Unbiased estimation of permutation entropy in EEG analysis for Alzheimer's disease classification

https://doi.org/10.1016/j.bspc.2017.08.012Get rights and content

Highlights

  • The EEG signal of Alzheimer's diseased and healthy patients is studied via permutation entropy.

  • The technique enables to use bigger window size and study influence of EEG sampling frequency.

  • Statistically significant decreasing of permutation entropy is observed for Alzheimer's disease.

Abstract

The EEG signal of healthy patient can be recognized as an output of a chaotic system. There are many measures of chaotic behaviour: Hurst and Lyapunov exponents, various dimensions of attractor, various entropy measures, etc. We prefer permutation entropy of equidistantly sampled data. The novelty of our approach is in bias reduction of permutation entropy estimates, memory decrease, and time complexities of permutation analysis. Therefore, we are not limited by the EEG signal and permutation sample lengths. This general method was used for channel by channel analysis of Alzheimer's diseased (AD) and healthy (CN) patients to point out the differences between AD and CN groups. Our technique also enables to study the influence of EEG sampling frequency in a wide range. The best results were obtained for sampling frequency 200 Hz, using at most window of length 10. In the case of Alzheimer's disease, we observed a statistically significant decrease in permutation entropy at all channels.

Introduction

The diagnosis of Alzheimer's disease (AD) is an up-to-date problem which is solved by various techniques. Non-invasive investigation of AD patients is preferred and both magnetic resonance imaging (MRI) and electroencephalography (EEG) are frequently used for image and signal processing related to the diagnosis. Being focused on EEG analysis, spectral analysis is one of the successful tools for AD investigation; however, the analysis of non-linear EEG dynamic offers a more complex signal analysis. Correlation dimension D2 is a good example of complexity measure offering lower values for AD patients [1]. The same trend is seen for the largest Lyapunov exponent λ. Unfortunately, the calculation of D2 and λ1 enforces the analysis of a very long time series [2]. In this study, we prefer entropy estimation from the EEG signal [3], [4] because of the relationship between the complexity of non-linear signals [5], [6] and entropy estimate. Dauwels et al. [7] and many other authors have shown that Alzheimer's disease increases power in the delta and theta-bands in frequency domain, but the power spectrum is a global characteristics of the EEG signal making it impossible to study and localise events in the signal. There are many possibilities how to organize the entropy evaluation. Our approach is based on permutation entropy [8], [9]. There are three reasons why to use permutation entropy for diagnosing AD. Human brain activity can be interpreted as behaviour of a complex chaotic system. Therefore, signals from EEG electrodes carry information about the chaos inside the scalp. The hypothesis of decreasing chaotic behaviour [3] during AD can be tested by using permutation entropy; the second reason for using it is its relative simplicity. This term is exactly defined for the windowed signal and is directly applicable to time series without any methodological difficulties. The time complexity of permutation entropy calculation is acceptable for small windows lengths as discussed in [8], [9]. However, the application to windows of size above 12 is impossible. Permutation entropy has not yet been applied to longer windows. We suppose there is a chance to obtain new and significant results if the window length were prolonged. Using hash table as a special data structure, it is possible to calculate permutation entropy for window lengths up to 30, which is the main novelty of our approach, as will be explained in following sections.

Section snippets

Shannon entropy and its estimation

Definition. Shannon entropy [3], [10] HS of a discrete random variable X with possible values x1, …, xm and probability mass function p(X) is defined asHS=i=1mpilnpi,where pi = p(xi).

