Early alterations of thalami- and hippocampi-cortical functional connectivity as biomarkers of seizures after traumatic brain injury

Abstract

The Epilepsy Bioinformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx, project 3) is a prospective multicenter clinical observational study to identify early biomarkers of epileptogenesis after moderate-to-severe traumatic brain injury (TBI). We used a seed-based approach applied to acute (i.e., ≤14 days) fMRI imaging data, directly testing the hypothesis that the presence of seizures up to two years following brain trauma is associated with functional changes within hippocampi and thalami-cortical networks. Additionally, we hypothesized that the network connectivity involving thalami and hippocampi circuits underlying early and late-onset seizures would differ. Approximately 30% of the initial dataset was deemed unusable due to MRI issues. Approximately 50% of the enrolled sample was lost to a 2-year follow-up. After preprocessing the fMRI data, approximately 40% of the follow-up sample had to be excluded from the analysis due to excessive in-scanner movements, as assessed by state-of-the-art quality control protocols. Only 37 patients provided data that was suitable for the seed-based analysis. Despite these challenges, the remaining, high-quality data returned noteworthy findings. We identified specific hippocampi and thalami biomarkers associated with both early and late seizures following TBI (p < .05, FWE-corrected at the cluster level). The predictive capability for the development of late seizures after TBI, when adding fMRI data to demographic and clinical data, provided 88% accuracy — an additional 8% improvement compared to using demographic and clinical data alone. Our findings highlight the potential of fMRI for uncovering, in hippocampal and thalamic cortical networks, biomarkers of early and late seizures following TBI. However, they also highlight the important challenges that need to be overcome in order for fMRI to become an effective biomarker and prognostic tool in the intensive care context.

1. Introduction

Posttraumatic epilepsy (PTE) is a common long-term health problem associated with traumatic brain injury (TBI) (Lowenstein, 2009), as well as a risk factor for long-term mortality after TBI (Uski et al., 2018). Increased incidence of seizures after TBI is related to injury severity and specific injury characteristics, such as penetrating injury, skull fracture, dural injury, hemorrhagic lesion, surgical treatment, and prolonged impaired consciousness (Asikainen et al., 1999; Frey, 2003; Temkin, 2003). The timing of seizures after TBI is variable and defined as early (within 7 days) and late (after 7 days) post-injury. The process of epileptogenesis occurs over the latent period and it remains unclear if there are viable biomarkers of this process (Agrawal, Timothy et al. 2006; Gupta, Sayed et al. 2014; Pitkänen and Immonen, 2014). However, the latency period represents an exceptional opportunity to discover novel biomarkers, including non-invasive imaging biomarkers, which capture or identify the epileptogenesis process early following TBI.

With this background in mind, the Epilepsy Bioinformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx) is a prospective observational study of moderate-to-severe TBI patients to identify biomarkers that may help identify and prevent seizure occurrence and better understand the mechanisms underlying PTE (Vespa, Shrestha et al. 2019). In this study, patients undergo continuous electroencephalography, MRI, and other serum collection starting in the first week after TBI and are followed up to 2 years. Given the extensive literature linking seizures with abnormalities in the hippocampi and thalami in animals models of TBI (Pitkänen et al., 2009; Vespa, McArthur et al. 2010; Immonen et al., 2013; Shultz et al., 2013; Golub and Reddy, 2022a, Golub and Reddy, 2022b), and TBI patients (Vespa, McArthur et al. 2010; Lutkenhoff et al., 2020), the EpiBioS4Rx project a priori hypothesized that acute abnormalities within hippocampi or thalami-cortical networks could indicate the presence of seizures in patients after moderate-to-severe TBI (Vespa, Shrestha et al. 2019).

Here, using a seed-based approach in acute post-injury fMRI data, we directly tested the hypothesis that the presence of early seizures following TBI is associated with functional changes within hippocampi and thalami-cortical networks. Additionally, we hypothesized that the pathological phenotypes—as reflected in differences in thalami and hippocampi networks connectivity, underlying the emergence of early versus late seizures after brain trauma would diverge, as previously shown to occur using anatomical data (Lutkenhoff et al., 2020b, Lutkenhoff et al., 2020a).

