Personalized biomarkers of multiscale functional alterations in temporal lobe epilepsy

Abstract

Temporal lobe epilepsy (TLE) is the most common pharmacoresistant epilepsy in adults, yet few patients receive curative surgery due to diagnostic and prognostic uncertainty. In a multicenter cohort, we analyzed multimodal MRI and clinical data from 282 TLE patients, 298 healthy controls, and 45 disease controls. Patient-specific deviations from typical lifespan trajectories of intrinsic brain function were mapped using normative modeling. Regional functional alterations were heterogeneous but overlapped most in the mesiotemporal cortex. Connectome-based simulations revealed abnormality spread followed structural network architecture, highlighting the hippocampus as well as paralimbic and medial default-mode regions as epicenters. Multimodal integration implicated superficial white-matter microstructural alterations as a key contributor. Supervised models achieved AUCs of 0.77 for distinguishing TLE from disease controls, 0.74 for lateralizing seizure focus, and 0.64 for predicting postsurgical seizure freedom; greater contralateral temporal deviations predicted poorer outcomes. These findings support individualized functional biomarkers for precision presurgical care in focal epilepsy.

Fig. 1. Overview of participants and analysis pipeline.

A Demographic information of the healthy control (n = 298) and TLE groups (n = 282). i Age distributions. Boxplots show the median line, 25th and 75th percentiles (lower and upper bounds), and minima and maxima 1.5 interquartile range whiskers. Dots represent individual participants. ii Proportion of TLE patients with left- or right-sided seizure focus in each dataset. iii Proportion of TLE patients with or without ipsilateral hippocampal sclerosis in each dataset. B Overview of methodology for calculating functional metrics from resting-state functional MRI (rs-fMRI) time series in each cortical region and subcortical structure: signal temporal variability, regional homogeneity, and node strength. C Schematic of individual functional deviation (W-score) estimation, with an example here for node strength (NS). i Building normative models in healthy individuals. Age- and sex-related variations in each rs-fMRI metric are modeled using a linear regression model in the healthy control group, yielding beta maps for intercept (β₀), age (β₁), sex (β₂), and standard deviation of residuals for each brain region. ii Estimating deviations in each patient. Predicted value (e.g., NS_pred) for a given patient’s age and sex is calculated as β₀ + β₁ × age + β₂ × sex. W-scores are defined as the normalized deviation of the observed values from the predicted values. iii Identifying extreme deviations. Brain regions with W-scores exceeding ±1.96 (i.e., | W-scores | ≥ 1.96) are classified as showing extreme deviations, corresponding to the upper and lower 2.5% of the normative distribution. Abbreviation: n = sample size; SD = standard deviation. Source data are provided as a Source Data file.

Fig. 2. Region-specific prevalence of extreme functional deviations in TLE patients.

A Proportion of TLE patients with extreme deviations (|W-score | ≥ 1.96) in each cortical region and subcortical structure for (i) signal variability (SV), (ii) regional homogeneity (ReHo), and (iii) node strength (NS). Spider plots show the mean proportion of extreme deviations in each cortical lobe defined on the HCPMMP1.0 atlas and in the subcortex. B i Proportion of TLE patients with multivariate W-scores (aggregating SV, ReHo and NS) exceeding ±1.96. ii Region-specific deviation prevalence distribution. Each dot represents an individual brain region, with its position indicating the prevalence values of the three metrics and its color denoting the composite deviation prevalence. iii Mean proportion of extreme deviations per lobe. iv Distribution of the number of extreme deviations per patient. v Proportion of TLE patients with extreme deviations in each subcortical structure. Abbreviation: ipsi = ipsilateral; contra = contralateral; F = frontal; P = parietal; T = temporal; O = occipital; I = insula; S = subcortex. Source data are provided as a Source Data file.

Fig. 3. Network-based spreading of regional functional deviations.

A i Group-level structural connectivity (SC) matrix from diffusion MRI of 100 unrelated healthy individuals. ii Schematic of functional deviations of a node (p_i) and its neighbors (${\overline{p}}_i$). If the regional deviation depends on SC network organization, regions connected to highly abnormal neighbors (i.e., high prevalence) will be more likely to be affected, whereas regions connected to healthy neighbors (i.e., low prevalence) will be less likely to be affected. iii Functional deviation of each node (x-axis) versus the mean deviation of its structurally connected neighbors (y-axis). Dots are brain regions; the line shows Spearman rank fit with a 95% CI (gray band). Correlation significance (i.e., P_spin) is assessed using spin permutation tests (5000 iterations; one-sided). B Schematic of disease epicenter identification. A region whose SC profile spatially strongly relates to the TLE-related functional deviation map in Fig. 2B is considered as a disease “epicenter”. Epicenter likelihood is defined as the Spearman correlation coefficient between two spatial maps. C, D SC-informed epicenter likelihood map. Statistical significance of the likelihood is determined using spin permutation tests (5000 iterations; P_spin < 0.05 against the null models, one-sided). Abbreviation: ipsi = ipsilateral; contra = contralateral. Source data are provided as a Source Data file.

Fig. 4. Spatial associations between brain structural and functional deviations.

A Multilinear regression model is used to predict regional functional deviation prevalence from deviation prevalence maps of cortical thickness (CT), fractional anisotropy (FA), and mean diffusivity (MD). B Empirical (x-axis) versus predicted (y-axis) functional deviation prevalence maps. Dots are brain regions; the line shows Spearman’s rank fit with a 95% CI (gray band). Correlation significance (i.e., P_spin) is assessed using spin permutation tests (5000 iterations; one-sided). C Dominance analysis quantifies the contribution of each feature (CT, FA and MD). Source data are provided as a Source Data file.

Fig. 5. Clinical utility of individual functional deviations.

Receiver operating curves showing the accuracy of functional deviations in (A) classifying TLE patients (n = 129) versus extratemporal FCD patients (n = 45), and (B) classifying left TLE (n = 147) versus right TLE (n = 135) patients. C Interhemispheric differences in the number of extreme deviations in seizure-free (SF, n = 74) and non-seizure-free (non-SF, n = 25) TLE patients. i T-values (paired t-test) showing the differences in the number of extreme deviations between the ipsilateral (ipsi) and contralateral (contra) hemispheres in SF and non-SF TLE patients in each cortical lobe, respectively. Positive/negative t-values indicate more/fewer extreme deviations in the ipsilateral hemisphere compared to the contralateral hemisphere. ii Patient-specific proportions of extreme deviations in the ipsilateral (red) and contralateral (green) temporal lobes for SF and non-SF TLE patients. Asterisks denote significance at P < 0.001 from one-sided paired t-tests after false discovery rate correction. Boxplots show the median line, 25th and 75th percentiles (lower and upper bounds), and minima and maxima, 1.5 interquartile range whiskers. Dots represent individual participants. Abbreviation: F = frontal; P = parietal; O = occipital; I = insula; T = temporal. Source data are provided as a Source Data file.