1 Introduction

Brain-computer interfaces (BCIs) have been developed to improve man–machine interaction. These interfaces associate patterns found in brain signals with pre-determined tasks in a computational system [1]. Accomplishing these tasks allows the user to interact with the surrounding environment. Brain signals can be acquired with electroencephalogram (EEG) electrodes. The characteristics of the signals can then be determined. One of these characteristics is the steady-state visually evoked potential (SSVEP). This potential is an evoked response to a flickering visual stimulation, with the flickering frequency appearing in the recorded EEG signal.

Steady-state visually evoked potential can be easily measured through electrodes located at occipital and parietal areas of the scalp [2, 3]. However, there is no consensus regarding the exact area of the scalp where SSVEP is most pronounced. This information is quite important because it allows the number of electrodes used in an SSVEP-based BCI to be reduced and can improve the performance of evoked response detection.

Some works have discussed the optimal positions of electrodes for a BCI based on mental tasks [4] and emotional state recognition [5]. Some methods [6–8] have been developed to perform electrode selection for an SSVEP-based BCI, with the selection evaluated based on databases with various electrode positions and volunteers. The present study quantifies the contribution of twelve EEG electrodes, located in the occipital and parieto-occipital regions of the scalp, in SSVEP detection. The recorded signals are analyzed in monopolar and bipolar configurations. Bipolar settings are obtained from the difference of the electrical potential between two monopolar electrodes [9]. Four EEG signal processing methods are analyzed and applied to a database.

This work ranks EEG electrodes either in monopolar or bipolar configuration according to their relevance in evoked response detection. The final result is the selection of a group of electrodes with the best performance regardless of BCI user. The rest of this paper is organized as follows. The methodology for EEG signal recording is shown in Sect. 2. The four signal processing methods are presented in Sect. 3. The results are reported in Sect. 4 and discussed in Sect. 5. Finally, some conclusions are highlighted in Sect. 6.

2 Materials and Methods

Electroencephalogram signals were recorded using the BrainNet36 (EMSA, Sao Paulo/Brazil) signal acquisition system. The sampling frequency was 600 Hz. The twelve electrodes were located at positions P7, PO7, P5, PO3, POz, PO4, P6, PO8, P8, O1, O2, and Oz, respectively, of the international 10–20 system, as illustrated in Fig. 1. The reference electrode was placed on the left ear, with the ground electrode on the forehead. The stimuli were generated by an FPGA subsystem [3].

Fig. 1
figure 1

EEG electrodes used in signal acquisition experiments

Full size image

The experiments were performed according to the rules of the Ethics Committee of UFES/Brazil, under registration number CEP-048/08, and the Helsinki Declaration of 1975. The experiment was performed in offline mode. The volunteer sat on a comfortable chair at a distance of 70 cm from a 17” (43.18 cm) video monitor, on which four flickering checkerboard stripes were exhibited on the top, bottom, left, and right, respectively. EEG signals were acquired under an established acquisition protocol composed of one trial of 7 min. In this trial, four stimuli were shown simultaneously and the user gazed at them clockwise, in sequence, starting from the top one. The trial consisted of two repetitions of each stimulus for 20 s, then four repetitions of each stimulus for 10 s, and finally five repetitions with each stripe for 5 s, as illustrated in Fig. 2. A voiced alarm was used to advise the subject to observe a different stripe. The top, right, bottom, and left stimuli flickered at 5.6, 6.4, 6.9, and 8.0 reversals per second (rps), respectively [10].

Fig. 2
figure 2

Diagram for stimuli observation in second experiment

Full size image

Eighteen volunteers participated in the experiment, as shown in Table 1. Among them, eight volunteers had disabilities, diversifying the profile of volunteers.

Table 1 Information on volunteers
Full size table

3 Signal Processing

The goal of processing the EEG signal recorded in an SSVEP-based BCI is to identify the frequency of the stimulus gazed by the volunteer. Four research groups accomplished this identification using different signal processing techniques, with signals recorded under distinct experiments. To standardize the application of these four methods, the EEG signal recorded in the experiment in Sect. 2 was used. The four techniques are described below.

