Abstracts

Rationale: Developing a quantitative algorithm for predicting seizure outcome (SO) following anterior temporal lobectomy (ATL) would constitute a significant advance for presurgical decision making of treating temporal lobe epilepsy (TLE). In this project, we tested the ability of topographic properties extracted from presurgical resting state functional connectivity (rsFC) to predict SO, using two separate machine learning classification methods (support vector machine, SVM, and random forest, RF). Methods: Fifty-six TLE patients (LTLE: 26; RTLE: 30) underwent a 5 minute rsfMRI scan prior to ATL. Based on their Engel Class (Engel, et al., 1993) at 1 year follow up they were divided into good (Class I) and poor outcome (Class II-V) groups (GO=35; PO=21). Groups were matched on 9 clinical characteristics: age, gender, handedness, laterality of TLE, epilepsy onset age and duration, seizure focality, interictal-spike type and the presence of hippocampal sclerosis (HS). After standard post-processing, we computed a rsFC cross-correlation matrix on 90 cerebral nodes, flipping right-to-left the matrices of the RTLE to maintain access to ictal versus non-ictal hemispheres. Five topographic properties were estimated (Rubinov, et al., 2010) from absolute binary matrices across a series of network densities (5~50%): global efficiency (Eglob), global clustering coefficient (CCglob), degree centrality (DC), betweenness centrality (BC) and eigenvector centrality (EC). Averaged over all the densities, this produced 272 variables [(2 (Eglobal, CCglob) + 3 (DC, BC, EC))?-90 nodes]. In addition, nine of the aforementioned clinical variables were utilized, yielding a total of 281 variables. For SVM, all the variables were first passed to an Elastic Net regularization method to select the optimal subset of variables. These variables were then used to train the SVM model. For RF, a stepwise variable selection was carried out. At beginning, all the variables were used to build a primary model and estimate the importance of each variable and the model's out of bag (OOB) error, completed for 100 cycles. Next, based on the order of average importance, the 10% least important variables were dropped out, with the remainder used to re-build and re-estimate an RF model, run 100 cycles at each step. The model with the lowest mean OOB error was deemed best. The prediction power of both SVM and RF models were tested by 3 cross validation methods: leave one out, split sample (1000 permutations), and 7-fold (1000 permutations). Results: The Elastic Net regularization selected DC and EC of bilateral thalamus as the optimal variables. The SVM classifier trained with these variables provided an average prediction accuracy of 75.51% for SO. The stepwise RF produced a model with DC of the non-ictal thalamus and EC of the non-ictal fusiform as the optimum variables. This model provided an average prediction accuracy of 78.80% for SO. None of the clinical variables noted above were selected as optimum by either method. When we trained the SVM and RF models with only these clinical variables, the average prediction accuracy values dropped (SVM=53.41%; RF=66.13%; see details in Table 1). Conclusions: In summary, these data show that, indeed, machine learning methods utilizing pre-surgical rsFC properties can reliably predict SO. The SVM and RF methods yield comparable prediction accuracy, with rsFC topographic measures clearly providing stronger predictive value than the standard clinical measures commonly used in surgical centers. References: Engel, et al. Outcome with respect to epileptic seizures. In Engel (Ed), Raven Press: New York; 1993:pp 609-621. Rubinov, et al. Complex network measures of brain connectivity: uses and interpretations. Neuroimage 2010;52:1059-1069. Funding: N.A.