Nonlinear dimension reduction for EEG-based epileptic seizure detection

Abstract

Approximately 0.1 percent of epileptic patients die from unexpected deaths. In general, for intractable seizures, it is crucial to have an algorithm to accurately and automatically detect the seizures and notify care-givers to assist patients. EEG signals are known as definitive diagnosis of seizure events. In this work, we utilize the frequency domain features (normalized in-band power spectral density) for the EEG channels. We applied a nonlinear data-embedding technique based on stochastic neighbor distance metric to capture the relationships among data elements in high dimension and improve the accuracy of seizure detection. This proposed data embedding technique not only makes it possible to visualize data in two or three dimensions, but also tackles the inherent difficulties regarding high dimensional data classification such as time complexity and memory requirement. We also applied a patient specific KNN classification to detect seizure and non-seizure events. The results indicate that our nonlinear technique provides significantly better visualization and classification efficiency (F-measure greater than 87%) compared to conventional dimension reduction approaches.