Biomarkers for Post-Traumatic Epilepsy: Advances in Imaging, Molecular Signatures, and AI-Assisted Prediction

Abstract

Early diagnosis of post-traumatic epilepsy (PTE) is crucial for timely intervention. However, it is hampered by the lack of reliable biomarkers. In this review, we provide a comprehensive summary of current advances in PTE biomarker research, drawing primarily on evidence from human cohort studies, with selective support from experimental animal models where mechanistic insights are required. We cover (i) neuroimaging, including CT, MRI, and EEG/qEEG, which reveal structural and functional alterations associated with epileptogenesis; (ii) molecular biomarkers, including RNAs, proteins, metabolites, and extracellular vesicle (EV)-derived molecules that reflect neuroinflammation, blood–brain barrier (BBB) dysfunction, neuronal injury, and synaptic remodeling; and (iii) artificial intelligence (AI)-assisted approaches, which integrate multimodal datasets to identify complex predictive patterns. While individual modalities offer valuable but incomplete prognostic information, AI-driven analytics hold the greatest promise for early predictive power by combining multimodal data. Future progress will depend on the integration of high-resolution imaging, multi-omics profiling, and rigorous validation to deliver clinically actionable biomarker panels and ultimately reduce the burden of PTE.

1. Introduction

Post-traumatic epilepsy (PTE) is a major long-term complication of traumatic brain injury (TBI) and represents an important neurological and public health burden. Epilepsy is a common neurological disorder characterized by recurrent, unprovoked seizures. It affects 65–70 million people worldwide, imposing substantial social and economic burdens [1,2,3]. Traumatic brain injury itself is highly prevalent, affecting an estimated 2.5 million individuals annually in both Europe and the United States, and the risk of epilepsy increases in parallel with injury severity, rising from about 2 to 4-fold after mild TBI to as high as 16-fold after severe TBI [4,5]. In penetrating TBI, the incidence is even greater, with epilepsy developing in up to 53% of patients [6,7]. Approximately 50% of epilepsy cases are linked to structural brain abnormalities, with TBI representing one of the major acquired causes [8]. Overall, PTE is estimated to account for approximately 5% of all epilepsies and up to 20% of structural or symptomatic epilepsies [1,3]. Among seizures occurring within the first 4 weeks after TBI, nearly 90% arise during the first week, underscoring the importance of early post-injury monitoring [9]. Given its association with chronic neurological disability and reduced quality of life, early identification of patients at greatest risk remains a critical clinical objective.

Early prediction of PTE remains difficult because epileptogenesis after TBI is biologically complex and clinically heterogeneous. A defining challenge is the latent interval between the initial traumatic insult and the later emergence of unprovoked seizures, during which progressive molecular, cellular, and network-level alterations evolve but often remain clinically silent [10]. Importantly, not all individuals with apparently similar injuries develop epilepsy, indicating that risk is shaped by multiple interacting factors rather than by injury severity alone. Clinical studies have shown that seizure risk is influenced by trauma type, lesion burden, injury location, severity, and the occurrence of early post-traumatic seizures [6,7]. At the mechanistic level, post-traumatic epileptogenesis is driven by interconnected processes that include neuroinflammation, blood–brain barrier (BBB) dysfunction, astrocytic and glial dysregulation, synaptic remodeling, and reorganization of neuronal networks [10,11]. Together, these features explain why conventional clinical variables alone are insufficient for precise individualized prediction of PTE.

Biomarker research has therefore emerged as a key strategy to improve risk stratification and to capture structural, functional, and molecular signatures of post-traumatic epileptogenesis. In principle, biomarkers can indicate pathogenic processes, disease susceptibility, progression, or response to intervention and may include molecular, radiologic, and physiologic measures [2,12]. In PTE, neuroimaging modalities such as computed tomography and magnetic resonance imaging can reveal lesion burden, hemorrhagic injury, microstructural damage, and other structural abnormalities associated with epileptogenic risk [4,13,14]. Electrophysiologic approaches, particularly EEG and quantitative EEG, may detect early disturbances in cortical excitability and abnormal network activity before overt epilepsy becomes established [15,16]. Molecular biomarkers measured in blood, cerebrospinal fluid, or extracellular vesicles may further reflect neuroinflammation, neuronal injury, BBB disruption, and maladaptive tissue remodeling [17,18]. However, each modality captures only part of the epileptogenic process, highlighting the need for integrative analytical approaches that combine complementary data sources rather than relying on any single marker class.

Artificial intelligence (AI) and machine learning are increasingly important in this context because they provide tools for integrating complex, high-dimensional, and multimodal datasets into clinically informative prediction models. Unlike conventional single-variable analyses, AI-based approaches can identify nonlinear relationships and combinatorial patterns across imaging, electrophysiologic, and molecular features that may not be apparent through standard statistical methods [19,20]. Although applications in PTE remain limited, early studies using machine learning on neuroimaging data have already demonstrated the feasibility of distinguishing patients with post-traumatic seizures from those without epilepsy and of identifying candidate network-level imaging features associated with risk [19]. These advances suggest that AI may help move the field from descriptive biomarker discovery toward individualized prediction and clinically actionable models.

In this review, we provide a comprehensive overview of current progress in PTE biomarker studies, focusing on: (1) neuroimaging biomarkers such as magnetic resonance imaging (MRI) and computed tomography (CT), which aim to identify structural brain alterations linked to epileptogenesis; (2) molecular biomarkers, including RNA and protein signatures that elucidate underlying mechanisms and identify therapeutic targets; and (3) AI-assisted models, which leverage neuroimaging and molecular data to improve predictive accuracy. We also discuss future directions of PTE biomarker research.

2. Literature Search Strategy

This review focuses on studies investigating biomarkers for post-traumatic epilepsy (PTE), including both human and animal cohorts. We included studies that explored neuroimaging biomarkers (CT, MRI, EEG/qEEG), molecular biomarkers (proteins, RNAs, metabolites), and artificial intelligence (AI)-assisted prediction models. The inclusion criteria encompassed studies that specifically assessed biomarkers related to post-traumatic epileptogenesis, excluding those focused solely on acute traumatic brain injury or unrelated neurological conditions. Data collection processes involved reviewing published literature from various databases (PubMed, Scopus, Web of Science), with a focus on peer-reviewed articles. For neuroimaging, studies using CT and MRI were considered, with particular attention to lesion volume, structural damage, and brain network disruptions. EEG/qEEG studies were included based on their investigation of early epileptiform activity and network disturbances post-TBI. Molecular biomarkers were measured from blood, plasma, cerebrospinal fluid (CSF), or extracellular vesicles, with standardized protocols for protein quantification (e.g., ELISA) and RNA measurement (e.g., qPCR). AI-based studies were reviewed for their application of machine learning techniques to predict PTE using multimodal data. For consistency and clarity, evidence was prioritized from human cohort studies, while findings from experimental animal models were selectively incorporated to support mechanistic interpretation and biological plausibility.

3. Neuroimaging and Electrophysiological Biomarkers of PTE

Neuroimaging and electrophysiological modalities provide complementary, non-invasive approaches for identifying biomarkers of post-traumatic epileptogenesis. Structural imaging techniques such as CT and MRI enable the characterization of acute hemorrhagic lesions, contusions, and microstructural brain alterations associated with epileptogenic processes [21,22,23]. In parallel, electroencephalography (EEG) captures functional network abnormalities, including epileptiform activity and background disturbances, which are linked to increased seizure susceptibility following traumatic brain injury [24]. These modalities differ in their sensitivity to structural versus functional changes, temporal applicability in the acute versus subacute phases, and overall predictive performance, with improved accuracy achieved when they are integrated into multimodal frameworks (

Table 1. Comparative overview of neuroimaging and electrophysiological modalities for PTE prediction.