2. Methods

2.1. Patient screening and enrollment

Patients admitted into the ICU after an acute moderate-severe TBI involving a frontal and/or temporal lobe hemorrhagic contusion were screened across 12 sites. Patients were eligible for enrollment in the EpiBioS4Rx project up to 72 h post-TBI. Criteria included ages 6 to 100 and Glasgow Coma Scale (GCS (Teasdale and Jennett, 1974), 3 to 13. Patients were excluded for isolated diffuse axonal injury, isolated epidural or subdural hemorrhages, isolated anoxic brain injury, pregnancy, incarceration, and pre-existing neurodegenerative or epileptic disorders (Lutkenhoff et al., 2020b, Lutkenhoff et al., 2020a). This work was approved by the UCLA Institutional Review Board (IRB# 16–001 576) and the local review boards at each EpiBioS4Rx Study Group institution. Assent and written consent were obtained from the legal representative as per state law.

Longitudinal assessment for PTE was obtained at discharge and on days 30, 90, 180 days, and 1- and 2 years post-injury with the Ottman PTE Questionnaire (Ottman et al., 2010). Patients were divided into three groups: patients who experienced no seizures (no seizure group, NS), patients who experienced at least one seizure starting during the first-week post-injury (early seizure group, ES), and patients who experienced at least one seizure starting after the first-week post-injury, up to two years (late seizure group, LS, (Lowenstein, 2009).

Patients received 24 h cEEG for 72 h minimum during the first 7 days after TBI. Scalp cEEG monitoring was performed at the patient's bedside using a 16–21 channel bipolar and referential composite montage (implemented according to each center's intensive care unit protocols). Mandatory parameters included a low-frequency filter at 0.1 Hz, a high-frequency filter at 50 Hz, Notch Filter, and a 200 Hz minimum sampling rate. Each site used its standard of care electrodes, including disk or needle scalp electrodes.

2.2. Data acquisition

High-resolution MRIs were acquired on 1.5 or 3 T MR systems, including anatomical T1-weighted and functional T2*-weighted echo planar images acquisitions (See Suppl. Table 1 and 2 for detailed parameter listing), performed up to 14 days (+4 days) post-injury.

2.3. Data processing

Processing of the fMRI data was performed using UF²C (http://www.lni.hc.unicamp.br/app/uf2c, https://www.nitrc.org/projects/uf2c (de Campos et al., 2020), a toolbox that runs in the MATLAB platform (MATLAB, 2020a, The MathWorks, Inc., Natick, Massachusetts, United States) with SPM12 (http://fil.ion.ucl.ac.uk/spm/). We performed the T1-weighted image coregistration with the fMRI mean image, tissue segmentation, and spatial normalization to the MNI-152 template. We preprocessed the functional images based on volumes realignment, normalization to the MNI-152 template, and smoothing at 6 × 6 × 6 mm³ FWHM. To remove variance attributable to head and respiration/cardiac-induced motion, we applied a pipeline previously shown to present the best performance in denoising TBI fMRI datasets (Weiler et al., 2021), which consists of regressing the six head motion parameters, the five components with greater eigenvalue for cerebrospinal fluid and white matter masks (aCompCor, (Behzadi et al., 2007), and any volume containing excessive motion (i.e., framewise displacement, FD) greater than 0.25 mm (Satterthwaite et al., 2013). Furthermore, participants were excluded if they presented less than 4 min of preprocessed data, mean FD greater than 0.25 mm, more than 20% of the volumes with FD greater than 0.2 mm, or any volume with FD greater than 5 mm (Satterthwaite et al., 2013). Additional preprocessing steps included detrending, band-pass filtering (0.008–0.1 Hz), and grey matter masking of functional images.

After preprocessing, we performed a voxel-wise seed-based functional connectivity analysis using the bilateral thalami (903 voxels, 7224 mm³) and hippocampi (1175 voxels, 9400 mm³) as seeds (MNI masks). Positive and negative correlation maps were estimated using Pearson's correlation coefficient, and then transformed to Fisher's Z estimates using Fisher's r-to-z transformation for subsequent harmonization of the data and statistical analysis.

Previous studies have reported the need to perform data harmonization in multisite MRI investigations (Chen et al., 2014). We first performed an exploratory analysis to confirm the existence of site effects in the data. Then, we controlled for site effects using the MATLAB version of ComBat v1.0.1 (Fortin et al., 2017; Fortin et al., 2018), https://github.com/Jfortin1/ComBatHarmonization), a software that removes inter-site technical variability even in small samples while preserving inter-site biological variability using a popular batch-effect correction tool used in genomics (Johnson et al., 2006). The Fisher's-z-transformed positive and negative functional connectivity maps were harmonized across sites, while disease group, age, injury severity (i.e., admission GCS), and postinjury day of the MRI session were biological covariates kept during the removal of site effects.

