A MOBILE EEG SYSTEM FOR PRACTICAL APPLICATIONS. Sciences, Beijing , China

Transcription

1 A MOBILE EEG SYSTEM FOR PRACTICAL APPLICATIONS Xiaoshan Huang 1,2 *, Erwei Yin 3 *, Yijun Wang 4, Rami Saab 1, Xiaorong Gao 1 1 Department of Biomedical Engineering, Tsinghua University, Beijing , China; 2 Neuracle Technology Co. Ltd, Beijing, , China; 3 National Key Laboratory of Human Factors Engineering, China Astronaut Research and Training Center, Beijing , China; 4 State Key Laboratory on Integrated Optoelectronics, Institute of Semiconductors, Chinese Academy of Sciences, Beijing , China ABSTRACT Translating brain-computer interfaces (BCIs) from research into real-life applications is a major challenge. A mobile electroencephalogram (EEG) system is a potential solution as it has practical advantages in flexibility, portability, and tolerance towards gross movements, however, it may compromise the EEG signal quality. In this paper, we present a new 64-channel mobile EEG system and evaluate the EEG signal quality by comparing it with a state-of-theart wired laboratory EEG system. The comparisons were validated with alpha wave and a steady-state visually evoed potential (SSVEP)-based BCI paradigm in standing, waling and running conditions. The experimental results indicated that both EEG systems performed equally well when used in EEG recordings in the standing and waling conditions. Furthermore, the mobile EEG system achieved significantly higher classification accuracies for SSVEP detection than the referenced wired EEG system in the running condition. These findings suggest that the mobile EEG system holds promise in real-life BCI applications. Index Terms electroencephalogram (EEG), amplifier, brain-computer interface (BCI), alpha wave, steady-state visually evoed potential (SSVEP) 1. INTRODUCTION A brain-computer interface (BCI) is a technology that offers access to communication and cognitive monitoring [1]. Two general classes of BCIs can be distinguished based on their techniques for brain signal acquisition, namely, invasive and noninvasive BCIs [2]. Noninvasive BCIs remain the most widely investigated and in particular, the *X. Huang and E. Yin contributed equally to this wor. This wor was supported in part by the National Natural Science Foundation of China, under Grant , and the Foundation of Key Laboratory of Science and Technology for National Defense, under Grant electroencephalogram (EEG) has been the most popular sensing modality in BCI research [3]. However, various ey problems still limit the practical applications of EEG-based BCIs. Over the past few years, BCI systems have became more powerful and robust under laboratory conditions, benefiting from the advancements in signal processing and paradigm design etc [4]. However, taing BCI systems into real-life applications is still one of the major challenges in the field [5-6]. Although conventional EEG systems are much more compact than other brain activity monitoring technologies, e.g., magnetic resonance imaging (MRI), magnetoencephalography (MEG) and near-infrared spectroscopy (NIRS) etc, they still weigh several ilograms and require wires for signal transmission [7]. Therefore, conventional EEG systems are cumbersome for applications outside the laboratory. Obviously, a small and wireless mobile EEG system has practical advantages as it provides the advantages such as flexibility, portability, and tolerance to gross movements, but it may compromise the EEG signal quality. In this study, we present a new 64-channel mobile EEG system (NeusenW, Neuracle Inc.), and compare it to a stateof-the-art wired laboratory EEG system and evaluate the EEG signal quality. Previous studies were only performed on seated participants in laboratory environments, and only a very limited number focus on motion conditions [8-9]. In this study, we instead implemented experiments in standing, waling and running conditions. To preliminarily evaluate the EEG quality recorded by both EEG systems, we firstly compared the alpha wave (8-13 Hz) frequency band using a spectral analysis approach, due to the fact that, while a EEG user has their eyes closed, the alpha wave power is the strongest spontaneous EEG signals. Steady-state visually evoed potential (SSVEP)-based BCI speller is one of the most popular BCI paradigms [10-11]. Here, SSVEP is a periodic response elicited by specific visual stimuli [12]. In our approach, we further validated the mobile EEG system /17/$ IEEE 995 GlobalSIP 2017

