Abstracts

Rationale: In temporal lobe epilepsy (TLE), precise predictions of clinical outcomes remain challenging, likely due to the dominant “one-size-fits-all” group-level analytical approaches that do not allow parsing inter-individual variations. Here, we estimated latent relations (or disease factors) from MRI features representing whole-brain patterns of structural pathology. We assessed the specificity of factors against healthy individuals and frontal lobe epilepsy (FLE) patients. Moreover, we evaluated the data-driven latent factor model for individualized prediction of drug-response, seizure outcome and cognitive dysfunction.

Methods: We studied 82 TLE patients (30 males, 35±10 yrs), 41 healthy controls (18 males, 32±8 yrs) and 29 FLE disease controls with type-II focal cortical dysplasia (29 males, 32±6 yrs) using T1-weighted, FLAIR and diffusion MRI at 3T. Among patients, 12 were pharmaco-responsive, and 43/57 who had surgery were seizure-free. For feature extraction (Fig. 1A), we generated surfaces running through cortical mantle, 2 mm below its boundary and central paths of hippocampal subfields. We sampled cortical thickness and hippocampal volume (to model atrophy), intracortical/hippocampal FLAIR intensity (gliosis), T1w/FLAIR (demyelination) and subcortical/hippocampal FA and MD (microstructural damage). We estimated patterns of alterations (or disease factors), expressed as posterior probability, and quantified their co-expression within each patient using the Latent Dirichlet Allocation algorithm (Fig. 1B).¹ Gradient Boosting classifiers predicted drug response (resistant vs controlled) and surgical outcome (Engel I vs Engel II-IV) as well as cognitive scores (verbal IQ, memory, sequential motor tapping), with 10-fold cross-validation repeated 100 times. Performance evaluation was based on balanced accuracy for clinical outcomes and linear correlations for cognitive scores.

1. Blei DM, Ng AY, Jordan MI. Latent Dirichlet allocation. J Mach Learn Res. 2003;3:993-1022.

Results: We identified four latent factors (Fig. 2A) characterized by ipsilateral hippocampal microstructural alterations, loss of myelin and atrophy (Factor-1), bilateral paralimbic and hippocampal gliosis (Factor-2), bilateral neocortical atrophy (Factor-3) and bilateral white matter microstructural alterations (Factor-4). Bootstrap analysis supported factors stability and robustness. While factors were variably co-expressed within each TLE patient, they were not expressed in healthy controls and only negligibly in disease controls, supporting specificity (Fig. 2B). Classifiers trained on latent disease factors accurately predicted patient-specific drug-response in 76±3% and postsurgical seizure outcome in 88±2% of patients as well as inter-patient variability in verbal IQ (r=0.40±0.03), memory (r=0.35±0.03) and sequential motor tapping (r=0.36±0.04), outperforming baseline learners (Fig. 2C).

Conclusions: Data-driven analysis of disease factors provides a novel appraisal of the continuum of interindividual variability, which is likely determined by multiple interacting pathological processes.

Funding: Please list any funding that was received in support of this abstract.: Canadian Institutes of Health Research, Epilepsy Canada, Savoy Epilepsy Foundation.