2.4. Data analysis

To investigate if thalami and hippocampi functional connectivity profiles could indicate the presence of seizures in patients after moderate-severe TBI, we performed a voxel-wise analysis of variance with the harmonized positive and negative functional connectivity maps separately as dependent variables, and group (NS, ES, LS) as the independent variable, regressing out the effects of age, injury severity, and postinjury day of the MRI session. Additionally, if the patient received any medication (classified as anti-seizure, anesthetic, benzodiazepine, anti-psychotic, or opioid) that could potentially affect the blood-oxygen-level-dependent signal during the MRI, a separate regressor was generated containing a value of 1 at that variable, and 0 at all others. In other words, if the patient received an anti-seizure medication during the MRI period, a value of 1 was added to the anti-seizure medication column, while a value of 0 was added to the anesthetic, benzodiazepine, anti-psychotic, and opioid columns. Group comparison was voxel-wise thresholded at p < .001 (uncorrected) and cluster-wise corrected using family-wise error (FWE) at p < .05 using SPM12.

In addition, to assess the relative importance of thalami and hippocampi networks in predicting vulnerability to seizures, we combined demographic, clinical, and MRI data in logistic regression models. Demographic information (age, sex, postinjury day of the MRI session), clinical data (admission GCS total), and the mean functional connectivity maps of thalami and hippocampi-cortical networks were entered in five binomial logistic regression models to distinguish patients who did not have seizures from patients that did (collapsing the ES and LS groups), as follows: Model 1: demographic data (age, sex, and postinjury day of the MRI session); Model 2: demographic data + clinical data; Model 3a: demographic data + clinical data + average thalami positive and negative functional connectivity; Model 3 b: demographic data + clinical data + average hippocampi positive and negative functional connectivity; Model 4: demographic data + clinical data + average thalami and hippocampi positive and negative functional connectivity.

Then, to compare the relative importance of each set of variables in predicting the occurrence of late seizures, we performed a multivariate receiver-operating characteristic curve (ROC) analysis using the same five models but considering only the LS group. Given that the differences across groups mostly involved positive correlations (see Fig. 1, Fig. 2), we decided not to include a model containing only negative maps. All non-imaging statistical analysis was carried out using the SPSS v28 package (IBM Corp. Released, 2022. IBM SPSS Statistics for Windows, Version 28.0. Armonk, NY: IBM Corp), and JASP software (Version 0.16.3, (2022) (JASP Team, 2022).

Fig. 1.

Pattern 1: early seizure biomarkers.

(a) Regions presenting higher positive connection with the thalami in the TBI early seizure group compared to the TBI no seizure group; (b) Regions presenting higher positive connection with the hippocampi in TBI early seizure group compared to the TBI no seizure group; (c) Regions presenting higher positive connection with the hippocampi in the TBI early seizure group compared the TBI late seizure group (p < .05, FWE-corrected at the cluster level). Scatter plots showing mean functional connectivity values for the regions showing significant differences between the groups. **p < .001, one sample t-test against zero.

Fig. 2.

Pattern 2: late seizure biomarkers.

(a) Regions presenting lower positive functional connectivity with the hippocampi in the TBI late seizure group compared to the TBI no seizure group; (b) Regions presenting increased negative functional connectivity with the hippocampi in the TBI late seizure group compared to the TBI no seizure group (p < .05, FWE-corrected at the cluster level). Scatter plots showing mean functional connectivity values for the regions showing significant differences between the groups. **p < .001, *p < .05, one sample t-test against zero.

