Improved motor imagery brain-computer interface performance via adaptive modulation filtering and two-stage classification

https://doi.org/10.1016/j.bspc.2019.101812Get rights and content

Highlights

  • EEG signal enhancement with novel modulation filtering technique.

  • Feature extraction with overlapping frequency bands.

  • Two-stage classification scheme for final four-class decision.

  • Outperformed the first-place winner of the BCI competition IV by 3.5 %.

Abstract

Electroencephalogram (EEG) based brain-computer interfaces (BCI) monitor neural activity and translate these signals into actions and/or decisions, with the final goal of enabling users to interact with a computer using only their thoughts. To this end, users must produce specific neural activity patterns that are used by the system as control signals. A common task used to elicit such signals is motor imagery (MI), where specific patterns are elicited in the sensorimotor cortex during imagination of movements (e.g., of the hands, arms, feet or tongue). The processing pipeline typically used in EEG-BCIs consists of three stages: pre-processing, feature extraction, and classification. Here, we propose innovations in pre-processing and classification and quantify the gains achieved on 4-class MI-based BCI performance. More specifically, for the pre-processing stage, we propose the concept of spectro-temporal filtering as we show that MI-elicited neural patterns have varying amplitude modulation variations relative to artifacts. For the classification stage, in turn, a two-step classification method is proposed. First, LDA classifiers are used to discriminate between different pair-wise MI tasks. Next, a naive Bayes classifier is used to predict the final task performed by the user based on the weighted outputs of the LDA classifiers. Experimental results showed that the proposed system outperformed the first-place winner of the BCI competition IV by 3.5 %.

Introduction

An EEG-based brain-computer interface (BCI) is a communication system that captures the electrical signals emitted by the cerebral cortex and translates these signals into actions and/or decisions, with the final goal of enabling the user to interact with a given equipment without using his/her muscles [1]. To operate a BCI, the user must produce different patterns of brain activity that will be identified and used by the system as control signals [2]. Motor imagery is one of the most widely used BCI paradigms and comprises of the use of imagined movements to elicit neural activity in the sensorimotor cortex, just as observed when one performs the movement [3]. The typical signal processing pipeline in current BCI systems consists of three stages: pre-processing, feature extraction, and classification [4].

Pre-processing aims at simplifying subsequent processing operations without losing relevant information. An important function of pre-processing is to improve signal quality by increasing the signal-to-noise ratio (SNR). A low SNR means that the brain patterns are buried in the signal (e.g. background EEG), which makes relevant patterns hard to detect [1]. A high SNR, on the other hand, tends to improve BCI classification accuracy. It is known that EEG signals are affected by various types of artifacts, such as power line interference, involuntary movements, muscular contractions, amongst others. For artifact removal, several manual, semi-automated or fully-automated methods have been proposed, such as time-frequency filtering, independent component analysis (ICA) and others [3]. Considering that EEG signals present non-stationary behavior, i.e., their spectral content changes over time, we explore the use of a recent technique called modulation filtering (also called spectro-temporal filtering) as a new and fully-automated EEG artifact removal method [5]. Modulation filtering is based on amplitude modulation analysis and requires the application of two transformation mappings: in the first, the EEG signal is transformed into a time-frequency representation (e.g., via short-time Fourier transforms or wavelet transforms). A second transform is then taken across the time dimension to result in a frequency-frequency representation termed “modulation spectrogram” [5]. Modulation filtering has been successfully employed in the past to blindly separate heart and lung sounds from breathing sound signals [6], to enhance noisy speech for hearing devices [7] and to enhance noisy electrocardiograms (ECG) [8]. To the best of our knowledge, this is the first time it is being explored within an EEG-based MI BCI paradigm [9].

In order to extract features for classification of motor-imagery tasks, we used the common spatial pattern (CSP), one of the most used and more effective methods for MI BCI. This technique provides spatial filters that allow to separate two conditions by maximizing differences in variance between them. Since variance of band-pass filtered signals is equivalent to band-power, CSP filters are well suited to discriminate mental states that are characterized by motor sensory rhythm effects [10]. Although the CSP algorithm is very efficient, it is noise sensitive for limited datasets [11]. To address such drawback, several CSP variants have been proposed to make it more robust. The interested reader is referred to [10] for more details. Here, we employed CSP combined with Tikhonov regularization [11].

