Epileptic Seizure Prediction Based On Multivariate EEG Analysis

For a long time, epileptic seizure prediction is a clinical problem which needs to solve. Seizures forecast early can provide the patients and doctors time to take appropriate protective measures, preventing the accidents caused by sudden seizures onset.In this paper, we do the research on epileptic seizure prediction mainly via multivariate EEG analysis, including EEG feature extraction and classification. In respect of EEG feature extraction, in order to overcome the limitations of classical phase synchronization analysis, a novel multivariate EEG feature extraction method based on phase analysis, which is called HHT-ELM for short, is proposed according to the mechanism of seizures. HHT-ELM uses Hilbert-Huang transform (HHT) to calculate the phases of EEG, and it extracts the phase interaction information among all the EEG channels by extreme learning machine (ELM) from the point of view of model identification. In respect of EEG classification, on the one hand, in order to make the sensitivity and false prediction rate both good at the same time, a probabilistic discriminative extreme learning machine (PDELM) is proposed. PDELM has a probability output rather than a simple class label, which is more in line with the medical habits. On the other hand, in order to improve the stability of epileptic seizure prediction system, a novel ensemble extreme learning machine (EELM) is proposed. In EELM, ensemble learning strategy is introduced to overcome the drawbacks of randomness and poor stability of machine learning. In addition to the above two respects, a dynamic update framework is also proposed, which makes the epileptic seizure prediction system being more flexible and adaptive to the patients’physiological status. Under the framework, the prediction model can be closer to the patients’real conditions, providing a theoretical reference for the development of portable seizures predictors. The results show that: HHT-ELM can extract effective EEG features to detect the seizures upcoming; PDELM and EELM can improve the performance of the epileptic seizure prediction system from different points; and the proposed dynamic update framework can make the prediction system being adaptive to patients.