Abstracts

Rationale:
Temporal lobe epilepsy (TLE) is the most common drug-resistant epilepsy, with hippocampal sclerosis as its hallmark. Previous MRI studies have shown gray (GM) and white matter (WM) alterations beyond the hippocampus. Yet, their individual variability and prognostic values remain unknown. Here, we applied Latent Dirichlet Allocation, a probabilistic model based on Bayesian inference originally introduced in text mining for unsupervised topic discovery. Our purpose was to extract hidden knowledge and relations from multimodal MRI features to model TLE neurobiology at an individual level. Unlike traditional clustering, this approach quantifies the expressions of multiple factors for each patient rather than assigning patients to a given category.
Method:
We studied 82 TLE patients (30 males, 35.2±9.7 yrs) and 41 age and sex-matched healthy controls using T1-weighted, FLAIR and diffusion MRI at 3T. Among patients, 11/82 responded to medication and 42/57 who had surgery were seizure-free. We generated surfaces running through cortical mantle, 2 mm below GM-WM boundary and central paths of hippocampal subfields. We sampled cortical thickness and hippocampal volume (to model atrophy), intracortical/hippocampal FLAIR intensity (gliosis), T1w/FLAIR ratios (demyelination) and subcortical/hippocampal FA and MD (microstructural damage).  Latent Dirichlet Allocation modeled each patient as a mixture of disease factors and each factor as a mixture of vertices quantifying imaging features. Linear models assessed associations of disease factors with local (fALFF, degree centrality) and global (clustering coefficient, global efficiency) function derived from resting-state fMRI. Extreme Gradient Boosting classifier predicted surgical outcome and medication response.
Results:
The algorithm identified 4 factors (Fig. 1). Factor 1: Ipsilateral hippocampal damage; Factor 2: Bilateral hippocampal and neocortical paralimbic gliosis; Factor 3: Bilateral GM thinning; Factor 4: Bilateral subcortical WM microstructural damage. Factors were variably expressed within and across individual patients. In relation to function (Fig. 2), Factors related to limbic pathology (1 and 2) were associated with abnormal function of limbic and somatomotor cortices, while Factor 3 representing neocortical thinning impacted sensorimotor cortices. Conversely, Factor 4 was associated with abnormal global topology. The classifier accurately predicted drug-resistance in 71±2% (pFDR=0.008, Factors 1 and 2 as main contributors) and Engel-1 outcome in 77±3% (pFDR=0.002; Factors 2 and 4) patients.
Conclusion:
MRI-based topic modeling uncovered four disease factors variably expressed within and across patients, with distinct impact on brain function, providing novel insights into the neurobiology of TLE. While bilateral paralimbic gliosis and ipsilateral hippocampal damage were most predictive of drug resistance, ipsilateral hippocampal and bilateral subcortical WM damage were predictive of favorable surgical outcome. MRI-based predictors may help reducing ineffective drug trials and accelerate referrals for pre-surgical investigation.
Funding:
:Savoy Epilepsy Foundation, Canadian Institutes of Health Research