Harnessing machine learning for functional connectivity-based feature discovery in post-traumatic epilepsy

Highlights

• •Leakage-free nested CV-RFE identifies a reproducible connectivity signature that characterizes the PTE-diseased state in our cohort.
• •Data-driven feature selection is essential for classification performance, with only four stable functional connections (from an original feature space of 4950) able to distinguish PTE from TBI with an AUC of 0.95.
• •We identify a specific connectivity phenotype in our PTE cohort which includes the Visual network as a hub of PTE-related changes, left-hemisphere overrepresentation and hyper-coupling, and right-hemisphere un-coupling.
• •The resting-state network disruptions identified by our ML paradigm went unrecognized by direct comparison between groups and traditional statistics, supporting the use of advanced data-driven methodologies for exploring functional connectivity in PTE.

1. Introduction

An estimated 64-74 million people worldwide suffer from traumatic brain injuries (TBIs) each year, with a significant subset developing post-traumatic epilepsy (PTE) (Dewan et al., 2019). Defined as two or more unprovoked seizures occurring more than one week after a TBI, PTE affects 2–50% of individuals with moderate to severe TBI and accounts for approximately 20% of all symptomatic epilepsies worldwide, underscoring the substantial proportion of patients with PTE who develop drug-resistant epilepsy (Schmidt and Löscher, 2005; Larkin et al., 2016; Hauser et al., 1990). Despite its prevalence and known burden to quality of life in patients with this disorder, the risk factors and pathogenesis of PTE remain incompletely understood. Established risk factors include injury severity, intracranial hemorrhage, early post-traumatic seizures, male sex, age between 40 and 79 years, and genetic predispositions; however a true “at-risk” population is difficult to define due to the limited treatment guidance of these risk factors and their unreliability at the individual level (Kazis et al., 2024; Karlander et al., 2021; DeGrauw et al., 2018). For example, although having early post-traumatic seizures is recognized as a risk factor, prophylactic prescription of antiepileptic drugs to control these seizures fails to reduce long-term PTE risk in patients with moderate to severe traumatic brain injury (Temkin et al., 1990; Wang et al., 2022).

A central challenge in reducing PTE risk is the inherent heterogeneity of TBI. Variations in injury mechanism, contusion location, severity, and individual host factors produce distinct phenotypes that may follow diverging biological pathways toward or away from epileptogenesis (Pitkänen and Lukasiuk, 2011). As epilepsy is increasingly understood as a disorder involving large-scale brain networks (Spencer, 2002; Laufs, 2012; Lehnertz et al., 2023)^–(Spencer, 2002; Laufs, 2012; Lehnertz et al., 2023), we investigate whether the development of PTE is associated with characteristic alterations in whole-brain functional organization, specifically in the pattern of resting-state functional connectivity (FC).

Resting-state functional MRI (rs-fMRI) is a neuroimaging tool that is used to characterize FC between brain regions, organizing the brain into resting-state networks (RSNs) (Biswal et al., 1995; Fox and Raichle, 2007). RSNs are defined as brain regions that exhibit synchronous low-frequency activity when the brain is in a functionally passive state and not engaged in a cognitive task. Converging evidence has demonstrated that RSN connectivity is systemically disrupted across a wide range of neurological and psychiatric disorders such as Alzheimer's and autism spectrum disorder, supporting RSNs as an emerging framework for understanding brain diseases (Fox and Greicius, 2010; Broyd et al., 2009). Due to the large dimensionality in an FC characterization of the brain, machine learning (ML) approaches are often adopted in its analysis (Lemm et al., 2011). Nonetheless, FC alterations specifically associated with PTE remain poorly characterized. Akrami et al. trained ML models on FC derived from a coarse 15-region parcellation combined with structural MRI features and achieved an area under the receiver operating characteristic curve (AUC) of 0.78 (Akrami et al., 2024); similarly, Garner et al. reported an identical peak performance AUC of 0.78 using just FC features (Garner et al., 2019a). These findings suggest that FC differences may encode diagnostically relevant information in PTE populations, though prior work has been limited by coarse atlases, small samples, and, critically, methodological vulnerabilities including feature-selection data leakage.

The present study addresses these limitations in three ways. First, to ensure unbiased performance assessment, we implemented a fully nested cross-validation recursive feature elimination (nested CV-RFE) pipeline in which feature selection is performed entirely within each outer training fold, eliminating data leakage from held-out subjects and combatting overfitting in our identification of key FC differences (Varma and Simon, 2006). Second, we center the primary classification analysis on the direct comparison between TBI and PTE subjects, while retaining a secondary Epilepsy-Free vs. PTE analysis to contextualize findings against a broader non-epileptic reference. Third, we leverage the Schaefer 100-region atlas with the Yeo seven-network RSN parcellation to achieve fine-grained spatial resolution and RSN-level interpretability across all 4950 FC pairs, substantially exceeding the anatomical granularity of prior studies (Schaefer et al., 2018; Yeo et al., 2011). Together, these analytic advances allow us to examine a candidate set of sparse, reproducible FC features that may distinguish PTE from TBI at the individual subject level, and to assess their discriminative value relative to univariate ranking and random selection.

2. Methods

2.1. Subjects

This study was performed in compliance with relevant laws and institutional guidelines and was approved by the Institutional Review Board of Rutgers University (IRB approval number: Pro2019002307, approved on 7/8/2020). All participants provided written informed consent prior to enrollment. Three cohorts were recruited: Healthy Control (HC), Traumatic Brain Injury (TBI), and Post-Traumatic Epilepsy (PTE). Subjects 18-65 years of age were eligible. HC subjects were recruited if they had no history of neurological or psychological diagnoses, prior TBI or craniotomy. Eligibility for the TBI group required that the individual's first traumatic brain injury demonstrate radiographic evidence of a cerebral contusion on initial hospital admission, as assessed by hospital radiologists. PTE subjects were eligible if they had a diagnosis of PTE by an epileptologist with documented video EEG demonstrating evidence of focal epilepsy after TBI. Subjects were excluded if they had any history of seizures prior to TBI, other pathological lesions on MRI such as brain tumors or vascular malformations, history of CVA, other neurological or psychological diseases, prior brain surgery, or psychogenic non-epileptic seizures. Additionally, PTE subjects with a history of surgical treatment of epilepsy including vagal nerve stimulator were excluded. Subjects with a history of alcohol or recreational drug abuse or any contraindication to undergoing MRI were excluded.

2.2. MRI acquisition and preprocessing

MRI data were acquired on 3T Magnetom Prisma scanners (Siemens Healthineers, Erlangen, Germany). Rs-fMRI was collected using a gradient-echo EPI sequence (TR = 720 ms, TE = 33 ms, flip angle = 52°). Images were acquired with 2.0 × 2.0 × 2.0 mm³ voxels, 72 interleaved slices, multiband acceleration factor = 8, and GRAPPA acceleration factor = 2, covering the whole brain (FOV = 208 mm). A total of 1200 vol were obtained over 14 min 37 s. High-resolution structural images were acquired using a T1-weighted 3D MPRAGE sequence (TR = 2300 ms, TE = 2.98 ms, TI = 900 ms, flip angle = 9°). Images were collected with 1.0 × 1.0 × 1.0 mm³ voxels, FOV = 256 × 240 × 208 mm³, 208 sagittal slices, bandwidth = 240 Hz/pixel, and GRAPPA acceleration factor = 2. Acquisition time was 5 min 12 s. Structural images were used as the reference for tissue segmentation, skull masking, and MNI spatial normalization.