3. Results

3.1. Demographic and clinical data

At the time of this manuscript preparation, 239 patients had been assessed for eligibility. A total of 118 patients were excluded, and 57 patients were lost to follow-up (details in Suppl. Fig. 1). Twenty-seven patients were excluded after fMRI preprocessing due to preprocessing error or excessive in-scanner motion (cf., (Weiler et al., 2022) rendering the data not suitable for seed-based analysis (Satterthwaite, Wolf et al. 2012) (Suppl. Fig. 1). Table 1 shows the demographic and clinical data of the 37 patients included in the final sample. We could not detect any statistically significant difference in sex distribution (χ²₍₂₎ = 1.8, p = .407), age (H₍₂₎ = 4.227, p = .121), admission GCS (H₍₂₎ = 3.850, p = .146) or postinjury day of the MRI session (H₍₂₎ = 0.695, p = .706). All patients who experienced seizures during the first week after TBI also continued to have seizures in the subsequent two years. In other words, all patients classified as having ES also exhibited seizures after the initial week. Conversely, patients classified as having LS did not experience any seizures during the first week, but rather developed seizures only after this initial period.

Table 1.

Clinical and demographic information of patients included in this analysis.

	No seizure	Early seizure	Late seizure
N	16	8	13
Age (years)	45 ± 23 (7–84)	47 ± 21 (19–72)	32 ± 17 (15–65)
Sex (female)	6 (37.5%)	2 (25%)	2 (15.3%)
MRI post-injury day	9.81 ± 5.63 (0–18)	8.38 ± 7.19 (1–17)	10.38 ± 5.85 (1–18)
Admission GCS	9.41 ± 4.25 (3–15)	8.13 ± 4.05 (3–14)	6.31 ± 3.37 (3–14)
Contusion (%)	75	50	100
Brain edema (%)	50	50	76.92
Skull fracture (%)	93.75	75	100

Mean ± stdev (range). GCS = Glasgow Coma Scale.

3.2. Harmonization of the data

To investigate and correct site effects, we assessed the z-transformed functional connectivity maps pre- and post-harmonization. We first performed an exploratory analysis to investigate any correlations between site and thalami and hippocampi-cortical connectivity maps. As shown in Suppl Fig 2, multiple brain regions exhibited unwanted significant correlation with site (p < .05, FWE-corrected at the cluster level) in the pre-harmonized images, confirming the need to perform cross-site data harmonization. These significant correlations were no longer observed after harmonization. Suppl. Figs 3-8 depict the pre- and post-harmonization distributions of functional connectivity values for thalami and hippocampi networks.

Fig. 3.

ROC plots for the five models tested to predict the occurrence of late seizure in TBI patients.

3.3. Pattern 1: increased positive connectivity in thalami and hippocampi networks as biomarkers for early seizure following TBI

Voxel-wise analysis revealed that the ES group presented higher positive functional connectivity in thalami and hippocampi networks than in the NS and LS groups, namely pattern 1.

In ES patients compared to the NS group, the thalami were more strongly connected to the right cuneus, and bilateral medial superior frontal regions (Fig. 1a–Table 2, p < .05, FWE-corrected at the cluster level), whereas the hippocampi were more strongly connected to the right supplementary motor area and right cuneus (Fig. 1b–Table 2, p < .05, FWE-corrected at the cluster level). After finding significant differences between the groups, we then explored if such a difference was mainly due to a new pattern of functional connectivity in the ES group (i.e., inexistent functional connectivity in the NS group between those regions), or if the functional connectivity between the regions was just stronger in the ES group. We then masked the regions that were significantly different between the groups and extracted the mean connectivity value for each patient. A one-sample t-test against zero (i.e., testing if that specific pattern of connectivity significantly differed from zero, meaning that the specific spatial connectivity existed within the group), was statistically significant for both groups (Thalami network, NS group, t ₍₁₅₎ = 9.354, p = < 0.001, 95% CI [0.10, 0.16], Hedges' correction = 2.21. ES group, t ₍₇₎ = 11.252, p = < 0.001, 95% CI [0.19, 0.29], Hedges' correction = 3.53. Hippocampi network, NS group, t ₍₁₅₎ = 10.023, p = < 0.001, 95% CI [0.10, 0.16], Hedges' correction = 2.38. ES group, t ₍₇₎ = 17.478, p = < 0.001, 95% CI [0.26, 0.34], Hedges' correction = 5.49), suggesting that these patterns of connectivity are present in both NS and ES groups, but the ES group presents stronger connectivity (scatter plots in Fig. 1a and b).

Table 2.

Brain regions showing different functional connectivity in the early, late, and no seizure groups.