If the probability function is unknown for an experimental data set, and the number of possible values is finite for random variable X, we estimate probability function pi by relative frequency pj,N and number of events kN aspj,N=njn,kN=nj>01k,where nj is the number of occurrences xi of random variable X, and n

Permutation analysis for large samples

The main disadvantage of the original procedure of permutation analysis [8] is in its memory and time complexities. The authors implemented permutation memory as a matrix of w columns and w! rows together with counter vector of length w!. This enables permutation analysis on a typical computer only for w<13. Traditional applications [8] of permutation entropy use a window of length w<8. The time complexity of single permutation counting is also w!, in the worst case. Therefore, for permutation

Application to EEG

Our study provided the EEG data of two patient groups: 10 patients with moderate dementia (MMSE score 10–19) and control group of 10 healthy patients having no memory or other cognitive impairments. The average MMSE was 16.2 (SD of 2.1) for the AD group and 30 for normals, of course. The age profile was similar: 69.4 ± 9.2 for AD and 68.7 ± 7.7 for CN. The AD group consisted of 5 men and 5 women, the CN group of 4 men and 6 women. All subjects were recorded under the same conditions, i.e. lying

Discussion

From the biomedical point of view, the results of statistical testing can be interpreted as follows. The chaotic behaviour of normal human brain is observable by using the standard EEG 10–20 system for sampling frequencies up to 200 Hz at any channel. The differences in permutation entropy between men and women are not statistically significant but the permutation entropy significantly decreases in the case of Alzheimer's disease. Therefore, the permutation entropy on any electrode signal is a

Conclusion

Using Miller approach instead of direct calculation of Shannon entropy from permutation frequencies, we have developed a novel method of EEG analysis via permutation entropy. The advantage of our method is in its very fast permutation analysis and low consumption of computer memory which enables an analysis of large time series with greater length of permutation patterns as well as parametric study of the sampling frequency role. The method was used to diagnose Alzheimer's disease from

Acknowledgement

This work was supported by the Czech Technical University in Prague, grant SGS14/208/OHK4/3T/14.

References (14)

  • S. Tang et al.

    Entropy analysis in interacting diffusion systems on complex networks

    Int. J. Math. Comput. Simul.

    (2011)
  • F.C. Morabito et al.

    Multivariate multi-scale permutation entropy for complexity analysis of Alzheimer's

    Entropy

    (2012)
  • D.E. Abasolo et al.

    Approximate entropy of EEG background activity in Alzheimer's disease patients

    Intell. Autom. Soft Comput.

    (2009)
  • D. Labate et al.

    Entropic measures of EEG complexity in Alzheimer's disease through a multivariate multiscale approach

    IEEE Sens. J.

    (2013)
  • S.M. Pincus

    Approximate entropy as a measure of system complexity

    Proc. Natl. Acad. Sci. U. S. A.

    (1991)
  • J.-H. Park et al.

    Multiscale entropy analysis of EEG from patients under different pathological conditions

    Fractals

    (2007)
  • J. Dauwels et al.

    Diagnosis of Alzheimer's disease from EEG signals: where are we standing?

    Curr. Alzheimer Res.

    (2010)
There are more references available in the full text version of this article.

Cited by (41)

  • Early detection of Alzheimer's disease from EEG signals using Hjorth parameters

    2021, Biomedical Signal Processing and Control
    Citation Excerpt :

    At this phase, it is also possible to reduce the number of features by a feature selection or feature reduction method, in order to reduce the feature space dimensions, to make the features independent and to reduce the computation volume. In recent years, a variety of linear and nonlinear features have been used to diagnose Alzheimer's disease, including mean, variance, skewness, kurtosis, individual alpha frequency, interquartile range, spectral centroid, entropy, spectral entropy, sample entropy, fuzzy entropy, auto-mutual information, Shannon entropy, multiscale entropy, approximate entropy, permutation entropy, Tsallis entropy, Lempel–Ziv complexity, box dimension, Katz’s fractal dimension, Higuchi’s fractal dimension, and Hurst exponent [1,3,5,7–13]. Hjorth parameters are a set of nonlinear features based on statistical calculations that have been used in many studies to effectively describe the characteristics of the EEG signal [14].

  • A new approach for automatic detection of focal EEG signals using wavelet packet decomposition and quad binary pattern method

    2021, Biomedical Signal Processing and Control
    Citation Excerpt :

    In PermEn, the neighbouring points are compared to compute the complexity of the EEG signals [25]. It is widely being used in the analysis of EEG signals for epileptic detection [26,27] and Alzheimer detection [28]. In FuzzyEn, the similarity of the time-series signal is measured based on the fuzzy function.

View all citing articles on Scopus
View full text