We followed a preprocessing and quality control pipeline adapted from Di and Biswal (Supplementary Fig. 1). (Di and Biswal, 2023) Briefly, each subject's rs-fMRI sequence was corrected for realignment and motion, and co-registered to MNI standard space using their 3D MPRAGE images. Subjects were excluded if the framewise displacement exceeded quality control thresholds: maximum translation or rotation >2 mm/2°, or mean translation or rotation >0.3 mm/0.3°. Nuisance regression was performed using a general linear model that included 24 motion-related regressors (six realignment parameters, their squares, and their previous-frame values and squares), together with mean white matter and cerebrospinal fluid signals and their first principal components, consistent with the established protocol. Following this, global signal was used only as a quality-control diagnostic rather than as a nuisance regressor, and motion control was implemented via subject-level exclusion rather than frame-level scrubbing/censoring.

2.3. FC computation

Resting-state fMRI signals were parcellated into 100 brain ROIs using the Schaefer 100-ROI atlas (Schaefer et al., 2018). Signal values within each ROI were averaged into a single time series. A 5th-order Butterworth bandpass filter (0.01–0.1 Hz) was applied to suppress non-neuronal fluctuations. FC between each ROI pair was obtained by calculating the Pearson's correlation coefficient between the two ROI time series. Altogether, 4950 (100*(100-1)/2) FC values were computed for each subject. The Fisher Z transform was then applied to normalize correlation values for all subsequent statistical analyses.

Each ROI on Schaefer's atlas is assigned to one of Yeo seven resting-state networks (RSN) based on spatial overlap with the Yeo network maps. Yeo's seven network atlas included the following RSNs: Visual (VIS), Somato-Motor (SM), Dorsal-Attention (DA), Ventral-Attention (VA), Limbic (LIM), Fronto-Parietal (FP), and Default mode networks (DMN) (Schaefer et al., 2018; Yeo et al., 2011). As the relative volume occupied by each RSN in the brain is different, the number of ROIs assigned to each RSN varied. The ROI counts by RSN and hemisphere are listed in Supplementary Table 3.

2.4. FC discriminability: d′ analysis

To quantify the discriminability of individual FCs between groups, a discriminability index (d′) was computed for each of the 4950 FCs as:

$$d^{\prime }=sign\left(K\right)\frac{\left|{\mu }_{Epileptic}-{\mu }_{Epilepsy-Free}\right|}{\sqrt{0.5\left({{\theta }_{Epileptic}^2+}_{Epilepsy-Free}^2\right)}}$$

$$K=\left|{\mu }_{Epileptic}\right|-\left|{\mu }_{Epilepsy-Free}\right|$$

where μ₀ and μ₁ are group means, ${\theta }_1^2$ and ${\theta }_0^2$ are group variances, and K encodes coupling direction. The d’ statistic is a standardized effect index adapted from Cohen's d that captures the separation between two distributions while accounting for unequal variance. (Cohen, 1988) A positive d′ (hyper-coupling, d′ ≥ 0.5) indicates that FC magnitude increases in group 1 relative to group 0; a negative d′ (un-coupling, d′ ≤ −0.5) indicates FC magnitude decreases. FCs with |d′| < 0.5 were classified as unchanged. The primary analysis compared TBI (group 0) against PTE (group 1), representing the clinically relevant comparison for identifying FC features associated with established PTE relative to TBI. Therefore, in this comparison schema, a hyper-coupled d′ represents FC that is closer to +1 or −1 in the PTE group than TBI, while an un-couped d′ represents FC that is closer to 0 in the PTE group. A secondary analysis pooled HC and TBI subjects into an Epilepsy-Free class (n = 76, group 0) compared against PTE (n = 20, group 1), to contextualize PTE FC against a broader non-epileptic reference.

2.5. Machine learning: nested cross-validation recursive feature elimination

To identify FC features that collectively discriminate PTE from TBI at the individual subject level, a nested cross-validation recursive feature elimination (nested CV-RFE) framework was implemented using a RF classifier with balanced class weighting. A linear Support Vector Machine (SVM) with balanced class weighting was evaluated in parallel and with results reported for transparency; however, all feature-level biological interpretation is restricted to the RF classifier.

The nested CV framework comprised an outer loop (5-fold stratified cross-validation repeated 5 times; 25 outer folds total) providing unbiased performance estimates on held-out test data, and an inner loop (5-fold stratified cross-validation within each outer training fold) performing feature selection. The outer test fold had no influence on any aspect of feature selection, ensuring complete separation between feature selection and performance estimation. Within each inner loop, RFE proceeded in two stages: coarse elimination from 4950 to 200 features (steps of 200), then fine elimination from 200 to 1 feature (steps of 5, then 1), yielding 80 evaluation points. Feature importance was computed from Gini impurity scores. The optimal feature count for each outer fold was defined as the count maximizing mean inner CV AUC, and a final RF model was refit on the complete outer training fold using the selected features.

Performance metrics, AUC, Area Under the Precision-Recall Curve (PR-AUC), accuracy, precision, recall, F1, were computed for each outer fold and reported as mean ± SD across 25 outer folds. A permutation test was conducted by randomly shuffling class labels 200 times, running the full nested CV-RFE pipeline on each permutation, and comparing observed AUC against the resulting null distribution. To contextualize the added value of feature selection, baseline RF and SVM classifiers using all 4950 FC features (no selection step) were additionally evaluated via 5-fold stratified cross-validation repeated 20 times (100 total folds) using identical classifier hyperparameters.

2.6. Feature selection stability and RSN overrepresentation

After all 25 outer folds, selection frequency was computed for each FC as the proportion of outer folds in which that feature appeared in the fold-specific optimal subset. Features were classified into the highest stability tier (Tier I) if selection frequency exceeded 60% and hereafter represent the primary interpretive set.

To characterize the network-level distribution of RF-selected features, a frequency-weighted RSN overrepresentation analysis was performed across all 28 unique RSN-to-RSN blocks. A weighted representation score was computed for each block as the sum of RF selection frequencies of all member features normalized by the total number of possible connections within that block.

2.7. Hemispheric analysis

To further characterize the RF-selected feature set, hemispheric and coupling analysis was implemented in conjunction to identify the presence of connectivity trends that emerged in our cohort. Hemispheric assignment was determined by Schaefer atlas naming convention: ROIs 1–50 (left hemisphere, LH) and ROIs 51–100 (right hemisphere, RH). Each FC was classified as left intra-hemispheric (LH–LH, n = 1225), right intra-hemispheric (RH–RH, n = 1225), or cross-hemispheric (Cross-Hemi, n = 2500). Frequency-weighted representation scores were computed per hemispheric category. The analysis was extended to a 14×14 half-network framework, splitting each RSN into left and right components, and frequency-weighted representation scores were computed for each of the 105 unique half-network blocks. Coupling direction (hyper-coupled vs. un-coupled) of all impacted FCs (|d′| ≥ 0.5) was cross-tabulated by hemispheric category.

2.8. Validation: stable features vs. alternative selection strategies

To validate the informativeness of the RF stable features, their classification performance was compared against: (1) the top four features ranked by univariate |d′| in the primary TBI vs. PTE analysis; and (2) randomly selected features (mean performance across 100 independent random draws of 4 features). All three feature sets were evaluated with RF classifiers using repeated stratified cross-validation (5-fold × 20 repeats, 100 total folds) without any additional feature selection within the CV loop, isolating the discriminative contribution of the feature identities themselves.

2.9. Statistical analysis

All analyses were performed in Python (version 3.12.7) using scipy, pandas, numpy, scikit-learn, statsmodels, and nilearn packages. The Tier 1 stable features (RF selection frequency ≥ 60%) were visualized as a connectome on the MNI glass brain using the Nilearn Python package (version 0.12.0). MRI preprocessing was performed in MATLAB R2023a using the Statistical Parametric Mapping (SPM) toolbox (Inc. TM, 2023; Friston et al., 1994).