Modality	Strengths	Limitations	Predictive Utility	Representative Biomarkers
CT	Rapid, widely available; detects acute hemorrhage, contusions, and mass effect	Limited sensitivity for diffuse and microstructural injury	Moderate (AUC ~0.60–0.75); improves when combined with clinical/EEG data	Contusion volume, SDH, SAH, midline shift, frontotemporal lesions [21]
MRI	High spatial resolution; detects microstructural damage, network disruption, and gliosis	Limited availability in acute phase; higher cost; processing challenges	Moderate to high (AUC ~0.75–0.89 in multimodal models)	FA/MD alterations, lesion volume, iron deposition, functional connectivity [22,23]
EEG/qEEG	Real-time functional assessment; detects epileptiform activity and network dysfunction	Limited monitoring duration; variability in acquisition and analysis	Moderate to high (AUC up to ~0.85); enhanced with quantitative metrics	Epileptiform discharges, suppression burden, delta variability, PSWEs [24]

CT, computed tomography; MRI, magnetic resonance imaging; EEG, electroencephalography; qEEG, quantitative electroencephalography; PTE, post-traumatic epilepsy; SDH, subdural hematoma; SAH, subarachnoid hemorrhage; FA, fractional anisotropy; MD, mean diffusivity; PSWEs, paroxysmal slow-wave events; AUC, area under the receiver operating characteristic curve.

).

Taken together, the available evidence suggests that CT, MRI, and EEG capture different but complementary stages of post-traumatic epileptogenesis [21,22,23,24]. CT-based markers appear most clinically translatable because they are available in the acute setting and have been validated in relatively large human cohorts, but their biological resolution is limited mainly to macroscopic structural injury. MRI provides greater sensitivity to microstructural and network-level abnormalities, particularly white matter disruption and altered functional connectivity, yet its predictive performance is derived largely from smaller research cohorts and is constrained by cost, timing, and processing variability. EEG and qEEG, in contrast, directly reflect functional network instability and may be more closely linked to imminent epileptogenic activity, but their interpretation remains limited by inconsistent acquisition protocols and restricted monitoring duration. Thus, current evidence does not support a single superior modality but rather indicates that structural burden, network dysfunction, and electrophysiologic instability represent distinct but interacting dimensions of PTE risk.

3.1. Computed Tomography (CT)

CT is the first-line neuroimaging modality for the evaluation of acute traumatic brain injury due to its rapid acquisition, wide availability, and high sensitivity for detecting hemorrhage, contusions, and mass effect [25]. Early CT findings provide important prognostic information, as focal cortical injuries, particularly in the temporal and frontal lobes, are consistently associated with an increased risk of post-traumatic epilepsy [26]. Higher injury severity, reflected by CT-based classifications such as the Marshall score, further correlates with elevated PTE risk, especially in pediatric populations [27]. Additional CT features, including subdural hematoma, subarachnoid hemorrhage, epidural hemorrhage, and skull fractures, have also been linked to increased epileptogenic potential [28].

Recent large-scale and prospective studies consistently demonstrate that CT-derived structural features are associated with the development of PTE across diverse populations [29,30,31,32,33,34,35]. Both quantitative and qualitative analyses highlight the importance of hemorrhagic injury burden, lesion distribution, and cortical involvement. Ayvaz et al. [29] demonstrated that increasing total contusion volume on early CT independently predicted PTE risk, with an adjusted hazard ratio of 1.17 per 5 cc increase (95% CI 1.12–1.23), with additional contribution from lesion topology, particularly in frontal and temporal regions (

Table 2. CT-based prediction models for PTE.

Study	Year	Population	Cohort/Data Source	N (Total/PTE)	Model Type	CT Predictors	Outcome	Performance
Ayvaz et al. [29]	2025	Adults	Retrospective cohort; single-center TBI registry	1017/61	Competing risk regression	Contusion volume	Late PTE (>7 days)	HR 1.17 per 5 cc increase (95% CI 1.12–1.23)
Gómez-Rosiles et al. [30]	2025	Children	Cross-sectional cohort; single-center hospital	101/43	Multivariable logistic regression	Subdural hematoma, subdural hygroma	PTE (>7 days)	OR 5.6 (95% CI 2.1–14.8); R2 = 0.30
Awan et al. Model 1 [31]	2024	Adults	Retrospective cohort; TBI Model Systems National Database	4126/916	Logistic regression	Intracranial fragments, traumatic hemorrhage	PTE within 2 years	AUROC 0.73; sensitivity 76%
Awan et al. Model 2 [31]	2024	Adults	Retrospective cohort; TBI Model Systems National Database	4126/0	Logistic regression	Intracranial fragments, traumatic hemorrhage	PTE within 2 years	AUROC 0.70; sensitivity 72%
Awan et al. Model 3 [31]	2024	Adults	Retrospective cohort; TBI Model Systems National Database	3788/501	Logistic regression	Intracranial fragments, traumatic hemorrhage	Seizures during year 2	AUROC 0.84; sensitivity 86%
Awan et al. Model 4 [31]	2024	Adults	Retrospective cohort; TBI Model Systems National Database	3333/~200	Logistic regression	Intracranial fragments, traumatic hemorrhage	Seizures during year 2	AUROC 0.67; sensitivity 73%
Lin et al. [32]	2022	Adults	Retrospective cohort; single-center hospital	457/71	Logistic regression	Contusion site, volume, skull fracture, subdural hematoma	PTE (>7 days)	OR 6.57 for contusion volume >13.5 mL; OR 14.31 for SDH, 95% CI 2.85–71.87; OR 78.88 for non-LPTS, 95% CI 13.73–453.28
Wang et al. [33]	2021	Adults	Retrospective cohort with external validation; multicenter hospital cohorts	1301/166	Cox proportional hazards model	Subdural hematoma, contusion location	PTE during follow-up	C-index 0.846 (training), 0.895 (validation)
Ritter et al. Acute model [34]	2016	Adults	Prospective cohort; TBI Model Systems National Database	796/98	Logistic regression	Contusion load	Acute post-traumatic seizures	C-statistic 0.599
Ritter et al. Year-1 model [34]	2016	Adults	Prospective cohort; TBI Model Systems National Database	796/95	Logistic regression	Contusion load, subdural hematoma	Seizures within 1 year	C-statistic 0.747
Ritter et al. Year-2 model [34]	2016	Adults	Prospective cohort; TBI Model Systems National Database	796/134	Logistic regression	Contusion load, subdural hematoma	Seizures within 2 years	C-statistic 0.716
Englander et al. [35]	2003	Adults	Prospective multicenter cohort (4 trauma centers)	647/66	Survival analysis	Midline shift, contusions, SDH, EDH, SAH, IVH, fragments	Late seizures (8 days–24 months)	24-month cumulative risk 13.8%

PTE, post-traumatic epilepsy; TBI, traumatic brain injury; HR, hazard ratio; OR, odds ratio; CI, confidence interval; AUROC, area under the receiver operating characteristic curve; SDH, subdural hematoma; EDH, epidural hematoma; SAH, subarachnoid hemorrhage; IVH, intraventricular hemorrhage.

). Similarly, Lin et al. [32] found that contusion volume >13.5 mL, subdural hematoma, severe GCS score, and non-late post-traumatic seizures were independently associated with PTE in multivariable logistic regression, although the estimates for subdural hematoma and non-late post-traumatic seizures had wide confidence intervals. Gómez-Rosiles et al. [30] further identified subdural lesions as a major risk factor in pediatric populations, with an adjusted odds ratio of aOR 5.6 (95% CI 2.1–14.8). In large cohort analyses, Awan et al. [31] and Wang et al. [33] confirmed that subdural hemorrhage (HR 2.20) and frontotemporal contusions (HR 2.46) significantly increase PTE risk, with predictive models achieving AUROC values of 0.70–0.84 and external validation C-index up to 0.895. Prospective studies by Ritter et al. [34] and Englander et al. [35] further demonstrated that subarachnoid hemorrhage (RR 2.06), higher contusion load (RR 2.17), and surgical evacuation (RR up to 3.05) are associated with increased late seizure incidence, while CT-defined features such as midline shift and multifocal cortical injury correspond to particularly high cumulative seizure risk. Taken together, these findings demonstrate that CT-derived structural parameters, particularly contusion volume, lesion location, and hemorrhagic burden, consistently provide clinically relevant indicators of epileptogenic risk following TBI.

Early prospective studies primarily relied on categorical variables, including lesion type and neurosurgical intervention, to define high-risk groups [34,35]. In contrast, more recent investigations incorporate quantitative imaging features such as contusion volume, contusion load, and lesion location, together with clinical variables including early seizures and injury severity [29,31,32,33]. Across these approaches, seizure risk is more strongly associated with the cumulative extent and complexity of injury rather than any single lesion type. In particular, cortical involvement, lesion multiplicity, and mass effect are repeatedly linked to increased epileptogenic potential. Frontal and temporal regions appear especially relevant, likely reflecting their role in large-scale neuronal networks and increased susceptibility to excitotoxic and inflammatory processes. However, variability in study design, outcome definitions, and follow-up duration contributes to differences in reported effect sizes, emphasizing the need for standardized CT-based metrics and harmonized outcome definitions in future studies [29,30,31,32,33,34,35].

