Authors :
Presenting Author: Arushi Munjal, BS in Mathematics and Computer Science – University of California, San Diego (UCSD)
Sneha Sairam, BS – University of California, San Diego (UCSD)
Donatello Arienzo, PhD – University of California, San Diego (UCSD)
Carrie McDonald, PhD – UC San Diego
Taha Gholipour, M.D. – University of California San Diego
Rationale:
Mesial temporal lobe epilepsy (TLE) is characterized by functional disruptions that often extend beyond the temporal lobe, affecting widespread cortical and subcortical regions. Functional Anomaly Mapping (FAM) is a machine learning-based approach that uses resting-state fMRI (rs-fMRI) to identify deviations in functional connectivity, and was shown to correlate with laterality of TLE. Patient-level utility and reproducibility of this method have not been established. In this study, we leverage the highly phenotyped Epilepsy Connectome Project (ECP) dataset to identify consistent functional aberrations associated with left and right TLE and to evaluate the reproducibility of FAM within individuals across sessions.
Methods:
T1-weighted and resting-state fMRI data from 103 TLE patients (66 left, 21 right, 15 bilateral TLE; 66.9% male, 33.1% female) and 77 controls (51.9% male, 48.1% female) from the ECP dataset were preprocessed using fMRIPrep, and time series were extracted using nilearn from 400 cortical and 32 subcortical regions defined by functionally-derived brain atlases. Each participant had multiple sessions, and multiple fMRI runs per session. Functional Anomaly Maps (FAMs) were generated using Support Vector Regression (SVR) implemented in scikit-learn, modeling aggregated deviations in each session relative to all controls. Individual FAMs were computed for each session, consisting of an relative anomaly score for each region (from 0 to 1, the highest estimated absolute SVR region weight). The prevalence of each region among the top 10% of individual patients were tabulated compared between left and right TLE groups. Reproducibility was evaluated by calculating Pearson’s correlation coefficient between FAM maps from different sessions of the same subject.
Results:
Individual FAMs were computed for 221 sessions from 103 patients and visualized on a brain template volume (Figure 1 illustrates the rs-fMRI processing pipeline). Aberrant patterns were predominantly non-limbic in both hemispheres, with right parietal and temporal regions—affiliated with dorsal attention, control, and default mode networks—showing the most frequent high anomaly. Somatomotor network regions were also frequently represented in the top 10% across patients, with some lateralization differences between left and right TLE. Within-subject reproducibility showed moderate variability (mean Pearson correlation: 0.25; median: 0.22; IQR: 0.12–0.32).
Conclusions:
We examined the functional anomalies associated with TLE and evaluated the reproducibility of individualized SVR-based FAMs using the ECP dataset. We observed session-to-session variability within individuals and across subjects with similar clinical phenotypes. These findings suggest that FAM captures both stable and state-dependent functional changes. Dynamic or time-resolved modeling may help disentangle disease-related signals from acquisition-related variability. This study provides a benchmark for improving the clinical utility of FAM in epilepsy, with potential implications for individualized treatment planning.
Funding:
NINDS Career Development Award 1K23NS135108 to TG.