2 using a SSVEP-based BCI paradigm. Finally, to evaluate the EEG signal quality, the online and offline classification accuracies were compared Participants 2. METHODS AND MATERIALS Six able-bodied volunteers (five males, one female) between the ages of 20 and 30 (mean 26.33years) participated in this study. All participants had normal or corrected-to-normal vision, and had no cognitive deficits. Only one of the participants had previous experience with SSVEP-based BCI. The participants were not allowed to drin any caffeine or alcohol four hours before each session. All participants provided written informed consent Data collection In the experiments, we compared the new mobile EEG (NeusenW, Neuracle Inc.) to a well-established, wired laboratory EEG amplifier (Synamps2, Neuroscan Inc.), using a repeated measurement (seen in Fig. 1). The mobile EEG amplifier weighted 77 g (size cm 3 ), was tightly attached to the cap, and had Wi-Fi wireless connectivity. Instead, the reference EEG wired amplifier was weighted 1500 g (size cm 3 ), was placed in a bacpac with the battery and A/D boxes, and connected with a cable (~ 200 cm length) to the recording PC SSVEP-based BCI speller This study employed a 12-target SSVEP-based BCI speller with a 3 4 matrix. The stimulation matrix was presented on a 23-inch LED screen with a resolution of 1,920 1, 080 pixels and a refresh rate of 60 Hz. Each cell in the speller flicered between white and blac at a unique, constant frequency. Twelve frequencies were employed in the design of a periodic stimulus mechanism (i.e Hz with an interval of 0.5 Hz) [13]. Moreover, we used the sampled sinusoidal stimulation method to present visual flicers [14] Experimental Procedures The experiments were performed in a normal laboratory. Each participant too part in two sessions, wearing two different EEG systems, on two different half days within a one-wee period. The experimental setup was identical between sessions, only the EEG system was switched. The sequence of the sessions was randomized to avoid order effects. Additional details of the experimental arrangement are illustrated in Fig. 2 and described as follows: Fig. 2. Experimental session structure. Fig. 1. The visualization of the setup for both EEG systems. (a) NeusenW, Neuracle Inc.; (b) Synamps2, Neuroscan Inc. EEG signals were recorded from the surface of the scalp via eight electrodes placed at PO5, PO3, POz, PO4, PO6, O1, Oz, and O2, referenced to CPz and grounded to FPz, based on the 64-channel extended international 10/20 system. Impedances were ept below 10 Ω prior to recording using conductive paste. The EEG signals were sampled at 250 Hz and filtered using a 50-Hz notch filter. One session consisted of three phases, comprising standing, waling and running conditions. Specifically, the participants were standing on a treadmill facing a computer screen in the first phase; after that, in the remaining two phases, the participants were waling and running on the treadmill at the speed of 3 and 5 m/h, respectively. To mitigate fatigue, each participant was given 5-min breas between phases. In each phase, the participants first ept their eyes focused on a fixation cross on the screen for 60 s and then ept their eyes closed also for 60 s. After that, the participants were ased to spell characters for five runs, using copy-spelling paradigm. Each run contained 24 trials corresponding to all 12 characters presented twice in a random order. Each trial started with a visual cue to promote a spelling tas. The participants were then ased to shift their gaze onto the cued item within the cue duration (1 s). 996

3 After the cue offset, all stimuli items started to flicer for 3 s and the participants was ased to maintain their gaze and avoided eye blins during the stimulus period. Following stimulus offset, the selected character was typed in the text input field on the top of the screen, and a 1-s brea was given to allow the participants to see the selected result. To reduce visual fatigue, a 0.5 to 2-min brea was given after every run Target detection In this study, we employed filter ban canonical correlation analysis (FBCCA) for target detection [15-16]. The bandwidth of the stimulation frequencies (i.e Hz) was 8 Hz. According to the analysis in [16], a frequency range of 8-88 Hz was selected for the filter bans (ten filter bans in total). In the implementation of bandpass filter, an additional bandwidth of 2 Hz was added to both sides of the passband for each sub-band. Thus, the n th sub-band started from the frequency at 8n 2Hz and ended at 90 Hz. In the FBCCA, the bandpass