Automated detection of dynamical change in EEG signals based on a new rhythm measure

https://doi.org/10.1016/j.artmed.2020.101920Get rights and content

Highlights

  • An effective and accurate EEG rhythm indicator extracted through graph-based modeling method.

  • A new method for monitoring changes in EEG signals which combines extracted EEG rhythm indicators and null hypothesis testing.

  • Promising results conducted on both simulated signals and real signals.

Abstract

Automated detection of dynamical change in EEG signals has been a long-standing problem in a wide range of clinic applications. It is essential to extract an effective and accurate EEG rhythm indicator that can reflect the dynamical behavior of a given EEG signal. Time-frequency analysis is a promising method to achieve this end, but existing methods still have limitations in real implementation making this kind of methods still progressive until the present day. In this paper, along the line of ongoing research on time-frequency methods, we present a new method based on graph-based modeling. By virtue of this method, an effective and accurate EEG rhythm indicator can be extracted to characterize the dynamical EEG time series. Together with the extracted EEG rhythm indicator, an automatic analysis of continuous monitoring of EEG signal, is developed by means of a null hypothesis testing to inspect whether an EEG change occurs or not during a monitoring period. The proposed framework is applied to both simulated data and real signals respectively to validate its effectiveness. Experimental results, together with theoretical interpretation and discussions, suggest its promising potentials in practice.

Introduction

Automated detection of dynamical changes (e.g., anomalies, switching and transit points) in Electroencephalogram (EEG) signals has attracted increasing attention in the field of EEG signal processing over past years (see [1], [2], [3], [4] for examples). This technique provides a preliminary analysis for EEG signals, which can be integrated with further analysis/methods for particular clinical applications. For example, Gajic et al. proposed a new technique for detection of abnormal activity in a given EEG signals [5], and as such the detected suspicious EEG segments is fed to an already-trained classifier for epilepsy identification [6]. A similar study can be also found in a recent article [7], which presented a unify framework for epileptic seizure detection and epilepsy diagnosis. In order to confirm the eye state, Saghafi et al. firstly used an automated logistic regression-based method to describe a possible change in EEG signal, and then fed the detected EEG segment into artificial neural network for classification [8].

EEG change detection can be regarded a novelty detection task which recognizes new inputs that differ in some way from the past that are usual under normal states [2], [4], [9]. The most challenge is to automatically extract an effective and accurate rhythm/waveform indicator from highly-noisy EEG signal that can reflect its dynamical behaviors [10], [11]. See Fig. 1. Once the rhythm indicator is extracted, one can model the EEG data, typically using a Gaussian model, such that potential changes could be detected via statistical analysis of input parameters with respect to the model. Here, it is worth noting that, although collected EEG signals often include multiple channels, this paper is focused on the analysis of single channel EEG data as it provides a preliminary basis for analyzing multi-channel data [12], [13], [14]. Therefore, additional discussions on multi-channel fusion/combination are not provided in this paper. The interested reader can refer to [15], [16], [17] for this issue.

EEG rhythm extraction is directed at processing time-domain EEG data collected from the monitored subject to produce an rhythm/waveform indicator over the previous time period. Current methods include time-domain, frequency-domain and time-frequency domain methods. Time-domain methods exploit the natural characteristics of EEG signal in the time domain [18], [19], and normally include mean, median, variation, kurtosis, skewness, etc. A summarization of them can be found in [20]. These features can effectively characterize the dynamical EEG status with an assumption of stationary signals. However, the actual EEG signal is normally considered as complex and non-stationary, and meanwhile often submerged by the strong noises caused by human brain activities [21], [22]. These situations bring uncertainties in statistics such as the mean and median, thus yielding inaccuracy in those methods that only utilize time-domain information for EEG rhythm extraction.

Frequency-domain methods, typically the Fourier-based analysis, might not be feasible for dynamic characterization of EEG state, due to the high amplitude fluctuations and short duration of irregular EEG signals at an incipient stage [23], [24]. Another limitation of such techniques is the insufficient capacity of dealing with non-stationary signals [25]. It is therefore difficult to detect and identify EEG changes directly from the Fourier spectrum especially considering that the changing information is often diluted across the whole basis.

By noticing the limitations of both time-domain and frequency-domain methods, there is an increasing interest on integrated time-frequency domain methods recently as seen from [26], [27], [28]. Representative time-frequency domain approaches include short-time Fourier transform (STFT), empirical model decomposition (EMD), continuous wavelet transform (CWT), discrete wavelet transform (DWT), etc. A comprehensive review of them can be found in [29]. Although many successful attempts have been reported based on these methods, it should be noted that existing integrated approaches still have limitations in practice as discussed in [30]. Continuous efforts have been therefore being devoted to this kind of approaches with various enhancements until the present day.

This paper aims at the extraction of EEG rhythm indicator, and presents a new and universal framework for EEG change detection. Graph, as a kind of spectrum analysis algorithm, can capture the structural and topological information hidden in the data [31]. Meanwhile, it is also robust to the noise. By the use of graph modeling, the given EEG signal can be converted from its raw space into graph space, such that the dynamical characteristics of EEG signal can be extracted. The graph has become an important tool in the analysis of EEG signals as seen in [12], [20], [32]. However, existing methods commonly construct the graph in the time domain. This paper suggests the correlation information, hidden in the frequency domain, is also an important indication for EEG state. To achieve this end, a new graph modeling method is developed in this paper, by which an effective and accurate EEG rhythm indicator can be extracted. Finally, together with the indicator, an automatic analysis of continuous monitoring of EEG signal by means of a null hypothesis testing, is developed to inspect whether an EEG change occurs or not. The framework is applied to simulated and real scenarios, respectively, to validate its effectiveness. Experimental results, together with theoretical interpretation and discussion, suggest its promising potentials in real usages.