3.1 Independent Component Analysis

Wang et al. [8, 11] used independent component analysis (ICA) to separate the SSVEP response (evoked response) from noise. The correlation coefficients between the evoked response signal and the noise were determined to facilitate the selection of the most suitable electrodes. The goal was to choose a single bipolar configuration to detect SSVEP. In the present paper, this technique is used to rank all bipolar settings.

According to a previous study [8], the electrode with the highest SSVEP amplitude, called the signal electrode, is that located on the visual cortex. The study selected an appropriated electrode to be used as a reference in a single bipolar setting. The right choice of reference electrode can highlight the evoked response and reduce noise influence for having low SSVEP amplitude. Furthermore, the reference electrode registers background activities similar to those registered by the signal electrode, which helps reduce background noise. Therefore, electrodes next to the signal electrode with a low SSVEP amplitude can be the reference electrode. Following this idea, this work determined the optimal signal and reference electrodes by testing all possible bipolar settings using twelve EEG electrodes. A previously developed method [8] was used to rank the bipolar settings.

A previous method [8, 11] uses the first four independent components (ICs) to recover the evoked response, with the other components considered to be noise. The procedure adopted for the signals recorded in the two experiments in Sect. 2 comprises the following steps:

  1. (1)

    The signal recorded by twelve EEG electrodes at a certain frequency is used to create the X matrix. This matrix is determined separately for each stimulus frequency (5.6, 6.4, 6.9, and 8.0 rps). Then, 12 ICs are calculated as sources S, that is, S = WX, where W is the demixing matrix.

  2. (2)

    The four ICs with the highest normalized amplitudes at each stimulus frequency are considered as containing signal activities (Ssignal) and the other eight ICs are considered as containing noise activity (Snoise). The three closest points to the stimulus frequency (closest, anterior, and posterior points) are considered in the frequency spectrum.

  3. (3)

    Using the equation X = W−1S, the X matrix is divided in two parts: Xsignal = W−1Ssignal and Xnoise = W−1Snoise, where Xsignal and Xnoise are the reconstructions of the SSVEP and noise activities, respectively.

  4. (4)

    The correlation coefficients are calculated for Xsignal and Xnoise between different pairs of electrodes, i.e., \(\rho_{ij}^{signal}\) and \(\rho_{ij}^{noise}\), as demonstrated in a previous study [8].

According to previous studies [8, 11], the ratio of the signal and noise correlation can be used as the basis of optimal electrode selection. These works found that the most significant electrodes can be selected by observing only the spatial distribution of the amplitude spectra on the scalp over the visual cortex. For setting up a more systematic ranking of the bipolar combinations, two correlation matrices were created from the correlation coefficients. The lines of these matrices were listed in descendent order and scaled from 1 to 12, where 1 corresponds to the lowest value of the coefficient in the line under analysis.

From the decomposition process, a bipolar combination with low signal relation and high noise relation indicates a combination that presents a distinction between the evoked response in the two involved electrodes (high and low SSVEP activities) and which electrodes have similar background activities, respectively. Based on an idea proposed in a previous study [8], bipolar combinations with low values of \(\rho_{ij}^{signal}\) and high values of \(\rho_{ij}^{noise}\) can highlight the evoked response and reduce noise activities. Thus, bipolar settings with a signal correlation coefficient of between 1 and 6 and a noise correlation coefficient of between 7 and 12 were considered.

Wang et al. [8] considered the power spectrum in only the first harmonic of the stimulus frequency, whereas the present work considers the three first harmonics. This is justified since in our previous work [12] it was proven that evoked response detection improves when more harmonics are considered.