CT provides a practical and immediately available framework for early structural risk assessment following traumatic brain injury. It enables direct evaluation of key features, including hemorrhagic burden, contusion volume, lesion distribution, and mass effect, all of which have been associated with subsequent seizure risk. Importantly, CT-derived parameters such as subdural hematoma, subarachnoid hemorrhage, contusion load, frontotemporal involvement, and midline shift capture both the severity and spatial organization of injury, which are central determinants of epileptogenic potential. Because CT is routinely performed in the acute setting, these features can be incorporated into predictive models without additional imaging burden. Recent studies demonstrate that CT-based variables can support individualized risk estimation with moderate-to-high discrimination when combined with clinical factors [29,31,32,33]. Overall, CT serves as a core modality for early identification of patients at increased risk of PTE and provides a structural foundation for developing clinically applicable prediction strategies.

A critical distinction across CT studies is that predictive performance improves when structural features are quantified rather than categorized. Earlier studies mainly identified broad lesion classes or surgical variables as risk markers, whereas more recent models show that contusion volume, lesion topology, and frontotemporal localization provide greater discriminatory value. However, the apparent strength of some newer models should be interpreted cautiously, because performance estimates vary according to cohort composition, seizure definition, and follow-up duration. In this context, CT is best viewed as a robust first-line stratification tool rather than a standalone predictor of epileptogenesis, with greatest value when integrated with clinical and electrophysiological variables.

3.2. Magnetic Resonance Imaging (MRI)

MRI offers critical insights into the structural and functional alterations of the brain associated with post-traumatic epileptogenesis (

Table 3. MRI-based biomarkers associated with PTE.

Study	Year	Population	N (Total/PTE)	MRI Modality	MRI Biomarker(s)	Imaging Timing	Outcome Definition	Main Findings	Effect Size/Performance
Messori et al. [38]	2005	Adults	135/20	T1/T2/GE T2*/FLAIR MRI	sSDH-C; IW/I-CW vs. CW gliosis–hemosiderin patterns	Serial MRI ≤ 2 years; follow-up 5–10 years	PTE, ≥2 late seizures > 4 weeks	Incomplete/evolving gliosis–hemosiderin patterns and surgically treated focal lesions were associated with higher PTE risk; CW lesions were lower risk	sSDH-C: HR 4.38; IW/I-CW vs. CW: HR 6.61
Smitherman et al. [39]	2015	Children	63/NR	FLAIR MRI	Lesion volume (HLVI); lesion location (zones A–C)	MRI ~6 days post-TBI; follow-up ~13 months	Functional outcome (GOS-E Peds)	Higher lesion volume and brainstem involvement (zone C) were associated with worse outcomes; lesions across all zones increased risk of unfavorable outcome	HLVI-total: r = 0.39; zone C: r = 0.37; A + B + C: OR 4.38
La Rocca et al. [40]	2020	Adults	53/NR	T1-weighted MRI	Structural network features (multiplex connectivity)	MRI ≤ 32 days post-TBI	Seizure occurrence (≥1 seizure)	MRI-derived network features differentiated seizure vs. non-seizure patients	Accuracy ~70%; AUC ~0.75
Lin et al. [32]	2022	Adults	457/NR	Structural MRI	Contusion site, volume, SDH	Acute	PTE (>7 days)	Combined structural predictors strongly associated with PTE	C-index 0.98
Gupta et al. [41]	2005	Adults	23/14	DTI + T2/FLAIR MRI	FA ↓; MD ↑; volume of microstructural damage	MRI 1–3 years post-TBI	PTE (clinical diagnosis)	Reduced FA and increased MD beyond visible lesions were associated with PTE; greater microstructural damage in PTE vs. non-PTE	FA: 0.57 vs. 0.68 (p < 0.05); MD ↑ (NS)
Li et al. [42]	2018	Adults	151/60	SWI + DKI MRI	Iron deposition around encephalomalacia; MK; MD; lesion volume	MRI at 1 year after discharge	PTE at 1-year follow-up	Higher perilesional iron deposition and altered MK were associated with epilepsy, whereas lesion volume and MD were not significantly different	MK: p = 0.035; SWI iron deposition: p = 0.002; lesion volume: p = 0.245; MD: p = 0.291
Akbar et al. [43]	2021	Adults	22/13	dMRI (DTI) + ML	FA in WM tracts (TBSS features); lesion-normalized FA features	MRI ~15 days post-TBI	Late seizures (>7 days post-TBI)	Lesion-normalized FA features improved seizure prediction; FA alterations (e.g., ALIC-R) identified as potential biomarkers	Accuracy: 0.819; AUC: 0.785
Weiler et al. [44]	2024	Adults	37/NR	fMRI	Thalamic and hippocampal functional connectivity	MRI ≤ 14 days post-TBI; follow-up up to 2 years	Early and late seizures (>7 days)	Altered thalamo-hippocampal connectivity differentiated early vs. late seizure phenotypes and predicted late seizures	AUC: 0.877; accuracy ~0.88; sensitivity 84.6%; specificity 86.7%
La Rocca et al. [37]	2023	Adults	75/28	rs-fMRI	Functional connectivity (graph metrics: integration, segregation, network topology)	MRI within ≤36 days post-TBI; follow-up up to 2 years	Late seizures (>7 days post-TBI)	Seizure group showed hyperconnectivity, hyperintegration, and hyposegregation with disrupted network topology	Significant group differences (p < 0.05 across network metrics); no AUC reported
Akbar et al. [45]	2024	Adults	48/17	Multimodal MRI (dMRI, fMRI ± FLAIR)	dMRI (ALIC-L FA), fMRI (right middle temporal gyrus), multimodal features	~14 days post-TBI; follow-up 2 years	Late post-traumatic seizures	Identified modality-specific biomarkers; multimodal fusion improves prediction	Best model AUC ≈ 0.79

TBI, traumatic brain injury; PTE, post-traumatic epilepsy; MRI, magnetic resonance imaging; DTI, diffusion tensor imaging; dMRI, diffusion-weighted MRI; SWI, susceptibility-weighted imaging; DKI, diffusion kurtosis imaging; rs-fMRI, resting-state functional MRI; FLAIR, fluid-attenuated inversion recovery; GE T2*, gradient-echo T2*; FA, fractional anisotropy; MD, mean diffusivity; MK, mean kurtosis; TBSS, tract-based spatial statistics; ALIC, anterior limb of the internal capsule; SDH, subdural hematoma; sSDH-C, surgically treated subdural hematoma with contusion; HLVI, hemorrhagic lesion volume index; GOS-E Peds, Glasgow Outcome Scale–Extended Pediatric; HR, hazard ratio; OR, odds ratio; AUC, area under the curve; C-index, concordance index; NS, not significant; CW, complete wall; IW, incomplete wall; I-CW, intermediate-to-complete wall; ↑, increase; ↓, decrease.

). Diffusion-based MRI techniques, including diffusion tensor imaging (DTI) and diffusion-weighted imaging (dMRI), are especially valuable for detecting microstructural white matter changes that extend beyond visible lesions, which are linked to increased seizure susceptibility in TBI [13,36]. Functional MRI (fMRI) adds an additional layer of analysis, allowing the investigation of brain network disruptions, such as abnormal hyperconnectivity and functional integration, which are often seen in patients who develop PTE [37].

Recent studies have confirmed that MRI-derived biomarkers, such as lesion volume, location, and white matter abnormalities, are strong predictors of PTE [32,38,39,40]. Diffusion MRI alterations, such as reduced fractional anisotropy (FA) and increased mean diffusivity (MD), indicate microstructural damage that correlates with epileptogenesis, even in the absence of visible lesions [41,42,43]. In addition to structural MRI, fMRI studies demonstrate that patients with PTE exhibit abnormal network organization, including disrupted thalamocortical connectivity and altered functional network topology, which contribute to increased seizure risk [37,44]. Integration of structural and functional MRI features enhances predictive accuracy, and machine learning models incorporating lesion characteristics and functional connectivity metrics have achieved AUC values up to approximately 0.79 [45].

