Content area
Abstract
A reliable method of seizure detection and prediction in people with epilepsy would provide a key strategy for assessing risk of sudden unexpected death in epilepsy (SUDEP) in those that suffer from uncontrollable seizures, and would guide research in the development of preventative interventions. This research proposes objective indications of seizure onset observed from electroencephalogram (EEG). The algorithms utilize scalp EEG that is minimally invasive and has shown promise for high sensitivity and specificity in seizure event forewarning. This approach considers the brain as a nonlinear dynamical system whose state can be derived through time delay embedding of the time-serial EEG data and characterized to determine change in brain dynamics related to the preictal state. This method targets specific phase-space graph properties as biomarkers for seizure detection and prediction. The data analysis approach efficiently processes individual windows of data, correlates historical degrees of change in the brain state from repeated measurements, and makes accurate forewarning of seizure onset. Specifically, we contribute to the field in the following ways:
Seizure Prediction. We provide three novel techniques for predicting seizures prior to onset: 1. Phase-Space Adjacency Spectrum: This method targets the spectrum of phase-space graph adjacency matrices as a biomarker for seizure prediction. The best results corresponded to an accuracy of 97% (58/60), a sensitivity of 100% (40/40), and a specificity of 90% (18/20) on training data. In out of sample testing, this method achieved an accuracy of 75%, specificity of 70% (7/10), and sensitivity of 80% (8/10). After minor adjustments to a single parameter, the out of sample test achieved an overall accuracy of 90% (18/20). 2. Phase-Space Laplacian Spectrum: This method targets the spectrum of phase-space graph laplacian matrices as a biomarker for seizure prediction. The best results corresponded to an accuracy of 93% (56/60), a sensitivity of 93% (37/40), and a specificity of 95% (19/20). Out of sample testing resulted in a specificity of 60% (6/10) and sensitivity of 70% (7/10). After minor adjustments to a single parameter, the out of sample test achieved an overall accuracy of 80%. 3. Hypergraph Analysis of Phase-Space Graphs: This method analyzes subsets of the edge set of phase-space graphs identified as hyperedges of a hypergraph as biomarkers for seizure prediction. The hypergraph is represented with a matrix which captures the essential structure and connectivity of the graph. The spectral features of the triangle matrix are used to predict seizure onset. This method yields a training accuracy of 93% and testing accuracy of 80%.
Seizure Detection. We provide a novel technique for a patient specific seizure detection algorithm. This method combines phase-space analysis and deep learning using convolutional neural networks (CNN) to indicate seizure onset. The output of the CNN is filtered using a combination of exponential time decay and a sliding mean window. This method achieved a 100% true positive and 99% true negative rate on four patients.