filters for extracting subband components ( Y, n 1, 2,..., N ) from the original SBn EEG signals Y were zero-phase Chebyshev type infinite impulse response (IIR) filters. After that, the standard CCA algorithm was applied to each of the sub-band components separately, resulting in correlation values between the subband components and the predefined reference signals corresponding to all stimulation frequencies X, 1, 2,..., 12 ) producing a correlation vector ( f consisting of N correlation values for each th reference signal. A weighted sum of squares of the correlation values corresponding to all sub-band components (i.e. 1 N,..., ) was calculated as the feature for target identification: N n 2 wn ( ) ( ) (1) n1 where n is the index of the sub-band. Furthermore, the weights for the sub-band components were defined as follows (see [16]): a w( n) n b, n [1 N] (2) where a and b are constants that maximize the classification performance. In practice, a and b can be determined using a grid search method during offline analysis. In our approach, the correlation coefficient is calculated using the function canoncorr() in the MATLAB toolbox [17], using the stimulus frequency ( X () t ) and the multidimensional EEG data ( Y SBn f ) as inputs. Here, the stimulus frequency is represented as sinusoidal signals that can be decomposed into a Fourier series of its harmonics: sin(2 ft) cos(2 ft) 1 2 M X f ( t), t,,..., (3) S S S sin(2 Nh ft) cos(2 Nh ft) where N h is the number of harmonics, M is the number of sampling points per selection, and S is the sampling rate. Finally, corresponding to all stimulation frequencies (i.e. 1,..., 12 ) was used to determine the frequency of the SSVEPs. The frequency of the reference signals with the maximal is considered to be the frequency of the SSVEP. 3. RESULTS 3.1. Copy-spelling performance Although the stimulus time of each trial was 3 s, the online spelling results were detected based on 2-s EEG signals as it is the mostly widely used time-period for SSVEP online detection [18]. As shown in Table 1, in the standing and waling conditions, the average online accuracies achieved by the two EEG systems were quite similar, both surpassing 95%. Interestingly, in the running condition, the average accuracy of the new mobile EEG system (73.14%) was significantly higher comparing to that of the referenced EEG system (56.95%). Moreover, we also compared the classification accuracies between the two EEG systems using the offline analysis approach (seen in Fig. 3). As shown in Fig. 3, the accuracies of both EEG systems were similar to each other in standing and waling conditions, while the accuracies of the mobile EEG system were consistently higher than those of the reference EEG system in running condition. Fig. 3. Comparisons of offline classification accuracies Spectral analysis 997

4 To further evaluate the signal quality, we compared the frequency spectrum in eyes open and closed conditions, by averaging the 2-40-Hz spectra of all selected channels over the six participants (seen in Fig. 4). We found that both EEG systems yielded similar alpha amplitudes in all the conditions. Moreover, the spectral power above 20 Hz was consistently higher in the reference EEG system. We further analyzed the frequency spectra of all conditions during SSVEP tass (seen in Fig. 5). The frequency response of the second and third harmonics are weaer with the increase of speeds, the SSVEP harmonics of referenced EEG system were obviously weaer than the mobile EEG system. These results indicated that the mobile EEG system may be more robust to high-frequency non-tas related components. 4. CONCLUSIONS AND DISCUSSIONS In this paper, we compared the EEG signal quality recorded by a new mobile EEG system to that of a state-of-the-art wired laboratory EEG system. Specifically, alpha wave and SSVEP signals were recorded during standing, waling and running conditions. In both standing and waling conditions, similar results were obtained with the testing and reference EEG systems, in terms of spectral amplitudes of alpha wave and classification accuracies of SSVEP-based BCI speller. However, an obviously better performance was achieved by the mobile EEG system compared with the reference wired EEG system in running condition. As we now, wired EEG systems limit the natural behavior of users during signal acquisition and thereby lead to heavily constrained recording conditions. Mobile EEG system on the other hand are quic to set up, better wearing comfort, and are more robust against gross