Compared with CT, MRI offers stronger mechanistic insight into post-traumatic epileptogenesis because it detects abnormalities beyond visible lesion boundaries, including diffuse white matter injury, iron-related tissue changes, and altered network organization. However, MRI studies remain less standardized and are often based on smaller cohorts with heterogeneous imaging timing, which makes direct comparison across studies difficult. Importantly, not all MRI-derived abnormalities appear equally informative: some markers, such as lesion volume, show inconsistent associations, whereas microstructural and connectivity-based features appear more closely linked to seizure development. This suggests that the predictive value of MRI may depend less on conventional lesion burden alone and more on its ability to capture distributed network vulnerability.

3.3. Electrophysiology

Electroencephalography (EEG) is a widely used clinical tool for diagnosing and predicting PTE in TBI populations. Although earlier studies in the 1970s did not establish EEG as a reliable predictor of PTE [46], more recent research has shown its predictive value (

Table 4. EEG-based biomarkers associated with PTE.

Study	Year	Population	N	Method	EEG Biomarker(s)	Main Findings
Kim et al. [47]	2018	Adults	50	EEG	EAs	Acute EAs independently predict first-year PTE
Oliveira et al. [53]	2025	Adults	73	EEG	Diffuse theta waves; abnormal bilateral sleep elements	EEG abnormalities and temporal lesions predict higher PTE risk
Serlin et al. [54]	2026	Adults	45	EEG	PSWEs	PSWEs predict PTE and disability
Rubinos et al. [55]	2022	Adults	NR	EEG	Slowing; epileptiform activity	EEG abnormalities at 1 month associated with PTE
Pease et al. [51]	2023	Adults	NR	qEEG	Delta/theta power; rhythmic spectrum	Acute qEEG predicts PTE (AUC 0.85)
Chen et al. [50]	2023	Adults	126	qEEG	Epileptiform burden; suppression; beta variability	Increased epileptiform burden associated with ~4.6× higher PTE risk
Hameren et al. [56]	2024	Adults	NR	EEG/ECoG	PSWEs	PSWEs validated as biomarker in humans

EEG, electroencephalography; qEEG, quantitative electroencephalography; EAs, epileptiform abnormalities; PSWEs, paroxysmal slow-wave events; AUC, area under the curve.

). Epileptiform abnormalities observed within the first five days after TBI can independently predict the development of PTE within one year, with an odds ratio (OR) of 3.16 [47]. Additionally, focal EEG abnormalities, particularly those present more than a month after TBI, are associated with a three-fold increase in the risk of PTE one year after injury [48]. Key EEG biomarkers for predicting PTE include non-seizure epileptiform activities such as interictal discharges, generalized and lateralized periodic discharges, and lateralized rhythmic delta activity [49]. Furthermore, quantitative EEG (qEEG) has demonstrated that altered background activity, in addition to epileptiform discharges, can provide significant prognostic value. Increased abnormalities in EEG-derived measures, including epileptiform activity, suppression patterns, and beta-band fluctuations, were associated with elevated risk of developing PTE during the first year following TBI [50]. Recent studies examining lateralization of post-traumatic lesions and breach artifacts on qEEG have found that pronounced and variable delta activity is predictive of PTE [51]. However, despite the advancements in EEG analysis, challenges remain, such as limited monitoring periods, a primary focus on early post-injury monitoring, and a lack of data on sleep abnormalities, which could further enhance understanding of epileptogenesis and improve predictive accuracy for PTE [52].

In contrast to CT and MRI, which primarily characterize structural substrates of risk, EEG provides a more direct readout of functional cortical instability. This may explain why epileptiform abnormalities, suppression burden, and quantitative background features often show relatively strong short-term predictive value. At the same time, EEG-based studies are methodologically more heterogeneous, particularly with respect to recording duration, timing after injury, and feature definition, which limits reproducibility. Therefore, EEG may be especially valuable for identifying active epileptogenic states, whereas imaging contributes more to defining structural susceptibility; their combined use is likely to be more informative than either modality alone.

4. Molecular Biomarkers

Molecular biomarkers play a crucial role in PTE by reflecting key pathophysiological processes such as neuroinflammation, BBB dysfunction, astrocyte dysregulation, and neuronal injury. These biomarkers include RNA-based markers like microRNAs (miRNAs), extracellular vesicle (EV)-derived molecules, proteins, and metabolites. miRNAs regulate inflammatory pathways and glial activation, while EVs carry a wide range of biomolecules that can provide a comprehensive snapshot of cellular changes (Figure 1).

4.1. RNA Biomarkers

MicroRNAs (miRNAs) have gained significant attention as potential biomarkers for PTE due to their crucial roles in various pathophysiological processes, including neuroinflammation, BBB dysfunction, and astrocyte dysregulation (

Table 5. miRNAs associated with PTE categorized by their roles in neuroinflammation, cell signaling, and epigenetic dysregulation.

Study	Year	Population	miRNA Type	Specific miRNA(s)	Category	Key Biomarker Role	Key Finding
Heiskanen et al. [63]	2025	Human (TBI, epilepsy)	microRNA	miR-434-3p, miR-9a-3p	Epigenetic dysregulation	Neuronally enriched miRNAs reflect injury severity but do not predict PTE	Neuronally enriched miRNAs indicate injury severity but not PTE prediction
Cinar [64]	2022	Human (TBI, epilepsy)	microRNA	10 commonly dysregulated miRNAs	Epigenetic dysregulation	Dysregulated miRNAs involved in apoptosis, inflammation, and cell cycle regulation	10 miRNAs linked to apoptosis, inflammation, and cell cycle
Meng et al. [65]	2025	Human (Astrocytes and microglia)	microRNA, lncRNA	Not specified	Epigenetic dysregulation	miRNA-lncRNA crosstalk in astrocytes and microglia triggers neuroinflammation in epilepsy	miRNA-lncRNA interactions drive neuroinflammation
Iori et al. [66]	2017	Animal, Human (Epilepsy models)	microRNA	miR-146a	Epigenetic dysregulation	Implicated in the regulation of the IL-1R1/TLR4 inflammatory axis	miR-146a controls IL-1R1/TLR4 axis in epileptogenesis
Zhang et al. [67]	2022	Rat (PTE model)	microRNA, mRNA	miR-98-5p, miR-1224	Cell Signaling	miR-98-5p–Slc17a6 and miR-1224–Slc25a22 interaction pairs in PTE	miRNA-mRNA interaction pairs in PTE development
Meng et al. [68]	2015	Rat (Epilepsy model)	microRNA	rno-miR-21-5p	Cell Signaling	miR-21-5p regulates neuronal calcium signaling pathways	miR-21-5p affects neuronal calcium signaling
Jia et al. [69]	2023	Rat (PTE model)	microRNA, mRNA	7 miRNAs (unspecified)	Cell Signaling	miRNAs associated with calcium channels form miRNA-mRNA pairs	Calcium channel miRNA-mRNA pairs in PTE
Puhakka et al. [70]	2022	Rat (Chronic TBI)	microRNA, tRNA-derived fragments	miR-146a-5p, miR-155-5p, miR-375-3p	Neuroinflammation	Upregulated miRNAs and tRNA fragments linked to neuroinflammation and behavioral outcomes	Chronic TBI-associated miRNAs and tRNA fragments in neuroinflammation

TBI, traumatic brain injury; PTE, post-traumatic epilepsy; miRNA, microRNA; lncRNA, long non-coding RNA; Slc, solute carrier; IL-1R1, interleukin-1 receptor type 1; TLR4, toll-like receptor 4; mRNA, messenger RNA.

). These miRNAs regulate inflammatory signaling pathways, cellular integrity, and glial activation, which are central to the development of epilepsy following TBI [57,58,59]. The neuroinflammatory response in TBI involves the activation of microglia and astrocytes, the release of pro-inflammatory cytokines, and the recruitment of immune cells, all of which contribute to epileptogenesis. Simultaneously, the breakdown of the BBB exacerbates the inflammatory process by allowing blood-derived molecules to infiltrate the brain parenchyma [60]. Several miRNAs, such as miR-21, miR-155, miR-142, and miR-146a, have been identified as key regulators in these processes, with both protective and pathogenic roles depending on the timing and context of their expression [57,58,59]. Additionally, miR-9-5p and miR-124-3p have been shown to play protective roles in maintaining BBB integrity and regulating neuroinflammation, underscoring their potential as therapeutic targets [61,62].