3.2 Lock-in Analyzer System

Electroencephalogram signal processing has been conducted using methods based on the lock-in analyzer system (LAS) [7, 12, 13]. The LAS is a tool for filtering small and repetitive signals inserted into noisy random signals or asynchronous signals of other frequencies [14]. The basic operation of a LAS is the multiplication of the EEG signal containing the SSVEP response by sine and cosine functions. These functions are called reference signals, and oscillate at the same frequency as that of the stimulus to be identified. The output of a LAS is a constant value that is proportional to the EEG signal synchronized with the reference signal. The phase between the two signals can also be determined.

For instance, if a BCI uses two stimulus frequencies, fT1 and fT2, and the EEG signal is recorded by two EEG electrodes, y1 and y2, the LAS would be composed as shown in Fig. 3.

Fig. 3
figure 3

LAS working with two stimulus frequencies and two EEG electrodes

Full size image

The number of LAS subsystems shown in Fig. 3 depends on the number of electrodes, the number of frequencies of stimulation, and the number of harmonics considered. In this work, 12 electrodes, 4 stimulus frequencies (5.6, 6.4, 6.9, and 8.0 rps), and three harmonics were considered. Thus, the number of LAS subsystems is 12 × 4 × 3 = 144. The output values of these systems are evaluated. The frequency with the highest output is chosen as the stimulus frequency detected by the BCI.

3.3 Multiple Channel Detection

The method presented in a previous study [6] is based on a technique called combined subspace detectors [15]. This detector is used to find distinct and known frequency components which compose the EEG signal containing the SSVEP response, as demonstrated in other studies [16, 17]. The multiple channel detection (MCD) method performs the evoked response detection using a combination of signals measured by different EEG electrodes. This combination is called a channel. A channel signal can be the same as an electrode signal but, in general, it consists of a set of electrode signals. The goal is to combine electrode signals into channel signals to increase the SSVEP response and cancel nuisance and noise signals.

Among the approaches used by Friman et al. [6], the most effective was the minimum energy combination. This method considers the EEG signal as being composed of three components: (i) the SSVEP response, which consists of sine functions with a frequency equal to the stimulus frequency and its harmonics; (ii) the nuisance signals, generated, for example, from external disturbances such as breathing artifacts and power-line interference; and (iii) the noise component. Power-line interference can be removed using a projection operation [6].

A channel signal is generated by a linear combination of EEG electrode signals using a weight vector. Several channels can be created from different combinations (different weights) of electrode signals, which compose a channel matrix. The optimal choice of the weight matrix depends on the signal nature and the nuisance and noise signals. The minimum energy combination method searches for combinations of electrode signals that cancel as much of the nuisance signals as possible. To achieve this, SSVEP components are removed from the electrode signals using a projection operation, with the resulting signal being just the nuisance signal. Then, the weight vector that minimizes the resulting energy of the electrode signal combination is found. The optimization problem is solved choosing the first channels with lower content of nuisance signals.

A statistical test is used to detect the SSVEP response. This test calculates the mean of the signal-to-noise ratio (SNR) for all harmonics and for all channel signals to determine how much greater the estimated SSVEP response is compared to the case where there is no visual stimulus. The stimulus frequency detected by the BCI is that with the highest value in this statistical test.

3.4 Spectral F Test for Phase-Locked Changes

The method developed in our previous work [18] for SSVEP response detection is based on a statistical test and a rule-based classifier. The statistical test, called spectral F test (SFT), uses the ratio between the power at a certain frequency and the average power in its neighbor frequencies [19]. This statistical test is used to determine whether the spectrum at a certain frequency is statistically distinct from its neighbors, considering that the spectrum in this neighborhood is white. Then, a null hypothesis is associated with the absence of an evoked response, which can be rejected from a critic value generated by an F-distribution. Therefore, the presence of an evoked response is considered for values in the spectrum above this critical value.

The critical value is independent of the length of the analyzed data segment, which leads to a robust threshold for identifying the evoked response. A rule-based classifier was developed to identify the stimulus frequency detected by the BCI. This classifier, a decision tree, has the frequencies rejected by the null hypothesis of the statistical test as its input, and uses the first three harmonics of the stimulus frequencies. The properties of the four methods used are summarized in Table 2.

