Classification of biological signals using linear and nonlinear features.

This paper investigates the characterization ability of linear and nonlinear features and proposes combining such features in order to improve the classification of biological signals, in particular single-trial electroencephalogram (EEG) and electrocardiogram (ECG) data. For this purpose, three data sets composed of ECG, epileptic EEG and finger-movement EEG were utilized. The characterization ability of seven nonlinear features namely the approximate entropy, largest Lyapunov exponents, correlation dimension, nonlinear prediction error, Hurst exponent, higher order autocovariance and asymmetry due to time reversal are compared with two linear features namely the autoregressive (AR) reflection coefficients and AR model coefficients. The features were tested by their ability to differentiate between different classes of data using a linear discriminant analysis (LDA) method with tenfold cross-validation. The class separability of combined linear and nonlinear features was assessed using sequential floating forward search with linear discriminant analysis method (SFFS-LDA). The results demonstrated that linear and nonlinear features on their own provided comparable results for the ECG data set and the finger-movement EEG data set whilst the linear features provided a better class separability compared to nonlinear features for the epileptic EEG data set. Combining linear and nonlinear features demonstrated a significant improvement in the class separability for all of the data sets where an average improvement of 20.56% was obtained with the ECG data set, 7.45% with finger-movement data set and 6.62% with the epileptic EEG data set. Overall results suggest that the use of combined linear and nonlinear feature sets would be a better approach for the characterization and classification of biological signals such as EEG and ECG.