CLASSIFICATION OF INDIVIDUALS WITH TEMPORAL LOBE EPILEPSY USING NETWORK ANALYSIS OF RESTING-STATE FUNCTIONAL MRI
Abstract number :
2.243
Submission category :
5. Neuro Imaging
Year :
2014
Submission ID :
1868325
Source :
www.aesnet.org
Presentation date :
12/6/2014 12:00:00 AM
Published date :
Sep 29, 2014, 05:33 AM
Authors :
David Vaughan, Mangor Pedersen, Chris Tailby and Graeme Jackson
Rationale: Identification of abnormal brain networks is a critical step in the diagnosis and treatment of individuals with focal epilepsy. Here we apply multivariate pattern analysis (MVPA) to network maps extracted from resting-state fMRI, in patients with temporal lobe epilepsy (TLE). Methods: Patients with unilateral medically-refractory TLE were recruited from the Austin Hospital pre-surgical Epilepsy Program. The epileptic side was confirmed by video-EEG monitoring of seizures. 15 patients had TLE with hippocampal sclerosis (TLE-HS; 9 right sided epilepsy; 10 female; ages 25-64 years) and 15 patients had MRI-negative TLE (TLE-LN; 9 right; 6 female; ages 18-63 years). Comparison was made to 15 healthy non-epileptic controls (8 female; age range 15-52 years). Resting-state fMRI was acquired at 3T in wakefulness with eyes closed (3mm3 voxels, TE 30ms, TR 3000ms, 210 volumes). Images were registered to a symmetrical template, and flipped left-right to align the epileptic side. Physiological noise was removed using automated rejection of independent components, regression of spurious time courses, censoring of high-motion volumes, and band pass filtering 0.01-0.1Hz. Network characteristics were extracted for each individual. An un-weighted undirected graph was formed, with a node representing each voxel, and an edge if the Pearson correlation between each voxel pair, r > threshold > 0. Thresholds were set adaptively to obtain an edge density of 0.5%. Connectivity maps showing degree, clustering coefficient and eigenvector centrality at each node were extracted. We applied MVPA to these voxel-wise functional connectivity maps, using a linear support vector machine, to classify individuals as TLE-HS, TLE-LN or controls1. We quantified the overall accuracy of the classifier using leave-one-out cross-validation. Results: Classification of TLE-HS versus controls, considering all grey-matter voxels, performed best on eigenvector centrality maps (p=0.04, accuracy 70%, AUC 0.69). Similar results were obtained with clustering coefficient and degree maps. Voxels of greatest importance in this classifier were at the ipsilateral entorhinal cortex, hippocampus, amygdala, piriform cortex and cerebellum, plus bilaterally in the insula, thalami, and brainstem, driven by increased connectivity at these sites in TLE-HS. Restricting the classifier to a subset of voxels, identified on a traditional univariate t-test, improved performance to correctly label 12/15 controls and 11/15 TLE-HS patients (p=0.01, accuracy 77%, AUC 0.79). Comparison of TLE-LN to both TLE-HS and controls did not produce statistically significant classifiers. Conclusions: A network of increased functional connectivity distinguishes individuals with TLE-HS from healthy controls. This result highlights dysfunction of ipsilateral entorhinal cortex and mesial temporal structures, in addition to associated changes in piriform-insular-thalamic and cerebellar-brainstem networks.
Neuroimaging