UC San Diego

Objective: We aimed to determine whether a novel, white matter (WM)

network-based approach (i.e. structural connectomes; SC) outperforms

traditional tract-based measures and hippocampal volume (HCV) for discriminating

temporal lobe epilepsy (TLE) patients with memory-impairment from those with normal

memory.

Methods: T1-and diffusion-weighted imaging (DWI), clinical, and

neuropsychological data were available for 81 patients with TLE and 61 healthy

controls. Five medial temporal lobe WM tracts were extracted from the DWI and

fractional anisotropy (FA) was calculated for each tract. Left and right HCVs were

derived using FreeSurfer. SCs were derived by performing probabilistic tractography

and measuring the number of connections among cortical regions. SCs were fed into

XGBoost, a robust tree-based classifier. The model was trained on 48 patients

from UCSD and tested on 33 patients from UCSF.

Results: Both SCs (76% accuracy) and tracts (73% accuracy) were better

classifiers of verbal memory performance than HCV (66\% accuracy) or

clinical variables (61% accuracy). Within the WM-based models, the SC

performed similar to or better than the tract-based models, with slightly

higher accuracy and positive predictive value (PPV; 0.81 for SC, 0.77 for

tracts). The SC appeared to provide more specific information regarding

abnormal connectivity contributing to verbal memory performance. Multivariate

models seemed to be most robust classifiers, providing the highest sensitivity

(0.95 for Tract+HCV) and specificity (0.67 for Connectome+HCV).

Conclusion: SCs and tract-based measures appear to outperform HCVs and clinical

variables for predicting memory impairment in TLE. However, network-based

models combined with HCVs may provide the best prediction accuracy.