Table 2 Information on four signal processing methods used in this study
Full size table

4 Results

This section presents the results of applying the four EEG signal processing techniques described in Sect. 3 to a signal database. A ranking of the twelve EEG electrodes in either monopolar or bipolar configuration is generated. A bipolar setting corresponds to the difference between the signal recorded by two electrodes. In addition, P7-O1 and O1-P7 settings, for instance, are considered similar.

In monopolar settings, the ranking is given by a line of twelve positions. The electrodes are scored from 1 to 12, where 12 corresponds to the position with the highest accuracy. In bipolar settings, the ranking is shown in an upper triangular 12 × 12 matrix with all 66 possible combinations \(\left( {\left( \begin{aligned} 2 \hfill \\ 12 \hfill \\ \end{aligned} \right) = \frac{12!}{{2!\left( {12 - 2} \right)!}} = 66} \right)\).

Each bipolar setting is scored from 1 to 66, where 66 corresponds to the most accurate combination in SSVEP response detection. Among the four methods, only the second and the fourth ones generated a ranking for monopolar settings, whereas the ranking for bipolar settings was obtained for all techniques, as shown in Table 3.

Table 3 Cases for each EEG signal processing method
Full size table

The entire signal recording was performed at once, to perform the signal processing for the first method. For the other three cases, the signal processing was applied incrementally. The method was applied to EEG signal segments of 2 s, with 50 % (1 s) overlap. These parameters are suitable for an online BCI [18]. The correct detection during a signal segment analysis contributes to increase the hit rate, in spite of the wrong detection. The hit rate is determined as the ratio of the number of correct detections to the total number of detections. The results for each method are given below.

4.1 ICA

ICA was used to decompose the signal and noise activities from the recorded EEG signal. Once this method allows only bipolar analysis, bipolar settings with a low signal correlation coefficient and a high noise correlation coefficient were selected. Thus, an upper triangular matrix was obtained considering three harmonics and four stimulus frequencies (5.6, 6.4, 6.9, and 8.0 rps), as shown in Table 4.

Table 4 Ranking of bipolar settings using ICA method considering three harmonics of stimulus frequency
Full size table

4.2 LAS

The recorded EEG signal contains the evoked response multiplied by the sine and cosine functions of frequencies equal the three first harmonics of the four stimulus frequencies (5.6, 6.4, 6.9, and 8.0 rps). Each frequency harmonic generates a lock-in system for each of the 12 EEG electrodes. The highest system output value among the frequencies and their harmonics determines the stimulus detected by the BCI.

The precision in SSVEP detection was determined for each stimulus frequency and the mean was taken among the volunteers for each EEG electrode (monopolar setting) and for each EEG electrode combination (bipolar setting). The ranking for the monopolar case is shown in Fig. 4, and the results for the bipolar setting are listed in Table 5.

Fig. 4
figure 4

Ranking of monopolar setting using LAS considering three harmonics of stimulus frequency. Higher scores correspond to higher precision in evoked response detection. The twelve best settings are highlighted

Full size image
Table 5 Ranking of bipolar combinations using LAS method considering three harmonics
Full size table

4.3 MCD

This method allows the evaluation of SSVEP response detection using sets of electrodes called channels. The lowest number of electrodes for a set is two, which corresponds to the bipolar setting used in the other methods. Thus, it is not possible to evaluate monopolar settings.

The precision of the evoked response detection was determined by the highest value from the statistical t test used in this method, as done in a previous study [6]. This technique was applied for recorded signals and three harmonics were considered. The results are shown in Table 6.

Table 6 Ranking of bipolar settings using MCD method considering three harmonics of stimulus frequency
Full size table

4.4 SFT

This method allows the evaluation of SSVEP response detection considering the three harmonics of the stimulus frequency due to the structure of the rule-based classifier used. The performance of both monopolar and bipolar settings can be evaluated. The classification process was applied individually for each electrode and bipolar combination.

