Implementing a fuzzy inference system in a multi-objective EEG channel selection model for imagined speech classification

Highlights

• It was searched the minimal subset of channels for imagined speech.

• Channel selection was approached as multi-objective to obtain a Pareto front.

• A fuzzy system inference was applied to find a promising solution from Pareto front.

• Channel selection had a statistically similar performance to the use of all channels.

• It was observed a dependence between features and classes of imagined speech.

Abstract

One of the main purposes of brain-computer interfaces (BCI) is to provide persons of an alternative communication channel. This objective was firstly focused on handicapped subjects but nowadays its scope has increased to healthy persons. Usually, BCIs record brain activity using electroencephalograms (EEG), according to four main neuro-paradigms (slow cortical potentials, motor imagery, P300 component and visual evoked potentials). These analytical paradigms are not intuitive and are difficult to implement. Accordingly, this work researches an alternative neuro-paradigm called imagined speech, which refers to the internal pronunciation of words without emitting sounds or doing facial movements. Specifically, the present research is focused on the recognition of five Spanish words corresponding to the English words “up,” “down,” “left,” “right” and “select”, with which a computer cursor could be controlled. We perform an offline computer automatic classification procedure of a dataset of EEG signals from 27 subjects. The method implements a channel selection composed of two stages; the first one obtains a Pareto front and is approached as a multi-objective optimization problem dealing with the error rate and the number of channels; the second stage selects a single solution (channel combination) from the front, applying a fuzzy inference system (FIS). We assess the method’s performance through a channel combination and a test set not used to generate the front. Several FIS configurations were explored to evaluate if a FIS is able to select channel combinations that improve or, at least, keep the obtained accuracies using all channels for each subject’s data. We found that a FIS configuration, FIS3×3 (three membership functions for both input variables: error rate and the number of channels), obtained the best trade-off between the number of fuzzy rules and its accuracy (68.18% using around 7 channels). Also, the FIS3×3 obtained a similar statistically accuracy compared to the use of all channels (70.33%). Results of our method demonstrate the feasibility of using a FIS to automatically select a solution from the Pareto front to select channels applied to a problem of imagined speech classification. The presented method outperforms previous works in accuracy and showed a dependence relationship between EEG data and imagined words.

Introduction

A BCI system tries to provide a new channel to the brain to transmit messages and commands to the external world (Wolpaw, Birbaumer, McFarland, Pfurtscheller, & Vaughan, 2002). In general, a BCI can be seen as a pattern recognition system where EEG is utilized as the primary source of raw information and machine-learning algorithms are employed to learn an inference function from EEG signals; after which the algorithms can decode the EEG signals into commands to be executed by an electro-mechanic device. Some applications of BCI research are targeted to aid persons with severe motor disabilities. However, there are other relevant applications that have been explored as, the semi-autonomous car driving (Göhring, Latotzky, Wang, & Rojas, 2013), the robot-assisted surgery, the military applications (like the U.S. Army project “Silent Talk” (Bogue, 2010)), the improvement of human-computer interaction (for example, in video game control), among others.

Based on the BCI definition by Wolpaw et al. (2002), the following question arises: “How can a BCI user generate the messages and commands to be transmitted to the exterior world?” The answer may lie in the neurological mechanisms or processes employed by the user to generate the control signals, called electrophysiological sources (neuroparadigms). The most used are: slow cortical potentials (SCP), P300 potentials, motor imagery (MI), and visual evoked potentials (VEP) (Bashashati, Fatourechi, Ward, Birch, 2007, Brumberg, Nieto-Castanon, Kennedy, Guenther, 2010, Wolpaw, Birbaumer, McFarland, Pfurtscheller, Vaughan, 2002). However, these electrophysiological sources depend on external stimuli (VEP and P300), or they are low (SCP). Even MI is not easy to generate due to it is related to the intention of movement, which is usually performed unconsciously during movement preparation (Jeannerod et al., 1994). This results in excessive training time needed by a person to use a BCI and low communication rates (Brumberg et al., 2010). The works, described in Brigham and Kumar (2010); DaSalla, Kambara, Sato, and Koike (2009); DÆZmura, Deng, Lappas, Thorpe, and Srinivasan (2009); Gonzalez-Castañeda (2015); Porbadnigk (2008); Suppes, Lu, and Han (1997); Torres-García, Reyes-García, Villaseñor-Pineda, 2012, Torres-García, Reyes-García, Villaseñor-Pineda, 2013; Wang, Zhang, Zhong, and Zhang (2013); Wester and Schultz (2006), have explored the use of imagined speech for trying to solve these problems.

According to Wester and Schultz (2006), imagined speech -sometimes called internal, inner or unspoken speech- refers to the action of internally pronouncing a word without moving any muscles or emitting any sounds. In this case, it is possible to record electroencephalograms (EEG) during a person’s imagined speech. Thus, the interpretation of EEG recorded during imagined speech is also of interest for the development of silent speech interfaces (SSI). According to Denby et al. (2010), SSIs are desirable when a subject cannot emit an intelligible sound. For example, when a person is handicapped, needs to keep a message private, or he/she is in a noisy environment.

Despite the complexity of the imagined speech recognition, a BCI/SSI based on this neuroparadigm offers a new kind of communication that opens many possibilities. Nevertheless, there are some challenges that avoid the use of these systems in real-life applications. An important issue is that several of the available algorithms are focused on analyzing and processing the information of multi-channel EEG (Wang, Gao, & Gao, 2006). Hence, the following question arises: How can the multi-channel nature of EEG be efficiently exploited?

Motor-imagery-based BCIs have followed three approaches for trying to solve the above problem. They are feature; selection, spatial filtering and channel selection. The feature selection approach aims to automatically select a subset of relevant features resulting in a dimensionality reduction. Whereas, the spatial filtering approach is followed on those methods that combine (generally, a weighted linear combination) several channels into a single one from which features will be extracted. Finally, channel selection approach is employed on those methods to automatically select a subset of relevant channels (electrodes).

Channel selection approach is relatively recent and aims to identify a more interpretable form of reducing the number of channels needed to achieve the same accuracy as a full channel configuration. This helps to reduce the processing time of a BCI system because a lower number of channels generates a lower amount of information to be learned for any BCI computational model.

According to Lal et al. (2004), the channel locations are known for the recording of motor imagery. However, these are unknown for other electrophysiological sources. This is the case of imagined speech where, although the major brain activity is present in the left hemisphere for most people, there is evidence that this activity could be located in another part of the brain depending on the person. Hence, channel selection methods may help in extracting the most relevant information from the EEG signals recorded during imagined speech.

Our research is focused on imagined speech because it is an electrophysiological source that allows an alternative natural communication channel (Brumberg et al., 2010). Specifically, the objective of this work is to develop a method for channel selection and clasification of EEG signals recorded during imagined speech.

Following, the remaining sections are briefly described. Section 2 provides a description of similar works. The main components of the proposed method to solve the problem are described in Section 3. Section 4 presents the dataset used, experiments done to assess the performance of our method, and a discussion about the results obtained. Finally, the main conclusions of this work, the managerial insights based on our experimental outcomes, a brief description of the strengths and weakness of the proposed method, a comparison regarding previous works and future work that could help to improve the performance of the proposed solution are presented in Section 5.