Demographic and clinical characteristics were compared across all three groups (HC, TBI, PTE) as well as pairwise tests restricted to the primary analytic comparison of interest (TBI vs. PTE). For age, the Kruskal-Wallis H-test was used for three-group comparisons and the Mann-Whitney U test for the TBI vs. PTE pairwise comparison. For categorical variables (gender, race, ethnicity, education, SES, and handedness), Pearson chi-square tests were applied for HC vs TBI vs PTE and again for TBI vs PTE. Contusion volume (mm³) and number of Yeo RSNs with contusion involvement were compared between TBI and PTE participants using the Mann-Whitney U test. PTE-specific clinical variables, age at seizure onset and number of antiseizure medications (ASMs) are reported descriptively as these variables are unique only to the PTE group. Mean framewise displacement (FD) was compared across the three cohorts using the Kruskal-Wallis H-test, with subsequent pairwise Mann-Whitney U tests (Bonferroni-corrected, k = 3) given its role as a potential confound across all cohorts.

For RSN-level group comparisons, one-way ANOVAs were performed across all 28 RSN-to-RSN blocks with cohort (HC, TBI, PTE) as the main factor, followed by pairwise independent-samples Tukey's post-hoc t-tests for all three cohort combinations (HC vs. TBI, HC vs. PTE, TBI vs. PTE). All p-values were corrected for multiple comparisons using the Benjamini-Hochberg false discovery rate (FDR) procedure, applied separately within each family of tests. For RSN overrepresentation and hemispheric analyses, statistical significance was assessed using permutation-based tests (10,000 permutations), with one-tailed p-values and FDR correction applied across all blocks within each analysis. A frequency-weighted overrepresentation score (Rep-Score) was calculated for each analysis unit (RSN-to-RSN or hemispheric block) as the sum of RF selection frequencies (the proportion of outer cross-validation folds in which each FC was selected) across all member FC features, normalized by the total number of possible FC connections within that block. Lastly, a permutation test (200 permutations of class labels) was used to assess whether nested CV-RFE classification performance exceeded chance.

3. Results

3.1. Cohort demographics

From June 2021 to January 2026, a total of 101 subjects were recruited for the study. Five subjects were excluded due to excessive head motion and framewise displacement exceeding quality control thresholds. We analyzed neuroimages from 41 healthy controls (HC), 35 individuals with traumatic brain injury (TBI), and 20 subjects with post-traumatic epilepsy (PTE), totaling 96 participants. Three-group comparisons revealed significant differences in age (Kruskal-Wallis H p = 0.020), sex distribution (χ² p = 0.002), and educational attainment (χ² p = 0.003). These effects were driven primarily by differences between the HC and injury (TBI and PTE) groups, as the main TBI vs PTE comparisons demonstrated no significant differences in these demographic categories. HC participants were younger on average (34.3 ± 12.5) relative to TBI (41.0 ± 12.8) and PTE (38.8 ± 9.7), and HC skewed more female (52.5% female) and more highly educated (46% graduate/professional degree) than the injury cohorts. Critically, the primary TBI vs PTE comparison had no significant differences observed between any demographic categories, indicating that the two injury groups were well-matched on measured demographic characteristics.

Contusion characteristics were available for a subset of TBI (n = 18) and PTE (n = 12) participants. Contusion volume did not differ significantly between TBI and PTE (median 8344 vs 19,512 mm³; Mann-Whitney U p = 0.122), though a trend toward larger contusions in the PTE group was observed. The number of Yeo seven-network parcels intersected by contusion also did not differ significantly (median 4 vs 5.5; Mann-Whitney U p = 0.256). Among PTE participants, age at seizure onset was available for 17 of 20 participants (32.6 ± 11.6 years) ASM data were available for 15 of 20 PTE participants; of these, 15 (100%) were receiving at least one ASM at the time of imaging (mean 1.7 ± 1.0). ASMs prescribed included levetiracetam, lacosamide, valproate/divalproex and lamotrigine. All demographic information is summarized in Supplementary Table 1.

Mean FD did not differ significantly across the three cohorts (HC 0.129 ± 0.060 mm, TBI 0.132 ± 0.061 mm, PTE 0.106 ± 0.034 mm; Kruskal-Wallis p = 0.120). The pairwise TBI vs. PTE comparison showed a nominal difference (Mann-Whitney U p = 0.032) that did not survive Bonferroni correction (p = 0.096), indicating motion is not a significant confound on the primary classification contrast. FD summaries for each cohort are displayed in Supplementary Table 2.

3.2. RSN-level FC differences across cohorts

Mean FC matrices derived from rs-fMRI for each cohort are plotted with 100×100 ROIs organized by RSN (Fig. 1A). While FC within individual RSNs was expectedly stronger than across RSNs, a diffuse further weakening in cross-RSN FCs was observed in both TBI and PTE relative to the HC cohorts (Fig. 1B), reflected by global reductions in positive correlations. After FDR correction, one-way ANOVA revealed four RSN blocks with statistically significant difference in average FC across cohorts: VIS-SM, VIS-DMN, LIM-DMN and DMN-DMN (Fig. 1Ci). Post-hoc pairwise t-tests across all three cohort combinations (Fig. 1Cii–iv) revealed that the majority of significant differences were attributable to HC vs. TBI (6 RSN blocks) and HC vs. PTE (1 RSN block). Critically, no RSN block reached statistical significance in the TBI vs. PTE comparison after FDR correction. In the FDR-significant RSN blocks, TBI and PTE subjects showed indistinguishable differences in connectivity but consistently reduced connectivity compared to HC (Fig. 1D). These findings indicate that while TBI and PTE share broad patterns of reduced RSN-level FC relative to healthy controls, group-level network differences do not statistically distinguish TBI from PTE, motivating the individual FC-level discrimination strength and machine learning analyses described below.

Fig. 1.

RSN-level FC across Cohorts. (A) FC matrices of 100x100 regions of interest (ROIs) delineated by resting-state networks (RSNs) for (i) Healthy Control (HC; n = 41), (ii) Traumatic Brain Injury (TBI; n = 35), and (iii) Post-Traumatic Epilepsy (PTE; n = 20) cohorts. FC values were computed using Pearson's correlation coefficients (ranging from −1 to 1). (B) Mean FC between RSN-to-RSN connections derived by averaging FC values within each RSN-RSN block. (C) P-value heatmap from one-way ANOVA of the mean RSN-to-RSN FC for (i) cohorts as the main factor, and (ii-iv) Tukey post-hoc contrasts. (D) Mean RSN-to-RSN FC for the four statistically significant RSN-to-RSN blocks identified by ANOVA in (Ci). Statistical significance is denoted as p < 0.05 (*), p < 0.01 (**), and p < 0.001 (***). Abbreviations: FC (Z) = Fisher Z transformed FC Pearson correlation, VIS = Visual; SM = Somato-Motor; DA = Dorsal-Attention; VA = Ventral-Attention; LIM = Limbic; FP = Fronto-Parietal; DMN = Default Mode.

3.3. FC discriminability between groups: d′ analysis

With RSN-level FC failing to discriminate TBI and PTE, we sought to then characterize individual-level FC differences. d′ was computed for each of the 4950 FCs, with positive hyper-coupled (d′ ≥ 0.5) or negative un-coupled (d′ ≤ −0.5) status recorded (examples of the most discriminative hyper and un-coupled FCs in Fig. 2A). We performed this analysis under two frameworks: (1) a primary analysis comparing TBI directly against PTE; and (2) a secondary analysis pooling HC and TBI into an Epilepsy-Free class (n = 76) compared against PTE.

