A0341
Title: Spectral functional classification for time series with an application to epileptic detection
Authors: Kun Chen - Southwestern University of Finance and Economics (China) [presenting]
Rui Huang - Nanjing University (China)
Chun Yip Yau - Chinese University of Hong Kong (Hong Kong)
Xingzuo He - Southwestern University of Finance and Economics (China)
Abstract: Statistical diagnosis of epileptic seizures based on electroencephalogram (EEG) signals is often challenging to implement due to time dependency and the possible existence of measurement errors. By reframing the seizure detection as a classification problem for stationary time series processes, a novel frequency domain functional approach is proposed to simultaneously account for time dependency and noise reduction. The procedure is based on the spectral theory of time series and identifies series whose spectral density shares similar shapes or oscillations. The main tool of the proposed method is the logarithm of the periodogram, which serves as an asymptotically consistent estimator for the logarithm of the spectral density. By employing the functional principal component analysis combined with a distance-based rule, similarities between smoothed log-periodogram functions and, thus, spectral density functions in logarithms can be identified. Under mild conditions, it is theoretically demonstrated that the misclassification rates tend to be zero. Based on extensive simulations in various scenarios and real applications to EEG data of epilepsy patients, the efficacy of the proposed method is proven.
