In the dissertation we seek to develop and validate reliable frameworks for human epileptic seizure prediction with electrocorticogram (ECoG) and intracranial electroencephalogram (iEEG). The long-term goal of the research is to develop and prototype an implantable device that can reliably provide alarms prior to a seizure in real-time. The specific objective is to develop a patient-specific algorithm that can predict seizures in ECoG/iEEG with high sensitivity and low false positive rate as well as low complexity. This dissertation starts by demonstrating that seizures can be predicted with linear features of spectral power, and it ultimately focuses on developing a reduced-complexity algorithm that can decode ECoG/iEEG for human epileptic seizure prediction with high sensitivity and acceptable low false positive rate. By contrast to prior prediction work, most of which focused on nonlinear measurements, we demonstrate that human epileptic seizures can be predicted with linear features of ECoG/iEEG in machine learning classification approach.
To begin with, a new patient-specific seizure prediction algorithm with ECoG/iEEG is proposed. It is novel in sense that it employs a set of linear features of spectral power from ECoG/iEEG for prediction and that predictive models are established and tested using cost-sensitive support vector machines (SVMs) using double cross-validation method. The proposed algorithm is tested over 433.2 hours of interictal recordings including 80 seizure events from 18 human epileptics in the Freiburg EEG database. It achieves high sensitivity of 97.5% (78/80), a low false alarm rate of 0.27 per hour (total 117 FPs), and total false prediction times of 13.0% (56.4-hour). Bipolar and/or time-differential preprocessing improves sensitivity and false positive rate.
For the seizure prediction algorithm to be practically feasible on an implantable device, we further propose a reduced-complexity prediction algorithm. We lower the complexity of the algorithm by investigating and using small numbers of essential features and by replacing nonlinear SVMs and the Kalman filter with linear SVMs and moving-average filters. The key features are determined using the RFE SVM (recursive feature elimination using SVMs). The proposed reduced-complexity algorithm significantly lowers the predictor's complexity and thus the power consumption, while producing high sensitivity as well as reasonable false positives. It is tested on 9 subjects selected from the Freiburg database that result in high prediction rate when the initial prediction algorithm is applied, and successfully demonstrates high sensitivity of 100.0% (38/38) as well as low false positive rate of 0.15 per hour (total 32 FPs) and false positive portion of 9.65% (21.0-hour) in the 217.5-hour interictal recordings with the selected six time-differential features. It has been observed that time-differential preprocessing improves the prediction rate significantly.
Additionally, we develop an enhanced approach for seizure onset and offset detection in rats' ECoG. This is an improved version of the automatic seizure detection and termination system in in-vivo rats' ECoG. We improve the system by using a specific frequency range of 14-22Hz, which has been observed to be more relevant to seizure onsets than other bands; by using spectral power instead of spectral amplitudes as a feature set; and by substituting the 2-point moving average filter with the Kalman filter. Furthermore, while the proposed algorithm provides better detection statistics, it also lowers the system's complexity by removing the fast Fourier transform computation and keeping a single structure though the proposed algorithm uses the two different spectral features for detecting onsets and offsets.