Fig. 2.

d′ Discriminability Analysis of TBI vs PTE and Epilepsy-Free vs PTE. (A) Two kernel density estimate plots of the most discriminative hyper-coupled FC (highest d′ ≥ +1.0, FC 230, left) and one showing the most discriminative un-coupled FC (lowest d′ ≤ −1.0, FC 965, right) in the primary TBI vs. PTE comparison. Arrows indicate direction of coupling. (B) Three d′ 100×100 matrices for the primary TBI vs. PTE comparison (i) full signed d′ matrix; (ii) sparse matrix showing only FCs with |d′| ≥ 0.5; (iii) sparse matrix showing only |d′| ≥ 1.0. (C) identical to (B) but for the secondary Epilepsy-free vs PTE comparison. (D) Comparisons between the TBI vs PTE and Epilepsy-Free vs PTE analysis for (i) proportion of unaffected to affected (|d′| ≥ 0.5) FCs (ii) proportion of hyper-coupled to un-coupled FCs with |d′| ≥ 0.5 (iii) same as (ii) but with FCs |d′| ≥ 1.0.

In the primary TBI vs. PTE comparison, 413 of 4950 FCs (8.3%) were classified as impacted (|d′| ≥ 0.5; Fig. 2Bii, Di). The distribution was near-balanced between coupling directions, with 204 (49.4%) of impacted FCs hyper-coupled and 209 (50.6%) un-coupled (Fig. 2Dii). Only five FCs reached |d′| ≥ 1.0 with all five of these strongly discriminative FCs being hyper-coupled (Fig. 2Diii), spanning three RSN blocks: three within VIS–FP, one within DA–LIM, and one within FP–FP (Fig. 2Biii).

In the secondary Epilepsy-Free vs. PTE comparison, 376 of 4950 FCs (7.6%) were classified as impacted (Fig. 2Cii, Di). In contrast to the primary analysis, un-coupling was the dominant alteration: 235 (62.5%) were un-coupled and 141 (37.5%) were hyper-coupled (Fig. 2Dii). Once again, five FCs reached |d′| ≥ 1.0, distributed across three RSN blocks: two within VIS–FP, two within VIS–SM, and one within VA–DMN (Fig. 2Ciii), though three of these were un-coupled and two were hyper-coupled (Fig. 2Diii). Partial convergence between the two analysis streams was demonstrated: two hyper-coupled FCs residing within the VIS–FP block were shared in both comparisons. Notably, these two shared features also carried the largest d′ magnitudes (d′ = +1.41 and + 1.22 in the primary; +1.20 and + 1.07 in the secondary).

3.4. Classifier performance

The RF classifier achieved a mean outer CV AUC of 0.79 ± 0.14 and accuracy of 0.68 ± 0.10 across 25 outer folds (Fig. 3A, Table 1). The mean inner CV AUC peaked at 0.96 at approximately 15 features (Fig. 3B). The SVM achieved a mean outer CV AUC of 0.64 ± 0.13 (Fig. 3A–Table 1); however, inspection of the inner RFE loop revealed that the SVM consistently selected optimal feature counts at a median of 2350 features (range 1750–3350) with inner CV AUC values of 1.00 (Fig. 3B). This pattern is characteristic of linear SVM-RFE failure in the high-dimensional low-sample regime (p ≫ n), where the model memorizes training data rather than generalizing (Guyon et al., 2002). The outer AUC of 0.64 confirms that the SVM's inner-loop performance does not reflect genuine discriminability; all subsequent feature-level analyses are therefore restricted to the RF classifier.

Fig. 3.

Nested CV-RFE Classifier Performance and Feature Stability. (A) (left) Mean ROC curves for RF (green) and SVM (navy) from the outer CV loop. Shaded bands ± 1 SD. Diagonal dashed line represents 50% chance. (right) Mean precision-recall curves for RF and SVM. Horizontal dashed line represents no-skill baseline (nPTE/(nPTE + nTBI) = 20/(20 + 35) = 36%). (B) Mean inner CV AUC vs. number of retained features (log x-axis), shown separately for RF (green) and SVM (navy). Shaded bands ± 1 SD across outer folds. (C) Histogram of RF selection frequencies across all 4950 FCs (or features selected in ≥1 fold). Four features appear above the 60% threshold, indicated by arrows. (D) Null distribution of outer CV AUC from 200 permutations of shuffled class labels for RF.

Table 1.

Nested Cross-Validation and Baseline Performance of RF and SVM for TBI vs PTE Classification.

Performance of nested cross-validated RF and SVM classifiers for TBI vs. PTE classification, compared to baseline models trained on the full feature set. Nested CV-RFE models (5-fold × 5-repeat outer loop, 5-fold inner loop) performed two-stage recursive feature elimination within each training fold to prevent data leakage; reported metrics reflect performance on held-out outer-fold test data only. Baseline models were trained on all 4950 FC features with no feature selection. Optimal n features reflects the median (range) number of features retained at peak inner-loop performance across outer folds; stable features denotes the number of individual FC features selected in ≥50% of outer folds. Values are mean ± SD across outer folds. AUC = Area Under the ROC Curve; PR-AUC = Precision-Recall AUC; RF = Random Forest; SVM = Support Vector Machine.

	Classification performance						RFE feature stability
	AUC	PR-AUC	Accuracy	Precision	Recall	F1	Optimal n features (median, range)	Tier I Stable Features (≥60% folds)
RF (nested CV, outer loop)	0.79 ± 0.14	0.74 ± 0.17	0.68 ± 0.10	0.51 ± 0.36	0.35 ± 0.27	0.39 ± 0.27	18 (3–90)	4
SVM (nested CV, outer loop)	0.64 ± 0.13	0.58 ± 0.13	0.62 ± 0.09	0.46 ± 0.26	0.39 ± 0.23	0.40 ± 0.21	2350 (1750–3350)	2197
RF (baseline, all 4950 features)	0.63 ± 0.16	0.58 ± 0.15	0.63 ± 0.03	0.04 ± 0.20	0.01 ± 0.05	0.02 ± 0.08	4950 (fixed)	-
SVM (baseline, all 4950 features)	0.53 ± 0.20	0.51 ± 0.17	0.61 ± 0.13	0.43 ± 0.32	0.34 ± 0.26	0.35 ± 0.24	4950 (fixed)	-

Permutation testing using the complete nested CV-RFE pipeline confirmed that RF classification performance significantly exceeded chance: no permutation AUC exceeded or even met the observed outer CV AUC of 0.79 across 200 permutations (permutation p < 0.005; Fig. 3D), confirming that the classification signal reflects genuine biological differences rather than artifact from overfitting or feature selection instability. Baseline classifiers trained on all 4950 features without any selection step performed markedly worse than their RFE-selected counterparts (RF: AUC 0.63 ± 0.16, PR-AUC 0.58 ± 0.15; SVM: AUC 0.53 ± 0.20, PR-AUC 0.51 ± 0.17; Table 1), indicating that the discriminative signal identified by nested CV-RFE is not attributable to the raw feature space alone and that the selection procedure meaningfully improves classification performance. Notably, baseline RF precision and recall were near zero (0.04 ± 0.20 and 0.01 ± 0.05, respectively), despite balanced class weighting, indicating that at default classification thresholds the model essentially never predicted the PTE class when confronted with all 4950 unselected features.

3.5. Stable functional connectivity signature

The optimal number of features selected by the RF inner loop varied substantially across outer folds (range 3–90, median 18, Table 1). We therefore characterize feature reliability using selection frequency: the proportion of the 25 outer folds in which each feature appeared in the fold-specific optimal subset (Fig. 3C).