Main contributions of this study are summarized as below.

  • Exploration and exploitation of a novel graph-based data modeling for automated extraction of EEG rhythm;

  • New framework for automated EEG change detection based on graph modeling coupled with a null hypothesis testing.

The rest of this paper is structured as follows. Section 2 provides details of graph-based EEG rhythm extraction. Section 3 describes a statistical analysis for decision making. Section 4 gives the algorithm of the proposed method. Section 5 shows experimental results. Conclusions are drawn finally in Section 6.

Section snippets

EEG rhythm extraction based on graph modeling

In the context of EEG signal processing, it is a common practice to extract explanatory parameters from raw data. A well-established methodology is first to confirm a specific sub-band under inspection, and subsequently calculates its statistics to extract a rhythm indicator. An effective and accurate indicator provides a precise understanding of the dynamic behavior of EEG time series, thus allowing for the detection and location of possible EEG changes. If a change happens, the extracted

Hypothesis testing for decision making

Detection of the change from EEG signals is much challenging mostly because the EEG signals are non-linear and non-stationary, and meanwhile immersed in noise [21], [22]. To adopt a threshold-based decision making is straightforward, and alternatively statistical methods are often used which can be grouped into statistical learning methods and statistical modeling methods [5], [8], [9]. The statistical learning methods rely heavily on the suspicious histories of the EEG signals and are suitable

Algorithm

In summary, the algorithm of the proposed EEG change detection framework is given in Algorithm 1. It operates continuously and periodically, such that the collected EEG data can be inspected with an on-line and self-conducted manner. This process will continue until a pre-defined stopping time or received a stopping order from user.

Algorithm 1

Algorithm of the proposed framework.

Step 1: Collect EEG data continuously from a monitored individual;
Step 2: Filter the raw data with a pre-defined sub-band pass

Experiment

In this section, the proposed method is validated on two different data sets respectively. It is also compared with competitors using typical time-domain features including mean, root mean square (RMS), kurtosis and skewness, as well as representative time-frequency domain features including STFT, EMD, CWT and DWT. We use the same 3σ criterion to make change decision for all methods for a fair comparison purpose. As illustrated in Fig. 7, the experiment is conducted based on the following steps:

Conclusion

The objective of this paper is to design an automated system that can detect the change in a given EEG time series. To achieve this end, a novel EEG data modeling is proposed based on graph modeling. The method has been validated based on two data sets. Comparison with state-of-the-art methods demonstrated its outperforming performance, suggesting its great potential in real applications.

The proposed method provides a preliminary analysis for EEG signals, i.e., it performs the detection of

Acknowledgments

This study is supported by the Shandong Provincial Natural Science Foundation, China (ZR2019MEE063) and the Fundamental Research Funds of Shandong University (2018JC010).

References (43)

  • S.W. Yahaya et al.

    A consensus novelty detection ensemble approach for anomaly detection in activities of daily living

    Appl Soft Comput

    (2019)
  • M. Winterhalder et al.

    Spatio-temporal patient-individual assessment of synchronization changes for epileptic seizure prediction

    Clin Neurophysiol

    (2006)
  • M.F. Mohamed Saaid et al.

    Change point detection of EEG signals based on particle swarm optimization

    5th Kuala Lumpur international conference on biomedical engineering 2011

    (2011)
  • F.E. Gunawan et al.

    Detecting the early drop of attention using EEG signal.

  • Z. Gao et al.

    Automatic change detection for real-time monitoring of EEG signals

    Front Physiol

    (2018)
  • D. Gajic et al.

    Detection of epileptiform activity in EEG signals based on time-frequency and non-linear analysis

    Front Comput Neurosci

    (2015)
  • D. Gajic et al.

    Classification of EEG signals for detection of epileptic seizures based on wavelets and statistical pattern recognition

    Biomed Eng: Appl Basis Commun

    (2014)
  • Z. Chen et al.

    A unified framework and method for EEG-based early epileptic seizure detection and epilepsy diagnosis

    IEEE Access

    (2020)
  • M. Ermes et al.

    Method for the automatic detection of epileptiform waveforms in sevoflurane-induced anesthesia EEG

  • P. Singh et al.

    Fourier-based feature extraction for classification of EEG signals using EEG rhythms

    Circuits Syst Signal Process

    (2016)
  • T. Tanaka et al.

    Rhythmic component extraction for multi-channel EEG data analysis.

  • Cited by (6)

    • Sleep stage classification for child patients using DeConvolutional Neural Network

      2020, Artificial Intelligence in Medicine
      Citation Excerpt :

      Many researchers exploit EEG to monitor changes in the sleeper's brain waves to predict sleep conditions. Depending on frequency analysis, PSG recordings exhibit different band characteristics, e.g., EEG data can be decomposed into functionally distinct frequency subbands, such as delta (0.5–3.9 Hz), theta (4–7.9 Hz), alpha (8–12 Hz) and beta (12–30 Hz) [4,5,22–33]. In other studies, the use of the EOG signal provides information about eye movements during sleep, chin EMG and ECG signals can track chin muscle movements and heartbeats changes, respectively [15,17,18,34].

    • Sensor-Based Sleep Stage Classification Using Deep Learning

      2023, Sensor-Based Sleep Stage Classification Using Deep Learning
    View full text