Abstracts

Rationale: Temporal lobe epilepsy (TLE) is the most common drug-resistant epilepsy, with hippocampal sclerosis (HS) as its hallmark (Blümcke et al., 2013). Predicting drug resistance and surgical outcome are major clinical challenges. Previous MRI studies have shown structural and microstructural alterations of the gray (GM) and white matter (WM) beyond the hippocampus (Bernhardt et al., 2013; Liu et al., 2016; Bernhardt et al., 2017; Adler et al., 2018). However, the inter-individual variability of these alterations and their prognostic values remain unknown. Here, we hypothesized that whole-brain disease factors variably expressed within and across patients are predictive of clinical outcomes. We applied Latent Dirichlet Allocation (Blei et al., 2003), a generative 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 expressed within individual patients. Unlike traditional clustering, this approach quantifies the expressions of multiple factors for each patient rather than assigning patients to a given category. Methods: We studied 82 TLE patients with DRE (11/82 responders) and histologically-verified HS (42/57 had Engel-I) using T1w, FLAIR and diffusion MRI at 3T. Patients were compared to 41 healthy controls. We generated surfaces running through the cortical mantle, 2 mm below GM-WM boundary and central paths of hippocampal subfields. We then sampled cortical thickness and hippocampal columnar volume (to model atrophy), intracortical/hippocampal FLAIR intensity (indexing gliosis), T1w/FLAIR ratios (modeling demyelination) and subcortical/hippocampal FA and MD (indexing 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. An Extreme Gradient Boosting classifier predicted individualized drug resistance and favorable surgical outcome. The classifier was validated using a 10-fold scheme repeated 100 times; performance relative to 10,000 random guesses determined statistical significance. Results: The algorithm identified 4 latent factors (Fig. 1A). Factor 1: Bilateral hippocampal and neocortical paralimbic gliosis; Factor 2: Ipsilateral hippocampal damage; Factor 3: Bilateral neocortical thinning; Factor 4: Bilateral subcortical WM microstructural damage. Factors were variably expressed within and across individual patients (Fig. 1B). The classifier accurately predicted drug resistance in 67.1 ± 3.3% (p=0.02) of patients, with Factor 1 as the main contributor (Fig. 2A). Favorable response to surgery was predicted in 76.4 ± 2.8% (p = 0.0004), with strongest contribution from Factors 2 and 4 (Fig. 2B). Conclusions: MRI-based factor modeling uncovered four disease factors, variably expressed within and across patients, providing novel insights into the neurobiology of TLE. While bilateral limbic gliosis was the strongest predictor of drug resistance, ipsilateral hippocampal and bilateral subcortical WM damage were predictive of favorable surgical outcome. The availability of objective MRI-based predictors may help reducing ineffective drug trials and accelerate referrals for pre-surgical investigation, thereby offering the chance of a seizure-free future to more patients. Funding: Savoy Epilepsy Foundation, Canadian Institutes of Health Research