The vast majority of features were selected in fewer than 10% of outer folds. However, a small number demonstrated robust, reproducible selection. Using a threshold of ≥60% of outer folds, four features met this criterion for RF (Fig. 3C): feature 230 (connecting left occipital VIS ROI to left precuneus ROI) was selected in 100% of outer folds, feature 260 (same left VIS ROI as 230 to right SM cortex ROI) was selected in 84% of folds, and both features 412 (separate left VIS ROI to left VA lateral prefrontal cortex) and 4805 (right FP lateral prefrontal cortex to right DMN temporal cortex) selected in 64% of folds. These four features map to distinct RSN-to-RSN connections spanning five RSNs (Table 2). Three of the four stable features involve the left VIS network as one endpoint, with the other endpoints including the left FP, left VA and right SM. The last non-VIS stable feature was an RH-RH FP–DMN connection. The dominant coupling pattern is un-coupling (three of four features), with the sole hyper-coupled feature representing the VIS–FP connection, once again emerging as the strongest discriminator by both selection frequency and d′ magnitude. These four stable features are visualized as a connectome on the MNI glass brain (Fig. 4).

Table 2.

Stable RF Features.

Tier 1 stable FC features identified by the RF nested CV-RFE model, defined as features selected in ≥60% of outer cross-validation folds. ROI 1/RSN 1 and ROI 2/RSN 2 denote the two Schaefer atlas regions comprising each FC edge and their corresponding Yeo resting-state network assignment. Hemisphere Type indicates whether the edge is intra-hemispheric (LH-LH, RH-RH) or inter-hemispheric (Cross-Hemi). Selection Frequency is the proportion of outer folds in which the feature was retained. d′ is the discriminability index between TBI and PTE. FC # = FC feature number as defined by Schaefer atlas; ROI = region of interest; RSN = resting-state network; LH = left hemisphere; RH = right hemisphere.

FC #	ROI 1	RSN 1	ROI 2	RSN 2	Hemisphere Type	Selection Frequency	d’
230	Vis_3	VIS	Cont_pCun_1	FP	LH-LH	1.00	+1.413
260	Vis_3	VIS	SomMot_8	SM	Cross-Hemi	0.84	−0.649
412	Vis_5	VIS	PFCl_1	VA	LH-LH	0.64	−0.648
4805	PFCl_1	FP	Temp_1	DMN	RH-RH	0.64	−0.782

Fig. 4.

Brain graph visualization of the Tier 1 stable FC signature. Three glass brain connectome panels side by side: sagittal (left hemisphere view), coronal (posterior view), and axial (top-down view). Four edges of Tier I features are drawn between ROI centroids in MNI space that are derived from the Schaefer 100-ROI atlas. Edge color encodes coupling direction: red = hyper-coupling (d′ > 0); blue = un-coupling (d′ < 0). Node size is proportional to the sum of RF selection frequencies of stable connections at that ROI. Node colors by Yeo RSN: dark red = VIS, dark orange = FP, royal blue = SM, hot pink = VA, violet = DMN.

3.6. Visual network connections are disproportionately represented in the RF-selected feature space

To characterize the network-level distribution of FC features driving RF classification, we computed a Rep-Score for each of the 28 unique RSN blocks, normalized by the number of possible connections per block and assessed against a permutation null distribution (10,000 permutations, FDR correction).

Four RSN blocks were significantly overrepresented after FDR correction: VIS–FP (pFDR < 0.001), VIS–VIS (pFDR = 0.001), and VIS–SM (pFDR = 0.002) and FP-DMN (pFDR = 0.049; Fig. 5). The Visual network was the unambiguous structural epicenter of RF feature selection: three out of four FDR-significant block had VIS as at least one endpoint and the VIS–VIS block carried the highest absolute weighted representation score (Rep-Score = 0.022). Interestingly, this VIS-VIS block was the only significant block that did not contain a Tier 1 stable feature.

Fig. 5.

Permutation-Based RSN-to-RSN Overrepresentation Analysis of RF-selected Features. Each panel in the upper triangle and diagonal of the 7×7 matrix represents one unique RSN-to-RSN block and displays the permutation null distribution of frequency-weighted representation scores (10,000 permutations of RF selection frequencies across features). The vertical black line in each panel indicates the observed Rep-Score for that block, defined as the sum of RF selection frequencies of all member features normalized by the number of possible connections within the block (annotated in upper right corner of each panel). Panel color indicates statistical significance after Benjamini-Hochberg FDR correction across all 28 RSN blocks: red panels survived FDR correction (p_FDR < 0.05); orange panels were nominally significant before but not after FDR correction (p_FDR ≥ 0.05); blue panels were not significant. Gold stars (★) indicate RSN blocks containing at least one Tier 1 stable feature.

The two highest-frequency Tier 1 stable features reside within the two next most significantly overrepresented blocks: Feature 230 (VIS–FP, Rep-Score = 0.017, pFDR < 0.001) and Feature 260 (VIS–SM, Rep-Score = 0.016, pFDR = 0.002). Feature 4805 (FP–DMN, Rep-Score = 0.011, pFDR = 0.049) also mapped to a RSN block that was significantly overrepresented. The last Tier I stable feature, Feature 412, mapped to a block with elevated, though a subthreshold, representation score (VIS–VA, Rep-Score = 0.010, pFDR = 0.27). Lastly, VIS-DMN, LIM-DMN and DMN-DMN blocks, which showed significant group-level differences in the RSN-level ANOVA (Fig. 1D) and features d′ ≥ 0.5 in the Epilepsy-free vs PTE comparison (Fig. 2Cii), were not shown to be significantly overrepresented in the RF-selected feature space.

3.7. RF-selected features show significant left-hemispheric lateralization and hyper-coupling

To examine whether the FC features driving RF classification showed a hemispheric bias, we analyzed the distribution of RF selection frequencies across LH-LH, RH-RH, and Cross-Hemi categories.

LH–LH connections were significantly overrepresented in the RF-selected feature space (Rep-Score = 0.0092, pFDR = 0.0003, Fig. 6Ai). Neither RH–RH nor cross-hemispheric connections showed significant overrepresentation (Rep-Score = 0.0061 and 0.0029, pFDR = 0.248 and 1.000, respectively). LH–LH connections accounted for a disproportionate share of total weighted RF selection frequency relative to their native baseline, while cross-hemispheric connections were underrepresented (Fig. 6Aii).

Fig. 6.

Hemispheric Lateralization and Coupling of RF-Selected Features. (A) (i) Bar chart comparing Rep-Score for LH–LH, RH–RH, and Cross-Hemi categories. Colored bars = observed scores; gray bars = permutation null mean ± SD. Significance annotations above bars (***p_FDR < 0.001). (ii) Comparing observed proportional RF selection frequency per hemispheric category (colored) against native baseline proportions from the full 4950-FC distribution (gray). (B) 14×14 hemispheric RSN matrix heatmap with cell values representing Rep-Score. Gold stars on blocks containing ≥1 Tier 1 stable feature. (C) Coupling direction by hemispheric category. (i) Stacked bar chart showing counts of hyper-coupled (red) vs. un-coupled (blue) FCs among all impacted FCs (|d′| ≥ 0.5, n = 413) for each hemispheric category (LH–LH, RH–RH, Cross). (ii) Equivalent chart weighted by RF selection frequency.

Decomposition into the 14×14 hemispheric RSN framework (Fig. 6B) identified L-VIS–L-FP and L-VIS–L-VIS as the second and third highest-scoring half-network blocks by Rep-Score (0.077 and 0.053 respectively), directly reflecting the concentration of left Visual network connections in the RF-selected feature space. In terms of Tier I stable features, Feature 230 is mapped to the above-identified L-VIS–L-FP block, while Feature 412 (L-VIS–L-VA, Rep-Score = 0.025) represents an additional LH-LH VIS connection. Features 4805 (R-FP–R-DMN, Rep-Score = 0.034) and 260 (L-VIS-R-SM, Rep-Score = 0.029) represent independent right-hemispheric and cross-hemispheric components of the stable signature. The highest-scoring block overall, R-LIM–R-LIM (Rep-Score = 0.08), did not contain Tier 1 stable features and is likely attributable to the small number of possible connections within that block in the Schaefer 100 parcellation rather than a true biological overrepresentation.

