This abstract will be presented during the Neuro Imaging Platform poster session
Rationale: Patients with medically-refractory mesial temporal lobe epilepsy (mTLE) often undergo extensive diagnostic workups to lateralize and locate epileptogenic foci. It is important to determine if the patient has bilateral seizure onset because this drastically affects treatment options. The gold standard to discern bilateral seizure onset is to perform invasive intracranial monitoring. Furthermore, nearly a third of patients with bilateral mTLE may require more than four weeks to demonstrate bilateral seizures (King et al. Epilepsia, 2015, 56(6): p. 959-967). A simpler, faster, and non-invasive technique to identify mTLE patients with bilateral epileptogenicity would be clinically valuable. Diffusion-weighted imaging (DWI) offers a non-invasive approach to study the white-matter of patients with mTLE. We aim to show that specific white-matter features can be extracted that help elucidate if a patient is likely to harbor bilateral seizure onset. Methods: Fifteen patients were selected with unilateral (n=10) or bilateral (n=5) mTLE confirmed with intracranial monitoring and improvement after resection or responsive neurostimulation surgery. Structural connectivity (SC) was elucidated by using MRTrix3 to conduct probabilistic tractography (mrtrix3.org) on DWI volumes. Tractography was run four to nine times per patient to balance the dataset. We ran an independent components analysis (ICA)-based feature extraction technique created by Kwak et al. (Springer, 2001, LNCS 2130, p. 568-573) repeatedly until 1000 candidate SC feature matrices were generated. Principal components analysis was then performed and the eigenvector of the first principal component was kept and dubbed the “Decoding Matrix”. We performed leave-one-out cross fold validation.The decoding matrix is multiplied element-wise to the patient’s SC matrix to obtain a prediction number for each patient to be used in a logistic regression. Beyond prediction, the decoding matrix can be directly interpreted with structural connections important for classification having high absolute values. Results: The decoding matrix (Figure 1A) was able to correctly classify unilateral vs bilateral mTLE with an area under the receiver operating characteristic curve (AUC) of 0.856 for leave-one-out validation (Figure 1B). Bilateral mTLE patient data was correctly identified 87% of the time (note: data was augmented so each patient had multiple predictive values). When taking the average prediction per patient, four of five bilateral mTLE patients were correctly classified.The vast majority of white-matter structural connections important to classify unilateral vs bilateral mTLE are cross-hemispheric connections (dark boxes in Figure 1A). Interestingly, the white-matter connections of the mesial structures (ROIs 40-42 and 47-49) were not strongly predictive. Finally, there is very strong predictive value of the bilateral connections of each hemisphere’s frontal pole (ROIs 31, 80), temporal pole (ROIs 32, 81), and transverse temporal region (ROIs 33, 82), as seen in the dark lines for these ROIs. See Table 1 for a full list of ROIs. Conclusions: This technique provides a preliminary non-invasive classification scheme to distinguish patients with unilateral vs. bilateral mTLE. Once improved and validated, this technique may have clinical value in identifying mTLE patients at high risk for harboring bilateral epileptogenicity, and thus may require prolonged intracranial EEG before making a surgical decision. Of further interest is that the decoding matrix can be directly interpreted to elucidate pathological motifs that are uniquely representative of disease phenotype. Funding: Please list any funding that was received in support of this abstract.: R00NS09761805, R01NS11225202, T32GM734740, R01NS110130, R01NS108445
.jpg)