Early seizure vs no seizure group
Seed	Anatomical location	p (FWE-corr)	cluster size (voxels)	T-value	Z-value	x (MNI)	y (MNI)	z (MNI)
Thalami – positive correlations	right cuneus	0.034	83	4.7	3.96	20	−68	28
	right middle occipital	0.031	85	4.62	3.91	34	−76	12
	left medial superior frontal	0.024	90	8.1	5.67	−10	64	8
	right medial superior frontal	0.012	104	5.72	4.56	14	62	16
	right cuneus	0.006	117	4.96	4.13	16	−70	40
Hippocampi – positive correlations	right supp motor area	0.050	67	7.57	5.45	8	−22	56
	right cuneus	0.035	75	6.46	4.94	10	−76	42
Early seizure vs late seizure group
Hippocampi - positive correlations	right cerebellum crus II	0.039	73	5.65	4.52	30	−72	−38
	right inferior parietal	0.007	103	5.03	4.17	44	−36	54
	right supp motor area	0.003	120	6.03	4.73	10	−22	54
	left middle frontal	0.000	216	7.82	5.56	−34	44	4
Hippocampi - positive correlations	right middle frontal	0.008	102	6.24	4.84	34	46	0
	left cerebellum crus I	0.000	159	8.21	5.72	−6	−78	−18
	left middle frontal	0.000	338	7.79	5.55	−40	44	0
Late seizure vs no seizure group
Hippocampi - negative correlations	left middle temporal	0.021	49	5.43	4.4	−50	−50	12

The ES group also presented higher positive functional connectivity in the hippocampi network when compared to the LS group. Regions showing stronger positive connectivity with the hippocampi included the right cerebellum, right inferior parietal, right supplementary motor area, and left middle frontal (Fig. 1c–Table 2, p < .05, FWE-corrected at the cluster level). A one-sample t-test against zero using the connectivity values within the masked results showed that this pattern of connectivity was statistically significant for both groups (ES group, t ₍₇₎ = 18.242, p = < 0.001, 95% CI [0.25, 0.33], Hedges' correction = 5.72; LS group, t ₍₁₂₎ = 16.514, p = < 0.001, 95% CI [0.11, 0.15], Hedges' correction = 4.28), suggesting that this pattern of connectivity is present in both the ES and LS groups, but the ES group presents higher connectivity values (scatter plots in Fig. 1c).

3.4. Pattern 2: decreased positive and increased negative connectivity in the hippocampi network as biomarkers for late seizure following TBI

In addition to the lower connectivity observed in the hippocampi network in the LS group compared to the ES group described above (Fig. 1c), hypoconnectivity was also observed when the LS group was compared to the NS group, namely pattern 2. More specifically, the voxel-wise analysis revealed that the LS group presented lower positive functional connectivity between the hippocampi and bilateral middle frontal and left cerebellum than the NS group (Fig. 2a–Table 2, p < .05, FWE-corrected at the cluster level). A one-sample t-test against zero using the connectivity values within the masked results showed that this pattern of connectivity was statistically significant for both groups (NS group, t ₍₁₅₎ = 10.466, p = < 0.001, 95% CI [0.19, 0.29], Hedges' correction = 2.48; LS group, t ₍₁₂₎ = 13.130, p = < 0.001, 95% CI [0.11, 0.15], Hedges' correction = 3.41), but the LS group presented lower connectivity values (scatter plots in Fig. 2a).

Regions presenting increased negative functional connectivity (anti-correlations) with the hippocampi in the LS group compared to the NS group included the left middle temporal region (Fig. 2b–Table 2, p < .05, FWE-corrected at the cluster level). A one-sample t-test against zero using the connectivity values within the masked results showed that this pattern of connectivity was statistically significant for only the LS groups (NS group, t ₍₁₅₎ = −1.668, p = .116, 95% CI [-0.002, 0.0002], Hedges' correction = −0.396; LS group, t ₍₁₂₎ = −3.335, p = .006, 95% CI [-0.091, −0.019], Hedges' correction = −0.866), suggesting that this pattern of anti-correlations is not present in the NS group, but only in the LS group (scatter plots in Fig. 2b).

3.5. Early dysfunction in thalami and hippocampi networks predicts the occurrence of late seizures

When we compared the NS group vs. seizure group, the latter combining ES and LS patients, all binomial logistic regression models that included demographic information, clinical data, and the mean functional connectivity maps of thalami and hippocampi networks as predictors, failed to show statistically significant results (Model 1, χ²₍₃₎ = 3.415, p = .332; Model 2, χ²₍₃₎ = 6.502, p = .165; Model 3a, χ²₍₃₎ = 6.718, p = .348; Model 3 b, χ²₍₃₎ = 7.432, p = .283; Model 4, χ²₍₃₎ = 6.538, p = .366; and Model 5, χ²₍₃₎ = 7.601, p = .473).