For classification, we proposed a novel two-stage design: the well-known and widely used linear discriminant analysis (LDA) algorithm [12] was used first as a binary classifier to generate weighed outputs of the six possible pairwise combinations among four motor-imagery tasks (right hand, left hand, foot and tongue) performed by the subjects. LDA classifier outputs were then directed as inputs to a naive Bayes classifier, which made the final four-class decision.

The remainder of this article is organized as follows. Section 2 describes the methods and materials, including the modulation representation, standard database used, feature extraction, classifiers, and performance figure of merit. Sections 3 and 4, in turn, present the experimental results and discusses them, respectively. Lastly, conclusions are drawn in Section 5.

Section snippets

Amplitude modulation analysis

A non-stationary signal, such as the EEG, can be modeled as the result of the interaction between two independent signals: a low-frequency signal (modulator) that changes or modulates the amplitude of a higher frequency (carrier) signal [5,7]. This corresponds to the well-known concept of amplitude modulation. In amplitude modulation analysis, a signal x(t) can be expressed as the product of a low-frequency modulating signal m(t), and a high-frequency carrier signal, c(t):xt=mtc(t),assuming:

Results

Each of the nine subjects has his/her own two-stage classifier trained with 288 trials from each of the four motor-imagery tasks. Each subject’s pair of classifiers (LDA followed by Naive Bayes) was then tested with other 288 trials (per MI task) from the same subject, recorded in a different day, as already pointed out in the previous section. In Table 2 we present the results of our three experiments, compared to the results obtained in BCIC IV. All experiments focus on EEG signal enhancement

Discussion

To the best of our knowledge, our reported classification results on the four tasks of BCIC IV dataset 2a were only slightly surpassed by three recent studies. In 2017, Davoudi, Ghidary and Sadatnejad [20] used dimensionality reduction based on distance preservation to local mean (DPLM) and got a global kappa of 0.6 (averaged over the nine subjects). Last year (2018), Gaur et al. achieved the same kappa (0.6) through multivariate empirical mode decomposition based filtering and Riemannian

Conclusions

Conventional EEG signal pre-processing approaches for BCI are based on frequency or spatial filtering techniques [29]. In order to increase the classification performance of motor-imagery tasks, we investigated a novel filtering technique, which is performed in the modulation frequency domain, to improve EEG SNR. Visual inspection of the modulation spectrogram was carried out to delimitate the modulation filtering regions, applying to EEG a similar method successfully used by Tobon and Falk for

Acknowledgments

This work was supported by the São Paulo Research Foundation (FAPESP, grant #2017/15243-7) through equipment and software and by the Brazilian Government through a graduate-student scholarship from the National Council for the Improvement of Higher Education (CAPES).

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (29)

  • D.P. Tobon et al.

    Adaptive spectro-temporal filtering for electrocardiogram signal enhancement

    IEEE J. Biomed. Health Inform.

    (2018)
  • D. Wang et al.

    Multi-class motor imagery EEG decoding for brain-computer interfaces

    Front. Neurosci.

    (2012)
  • F. Lotte et al.

    Regularizing common spatial patterns to improve BCI designs: unified theory and new algorithms

    IEEE Trans. Biomed. Eng.

    (2011)
  • A. Meinel

    Tikhonov Regularization Enhances EEG-Based Spatial Filtering for Single Trial Regression to Cite this Version: HAL Id: hal-01655755

    (2017)
  • Cited by (20)

    • Comparative analysis of spectral and temporal combinations in CSP-based methods for decoding hand motor imagery tasks

      2022, Journal of Neuroscience Methods
      Citation Excerpt :

      Thus, algorithms for optimizing a spectral filter along with the spatial filter are required. In literature, some methods have been proposed to target such problem, including filter bank common spatial pattern (FBCSP) (Ang et al., 2008), common spatio-spectral pattern (CSSP) (Thomas et al., 2017),(Yuksel and Olmez, 2016), sub-band common spatial patterns (SBCSP) (Novi et al., 2007), among others (Zahid et al., 2020; Zhang et al., 2015; Kumar et al., 2019; Xygonakis et al., 2018; Dos Santos et al., 2020). In the filter bank common spatial patterns (FBCSP) and sub-band common spatial patterns (SBCSP) methods, a filter bank is implemented for decomposing the EEG signal into multiple bands; after that, a spatial filtering using CSP is carried out for each band (Ang et al., 2008); Novi et al., 2007).

    • A new approach to feature extraction in MI-based BCI systems

      2022, Artificial Intelligence-Based Brain-Computer Interface
    View all citing articles on Scopus
    View full text