The monopolar settings were scored from 1 to 12 whereas the bipolar settings were scored from 1 to 66, where a higher score corresponds to a higher precision in the evoked response detection. The results for monopolar settings are shown in Fig. 5, and the results for bipolar settings are listed in Table 7.

Fig. 5
figure 5

Ranking of monopolar settings using SFT method. Higher scores correspond to higher precision in evoked response detection

Full size image
Table 7 Ranking of bipolar settings using SFT method considering three harmonics of stimulus frequency
Full size table

5 Discussion

The techniques for EEG processing presented in Sect. 3 exhibit some similar characteristics. These methods all seek to highlight the SSVEP response (or evoked response) and simultaneously reduce noise activity. The analysis of monopolar settings is limited because the importance of each electrode is evaluated individually. However, the results of monopolar settings are important to corroborate the bipolar setting analysis. It is expected that the most suitable bipolar settings are those composed of two electrodes with high and low evoked responses, respectively, in the monopolar analysis. At the same time, these two electrodes should have similar background activities.

In monopolar analysis, the electrode ranking is obtained by summing the scores from Figs. 4 and 5, and ranking the scores from 1 to 12, as shown in Fig. 6.

Fig. 6
figure 6

Final ranking of EEG electrodes in monopolar setting considering three harmonics of stimulus frequency. Higher scores correspond to higher precision in evoked response detection

Full size image

The results in Fig. 6 show that the electrodes with more pronounced evoked responses are the three occipital electrodes, O1, O2, and Oz, followed by three parieto-occipital electrodes, PO3, PO4, and POz, as illustrated in Fig. 7. The six remaining electrodes have lower signal activity and, consequently, higher noise activity. The monopolar analysis shows that PO3, PO4, and POz electrodes have SSVEP activities less pronounced than those of O1, O2, and Oz electrodes but background activities similar to those of the three occipital electrodes.

Fig. 7
figure 7

Electrodes with highest SSVEP activity (highlighted) according to monopolar setting analysis

Full size image

All four methods were used to evaluate bipolar settings. Therefore, the scores from Tables 4, 5, 6, and 7, were summed. The scores were ranked from 1 to 66, as shown in Table 8.

Table 8 Final ranking of bipolar setting considering the first three harmonics of stimulus frequency
Full size table

6 Conclusion

This study used four EEG signal processing algorithms to rank EEG electrodes in monopolar and bipolar settings. 18 volunteers, 8 of whom had disabilities, participated in this study. The most suitable bipolar settings were determined. The O1–O2 bipolar setting is not the most suitable, in agreement with a previous study [7]. This combination is in the 10° position [7] (Table 5) and has the 39° position in the final ranking of this study (Table 8). The Oz electrode has an important role in an SSVEP-based BCI [7]. It is included in 6 of top 12 ranked bipolar settings and is the most suitable electrode in the monopolar analysis (Fig. 6). Although the POz electrode is not suitable in the final ranking of bipolar settings (Table 8), it participates significantly when the first method is used. This electrode is in 4 of the most suitable settings. This result agrees with a previous study [8] that uses this electrode in the optimal bipolar setting. The monopolar analysis is in agreement with the bipolar analysis since the first analysis indicates the most significant electrodes as participant in the most suitable settings of the second analysis. Considering the use of a few electrodes, it can be chosen some bipolar settings using the sixth most suitable monopolar settings. The results indicate that the bipolar settings POz-Oz, POz-O1, POZ-O2, PO3-Oz, and PO4-Oz are suitable for SSVEP-based BCIs.

This study quantified the contribution of twelve EEG electrodes in the detection of SSVEP response. The EEG signals from a database recorded for 18 volunteers (8 with disabilities) were analyzed in monopolar and bipolar settings using four techniques of signal processing. From the results, the occipital and parieto-occipital electrodes are the most suitable for SSVEP-based BCIs. More specifically, the five most suitable bipolar settings are POz-Oz, POz-O1, POz-O2, PO3-Oz, and PO4-Oz.