Abstracts

Rationale: Introduction: Although gray matter atrophy in temporal lobe epilepsy (TLE) has been extensively studied, advanced in-vivo MRI analyses of cortical gyrification may aggregate information about complex pathological processes. Mota et al.⁽¹⁾ validated a cortical folding model (in 55 mammals) based on metrics of cortical thickness (T), total surface area (At) and exposed surface area (Ae), describing a universal scaling law: At*T^1/2=k*Ae^5/4. Wang et al.⁽²⁾ advanced the model to a vector base space describing the human cortex with three orthogonal variables: axonal tension (K), cortical shape complexity (S) and brain isometric volume (I) (Figure1). So far, little is known about the pathological changes of these cortical variables in TLE and their capacity to differentiate subgroups.

Objective: To compare the 3 independent vector-based components between controls and right (RTLE) and left TLE (LTLE) subgroups. Secondly, we used machine learning models to evaluate their ability to distinguish the subgroups.


Methods: Methods: We extracted cortical measures (T, At, and Ae) from 263 controls (median age of 32 years) and 167 epilepsy patients (median age of 45 years; 96 LTLE, 71 RTLE; median duration of epilepsy of 35 years) with Freesurfer7.4.1. We computed hemispheric metrics I, K, and S as scalar metrics of vectors in a log10-normalized space, as shown in Figure1. The variables were adjusted for age and sex and standardized (z-score) relative to the controls. We used repeated measures ANOVA to compare TLE patients (total group and subgroups) and controls, with partial eta squared (η_p²) to report effect size and Bonferroni’s correction for multiple comparisons.

We applied a support vector machine (SVM, built using scikit-learn and Python 3) as a machine learning classification model (ML) to investigate whether cortical measures (raw FS and new cortical folding measures) can discriminate patients from controls.


Results: Results: We observed a reduction in the ipsilateral I in TLE compared to controls (brain isometric volume, η_p²=0.103; p=0.006, Bonferroni corrected) and a trend for bilateral increases of K and S. For subgroups, ipsilateral I was reduced in LTLE (η_p²=0.070; p=0.012) and RTLE (η_p²=0.085; p=0.037) compared to the controls. Contralateral K was higher (η_p²=0.005, p=0.023) in LTLE, and contralateral S was higher in RTLE (η_p²=0.014; p=0.032). The SVM analyses separated TLE and controls using IKS (AUC=0.61±0.11) or raw Freesurfer measures (AUC=0.60±0.11). Further, the ML models distinguished LTLE from RTLE using IKS (AUC=0.78±0.20) and raw FS measures (AUC=0.87±0.12) (Table 1).


Conclusions: Conclusions: While we identified pathological brain size reduction on TLE (specifically cortical thinning) using raw Freesurfer measures, the new morphological measures revealed more information about subgroups. Further exploration with higher granularity (lobar analyses and laterality) may advance the understanding of pathological cortical folding in epilepsy and improve artificial intelligence models.


Funding: This study was funded by FAPESP-BRAINN (2013/07559-3), FAPESP-Monash (2022/03894-1), FAPESP (2022/11786-4) and CAPES (Coordination of Superior Level Staff Improvement)