Abstracts

Rationale: Focal epilepsy disrupts normal brain networks. Resting state functional MRI (rs-fMRI) connectivity analyses have demonstrated changes in degree of connectivity at group level. There are no reliable methods to implement rs-fMRI data for individual patients in order to develop a clinical tool. Here we present an innovative machine learning method for identifying regions and patterns of rs-fMRI abnormalities in individuals with temporal lobe epilepsy (TLE). Methods: Anatomical T1-weighted image and resting state BOLD images from patients with temporal lobe epilepsy and controls were pre-processed using FreeSurfer. Surface-based anatomical atlas plus subcortical segmentation was used to create a 4D average functional time course in each parcel in brain's native space (42 parcels+segments per hemisphere). A support vector regression model (RSVM) derived from comparison group (contralateral TLE) was fitted to each patient data using Matlab. This is akin to multivariate regression but uses a machine learning approach to identify abnormalities in each parcel for single subjects. The model's predictive weight for each parcel (derived from the model's support vectors and its classification coefficient) were used to create 3D functional distinction maps (FDM) with and without normalization, then resampled into individual native brain space for visual inspection. Values were also extracted and differences between groups were statistically compared. Results: Data from 23 patients with left (LTLE) and 12 with right TLE (RTLE) were used. All patients were right-handed and left language dominant. For Each LTLE patient, RTLE group was used for model development. The resulting FDM included parcels in the left mesial and lateral temporal structures which showed high predictive weights in raw and normalized maps (figure 1). In analysis of asymmetry, cingulate and thalamic parcels showed the most common leftward shift, while lateral occipital and inferior temporal cortices had the most rightward shift. The laterality index across LTLE patients showed a wide range and a skewed distribution. For each RTLE patient fitted by LTLE-based comparison model, FDM showed larger changes in right mesial and lateral temporal, as well as medial frontal cortex. The entorhinal, inferior temporal, precentral, and precuneus parcels showed the most common leftward shift, while pars orbitalis and amygdala were most prominent on right hemisphere of RTLE patients.In group comparison of LTLE to RTLE, for left hemisphere the thalamus (weightx100 for LTLE -1.62, for RTLE +1.63), cingulate, lateral occipital, pars opercularis and postcentral gyrus showed the largest differences (all p<0.001). For the right hemisphere, inferior parietal (LTLE:-2.19; RTLE +1.69; p<0.001) and pars opercularis (LTLE: -1.53; RTLE: +.96; p<0.005) showed the largest differences in mean values between the two groups. Conclusions: Regions of distinctive functional connectivity can be mapped for individual patient using a RSVM machine learning approach and used to lateralize and localize the seizure focus. At group level, differences between RTLE and LTLE are demonstrated, in line with prior studies showing significant differences in rs-fMRI in regions distant from the seizure focus. The asymmetry and the predictive weight for each parcel is variable from patient to patient, and normal brain asymmetry might affect group level analysis. This variability can be used to identify individual phenotypes of focal epilepsy. Patient FDM, once validated on larger group of patients and controls and in prospective studies, has the potential to become a clinical imaging tool based on rs-fMRI. Funding: NIH NCATS UL1TR001876 / KL2TR001877 (Taha Gholipour)