2162-2248/15©2015IEEE34 IEEE ConsumEr ElECtronICs magazInE ^ july 2015

M ajor thrusts combining neuroscience, sen- sor chip design, and software development have already shown remarkable advancement, regardless of many uncertainties and challeng- es. Brain–computer interface (BCI) headsets,

which are injecting new break points in games and entertain- ment, deliver desirable special effects, aiding wellness train- ing and rehabilitation. This article summarizes the brain signal processing algorithmic approach in achieving the level of interpretability of brain signals to date. The emergence of imprecise brainwave headsets in the commercial world is illustrated. Current tools for research and future development are discussed, with a recommendation to standardize the brain signal databank, anticipating its reach to big data.

Sheryl Flynn and I delivered a tutorial “Therapeutic Neuro- game Application Development for Healthcare/Wellness” at the 2015 International Conference on Consumer Electronics

(ICCE). Dr. Flynn, founder and chief executive officer of Blue Marble Game Company, is a world-renowned expert with a unique perspective on therapeutic neurogaming products. Based on her experience and the feedback from the audience, it behooves me to elaborate on the state of the art of BCI technol- ogy and products.

BACKGROUND AND ROAD MAP In 1929, neuroscientists started to observe the primary cur- rents generated by synchronous firing of large populations of neurons in the brain. The secondary currents induced as the extracellular return currents are measurable as extracranial electric potentials by electroencephalography (EEG), reflect- ing human cognizance. These research activities intensified after the 1970s. The Wadsworth Center in New York [15] cre- ated a BCI system that incorporated electronic signals from the brain into a novel communication-and-control device in 1991, using as many as 75 sensors. However, this device required institutional electronic expertise, and it was very costly. It took the next 20 or so years for a consumer-grade product to be introduced. Since 2012, BCI headsets using one

Brain–Computer Interface Technology

and Development

By Narisa N.Y. Chu

The emergence of imprecise brainwave headsets

in the commercial world.

Digital Object Identifier 10.1109/MCE.2015.2421551

Date of publication: 15 July 2015

july 2015 ^ IEEE ConsumEr ElECtronICs magazInE 35

Im a g

e lIc

e n

s e d

by In

g r

a m

p u

b lIs

h In

g

to nine sensors have entered the market, with some examples demonstrated in Figures 1 and 2, where sensor placements have also been indicated for various settings based on standard P300 EEG detection locations on the scalp. A road map that highlights the BCI technology development is shown in Figure 3, contrasting with the original research focus on brain signal processing algorithms with the advent of many products’ introduction of headsets.

Due to the small signal induced by neurons, event-revoked potentials (ERPs) of the brain activities, measured from the outside of the scalp via sensors, require intricate interpreta- tion. More than 50 digital processing algorithms have been

developed since 1981 to dwarf the noise, overcome attenua- tion, and discriminate from physiological interferences due to tissue thickness and wetness; chemical change; and stimulus of visual, audio, and muscle movements from eye to toe. The major schools in developing these algorithms are represented by the five blue arrows in Figure 3. The reconstructed brain features reflect many assumptions, sometimes taking into consideration a priori knowledge, however, leaving plenty of room to guess about the real intention of the brain. These so- called “inverse problem solutions,” similar to “reverse engi- neering” effort, demonstrated accuracy between 60% to 90% in various controlled environments. The performance so far

36 IEEE ConsumEr ElECtronICs magazInE ^ july 2015

has significantly limited BCI usage in real life. Regardless of the range of uncertainty, innovative applications have been tri- aled recently, and BCI headsets (a sample product is shown at the bottom center in Figure 3, with no commercial endorse- ment) designed for mass consumption have made inroads for rehabilitation and learning fun. This accelerated product intro- duction has been quite strategic in building around an impre- cise BCI measurement while creating special effects for entertainment such as in games and videos. However, before more headsets and applications become popular, it is helpful to establish a standardized brainwave databank for identifying and sharing brain signals from many denominations with the aim to facilitate an intelligent search down the road. This stan- dardization effort would represent a broader collaboration between software and neuroscience leading to treatments for brain-related sickness, rehabilitation, and wellness, and proba- bly furthering brain-triggered marketing and privacy protec- tion. The new approach to standardization is illustrated in the gold stream in Figure 3.