4.1.1. miRNAs Associated with Neuroinflammation, Blood–Brain Barrier (BBB) Dysfunction, and Astrocyte Dysregulation

A growing body of evidence has identified specific microRNAs (miRNAs) that modulate these processes, thereby influencing epileptogenesis. Experimental studies suggest that miR-9-5p may attenuate BBB disruption and inflammatory responses after TBI through modulation of Hedgehog-related signaling and suppression of NF-κB/MMP-9–associated pathways [57]. Similarly, miR-29a-5p has been implicated in regulating BBB permeability in endothelial cells, providing protective effects against TBI-induced dysfunction [58]. However, while these miRNAs offer significant therapeutic promise, the acute inflammatory responses they regulate are not always straightforward. miR-155 and miR-142, for example, have been linked to promoting neuroinflammation and astrogliosis following TBI, raising concerns about their dual roles in both promoting injury repair and exacerbating epileptogenesis [60]. In the case of miR-21, it is involved in both inflammation and BBB permeability regulation, but its role is context-dependent, with different isoforms exhibiting varying effects depending on the cellular compartment and stage of injury [59,71]. These findings emphasize the complex, bidirectional relationship between miRNAs and neuroinflammation in PTE, highlighting the importance of time-sensitive interventions targeting specific miRNA pathways.

Furthermore, astrocyte-specific mechanisms represent an additional layer of post-traumatic epileptogenesis, with several miRNAs contributing to this process. miR-124-3p, delivered via microglial exosomes, has shown protective effects on neuronal repair and glial modulation, suggesting its potential as a therapeutic agent in mitigating the long-term consequences of TBI [62]. In contrast, miR-155 and miR-142 have been associated with the activation of astrocytes and the promotion of inflammation, which can contribute to secondary brain injury and subsequent epileptogenesis [60]. Notably, tRNA-derived fragments, such as miR-146a-5p, have been identified as key regulators in chronic neuroinflammation following TBI, further illustrating the multifaceted roles that different RNA species play in regulating neuroinflammatory responses [70]. These findings indicate that astrocyte-related RNA pathways are particularly relevant for sustaining long-term neuroinflammatory states and network remodeling. Despite their potential as biomarkers, the clinical application of these miRNAs remains challenging due to their pleiotropic effects and the need for precise timing in their modulation. Future research should aim to clarify the temporal dynamics of miRNA expression and develop targeted interventions that can modulate their activity to prevent the onset of PTE.

4.1.2. miRNAs Associated with Cell Signaling

Cell signaling pathways, particularly those involving ion channels, neurotransmitter receptors, and calcium homeostasis, are essential for neuronal excitability and seizure susceptibility in PTE. Several miRNAs have been identified as regulators of these pathways, highlighting their potential as biomarkers for epileptogenesis. miR-98-5p, for example, plays a pivotal role in regulating the glutamatergic system by targeting glutamate transporters such as Slc17a6, as well as potassium channels like KCNH2. These interactions suggest that miR-98-5p may act as a hub regulator of neuronal excitability, influencing the transition from normal activity to hyperexcitability and seizure generation [67,72]. Additionally, miR-21-5p has been shown to regulate neuronal calcium signaling pathways, a critical component in the development of epilepsy. The identification of these miRNA-mRNA interaction pairs underscores the importance of miRNAs in the regulation of ion channels and neurotransmitter systems, which are central to the pathological changes observed in PTE. However, while these findings provide mechanistic insights, the complex interplay of multiple miRNAs in signaling pathways suggests that future studies should focus on validating the functional relevance of these interactions in vivo and their potential as diagnostic biomarkers for PTE.

Moreover, the complexity of miRNA-mediated regulation in cell signaling is further highlighted by studies identifying a wide array of differentially expressed miRNAs associated with calcium signaling pathways in PTE models. In particular, Jia et al. identified seven miRNAs that form negatively correlated miRNA-mRNA pairs, influencing calcium channels and providing insights into the molecular mechanisms driving epileptogenesis [69]. The large number of differentially expressed miRNAs, including rno-miR-21-5p, indicates the highly dynamic nature of miRNA regulation in epileptic conditions [68]. These miRNAs affect multiple signaling pathways, which suggests that a multi-target approach may be necessary for effective PTE prediction and therapeutic intervention [70]. However, despite their potential, the specificity of these miRNAs to PTE remains uncertain, as similar miRNAs are often dysregulated in various neurological conditions. Thus, the challenge remains to define specific miRNA panels that can accurately predict the risk of PTE while distinguishing it from other brain pathologies. Further studies should aim to address these concerns by incorporating longitudinal data and functional validation of miRNA-mRNA interactions across different stages of epileptogenesis.

4.1.3. miRNAs Associated with Epigenetic Dysregulation

Epigenetic mechanisms play a critical role in the transition from transient injury-induced changes to persistent alterations in gene expression that promote epileptogenesis. miRNAs, which are both regulators and targets of epigenetic modifications, are central to these processes. Following TBI, miRNA expression profiles undergo sustained changes, some of which correlate with the development of PTE [73,74]. Specifically, miR-146a has been identified as an important epigenetic regulator involved in controlling the IL-1R1/TLR4 inflammatory axis, suggesting a direct link between inflammation and long-term gene expression changes associated with epileptogenesis [66]. Large-scale profiling studies have shown that several miRNAs are dysregulated in both TBI and epilepsy, including miR-21, miR-181a, and miR-155, which are involved in inflammation and cell cycle regulation [64,65]. These findings point to a shared epigenetic mechanism between TBI and epilepsy, but it remains unclear whether these miRNAs are biomarkers of injury severity or of the transition to epileptogenesis itself. The identification of these dysregulated miRNAs opens up potential therapeutic avenues, but their ability to predict the development of epilepsy from TBI remains inconsistent and requires further investigation.

The temporal dynamics of miRNA expression are essential for understanding the role of epigenetic dysregulation in PTE. Different miRNA signatures have been identified at various post-injury time points, reflecting evolving molecular processes associated with epileptogenesis [75,76]. Chronic downregulation of miR-124-3p in the perilesional cortex after TBI further suggests persistent alterations in neuronal gene regulation that may contribute to seizure susceptibility and long-term network remodeling [77]. Collectively, these findings highlight the importance of temporal and context-dependent epigenetic regulation in PTE and suggest that miRNA profiles should be interpreted within dynamic post-injury frameworks rather than as static biomarkers.

4.2. Protein Biomarkers

Most of the studies on protein biomarkers associated with PTE have identified two main categories: proteins involved in astrocyte dysregulation and neuroinflammation, and proteins indicative of neuronal damage, synaptic function, and neurovascular compromise. Astrocytic proteins, such as GFAP, S100B, and IL-1β, are consistently found to play significant roles in the neuroinflammatory response following TBI. These proteins are involved in astrocyte activation and the inflammatory processes that contribute to epileptogenesis [78,79,80,81,82]. In addition to classical inflammatory mediators, emerging evidence highlights the role of complement-mediated synaptic remodeling in post-traumatic epileptogenesis, where activation of complement components such as C1q and C3 contributes to aberrant synaptic pruning and network hyperexcitability. Furthermore, inflammasome signaling pathways, particularly NLRP3 inflammasome activation, have been implicated in the maturation of pro-inflammatory cytokines such as IL-1β and IL-18, amplifying neuroinflammation and promoting epileptogenic processes. In contrast, proteins like NfL, Tau, and Beta-synuclein are implicated in neuronal injury, axonal damage, and long-term neurodegeneration, reflecting the neuronal damage and synaptic dysfunction underlying PTE [83,84,85,86,87,88]. These biomarkers not only provide insight into the biological mechanisms of epilepsy development but also hold promise for early detection, prognosis, and targeted therapeutic interventions for TBI patients at risk of developing PTE.

4.2.1. Proteins Associated with Astrocyte Dysregulation and Neuroinflammation

Astrocytic biomarkers, particularly GFAP and S100B, have been extensively studied in the context of TBI and epilepsy (

Table 6. Proteins associated with astrocyte dysregulation and neuroinflammation in PTE.

