Abstracts

Rationale: Responsive electrical neurostimulation is an emerging therapy for epilepsy. This technology works by detecting that a seizure is imminent or has just begun, and then electrically stimulating the brain to prevent or abort the seizure. Today’s commercially available devices are effective at reducing seizure frequency but are very unlikely to produce seizure freedom, which is critical to significantly improving patient quality of life. Current devices use very simple algorithms for seizure anticipation/detection in order to minimize device energy demands. They could likely be more effective with more sophisticated detectors. The goal of this project was to develop a next-generation seizure anticipation algorithm that uses more sophisticated pattern classification techniques, while still being simple enough to meet the physical constraints of an implantable device. In contrast to much previous research in this field, the algorithm was also designed to work on any patient, without the need of patient-specific training data provided by a patient’s physician. Methods: We trained a support vector machine classifier on archived intracranial electroencephalographic (iEEG) seizure recordings from 14 patients (10/4 training/test split). The input features consisted of bandpassed signal magnitude in conventional frequency bands, with five different degrees of temporal smoothing. The classifier was trained to detect any time point between -4 and +9 seconds of the clinician defined seizure onset. Results: Algorithm balanced accuracy performance was 80% on the test data. All seizures were detected within 3 seconds of onset on average but with frequent false positives. Conclusions: Our current algorithm achieves excellent sensitivity but with mediocre specificity. Future work will attempt to improve specificity by adding additional features and training on more data. However the current success rate demonstrates that there are enough regularities in seizure onset signatures across patients for a low-complexity, seizure anticipation algorithm to be feasible. Moreover, we speculate that such an algorithm could potentially be useful in identifying the seizure onset zone in iEEG recordings from epilepsy surgery candidates. Funding: This research was supported by funding from the Ontario Brain Institute.