Abstracts

Rationale: In patients with intractable epilepsy, predicting seizures with clinically acceptable levels of sensitivity and specificity opens the doors to a variety of therapeutic solutions. This goal however, is not reached and clinically acceptable results have yet to be demonstrated. Many studies showed that preictal changes are detectable from the EEG using different measures and that seizure prediction is possible. When tested on long-term recordings, the performance consistency of the prediction algorithms and their superiority to random predictors remain however questionable. In this study, an intracranial EEG-based seizure prediction method is proposed. It is evaluated on patients with temporal lobe epilepsy (TLE). Methods: The method is based on the study by Gadhoumi et al. (Clin Neurophysiol 2012). 1565 hours of continuous intracranial EEG data (filtered at 500Hz and sampled at 2000Hz) from 17 patients with TLE were investigated. The recordings included 127 seizures. Bipolar channels from the 4 deepest contacts of bilaterally implanted electrodes in the amygdala, hippocampus and parahippocampus were analysed. In each patient the data was split into a training set and a testing set. During training, the EEG segments were analysed in 4 frequency bands (between 50Hz and 450Hz) using continuous wavelet transform. A reference state was defined in the 90s immediate preictal data and used to derive 3 features quantifying the discrimination between preictal and interictal data. A discriminant analysis based classifier was then trained in the feature space. Its performance was evaluated using cross validation applied to training data. The 3 channels and the frequency band yielding the best classification performance were selected for the testing procedure. The performance of the classifier is assessed for a range of Seizure Prediction Horizons (SPH) between 5 and 60 min and compared with a random Poisson predictor for statistical validation. Results: For SPHs between 30 and 60 min, better than random prediction performance is achieved in 7 of 17 patients (fig. 1). For this set of SPHs and for these patients, the sensitivity is higher than 85%, the false prediction rate is less than 0.1/h and the portion of time under false alarm is less than 9.2%. The median prediction time is ~ 45 min (fig. 2). EEG channels selected in the training phase as best performing channels were mostly bilateral. They were not all located in the hemisphere were neurologist indicated a seizure onset predominance. Most of these patients had bilateral epilepsy however. Conclusions: A seizure prediction method has been developed and evaluated on patients with TLE. Seizures were predictable above chance (p < 0.05) in 41% of patients. For these patients, the method yields sensitivity and specificity levels potentially interesting for applications of a closed-loop seizure control where intervention horizons above 30 min are desirable. The study shows that it may be possible to find a subset of patients in whom seizures can be reliably predicted. Supported by CIHR MOP-10189, RSC-NSERC CHRPJ 323490-06.