Spatial Feature Reduction in Long-term EEG for Patient-specific Epileptic Seizure Event Detection

Abstract

Seizure is a common phenomenon among patients with epilepsy. Detection of seizure at the onset enables immediate actions to be taken. However, real-time seizure event and onset detection in an ambulatory epilepsy patient using continuous scalp EEG monitoring proves to be difficult with the hardware associated with signal acquisition. Furthermore, developing a universal seizure detection algorithm is not applicable due to the patient dependence. The prominent challenge in working with EEG its non-stationarity. Hence, time-frequency methods such as wavelet analysis have been used in recent research, to develop algorithms. In this study, we develop patient specific seizure detection system in which, Discrete Wavelet Packet Transform (DWPT) is used to extract wavelet packet energy and use a single EEG channel to train an offline SVM classifier. This study is conducted to prove that accurate seizure prediction can be performed using at most two channels for most patients and develop a system for real-time monitoring and seizure event detection. The objective is to maintain a high level of accuracy despite the drastic reduction in spatial features and achieve a low latency of detection so that this algorithm could be used in seizure alert devices with less hardware complexity. CHB-MIT Scalp EEG database has been used in this study and an average sensitivity of detection of 90.3% an average specificity of 99% with an average latency of 3.41s have been achieved for 60% of the epileptic seizure subjects considered.