Entropy-based nonlinear analysis for electrophysiological recordings of brain activity in Alzheimer’s disease
Item statusRestricted Access
Embargo end date04/07/2019
Alzheimer’s disease (AD) is a neurodegenerative disorder in which the death of brain cells causes memory loss and cognitive decline. As AD progresses, changes in the electrophysiological brain activity take place. Such changes can be recorded by the electroencephalography (EEG) and magnetoencephalography (MEG) techniques. These are the only two neurophysiologic approaches able to directly measure the activity of the brain cortex. Since EEGs and MEGs are considered as the outputs of a nonlinear system (i.e., brain), there has been an interest in nonlinear methods for the analysis of EEGs and MEGs. One of the most powerful nonlinear metrics used to assess the dynamical characteristics of signals is that of entropy. The aim of this thesis is to develop entropy-based approaches for characterization of EEGs and MEGs paying close attention to AD. Recent developments in the field of entropy for the characterization of physiological signals have tried: 1) to improve the stability and reliability of entropy-based results for short and long signals; and 2) to extend the univariate entropy methods to their multivariate cases to be able to reveal the patterns across channels. To enhance the stability of entropy-based values for short univariate signals, refined composite multiscale fuzzy entropy (MFE - RCMFE) is developed. To decrease the running time and increase the stability of the existing multivariate MFE (mvMFE) while keeping its benefits, the refined composite mvMFE (RCmvMFE) with a new fuzzy membership function is developed here as well. In spite of the interesting results obtained by these improvements, fuzzy entropy (FuzEn), RCMFE, and RCmvMFE may still lead to unreliable results for short signals and are not fast enough for real-time applications. To address these shortcomings, dispersion entropy (DispEn) and frequency-based DispEn (FDispEn), which are based on our introduced dispersion patterns and the Shannon’s definition of entropy, are developed. The computational cost of DispEn and FDispEn is O(N) – where N is the signal length –, compared with the O(N2) for popular sample entropy (SampEn) and FuzEn. DispEn and FDispEn also overcome the problem of equal values for embedded vectors and discarding some information with regard to the signal amplitudes encountered in permutation entropy (PerEn). Moreover, unlike PerEn, DispEn and FDispEn are relatively insensitive to noise. As extensions of our developed DispEn, multiscale DispEn (MDE) and multivariate MDE (mvMDE) are introduced to quantify the complexity of univariate and multivariate signals, respectively. MDE and mvMDE have the following advantages over the existing univariate and multivariate multiscale methods: 1) they are noticeably faster; 2) MDE and mvMDE result in smaller coefficient of variations for synthetic and real signals showing more stable profiles; 3) they better distinguish various states of biomedical signals; 4) MDE and mvMDE do not result in undefined values for short time series; and 5) mvMDE, compared with multivariate multiscale SampEn (mvMSE) and mvMFE, needs to store a considerably smaller number of elements. In this Thesis, two restating-state electrophysiological datasets related to AD are analyzed: 1) 148-channel MEGs recorded from 62 subjects (36 AD patients vs. 26 age-matched controls); and 2) 16-channel EEGs recorded from 22 subjects (11 AD patients vs. 11 age-matched controls). The results obtained by MDE and mvMDE suggest that the controls’ signals are more and less complex at respectively short (scales between 1 to 4) and longer (scales between 5 to 12) scale factors than AD patients’ recordings for both the EEG and MEG datasets. The p-values based on Mann-Whitney U-test for AD patients vs. controls show that the MDE and mvMDE, compared with the existing complexity techniques, significantly discriminate the controls from subjects with AD at a larger number of scale factors for both the EEG and MEG datasets. Moreover, the smallest p-values are achieved by MDE (e.g., 0.0010 and 0.0181 for respectively MDE and MFE using EEG dataset) and mvMDE (e.g., 0.0086 and 0.2372 for respectively mvMDE and mvMFE using EEG dataset) for both the EEG and MEG datasets, illustrating the superiority of these developed entropy-based techniques over the state-of-the-art univariate and multivariate entropy approaches. Overall, the introduced FDispEn, DispEn, MDE, and mvMDE methods are expected to be useful for the analysis of physiological signals due to their ability to distinguish different types of time series with a low computation time.