Social network theory applied to resting-state fMRI connectivity data in the analysis of epilepsy networks

Abstract

Rationale: Epilepsy networks tend to be persistently abnormal, and should therefore be defined by the extent and strength of their components in the interictal state. Based on this hypothesis, we used a combination of functional MRI connectivity data and social network analysis to determine if network properties could identify differences between controls and MTL patients. If so this would provide motivation for further assessment of network theory in this context. Resting-state fMRI connectivity studies may be helpful in localizing abnormal networks. For this work the network properties of various anatomically defined regions of interest in the interictal state were examined. The abnormal ROIs found either in individual patient or group level compared with normal controls could potentially provide insight into the understanding of the epilepsy networks and ultimately better identify targets for surgical intervention. Methods: Applying social network theory methodology, we looked for abnormal network properties at the group level. A classifier algorithm was tested to determine if resting-state network data allowed separation of medial temporal lobe epilepsy (MTLE) patients from normal control subjects. Five social network properties were applied to brain data including: Degree, Strength, Closeness, Clustering coefficient and Betweenness centrality, were selected for brain network analysis. FMRI Data of 52 control subjects and 16 patients who suffered from intractable MTLE were imaged on a 3T Siemens Trio scanner at the Yale MRRC. Resting state functional data was obtained using a gradient echo T2*-weighted EPI sequence with TR=1550ms, TE=30ms, flip angle=80, FOV=22 22cm, matrix size 64x64, 25 slices, functional voxel size 3.4mm 3.4mm 6mm. 3-8 runs of resting state data were collected with 229 volumes per run. 36 ROIs were defined in MNI space for analysis of network properties and this included a number of limbic regions. The five local network properties obtained from 36 anatomically defined volumes of interest (ROIs) were served as features to a classifier. Results: Significant (p<0.05) abnormal network properties for a number of ROIs were identified in the patient group. A feature selection strategy was proposed to further improve the classification accuracy. An average sensitivity of 77.2% and specificity of 83.86% were achieved via 'leave one out' cross validation. Conclusions: The finding of significantly abnormal ROIs in group level data confirms our initial hypothesis that network properties measured from resting-state fMRI data may reveal abnormal nodes involved in the epileptogenic network. Social network theory provides measures such as Degree, Strength, Closeness, Clustering coefficient and Betweenness centrality, that appear to serve as efficient features that can distinguish healthy volunteers from MTLE patients, even when these measures are based on data collected in the interictal state.