However, when we aimed to compare the relative importance of each set of variables in predicting the occurrence of late seizures (LS group vs. NS group) in the multivariate ROC analysis, models were statistically significant. Amongst the measures calculated to evaluate the models, we describe below the Nagelkerke Pseudo-R² (a logarithmic scoring rule; the higher the score the higher the fitness of the predictive model); specificity (the percentage of TBI patients that did not have seizures and were correctly predicted by the model—i.e., true negatives); sensitivity (the percentage of TBI patients that had late seizures and were correctly predicted by the model—i.e., true positives); area under the curve (AUC, evaluates the model predictions using a probabilistic framework, showing the relationship between false positive and true positive rates for different probability thresholds of model predictions). As shown in Table 3 and Fig. 3, the model that included thalami and hippocampi functional connectivity values presented substantially higher metrics than the models that included demographic and clinical data only. Specifically, the AUC for model 4 was 87.7, which indicates an approximate 88% chance that the neurologist will correctly distinguish a TBI patient that will have late seizures (compared to around 74% chance if only using demographic data). The specificity for model 4 was 86.7, which indicates approximately 87% chance of correctly predicting an absence of seizures following a TBI, and sensitivity was 84.6, indicating an approximate 85% chance of correctly predicting the occurrence of seizures in 2 years following a TBI.

Table 3.

Metrics evaluating the four models used to predict the occurrence of late seizures in TBI patients.

Model	df	χ2	pa	Nagelkerke R2	Specificity	Sensitivity	AUC
Model 1	25	4.96	0.17	0.21	68.8	61.5	73.6
Model 2	23	8.49	0.07	0.35	80	69.2	79
Model 3a	21	11.05	0.08	0.43	73.3	76.9	82.6
Model 3 b	21	12.67	0.04	0.48	66.7	76.9	83.1
Model 4	19	15.07	0.05	0.55	86.7	84.6	87.7

^aOmnibus Tests of Model Coefficients were used to test the model fit. If the Model is significant, this shows that there is a significant improvement in fit as compared to the null model. Model 1: demographic data (age, sex, and postinjury day of the MRI session); Model 2: demographic data and clinical data; Model 3a: demographic data, clinical data, average thalami positive and negative functional connectivity; Model 3 b: demographic data, clinical data, average hippocampi positive and negative functional connectivity; Model 4: demographic data, clinical data, average thalami and hippocampi positive and negative functional connectivity.

Model 1: age + sex + postinjury day of the MRI session; Model 2: age + sex + postinjury day of the MRI session + injury severity (admission GCS); Model 3a: age + sex + postinjury day of the MRI session + injury severity + positive and negative thalami functional connectivity; Model 3 b: age + sex + postinjury day of the MRI session + injury severity + positive and negative hippocampi functional connectivity; Model 4: age + sex + postinjury day of the MRI session + injury severity + positive and negative thalami and hippocampi functional connectivity.

4. Discussion

In our study, we employed a seed-based approach utilizing fMRI data to investigate the potential of identifying biomarkers for seizure occurrence following a TBI. Specifically, we tested the hypothesis that (i) the presence of seizures following TBI is associated with changes in brain-wide thalami and hippocampi-cortical connectivity and that (ii) connectivity changes are different for patients developing seizures shortly after injury (i.e., <7 days post-injury) as compared to patients developing seizures at a later time. As hypothesized, significant differences in thalami and hippocampi-cortical connectivity were observed across groups – with early and late seizure groups showing very different connectivity patterns compared with the NS group as well as each other. Specifically, we found that TBI patients presenting early seizures exhibited hyper-positive connectivity in thalami and hippocampi-cortical networks compared with TBI patients who presented no seizures and with TBI patients who presented late seizures (pattern 1). On the contrary, TBI patients who presented late seizures presented lower positive but higher negative connectivity in the hippocampi network when compared with the TBI group that had no seizures (pattern 2). In addition, models that included dysfunctions in thalami and hippocampi networks were significantly able to predict the presence of late seizures following a TBI, whereas models that used only demographic and clinical data could not. Dysfunctions in thalami and hippocampi networks were able to predict with approximately 88% chance of the occurrence of late seizures following a TBI.