Study	Year	Specific Protein Biomarker	Biomarker Source	Study Population	Key Finding
Ilaria [78]	2025	GFAP	Thalamic astrocytic alterations, contralateral spiking activity	Murine TBI model with 50% PTE incidence by 5 months	Increased GFAP aggregation and astrocytic morphological alterations in the ipsilateral thalamus of high-risk mice contribute to epileptogenesis. Targeting GFAP improved TBI outcomes, suggesting its relevance to PTE.
Wanner et al. [79]	2025	GFAP	CSF, serum, and patient biofluids	TBI patients and a human trauma culture model	Distinct CSF patterns of GFAP-derived fragments were observed between favorable and unfavorable TBI outcomes, with persistent elevation of specific calpain-associated BDP isoforms during injury progression.
Hsu et al. [80]	2025	GFAP, S100B	Fluid biomarkers derived from astrocytes	Epilepsy in children and adults; clinical and experimental findings	The review focuses on astrocytic biomarkers like GFAP and S100B, examining their roles in assessing seizure burden, temporal dynamics, and potential in distinguishing seizure types. It discusses their therapeutic, prognostic, and mechanistic implications in epileptic disorders.
Sitovskaya [81]	2024	GFAP, S100B	Temporal lobe tissue	Pediatric patients with drug-resistant epilepsy	Changes in GFAP and S100 immunoreactivity were observed in the temporal lobe of pediatric patients with drug-resistant epilepsy.
Oris et al. [82]	2023	S100B	Blood	Children, adults, and athletes with mTBI	S100B protein is the most widely studied and used diagnostic biomarker for clinical decision-making in mTBI patients, also playing an active role in acute brain injury processes and serving as a potential therapeutic target.
Mochol et al. [83]	2023	GFAP, S100B, NSE	Serum	119 epilepsy patients and 80 healthy controls	Elevated serum levels of GFAP were found in epilepsy patients. However, none of the markers, including GFAP, were significantly associated with epilepsy duration, seizure type or severity, or recent seizures.
Hanin et al. [84]	2020	S100B, NSE	Cerebrospinal fluid and blood	Patients and animal models of status epilepticus	New biomarkers aim to prospectively identify the severity and development of disability and subsequent epilepsy of patients with status epilepticus. These include increased S100B and High Mobility Group Box 1 for gliosis/inflammation.
Michetti et al. [85]	2023	S100B	Biological fluids, nervous tissue	Patients and/or experimental models of different neural disorders	S100B levels are recognized as a reliable biomarker of active neural distress and correlate with clinical/toxic parameters in epilepsy and traumatic neural injury. Overexpression of S100B worsens clinical presentation, suggesting it is a pathogenic factor.
Bulduk et al. [86]	2018	S100B, GFAP	Serum and cardiac blood	Lithium–pilocarpine induced status epilepticus in rat model	Serum S100B is a candidate biomarker for monitoring neuroinflammation, with highly positive correlations found between S100B levels and microglial activation in CA1. It may also help predict diagnosis and prognosis.
Langeh et al. [87]	2020	S100B, TNF-α	Not specified	Various neurological disorders, including Alzheimer’s, Parkinson’s, multiple sclerosis, schizophrenia, and epilepsy	Increased S100B expression is associated with epilepsy and plays a crucial role in various neurological disorders through neuroinflammation. Elevated S100B levels are useful for assessing inflammatory markers and excitotoxicity-dependent neuronal loss.
Yates [88]	2011	GFAP, S100B	Serum	Patients with TBI	Serum levels of GFAP and S100B predict outcomes in TBI. These biochemical markers are adjuncts to assessing brain damage and enhance prognoses for TBI patients.
Pelinka et al. [89]	2004	GFAP, S100B	Serum	92 patients admitted after traumatic brain injury	Both GFAP and S100B measurements are useful non-invasive means for identifying brain damage after TBI, with differences based on TBI pattern and accompanying trauma/shock. They also have predictive value for mortality after TBI.
Singh et al. [90]	2024	GFAP, S100B	Serum	212 moderate and severe TBI patients in India	S100B is a better marker for TBI severity and outcome assessment than GFAP. S100B showed 66% sensitivity and specificity for disease severity, and performed better for early mortality prediction.
Komiotis et al. [91]	2024	IL-1β, NSE	Not specified	Pediatric traumatic brain injury patients	IL-1β appears to have the strongest correlation with PTE among inflammatory cytokines examined.

TBI, traumatic brain injury; mTBI, mild traumatic brain injury; PTE, post-traumatic epilepsy; GFAP, glial fibrillary acidic protein; S100B, S100 calcium-binding protein B; IL-1β, interleukin 1 beta; NSE, neuron-specific enolase; TNF-α, tumor necrosis factor-alpha.

). GFAP holds promise as both a diagnostic and prognostic marker, with recent studies indicating that GFAP breakdown products and aggregation patterns in specific brain regions, notably the thalamus, are linked to epileptogenesis [78,79]. The temporal dynamics of GFAP expression are crucial, as persistent elevations and specific fragment profiles correlate with poor clinical outcomes [79]. S100B has shown strong diagnostic utility in mild TBI, with associations to neuroinflammation and seizure activity [80,81,82,83,84,85,86,87,88,89,90]. Moreover, S100B may act not only as a passive biomarker but also as an active pathogenic factor, with overexpression contributing to worse clinical outcomes [85]. The combined analysis of GFAP and S100B provides complementary insights, with S100B often being more effective for early severity assessment, while GFAP is more valuable for prognostic purposes [90]. Among inflammatory cytokines, IL-1β has been found to have the strongest correlation with PTE development in pediatric populations [91]. However, the role of specific inflammatory mediators, such as TNF-α, CD40, and TNFSF-14, in predicting PTE remains underexplored. These astrocytic and neuroinflammatory biomarkers collectively highlight the reactive gliosis and inflammatory cascades central to epileptogenesis following brain injury.

4.2.2. Proteins Associated with Neuronal Damage and Neurovascular Function

Over the last two decades, numerous experimental and clinical studies have investigated protein biomarkers associated with neuronal damage and neurovascular dysfunction following traumatic brain injury, with particular attention to neurofilament light chain (NfL), glial fibrillary acidic protein (GFAP), tau protein (Tau), neuron-specific enolase (NSE), S100 calcium-binding protein B (S100B), and beta-synuclein as potential indicators of injury severity, prognosis, and post-traumatic outcomes (

Table 7. Biomarkers of neuronal damage and neurovascular function.

Study	Year	Specific Protein Biomarker	Biomarker Source	Study Population	Key Finding
Heiskanen et al. [92]	2022	NfL	Plasma	Rat TBI model with lateral fluid percussion injury	Plasma NF-L levels were significantly elevated after TBI, prognostic for cortical damage but not cognitive impairment or epileptogenesis.
Mondello et al. [94]	2020	GFAP, Tau, NfL	Serum, blood	21 patients with moderate-to-severe TBI	Exosomal UCH-L1 profile with acutely elevated values and secondary steep rise associated with early mortality. Diffuse injury patients showed higher acute exosomal NFL and GFAP.
Thelin et al. [95]	2017	GFAP, S100B, NfL, NSE	Serum	Human traumatic brain injury (TBI) patients	Serial sampling reveals different temporal trajectories with persisting high serum levels or secondary peaks associated with unfavorable outcomes or secondary events.
Sanabria et al. [96]	2026	Tau, NfL	Plasma	Adult male Wistar rats in a lateral fluid percussion injury (LFPI) model	TBI significantly increased plasma NfL levels, indicating neuronal damage. Both TBI groups showed higher seizure susceptibility, with biperiden slightly reducing seizure intensity.
Shahim [97]	2015	NfL	Blood, serum, cerebrospinal fluid	Ice hockey players and patients with mild to severe TBI	mTBI associated with altered serum levels of biomarkers related to neuronal injury, which correlated with return-to-play time. In sTBI, NFL levels showed high diagnostic accuracy.
Rubenstein et al. [98]	2023	Tau	CSF, serum (blood)	Moderate-to-severe TBI patients and healthy controls	Higher chronic (1–6 months) P-tau levels and P-tau:T-tau ratio associated with greater disability and worse global outcomes 12 months post-TBI.
Halbgebauer et al. [99]	2022	GFAP, Beta-synuclein, NfL	Plasma, blood	Polytraumatized patients with and without TBI, and healthy volunteers	Plasma NfL, beta-synuclein, and GFAP were significantly increased after polytrauma, predicting fatal outcome. A combined analysis discriminated TBI patients.
Shahim et al. [100]	2020	NfL	Serum, CSF	Hockey players, clinic-based TBI patients, and controls	Increased serum NfL concentrations distinguish TBI from controls, showing promise as a biomarker for acute, repetitive, subacute, and chronic TBI.
Whitehouse et al. [101]	2022	GFAP, Tau, NfL, NSE	Blood	Traumatic brain injury patients (human)	Blood proteomic biomarker levels related to lesion type and lesion burden in TBI patients.
Shahim et al. [102]	2018	NfL, Tau, S100B, NSE	Blood	Acute sports-related concussion (SRC) patients	Serum NfL outperformed tau, S100B, and NSE as a biomarker for SRC, identifying individuals at risk of prolonged PCS.
Siman et al. [103]	2009	NSE	CSF, serum	Severe TBI in humans	Identified neuron-enriched proteins as potential markers for severe TBI detection, management, and evaluation.
Zemlan et al. [104]	2002	Tau	CSF	Severe brain-injured patients	C-tau is a biomarker of neuronal damage in severe brain-injured patients, associated with elevated intracranial pressure and clinical outcome.
Do et al. [105]	2026	GFAP, S100B, Tau, NfL, NSE	Serum	373 CT-positive ICU traumatic brain injury patients (256 GCS 3-12)	Serum biomarker trajectory clusters predict functional outcome and quality of life for TBI patients.

