Classification of epileptic and non-epileptic events using tensor decomposition

Abstract

Misdiagnosis of epilepsy, even by experienced clinicians, can cause exposure of patients to medical procedures and treatments with potential complications. Moreover, diagnostic delays (for 7 to 10 years on average) impose economic burden at individual and population levels. In this paper, a seizure classification framework of epileptic and non-epileptic events from multi-channel EEG data is proposed. In contrast to relevant studies found in the literature, in this study, the non-epileptic class consists of two types of paroxysmal episodes of loss of consciousness, namely the psychogenic non-epileptic seizure (PNES) and the vasovagal syncope (VVS). EEG signals are represented in the spectral-spatial-temporal domain. A tensor-based approach is employed to extract signature features to feed the classification models. TUCKER decomposition is applied to learn the essence of original, high-dimensional domain of feature space and extract a multilinear discriminative subspace. The classification models were evaluated on EEG epochs from 11 subjects in an inter-subject cross-validation setting and achieved an accuracy of 96%.