How our fMRI findings as discussed here link to the molecular mechanisms of epileptogenesis is unclear and yet to be fully understood. However, the pattern 1 dysfunction of the ES group could be speculated as a result of the excitotoxic environment and inflammation known to occur acutely following TBI (Lucke-Wold et al., 2015). Indeed, the imbalance between excitation and inhibition, which relates to increased extracellular glutamate in the brain and/or reduction in GABA concentrations, has been a longstanding proposed mechanism regarding ictogenesis and epileptogenesis (Huusko et al., 2015). The contribution of glutamate excitotoxicity to early-onset seizure has been demonstrated in many rat models of TBI, especially in hippocampal (Lowenstein, Thomas et al. 1992; Zanier et al., 2003; Drexel et al., 2015) and thalamic regions (Immonen et al., 2019, Sowers et al., 2021). In this sense, immediate and early epileptogenesis are not considered to be “epileptic” and are thought to be a direct product of the injury itself (Golub and Reddy, 2022a, Golub and Reddy, 2022b). This excitotoxic environment could lead to the structural alterations often observed as atrophy in TBI (Vespa, McArthur et al. 2010; Shultz et al., 2013; Lutkenhoff et al., 2020) which, in turn, correlates with the functional outcome of patients (Lutkenhoff et al., 2020a, Lutkenhoff et al., 2020b) and represent a risk factor for the development of PTE (Tubi et al., 2019).

Early seizures are believed to have a different underlying pathogenesis than late seizures following a TBI (Agrawal, Timothy et al. 2006). Consistent with this view, we found the two groups to have different patterns of brain-wide thalami and hippocampi connectivity, as compared with each other and with the no seizure group. The divergent profiles of patterns 1 and 2 also suggest that the neurobiological underpinnings implicated in the development of late seizures are already present acutely, well in advance of any observable behavioral or manifestation. In addition, logistic regression models aiming to predict the occurrence of any seizures following a TBI (i.e., combining the early and late seizure groups) failed to find statistically significant biomarkers, further corroborating the hypothesis that early and late seizures present different underlying pathogenesis. Conversely, when we differentiated the two groups, the models could significantly predict which patient experienced, over the following 24 months, late seizures. Importantly here, only the models that included hippocampi and thalami functional connectivity data could correctly classify patients into no seizure or late seizure, whereas the models that included only demographic and clinical data could not. It should be noted that all the patients with an early seizure also presented late seizure, indicating that the occurrence of early seizure is also a strong risk factor for the development of late ones, as previously reported (Frey, 2003; Temkin, 2003). How exactly the physiopathology giving rise to early seizures plays into developing seizures in the long term, and the degree to which this group is separable from patients developing late seizures in the absence of early ictal episodes, remains to be understood.

Interestingly, in previous work conducted in our laboratory that used similar models, the authors found that injury severity, the left temporal pole, and left frontal pole were significantly associated with the probability of seizures after TBI (Tubi et al., 2019, Lutkenhoff et al., 2020). In their models, early and late seizure patients were combined into one group, suggesting that structural abnormalities in such regions could play a role in predicting the occurrence of seizures, either early or late. In another study, the authors used an innovative multiplex network approach to find informative complex network features to distinguish seizure-free subjects and seizure-affected subjects with an accuracy of 70% and an AUC of 76% by using T1-weighted MRI data (La Rocca et al., 2020). In the present work, functional abnormalities in thalami and hippocampi-cortical networks could not significantly predict the occurrence of seizures or not, but they could correctly distinguish a TBI patient that will have late seizures with an approximate 88% chance. In addition, this model presented an approximately 87% chance of correctly predicting the absence of seizures and an approximately 85% chance of correctly predicting the occurrence of seizures in 2 years following a TBI. In essence, our findings, combined with those previous reports, suggest that whereas both early and late seizure patients present similar anatomical abnormalities early on, they present functional alterations with different etiologies and neurobiological underpinnings that can already be detected in a short period following the injury. It is therefore a speculative yet not improbable conjecture to propose that the pattern of thalami and hippocampi-cortical connectivities during the initial period following the TBI in LS patients, become more similar to the pattern of ES patients, later when they develop seizures. However, longitudinal MRI acquisitions would be necessary to test this hypothesis.