movement. Therefore, we believe that the mobile EEG system could be employed in a wider range of environments, especially for out-door motion scenes, which could facilitate the transfer of BCI applications from the laboratory to real-life environments. To understand and explain why the new mobile EEG system can provide high-quality EEG acquisition and outperform wired EEG systems in running condition, we hypothesize that the following four factors contributed to the performance improvement: (1) signal transmission is based on a precise wireless protocol, which achieves a reliable and accurate performance on event synchronization; (2) the amplifier adopts wide dynamic range and DC-coupled technologies, to prevent saturation of the amplifier induced by electrode offset voltage and artifacts; (3) the amplifier is lightweight and head-mounted, which ensures all relevant parts remain together when individuals perform any gross movements; (4) long and isolated cables are avoided, which reduces the strong electromagnetic interferences induced by the movements of electrodes cables. Table 1. Online accuracy (%) for the copy-spelling tas. Participant Standing Waling Running Neuracle Neuroscan Neuracle Neuroscan Neuracle Neuroscan S S S S S S AVG T-Test H: 0, P: H: 0, P: H: 1, P: Fig. 4. Frequency spectra of the EEG signals recorded during the eyes open/eyes closed tass. Fig. 5. Frequency spectra of the EEG signals recorded during the SSVEP tass. 998

5 12. REFERENCES [1] B.Z Allison, E.W. Wolpaw, and J.R. Wolpaw, Braincomputer interface systems: progress and prospects, Expert Rev. Med. Devices, vol. 4, no. 4, pp , [2] A. Ortiz-Rosario, and H. Adeli, Brain-computer interface technologies: From signal to action, Rev. Neurosci., vol. 24, no. 5, pp , [3] S. Moghimi, A. Kushi, A.M. Guerguerian, and T. Chau, A review of EEG-based brain-computer interfaces as access pathways for individuals with severe disabilities, Assist. Technol, vol. 25, pp , [4] G. Muller-Putz, J.D. Millan, G. Schal, and K.R. Muller, The plurality of human brain-computer interfacing, Proc. IEEE, vol. 103, no. 6, pp , [5] C. Grozea, C.D. Voinescu, and S. Fazli, Bristle-sensors lowcost flexible passive dry EEG electrodes for neurofeedbac and BCI applications, J. Neural Eng., vol. 8, no. 2, pp , [14] Y. Wang, Y. Wang, and T. Jung, Visual stimulus design for high-rate SSVEP BCI, Electron. Lett., vol. 46, no. 15, pp , [15] X. Chen, Y. Wang, M. Naanishi, X. Gao, T. Jung, and S. Gao, High-speed spelling with a noninvasive brain-computer interface, Proc. Nati. Acad. Sci. USA, vol. 112, no. 44, pp. E6058-E6067, [16] X. Chen, Y. Wang, S. Gao, T. Jung, and X. Gao, Filter ban canonical correlation analysis for implementing a high-speed SSVEP-based brain-computer interface, J. Neural Eng., vol. 12, pp , [17] W.J. Krzanowsi, Principles of Multivariate Analysis: A User's Perspective, Oxford Univ. Press, New Yor, NY, USA, [18] G. Bin, X. Gao, Z. Yan, B. Hong, and S. Gao, An online multi-channel SSVEP-based brain-computer interface using a canonical correlation analysis method, J. Neural Eng., vol. 6, pp , [6] M. Bandad, H. Zarshenas, and M.A. Auais, Application of BCI systems in neurorehabilitation: a scoping review, Disabil. Rehabil. Assist. Technol., vol. 10, no. 5, pp , [7] S. Debener, R. Emes, M. De Vos, and M. Bleichner, Unobtrusive ambulatory EEG using a smartphone and flexible printed electrodes around the ear, Scientific Reports, vol. 5, pp , [8] Y. Lin, Y. Wang, and T. Jung, A mobile SSVEP-based braincomputer interface for freely moving humans: the robustness of canonical correlation analysis to motion artifacts, 35th Annual International Conference of the IEEE EMBS, pp , [9] J.E. Kline, H.J. Huang, K.L. Snyder, and D.P. Ferris, Isolating gait-related movement artifacts in electroencephalography during human waling, J. Neural Eng., vol. 12, pp , [10] E. Yin, Z. Zhou, J. Jiang, Y. Yu, and D. Hu, A dynamically optimized SSVEP brain-computer interface (BCI) speller, IEEE Trans. Biomed. Eng., vol. 62, no. 6, pp , [11] M. Naanishi, Y. Wang, Y. Wang, T. Jung, A comparison study of canonical correlation analysis based methods for detecting steady-state visual evoed potentials, Plos One, vol. 10, no. 10, pp. e , [12] X. Gao, D. Xu, M. Cheng, and S. Gao, A BCI-based environmental controller for the motion-disabled, IEEE Trans. Neural Syst. Rehabil. Eng., vol. 11, no. 2, pp , [13] P. Yuan, X. Chen, Y. Wang, X. Gao, and S. Gao, Enhancing performances of SSVEP-based brain-computer interfaces via exploiting inter-subject information, J. Neural Eng., vol. 12, pp ,