.1111/epi.18210

DATA AVAILABILITY STATEMENT

Data is available upon request to the corresponding author.