TBI, traumatic brain injury; NfL, neurofilament light chain; GFAP, glial fibrillary acidic protein; Tau, tau protein; NSE, neuron-specific enolase; P-tau, phosphorylated tau; CSF, cerebrospinal fluid.

) [92,93,94,95,96,97,98,99,100,101,102,103,104,105]. NfL is emerging as a highly promising biomarker for neuronal and axonal injury following TBI, with consistent elevation across multiple studies in both experimental models and clinical populations, either alone or as part of multimarker biomarker panels [92,93,94,95,96,97]. NfL demonstrates strong diagnostic accuracy for TBI severity, correlates with functional outcomes and brain atrophy, and shows utility across the spectrum from mild to severe injury [97,100]. Importantly, while NfL is an excellent marker of cortical damage evolution and somatomotor impairment, multiple studies indicate it does not directly predict epileptogenesis or cognitive impairment [92,93], suggesting that axonal injury alone may not be sufficient for PTE development. Tau protein, particularly in its phosphorylated forms, shows strong associations with TBI severity, elevated intracranial pressure, and long-term disability [98,104]. The P-tau:T-tau ratio appears particularly informative for prognostic assessment, with higher chronic ratios predicting worse outcomes at 12 months post-injury [97]. Beta-synuclein, though less extensively studied, demonstrates value when combined with other biomarkers (GFAP and NfL) for discriminating TBI patients and predicting fatal outcomes [98]. Neuron-specific enolase (NSE) has been evaluated as part of multi-marker panels but shows variable performance, with some studies suggesting it is outperformed by NfL for concussion assessment [102,103]. The temporal trajectories of these biomarkers are critical, with serial sampling revealing that persistent elevation or secondary peaks correlate with unfavorable outcomes and secondary injury events [98,105]. Multi-biomarker approaches using trajectory clustering show promise for predicting functional outcomes and quality of life [105]. Notably, the evidence base for VEGFα in PTE prediction is limited in the current literature, representing a gap in understanding neurovascular contributions to epileptogenesis.

4.3. Metabolite Biomarkers

Metabolite biomarkers play a significant role in understanding the pathophysiology of PTE and TBI, as they reflect both systemic and cerebral alterations that occur during and after injury. One critical metabolite is glucose, which is a key player in glycolysis. Acute dysregulation in glucose metabolism indicates a systemic and cerebral energy crisis, which is often correlated with injury severity and long-term outcomes in TBI patients. Dysregulated glucose utilization has also been implicated in epileptogenesis, with glycolysis emerging as a targetable node for future therapeutic interventions aimed at preventing PTE. Commonly used sample types for glucose metabolism include serum, plasma, and brain microdialysates, and platforms such as LC/GC-MS and microdialysis are utilized to measure glucose levels. The most informative time window for glucose-based biomarkers is typically within hours to days post-injury [106,107].

Another important metabolite in the context of PTE is β-hydroxybutyrate (BHB), a ketone body that serves as an alternative fuel source. After injury, there is a shift toward ketone metabolism, reflecting altered substrate utilization in the brain. BHB has emerged as a central mediator of the ketogenic effects that contribute to seizure control, making it relevant in the context of post-traumatic epileptogenesis prevention. Research on BHB in PTE is gaining traction, with BHB serving as both a biomarker and a potential therapeutic target for modulating neuroexcitability. Sample types commonly used for BHB analysis include serum, plasma, and brain tissue, and key platforms for detection are LC/GC-MS, targeted ketone assays, and tracer studies. BHB levels are particularly informative in acute injury phases as well as during chronic ketogenic therapy [108,109].

Lactate and pyruvate, both key metabolites in glycolysis and redox reactions, are frequently perturbed in conditions such as hypoxia, mitochondrial dysfunction, and metabolic stress. Elevated lactate levels often signal hypometabolism or ischemia, and their dysregulation is closely linked to excitotoxicity, contributing to the pathogenesis of both TBI and PTE. The lactate-to-pyruvate ratio (LPR) serves as an important redox index, stratifying metabolic crisis following TBI and integrating ischemic and mitochondrial dysfunction signals. Glutamate and glutamine, involved in the excitatory neurotransmission and the glutamate-glutamine cycle, are also critical biomarkers for PTE. Dysregulated glutamate signaling is a key driver of excitotoxic injury, which is central to the development of hyperexcitability and epileptogenesis. In PTE, elevated glutamate levels contribute to excitotoxic injury cascades, while altered glutamine metabolism reflects astrocyte dysfunction and excitotoxic stress. These metabolites are typically measured in brain microdialysates, CSF, serum, and plasma, with NMR and LC/GC-MS providing the necessary platforms for analysis. The temporal window for these biomarkers spans from acute to subacute and chronic stages following TBI [107,110,111].

4.4. Extracellular Vesicle (EV)-Derived Biomarkers

Extracellular vesicle (EV)-derived molecules, including proteins, RNAs, and metabolites, are promising candidates as biomarkers for PTE due to their ability to cross the BBB into accessible biofluids (e.g., CSF and blood). EVs are nano- to microscale, membrane-bound particles secreted by virtually all cell types, including neurons and glial cells. They include different subtypes, including exosomes (30–150 nm) and microvesicles (100–1000 nm). Compared with other molecules directly derived from biofluids, EVs offer the unique advantage of protecting their cargo from degradation, making them particularly suited for early, minimally invasive detection of PTE-related molecular changes [52]. Although no EV-derived biomarkers have yet been identified specifically in PTE patients, multiple studies have investigated EV cargo in TBI [53,54]. A recent study analyzed data from 18 animal and 19 human studies and found that miRNAs were the most frequently reported EV components [53]. Among these, miR-124-3p was the most consistently upregulated and was associated with neuroprotective effects, particularly through the modulation of neuroinflammation via neuronal and microglial EVs. miR-21 was also commonly elevated. Frequently identified EV proteins included GFAP, UCH-L1, and tau, as described above.

Beyond their role as candidate biomarkers, EV cargo composition enables the characterization of key pathological processes such as neuroinflammation, neuronal injury, and epileptogenesis [101,102]. EV-mediated intercellular communication further suggests a functional role in disease propagation and modulation rather than passive biomarker release [107]. However, the clinical translation of EV-based biomarkers remains limited by the lack of standardized isolation protocols, variability in EV yield and purity, and heterogeneity of cargo depending on the cell of origin and sampling time. In addition, substantial variability across studies in EV isolation methods, analytical platforms, and cohort characteristics complicates direct comparison of findings and limits reproducibility. Furthermore, most current evidence is derived from TBI studies rather than PTE-specific cohorts, which restricts their specificity for epileptogenesis. Importantly, it remains unclear whether EV-derived signatures reflect causal mechanisms of epileptogenesis or represent secondary byproducts of injury-related processes, which challenges their interpretability as predictive biomarkers. These features highlight the utility of EVs not only for biomarker discovery but also for advancing mechanistic understanding and identifying novel therapeutic targets in PTE.