The hemispheric lateralization finding was accompanied by a directional asymmetry in FC coupling across hemispheric categories (Fig. 6C). Among all impacted FCs (|d′| ≥ 0.5, n = 413) in the primary TBI vs. PTE comparison, LH–LH connections were predominantly hyper-coupled in PTE (61.7% hyper-coupled, 38.3% un-coupled), while RH–RH connections showed the opposite pattern (59.5% un-coupled, 40.5% hyper-coupled). Cross-hemispheric connections were near-balanced (48.6% vs. 51.4%). This directional lateralization was consistent across both impacted FC counts (Fig. 6Ci) and the RF frequency-weighted analysis (Fig. 6Cii), and is recapitulated in the Tier 1 stable features: the two LH–LH stable features include the only hyper-coupled Tier 1 feature (Feature 230, d′ = +1.41), while the RH–RH stable feature (Feature 4805, d′ = −0.78) is un-coupled.

3.8. The stable FC signature provides superior discriminability relative to alternative feature selection strategies

To validate the informativeness of the four RF stable features, their classification performance was compared against the top four features ranked by univariate |d′| and randomly selected features using identical repeated stratified cross-validation (5-fold × 20 repeats, 100 total folds) with RF classifiers and no additional feature selection within the CV loop.

The four RF stable features achieved a mean AUC of 0.95 ± 0.05 and accuracy of 0.88 ± 0.08 (Fig. 7, Table 3). This substantially exceeded the top-4 d′-ranked features (AUC = 0.92 ± 0.08), random features (AUC = 0.51 ± 0.09) and non-feature selected model with all 4950 features (AUC = 0.63 ± 0.16).

Fig. 7.

Validation comparison: RF stable features vs. d′-ranked, random and all features. (A) Mean ROC curve panel. Five overlaid curves: RF stable features (green solid, AUC = 0.95 ± 0.05); top-4 d′-ranked features (purple dashed, AUC = 0.92 ± 0.08); random 4-feature sets (gray dotted, mean of 100 draws ± SD, AUC ≈ 0.51); all 4950 FC features (orange dashed, AUC = 0.63 ± 0.16) nested CV outer loop (black semi-transparent, AUC = 0.79, for reference). Shaded bands ± 1 SD. Diagonal dashed line = chance. AUC ± SD values in legend. (B) Bar chart comparing mean AUC ± SD across all five feature sets. Bars colored to match ROC curve colors. Red dashed horizontal line at AUC = 0.5 (chance). Error bars = ± 1 SD.

Table 3.

Validation Comparison: RF Stable Features vs d′-Ranked and Random Feature Sets.

All feature sets evaluated using identical repeated stratified cross-validation (5-fold × 20 repeats) with RF classifiers on fixed, pre-identified feature subsets (no nested feature selection). The RF Nested CV outer loop result is included as a reference honest performance estimate. Accuracy and F1 not reported for the random feature baseline given near-chance performance. Values represent mean ± SD across cross-validation folds. d′ = discriminability index (see Methods). AUC = Area Under the ROC Curve; PR-AUC = Precision-Recall AUC.

	Classification Performance (Fixed Feature Set CV)				Notes
	AUC	PR-AUC	Accuracy	F1
RF Nested CV (outer loop)	0.79 ± 0.14	0.74 ± 0.17	0.68 ± 0.10	0.39 ± 0.27	Reference (leakage-free)
RF Stable (n=4 features)	0.95 ± 0.05	0.94 ± 0.07	0.88 ± 0.08	0.82 ± 0.13	≥50% fold stability
Top-4 d′ (n=4 features)	0.92 ± 0.08	0.89 ± 0.11	0.86 ± 0.09	0.79 ± 0.16	Univariate ranking
Random (n=4 features)	0.51 ± 0.09	0.49	—	—	Mean of 100 draws
All Features (n = 4950 features)	0.63 ± 0.16	0.58 ± 0.15	0.63 ± 0.03	0.04 ± 0.20	No feature-selection

The two feature sets (RF-selected and d’) share only one member: Feature 230, which is simultaneously the most stable RF feature (selection frequency = 1.00) and the highest |d′| feature (d′ = +1.41). The three additional RF stable features are not among the top-4 by d′ yet achieve superior combined performance, demonstrating that multivariate resampling-based selection identifies features with complementary discriminative information not captured by univariate ranking alone.

4. Discussion

TBI and PTE are conditions that lead to dramatic declines in quality of life, imposing long-term cognitive, psychiatric, and functional burdens on affected individuals; yet no validated neuroimaging biomarker exists to identify at-risk patients, and no prophylactic intervention has demonstrated efficacy (Kazis et al., 2024; Karlander et al., 2021; Burke et al., 2021; Lowenstein, 2009; Fang et al., 2023; Verellen and Cavazos, 2010). We developed a ML pipeline to analyze rs-fMRI data and characterize FC and RSN disruptions in HC, TBI, and PTE subjects, which we hope will enable design of future longitudinal studies that track the trajectory of PTE risk. In doing so, we place particular emphasis on a leakage-free feature selection framework and subsequent feature analysis as the central methodological contribution of this work, one that is designed to generalize beyond our own cohort to larger prospectively collected datasets.

A key finding of our study is that TBI and PTE are indistinguishable at the group level by classical network statistics: our RSN-level ANOVA identified no FDR-corrected significant differences between TBI and PTE across any of the 28 RSN blocks tested, mirroring prior reports that group-mean FC differences between injured cohorts with and without epilepsy are subtle or absent (Fig. 1Civ). (Weiler et al., 2024; Garner et al., 2019b) Building off this null finding, our ML pipeline identified a sparse subset of FCs that carries individual-level discriminative signal between TBI and PTE and use these to reveal several network FC patterns that emerged in our cohort, and which we reinforce through rigorous statistical validation and permutation testing. We present three principal contributions: (1) a methodological advance in leakage-free feature selection via nested CV-RFE, which recovered a discriminative signal invisible to baseline models trained on all FC features and which offers an approach directly transferable to larger, higher-dimensional FC datasets; (2), identification of a sparse, reproducible FC signature that discriminates established PTE from TBI at the individual level within this dataset at a high performance; and (3) convergent multi-framework evidence in this cohort, implicating the Visual network as a central hub of the FC differences associated with established PTE, characterized by bidirectional hemispheric coupling asymmetry.

4.1. Leakage-corrected honest performance of feature selection

Previous studies using FC-based classification, particularly those employing SVM-RFE on neuroimaging data, often overestimate classifier performance by evaluating it on data used for feature selection (Guyon et al., 2002; Cawley and Talbot). This study employed a nested cross-validation (CV) approach ensuring separation between feature selection and performance estimation, which resulted in a decrease in the estimated RF AUC from 0.96 (inner CV) to 0.79 (outer CV). This AUC of 0.79 remains competitive, reaching and surpassing previous reports, including an AUC of 0.78 by Akrami et al. (2024) and values between 0.69 and 0.78 reported by Garner et al. (2019a) Furthermore, our method uses only FC correlations as features on the more restrictive and clinically relevant TBI-only comparator, where the prior two studies integrated structural MRI features or used an Epilepsy-Free comparator. Our observed standard deviation of 0.14 across outer folds reflects the irreducible uncertainty of evaluating classification on approximately four PTE subjects per outer test fold, an inherent constraint of the sample size of our dataset.

