Abstracts

Rationale: Surgical resection of seizure generating tissue gives patients with drug resistant focal epilepsy the best chance for seizure freedom. However, even after resection, many patients suffer seizure recurrence, likely due to residual pathological tissue remaining after the surgery. Seizure localization model created by machine learning algorithms utilizing information from multiple diagnostic modalities could improve delineation of epileptogenic tissue and provide further guidance for resective surgery. Methods: Thirty patients underwent implantation of intracranial electrodes as part of their evaluation for drug resistant focal epilepsy and had excellent outcome after epilepsy surgery (Engel I, 1 year follow up). Intracranial EEG (iEEG) sampled at 5 kHz were used for automated detection (detector available at – http//:msel.mayo.edu) of high frequency oscillation (HFO) in randomly selected continuous 2-hour segments of interictal iEEG (patients hospitalized at ICU). Fusion of pre-implantation MRI scans with post-implantation CT scans were used to localize electrodes. Surgical records and post-operative MRI were used to determine whether the contacts were implanted in the resected or non-resected tissue. First, per-electrode iEEG features (mean HFO counts and amplitudes) and contact type (depth vs. subdural) were used, second, MNI coordinates of electrodes were added as features (subgroup of 28 patients), third, MNI coordinates and lesions from MRI were added as extra features (subgroup of 24 patients). Selected features entered a linear kernel support vector machine (SVM) classifier to classify tissue in place of electrode into epileptogenic or non-epileptogenic. A leave-one-out cross validation was performed to train and test SVM. Data of one patient were left out for testing, while SVM model was trained on remaining patients’ data. The probability estimate of correct class assignment was used to produce receiver operating curves and their corresponding areas under the curve (AUC). Statistical significance of AUC differences from chance (AUC=0.5) were tested by Hanley-McNeil test. Results: SVM classification using iEEG data and electrode type resulted in AUC=0.62, with optimal point at 50.6% sensitivity, 73.7% specificity and AUC significantly greater than random AUC (p < 0.0001). Addition of electrode MNI coordinates as a feature yielded AUC of 0.67, with optimal point at 65.4% sensitivity and 68.7% specificity. Finally, the addition of MRI-identifiable lesion resulted in AUC=0.71, with optimal point at 64.9% sensitivity and 70.7% specificity. Conclusions: This study showed that model trained by single machine learning approach using features from several diagnostic modalities to classify epileptogenicity of brain tissue that surround electrode contacts improves significantly and gradually with addition of each relevant non-iEEG feature. This shows a promise for future guidance of pathological tissue resection. Funding: Supported by project no. LH15047 (MEYS CR, KONTAKT II), project no. LQ1605 (MEYS CR, NPU II) and NIH R01 NS092882-03.