It is important to consider the specific context and outcome being investigated, since our study also revealed the challenges associated with utilizing fMRI data to identify biomarkers for seizures following brain trauma. The difficulties encountered in our research shed light on two primary issues that need to be addressed when using fMRI as biomarkers for post-TBI seizures: ensuring timely access to MRI for patients early on after the injury, and mitigating head movements during the scanning process.

Ensuring timely access to MRI for patients soon after a brain injury is undeniably a significant issue and was a major limitation in this study, exemplified by the loss of 32% (78 out of 239) of our initial sample. In many cases, the patient's condition must be carefully assessed to determine if they are stable enough to be transported from the ICU for an extended period. Unfortunately, this process introduces potential risks, as approximately one-third of patients who are transferred from the ICU to an imaging suite may experience adverse events (McLean and Thompson, 2023). While the use of portable MRI technology offers a promising solution to the challenges associated with timely access to imaging in the ICU, it is important to acknowledge that the widespread implementation of this technology in hospitals is still limited.

Furthermore, head movements during fMRI scanning pose a significant challenge in obtaining high-quality data. In our study, a considerable number of participants were excluded due to excessive in-scanner movements, which can introduce artifacts and compromise the accuracy of the results. Developing further strategies to mitigate head movements, such as improving participant comfort and optimizing imaging protocols, is crucial for obtaining reliable fMRI data in this population. Collaborative efforts between clinicians, researchers, and imaging specialists are necessary to ensure early access to MRI services for TBI patients and to refine imaging protocols to minimize motion artifacts.

Lastly, in interpreting our findings, it is important to be mindful of some limitations. Above all, the small sample size could be considered a significant weakness of this study. Many patients were dropped from the study because of the high-quality control standard adopted here (such as the low threshold for in-scanner head movement), and an important development in this area for future studies is to adopt techniques that diminish as much as possible in-scanner head movement. Relatedly, due to the limited sample size in our study, we were unable to perform a proper split into training/development and test cohorts. Consequently, the risk of overfitting remains a concern, and our findings should be interpreted with caution. Future studies with larger sample sizes should consider incorporating this crucial step to ensure more reliable and generalizable results. Second, although we tried to statistically control for the effects of any medication that could potentially affect the blood oxygenation level dependent signal during the MRI, their neural effects are yet to be unveiled.

To sum up, the delay in the emergence of chronic epileptogenesis after the initial injury represents an exceptional opportunity for intervention with antiepileptogenesis therapies once they have been developed. In this sense, there is a great need for the identification of biomarkers that provide quantitative measures of the process of post-traumatic epileptogenesis. However, while there is extensive reporting on functional changes following TBI (Johnson et al., 2012; Iraji et al., 2015; Sours et al., 2015), studies focusing specifically on the epileptogenic process following brain trauma are still scarce. In the current work, we found that early and late seizure patients have different phenotypes of brain-wide thalami and hippocampi connectivity, as compared with each other and with the no seizure group, suggesting that acute hippocampi and thalami-cortical network profiles are potential biomarkers for early and late seizure following a TBI. In addition, using functional connectivity data of thalami and hippocampi networks, we could distinguish, with almost 88% precision, which patients presented late seizures after TBI from those who did not. Our model was able to correctly predict with an 87% chance the absence of seizures following a TBI and correctly predict with an 85% chance the occurrence of a seizure in 2 years following a TBI. However, for fMRI to be effectively utilized as biomarkers in this context and to enhance the prediction of late seizures after TBI, several methodological challenges need to be addressed, and mitigating strategies should be developed.

CRediT authorship contribution statement

Marina Weiler: Writing – review & editing, Writing – original draft, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Evan S. Lutkenhoff: Writing – review & editing, Writing – original draft, Formal analysis, Data curation. Brunno M. de Campos: Writing – review & editing, Writing – original draft, Methodology, Formal analysis. Raphael F. Casseb: Writing – review & editing, Writing – original draft, Methodology, Formal analysis. Paul M. Vespa: Writing – review & editing, Writing – original draft, Supervision, Project administration, Investigation, Funding acquisition, Conceptualization. Martin M. Monti: Writing – review & editing, Writing – original draft, Supervision, Funding acquisition, Formal analysis, Conceptualization.

Declaration of competing interest

None.

Appendix A. Supplementary data

The following is the Supplementary data to this article:

Data availability

Data will be made available on request.