The permutation test offered robust validation, showing a null AUC distribution centered at 0.50, while our observed AUC of 0.79 surpassed all permutations and yielded a p-value of p < 0.005 (Fig. 3D). This indicates that the classification signal reflects genuine differences rather than overfitting or leakage artifacts. As a further test of whether this signal depends on feature selection itself, baseline RF and SVM classifiers trained on all 4950 unselected FC features performed substantially worse than their nested CV-RFE counterparts, with baseline RF in particular showing near-complete failure to identify PTE cases at default thresholds despite balanced class weighting (Table 1). This gap demonstrates that the discriminative signal recovered by nested CV-RFE is not an artifact of raw FC feature richness, but rather reflects the selection procedure's ability to isolate a genuinely informative subspace from a feature set otherwise dominated by noise. Meanwhile, a linear SVM classifier yielded an outer CV AUC of 0.64 and that it selected a large number of features, indicative of a failure mode in the p ≫ n regime (Guyon et al., 2002). The outer AUC confirms that the SVM inner-loop performance does not generalize. This highlights the need for caution when applying linear SVM-RFE in neuroimaging studies with fewer than 100 samples without explicit safeguards; we therefore limit all subsequent biological interpretations to the RF classifier results.

4.2. Feature selection stability and the sparse FC signature

With the optimal number of features selected by the RF inner loop varying substantially across outer folds (Table 1), we characterize feature reliability using selection frequency. The vast majority of the 4950 FCs being selected in fewer than 10% of outer folds is consistent with a large pool of weakly and interchangeably discriminative connections (Fig. 3C). This exchangeability of features is itself informative: it is the signature of a condition in which PTE connectivity disruption is diffuse and multivariate rather than focal, with various feature subsets carrying comparable discriminative information (Tracy and Doucet, 2015). Four FCs emerged as Tier 1 stable features, selected in 60% or more of outer folds. All four stable features had |d'| ≥ 0.5 in the primary TBI versus PTE discriminability analysis, confirming convergence between univariate discriminability and multivariate resampling stability (Table 2). The validation analysis demonstrated that these four features, when evaluated on a fixed-feature basis, achieved a performance exceeding the top-4 d'-ranked features and that was far superior to random and all features (Fig. 7). The superiority of the resampling-stable set demonstrates that multivariate selection identifies features with complementary discriminative information not captured by individual effect sizes alone (Wang et al., 2019).

A critical observation is that feature instability at the individual FC level does not preclude reproducible interpretation at the network level. The RSN overrepresentation analysis demonstrated that VIS-containing blocks had the highest representation in RF-selected features (Fig. 5), even while FCs with high selection frequency remained rare (Fig. 3C). In other words, the network-level pattern of Visual network overrepresentation remained stable across resampling iterations even when the selection specific FCs vary. This dissociation between individual feature instability and network-level stability is an important methodological finding: it justifies RSN-level biological conclusions even in settings where single-feature stability is limited.

4.3. The Visual network as a hub of FC differences associated with established PTE

The most robust finding of this study is the convergence of three independent analytical frameworks on the Visual network as the central hub of PTE-discriminative FC. The RSN-level ANOVA identified VIS-SM and VIS-DMN among the FDR-corrected significant blocks for group differences across HC, TBI, and PTE (Fig. 1C). The d' analysis found Visual network ROIs contributing the largest concentration of impacted FCs in both the primary TBI versus PTE and secondary Epilepsy-Free versus PTE comparisons, as well as the most discriminative by d’ (Fig. 2B and C). The Rep-Score analysis confirmed that VIS-containing RSN blocks make up the majority of significantly overrepresented RSN blocks in RF-selected features (Fig. 5). The convergence of classical group statistics, univariate discriminability indexing, and machine learning feature selection stability on the same set of VIS-centered RSN blocks constitutes strong and multi-level evidence that Visual network connectivity disruption is a core and reproducible feature of the PTE FC profile relative to TBI.

VIS-FP hyper-coupling, represented by Feature 230 connecting left occipital Visual cortex (Vis_3) to left posterior parietal Fronto-Parietal cortex (Cont_pCun_1, precuneus region), emerges as the strongest individual discriminator in the dataset by both selection frequency (1.00, Fig. 3C) and d' magnitude (+1.41, Table 2). The precuneus is a posterior cortical hub at the intersection of the FP and default mode networks, playing a central role in visuospatial attention, episodic memory retrieval, and self-referential processing (Cavanna and Trimble, 2006). Its abnormal strengthening with occipital Visual cortex in PTE relative to TBI may reflect maladaptive visuocortical-parietal synchrony linked to cortical hyperexcitability, a pattern that has been proposed as a mechanistic contributor to epileptogenic network reorganization in posterior cortical epilepsies (Spencer, 2002; Laufs, 2012). Notably, VIS-FP hyper-coupling was the only finding shared across both d' analyses (Fig. 2Biii, Ciii).

The within-Visual network overrepresentation (Fig. 5) is a novel finding suggesting that internal VIS network organization, not only inter-network connectivity, is preferentially disrupted in PTE. This observation extends prior reports that VIS network alterations in epilepsy are not confined to the epileptic focus but represent network-level consequences (Laufs, 2012). The VIS-VIS block did not contain a single Tier 1 stable feature, indicating that this within-network overrepresentation is driven by a distributed pool of individually unstable but collectively overrepresented connections, consistent with diffuse local reorganization of Visual network circuitry rather than disruption of a focal within-network pathway.

VIS-SM overrepresentation (Fig. 5) replicates the most consistently implicated RSN block from the RSN-level ANOVA (Fig. 1C) and confirms that visuomotor FC disruption survives both classical group statistics and machine learning feature selection when differentiating PTE. These findings are consistent with prior work implicating SM RSNs in FC-based PTE classification (Garner et al., 2019a) and with reports that chronic TBI is characterized by disproportionately large reductions in between-network functional connections (Han et al., 2016).

Equally informative is what was absent from the RF-selected feature space. DMN-DMN, SM-DMN, LIM-DMN, and DA-DMN blocks, which showed FDR-significant group differences in the RSN-level ANOVA (Fig. 1C) and elevated d' values in the secondary Epilepsy-Free versus PTE analysis (Fig. 2C), were not overrepresented in the primary TBI versus PTE ML analysis. This selectivity directly validates the revised comparison framework: these blocks reflect shared TBI-related FC disruption that is common to both TBI and PTE groups, and their absence from the RF-selected feature space supports a shared TBI-related disruption, rather than reorganization specific to established PTE in our cohort.

4.4. Bidirectional and lateralized hemispheric reorganization in PTE

The hemispheric analysis revealed that LH-LH connections are significantly overrepresented in the RF-selected feature space (Fig. 6A), driven by the concentration of left Visual network connections (Fig. 6B). The demonstrated pattern of dominant coupling per hemisphere, hyper-coupling in LH-LH and un-coupling in RH-RH, was consistent across raw FC counts (Fig. 6Ci) and RF selection frequency-weighted analyses (Fig. 6Cii). This pattern is further recapitulated in the Tier 1 stable features: the two LH-LH stable features include the only hyper-coupled Tier 1 feature, while the RH-RH stable feature is un-coupled (Table 2).

The overall near-equal balance between hyper-coupling (49.4%) and un-coupling (50.6%) across all impacted FCs in the primary TBI vs. PTE comparison reflects a true bidirectional FC reorganization signature (Fig. 2Dii). This stands in contrast to the predominance of un-coupling in the secondary Epilepsy-Free vs. PTE comparison (62.5% un-coupled, Fig. 2Dii), where inclusion of HC subjects in the reference class creates an elevated reference distribution that biases toward apparent un-coupling in PTE. When the comparison is restricted to TBI vs. PTE, this bias is removed, revealing that the established PTE state is associated with opposing connectivity remodeling across hemispheric compartments, consistent with the emerging view of epilepsy as a disorder of network remodeling rather than uniform connectivity loss (Laufs, 2012; Lehnertz et al., 2023).