Across molecular biomarker studies, the main challenge is that mechanistic relevance does not necessarily translate into predictive specificity. Many candidate miRNAs, proteins, metabolites, and EV-derived signals are clearly linked to neuroinflammation, neuronal injury, and BBB dysfunction, but these same pathways are also common to TBI more broadly and are not unique to PTE. As a result, the strongest molecular candidates are often biologically plausible but clinically nonspecific. This distinction is important: biomarkers with high mechanistic value may still have limited standalone predictive utility unless they are validated longitudinally and interpreted in combination with imaging or electrophysiologic data.

5. AI-Assisted Predictive Biomarker Discovery for PTE

Artificial intelligence (AI) approaches are increasingly being explored for PTE prediction, particularly for integrating heterogeneous biomarker data and identifying complex patterns associated with epileptogenesis. Machine learning models have been applied to neuroimaging modalities, including structural MRI, CT, and EEG, to improve risk stratification after TBI. Using MRI-based lesion mapping and statistical modeling, Akrami et al. identified several brain regions, including the temporal lobe, cerebellum, and occipital cortex, that were associated with increased seizure susceptibility following TBI [23]. In a separate pediatric study, Jin et al. applied random forest modeling to clinical and limited imaging variables to predict post-traumatic seizures after repetitive mild TBI, highlighting the contribution of combined clinical-imaging approaches for seizure risk estimation [112]. Despite these advances, current AI-based studies remain constrained by limited external validation, heterogeneous datasets, and insufficient integration of multimodal biomarkers, which restricts their broader clinical applicability.

Moreover, AI techniques have been applied to electronic health records (EHR) and multimodal imaging data to improve PTE prediction accuracy. Rama-murthy et al. applied a graph-based attention network to longitudinal EHR data, revealing the potential of using graph embeddings to predict PTE risk more effectively than conventional ML models. While this approach outperformed traditional methods, it lacked imaging or EEG biomarkers, and external validation was not conducted [113]. Additionally, Wang et al. developed a prognostic model using Cox regression to predict PTE risk based on clinical and CT data, achieving an impressive C-index of 0.89. While this tool is interpretable and practical for clinical use, it did not incorporate advanced ML techniques or imaging beyond CT, limiting its potential for a more nuanced understanding of PTE [33].

AI-assisted approaches that combine multiple biomarker modalities, including neuroimaging and clinical data, show promise for improving PTE prediction [33]. The prospective EpiBioS4Rx study demonstrated that structural MRI features, including cortical thinning and regional volume deficits, were associated with seizure development after TBI, with combined imaging and clinical variables achieving up to 89% classification accuracy [22]. However, the study primarily evaluated early and late post-traumatic seizures rather than long-term PTE, limiting its direct applicability to epileptogenesis prediction. Similarly, La Rocca et al. used structural MRI network-scale features to predict post-traumatic seizures with a random forest algorithm, achieving 70% accuracy. While this approach offers insight into seizure prediction, its reliance on MRI data limits its generalizability to broader clinical populations [40]. These studies demonstrate that while AI holds significant promise in advancing PTE prediction, further refinement of algorithms, inclusion of additional data types, and external validation are needed to enhance predictive accuracy and clinical applicability.

Recent studies have expanded AI-assisted prediction by incorporating multimodal fusion frameworks that integrate structural MRI, diffusion MRI, functional MRI, EEG, and biofluid biomarkers within unified analytical pipelines. These approaches aim to capture complementary structural, functional, and molecular features associated with epileptogenesis. Fusion strategies based on information decomposition and feature selection methods, including IDSF, have shown improved classification performance compared with single-modality models in small cross-validation cohorts [46]. Nevertheless, the absence of large independent validation datasets continues to limit confidence in model generalizability and reproducibility.

Among available modalities, EEG-derived data remain particularly promising for AI-assisted biomarker discovery because they directly reflect network instability and epileptiform activity. Quantitative EEG features, including epileptiform abnormalities, multifractal signal properties, and paroxysmal slow-wave events, have demonstrated high predictive performance in experimental studies, with reported accuracies approaching 95% and AUC values near 0.98 [23,114]. Deep learning methods, including recurrent neural networks, have also been applied for automated detection of epileptiform discharges, enabling scalable analysis of large electrophysiological datasets [115]. However, differences in acquisition protocols, recording duration, and feature extraction methods continue to hinder standardization across studies.

AI-assisted biomarker discovery has also expanded into molecular domains, where circulating microRNAs and inflammatory mediators have emerged as potential predictors of epileptogenesis. Analyses from the EpiBioS4Rx project identified dynamically regulated plasma miRNAs associated with post-traumatic epileptogenic processes, while elevated IL-6 levels have been linked to increased seizure risk in prospective cohorts [63,116]. At present, however, these molecular candidates remain exploratory because of limited longitudinal validation and variability in sampling methodologies.

Another important development is the use of explainability approaches, including Shapley-value analysis and region-wise importance mapping, which improve biological interpretability of AI-generated predictions. These methods have identified temporal lobe structures, cerebellar regions, and white matter pathways as potential contributors to PTE risk [46].

Importantly, the apparent performance of many AI-assisted prediction models should be interpreted cautiously. Studies reporting very high accuracies or AUC values are frequently based on small experimental cohorts, cross-validation frameworks, or highly selected datasets, which increases the risk of overfitting and limits generalizability. At present, clinically translatable prediction remains largely dependent on routinely available variables, particularly CT-derived structural features, basic EEG findings, and established clinical parameters. In contrast, multimodal fusion models, advanced qEEG metrics, circulating molecular biomarkers, and deep learning approaches remain predominantly exploratory despite their mechanistic and computational promise. Thus, the primary challenge in the field is no longer identification of candidate biomarkers, but rather large-scale validation, harmonization, and demonstration of reproducible clinical utility across independent cohorts.

Despite substantial progress, current AI-based studies remain dominated by relatively small single-center cohorts and heterogeneous analytical frameworks. Future advances will therefore depend on standardized multimodal datasets, external validation, and prospective multicenter studies capable of supporting clinically reliable prediction models.

6. Conclusions

Early diagnosis of PTE remains a significant clinical challenge. While advances in neuroimaging, molecular biomarker discovery, and computational approaches have deepened our understanding of epileptogenesis, no single modality provides sufficient predictive power across heterogeneous patient populations. Neuroimaging techniques (CT, MRI, EEG) capture structural and functional alterations, whereas molecular biomarkers reflect dynamic biological processes. Rather than reiterating modality-specific advantages, current evidence indicates that clinically meaningful prediction will depend on the integration of complementary data types within standardized multimodal frameworks. AI-based approaches further support this direction by enabling the analysis of complex, high-dimensional datasets and identifying nonlinear relationships across biomarker domains. Thus, the primary focus of the field is shifting from individual biomarker discovery toward integrated, data-driven risk stratification models.

Despite substantial progress, no biomarker has yet achieved routine clinical implementation for PTE prediction. Among available approaches, CT-based features combined with clinical variables remain the most immediately translatable due to their accessibility and validation in large human cohorts. In contrast, EEG/qEEG and MRI provide higher-resolution functional and network-level insights but are limited by variability in acquisition protocols, cost, and standardization challenges. These modalities are therefore best viewed as complementary components within multimodal prediction strategies rather than standalone solutions.

Molecular biomarkers, including miRNAs, proteins, metabolites, and EV-derived molecules, remain largely exploratory. Although they offer important mechanistic insights into neuroinflammation, neuronal injury, and BBB dysfunction, their predictive performance is inconsistent and often lacks specificity for PTE compared with general brain injury. Additionally, their translational potential is constrained by small cohort sizes, methodological heterogeneity, and limited validation in longitudinal human studies.

Collectively, the key barriers to clinical translation include biological heterogeneity of TBI, lack of standardized measurement protocols, variability in study design and outcome definitions, and insufficient external and longitudinal validation. Addressing these limitations will require large-scale, prospective, and multimodal studies with harmonized methodologies. Future progress should prioritize reproducibility, cross-cohort validation, and integration of imaging, molecular, and computational data to enable robust and clinically actionable prediction of PTE.