Abstracts

Rationale: Anterior temporal lobectomy (ATL) delivers seizure freedom in ~70% of the patients with temporal lobe epilepsy (TLE) 1. While insights about surgical outcome can be obtained from the presurgical MRI 2, knowledge regarding the postoperative brain organization, i.e., the actual substrates governing either seizure control or recurrence, may provide added value to the outcome prediction. Here we characterize the mesoscale functional organization of the brain through longitudinal resting-state fMRI (rsfMRI), and test whether its change pre- to post-surgery predicts surgical outcome. Methods: Forty-three TLE patients underwent two 5-min rsfMRI scans prior and ~1 year after their ATLs. Patients were assigned into seizure free (Engel I, L/R=13/15), and seizure recurrence (Engel II~IV, L/R=7/8) groups based on their surgical outcome at the time of the second scan. After standard post-processing, we estimated average BOLD time-series from 200 cortical and 15 subcortical regions, and built a functional correlation matrix for each scan and each patient. Regions inside the resection lacuna were retained in the postoperative scan matrix but had no functional connections to the rest of the network. We then stacked all 86 matrices and applied a multilayer community detection algorithm 3 to assign each region to different functional modules at each layer. By comparing the modular structure of each scan with known resting state networks (RSN), we characterized the probability of inter- vs. intra- RSN communication with two previously developed statistics: integration and recruitment 4 (Fig. 1). Results: Via paired t-tests (FDR-corrected for multiple comparisons), we found altered communication patterns in multiple RSNs postoperatively. The dorsal attention, limbic, fronto-parietal control, default mode, and subcortical RSNs presented reduced integration; the limbic, control, and subcortical RSNs presented reduced recruitment (Ps<0.05, Fig2AC). We also observed significant session by outcome interactions at limbic, control, and subcortical RSNs for integration (Fs>4.33, Ps<0.04, Fig2B) but not recruitment (Ps>0.05). Interestingly, random forest (RF) models using presurgical RSN integration values as predictors can only predict outcome with 60.5% accuracy while models built with RSN integration reorganization values (post–pre) can predict outcome with 72.1% accuracy (using leave-one-out cross-validation). In practice, although the second model can be trained retrospectively, the reorganization of the testing sample, e.g., a new patient, is not available before the surgery. To address this point, we built another RF model to predict the reorganization for the testing sample first, and then used the predicted reorganization values as input for the second model. Compared to the first model, this solution still produced better performance, reaching an accuracy of 65.1%. Conclusions: Our results suggest that a broad reorganization of the brain’s modular network architecture takes place after ATL. Specifically, selective reorganization of inter-network communication is related to surgical outcome, which may partially explain the obstacles to obtain optimal prediction with pure presurgical data. As a proof of concept, we demonstrate the added value of such reorganization patterns in outcome prediction, and provide a potential solution to overcome the absence of such information prospectively.References:1. de Tisi, J et al. (2011). Lancet 378, 1388–952. He, X et al. (2017). Neurology 88, 2285–933. Mucha, P et al. (2010). Science. 328, 876–84. Bassett, D et al. (2015). Nat. Neurosci. 18, 744–51 Funding: Postdoctoral Fellowship from the American Epilepsy Society Lennox & Lombroso Trust for Research & TrainingNINDS R01-NS099348-01NIMH R01-MH104606