The hemispheric dissociation between coupling directions is interpretable in light of prior work. Rigon et al. reported that TBI patients exhibit significantly reduced inter-hemispheric FC in frontoparietal and executive control networks, with homologous LH-RH couplings disproportionately weakened, suggesting that large-scale interhemispheric disconnection is a chronic sequela of injury (Rigon et al., 2016). Consistent with this, cross-hemispheric connections in our dataset were not selectively overrepresented in the TBI vs. PTE ML analysis, supporting the interpretation that interhemispheric disconnection is a shared feature of TBI-related injury rather than epilepsy-specific. The lateralized reorganization we observe, left-hemispheric hyper-coupling and right-hemispheric un-coupling therefore represents a pattern distinguishing established PTE from TBI in our cohort.

With respect to the right hemisphere specifically, Hüsser et al. reported marked hemispheric asymmetry in children with focal epilepsy, with increased RH-RH FC and local network efficiency, suggesting that the right hemisphere may assume a disproportionate burden of network reorganization in response to epileptogenic processes (Hüsser et al., 2023). In contrast, the RH-RH un-coupling pattern we observe in adult PTE relative to TBI, particularly the FP-DMN connection represented by Feature 4805 (Table 2), suggests progressive disconnection rather than compensatory strengthening. This distinction likely reflects the difference between pediatric focal epilepsy with active compensatory recruitment and adult post-traumatic epilepsy with chronic resting-state network fragmentation. Viewed together, both studies support the notion that the right hemisphere is a preferential locus of large-scale network reorganization in epilepsy, with the direction of that reorganization dependent on the nature, timing, and context of the epileptic process.

This feature 4805, which connects the right lateral prefrontal FP cortex to right temporal DMN cortex, represents an independent right-hemispheric component of the stable FC signature that is anatomically and functionally dissociable from the left Visual hub. The FP-DMN interaction is among the most studied large-scale network couplings in cognitive neuroscience, mediating the balance between goal-directed attention and internally directed thought (Fox and Greicius, 2010). Its disruption in PTE is consistent with frontoparietal-DMN FC reductions documented in chronic TBI and with the executive and memory impairments that characterize PTE patients relative to TBI survivors without seizures (Burke et al., 2021; Han et al., 2016). Studies in mild TBI have also documented both acute FC reductions in DMN, FP, and sensorimotor networks and transient hyper-coupling in some FCs attributed to compensatory mechanisms (Hillary et al., 2014; Iraji et al., 2016), with the longer-term trajectory appearing to depend on structural remodeling, glial activation, and neurotransmitter imbalances that may entrench maladaptive network states conducive to seizure development (Fang et al., 2023). Our finding that the VIS-FP block represents the singular convergence point of hyper-coupling in the stable feature set and the highest d' magnitudes across both comparisons implicates this specific visuoparietal pathway as a locus of maladaptive synchronization rather than compensatory recruitment, a distinction that suggests particular relevance to seizure network dynamics.

4.5. Limitations and future directions

Localized recruitment has limited our dataset to 20 PTE subjects and a slightly larger number of TBI and HC subjects. The primary consequence is the high variance of outer CV fold performance estimates (AUC SD = 0.14), arising from evaluation on approximately four PTE subjects per test fold. We report this variance transparently, as it accurately reflects the uncertainty of generalization estimates at this sample size. The permutation test and nested CV design provide internal validation that the signal is real, but they cannot substitute for external replication in an independent cohort.

A fundamental limitation of this cross-sectional design is that the FC features identified characterize established PTE rather than the epileptogenic process in evolution. Whether these connections are disrupted prior to the first seizure in at-risk TBI patients, and whether they could serve as prospective biomarkers of epileptogenesis, cannot be determined from the present data. The critical next step is a longitudinal prospective study following TBI patients post-injury and tracking FC trajectories in patients who go on to develop PTE to determine whether the VIS-FP hyper-coupling and the broader stable FC signature precede seizure.

The inherent clinical heterogeneity of TBI and PTE, including variability in injury mechanism, severity, anatomical lesion location, time since injury, seizure phenotype, anti-seizure medication exposure, and seizure control status, continues to complicate the explicit characteristic of PTE-related FC changes. Future work should stratify analyses by TBI severity classification, lesion location relative to the Visual network, and ASM class to determine whether the FC signature identified here generalizes across these clinical subgroups or is driven by a specific subset. Post quality control FD did not differ significantly across cohorts, and the nominal TBI vs. PTE difference did not survive multiple-comparison correction, arguing against motion as a primary driver of the classification signal. However, global signal regression and volume-level scrubbing were not performed in this pipeline; residual motion-related variance below the subject-exclusion threshold cannot be fully excluded, and future work should incorporate FD as a covariate or apply scrubbing as a sensitivity analysis.

5. Conclusion

In conclusion, this study contributes to a growing body of evidence that PTE is characterized by significant and spatially organized disruptions in the brain's network architecture, reinforcing the emerging view that epilepsy is fundamentally a network disorder and not merely a focal pathology. We introduce a novel ML and feature selection framework that is leakage-free and permutation tested, revealing specific network FC changes centered on the Visual network and organized by a bidirectional hemispheric coupling asymmetry. This represents a validated cross-sectional FC signature of the established epileptic state that is reproducible across resampling iterations and robust to the methodological corrections required by leakage-free evaluation. Larger prospective datasets with longitudinal tracking of TBI patients through PTE onset are essential for determining whether the FC differences identified here generalize and carry predictive value. The leakage-free feature selection framework applied in this study offers a directly transferable methodological foundation for such future work.

CRediT authorship contribution statement

Alan Ho: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. Daniel J. Valdivia: Data curation, Formal analysis, Investigation, Methodology, Project administration, Writing – original draft, Writing – review & editing. Henry Noren: Formal analysis, Investigation, Methodology, Software, Validation, Writing – original draft, Writing – review & editing. Juliana Cantarutti: Formal analysis, Investigation, Methodology, Writing – original draft, Writing – review & editing. Pratik Jain: Formal analysis, Methodology, Software, Validation, Writing – review & editing. Taylor Zink: Data curation, Formal analysis, Validation, Writing – review & editing. Koray Ercan: Data curation, Formal analysis, Writing – review & editing. Karthik Narra: Methodology, Software, Validation, Visualization. Christine Yohn: Project administration, Supervision, Validation, Writing – review & editing. David M. Scarisbrick: Investigation, Methodology, Supervision, Validation, Writing – original draft, Writing – review & editing. Bharat Biswal: Conceptualization, Data curation, Formal analysis, Funding acquisition, Project administration, Software, Supervision, Validation, Writing – review & editing. Spencer Chen: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Software, Supervision, Validation, Writing – original draft, Writing – review & editing. Hai Sun: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Supervision, Validation, Writing – review & editing.

Data and code availability statement

All analysis code used in this study will be made available via a publicly accessible zipped package on Mendeley Data (https://data.mendeley.com/preview/pmhpbx88sh?a=154fa612-ddf6-4b86-b2ad-9ba2c71af611). This package includes the scripts required to reproduce the analyses as well as the extracted functional connectivity (FC) correlation coefficients for each subject used in the reported analyses.

Raw neuroimaging data are not distributed directly with this manuscript and are instead shared through the Federal Interagency Traumatic Brain Injury Research (FITBIR) Informatics System, in accordance with U.S. Department of Defense (DoD) data-sharing requirements. Access to these data is subject to FITBIR data use agreements and approval procedures.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work the author(s) used ChatGPT (OpenAI) and Claude (Anthropic) in order to support code generation and creation, and preliminary literature review. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the published article.

Declaration of competing interests

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Hai Sun reports financial support was provided by US Department of Defense. Hai Sun reports financial support was provided by US Office of Congressionally Directed Medical Research Programs. Hai Sun reports financial support was provided by National Institute of Neurological Disorders and Stroke. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the Supplementary data to this article.

Data availability

I have shared the link to the data/code used in this manuscript via Mendeley in the Attach File step under Research Data.

• Mendeley DataPTE_FC (Original data)