The conventional digital signal processing algorithms based on statistics, complex mathematical transformation, and estimation tend to be agnostic, reaching a performance limit. Entrepreneurs have not been shy with this technology

Nz

Fp1 Fpz Fp2

AF7

F9

FT9

A1

TP9

P9

TP7

P7

PO7 PO3 PO4

PO8

O2OZ

lZ

O1

POZ

CP5

P5 P3 P1 P2 P4 P6 P8

P10

PZ

CP3 CP1 CPZ CP2 CP4 CP6 TP8 TP10

T9 T7 C5 C1 C2 C6

FT7 FC5 FC3 FC1 FCZ FC2 FC4 FC6 FC8

FT10

T10T8 A2

F7 F5 F3 F1 FZ F2 F4 F6

F8 F10

AF3 AFZ AF4 AF8

C3 CZ C4

MUSE (Four Sensors)

Staalhemel (Eight Sensors)

• Relaxation • Focus

Motor Imagery (MI):

C3–Right-Hand MI

CZ –Foot MI

C4–Left-Hand MI

Examples:

EPOC (4 + 9 = 14 Sensors):

AF3, AF4, F3, F4, FC5, FC6, F7, F8 T7, T8, P7, P8, O1, O2 - • Instantaneous Excitement • Long-Term Excitement • Frustration • Engagement

INSIGHT (Five Sensors):

AF3, AF4, T7, T8, Pz - • Instantaneous Excitement • Long-Term Excitement • Stress

• Meditation

• Engagement • Relaxation • Interest • Focus

TP9, TP10, Fp1, Fp2

Fp1, Fp2, F7, F8, C3, C4, O1, O2

eMotiv

� EPOC (2014) � INSIGHT (2015)

Interaxon NeuroSky

� Muse (2014)

� MindWave Mobile (2012)

FIGURE 1. BCI sensor placement to detect EEG based on the international P300 standard.

FIGURE 2. BCI products entering the market.

july 2015 ^ IEEE ConsumEr ElECtronICs magazInE 37

imprecision and have developed products with promising benefits to eHealth, testing out the brain signals of the young and the sick. Although growth has been proliferating where selective parameters were chosen to represent brainwaves within private groups, it is extremely difficult for reliable pat- tern matching and self-learning algorithms to be incorporated universally. With standardization, brainwave data from all sources and reactions can be intelligently accumulated and used. This new approach is facilitated by the understanding of the threshold (TH) between the brain’s actual and inter- preted meaning, as denoted in Figure 1. Applications can be triggered by this threshold and go on to another proven activ- ity such as games.

The databank should include not only the brainwave diagram but the processing and search algorithms associated with the brainwave, plus all prior knowledge. This databank will be well positioned to blend fuzzy logic inference and pattern recognition related to big data evolution.

The sections that follow will illustrate the paths taken in achieving the level of interpretability of brain signals to date. The emergence of imprecise brainwave headsets in the com- mercial world is illustrated. The current tools for research and future development are discussed, with a recommendation to standardize the brain-signal databank, anticipating its reach to big data and, perhaps, cloud computing.

BRAIN SIGNAL FREQUENCY BANDS Brain waves can be represented by six typical bands based on the frequency range between 1 and 100 Hz, designated as the

, , , , ,Ta b c n and i bands. These band frequency ranges are shown with a collective interpretation as follows [3]–[5].

▼ Frequency 1–4 Hz: T band, symbolizing high emotional conditions or in a sleep stage.

▼ Frequency 4–7 Hz: i band, similar to the T band, also symbolizing a calm and relaxed mood.

▼ Frequency 8–12 Hz: a band, symbolizing smooth pat- terns: awaken, calm, and eyes closed in a relaxed mood.

▼ Frequency 8–13 Hz: n band, desires from the sensorimotor cortex.

▼ Frequency 12–30 Hz: b band, for desynchronized—nor- mal awaken, open eyes, busy, churning, and concentrating.

▼ Frequency 25–100 Hz: c band, desires from somatosenso- ry cortex for touch—busy, churning, and concentrating. At these bands, typically a very low signal is collected by the

noninvasive sensor (in the range of 5–10 nV), while interfering noise of 10–20 times stronger than the brain signal is measured on the scalp. The correlation of these band signals with respect

Common Spatial Patterns

Blind Source Separation

Band Power Decomposition

Assumptions and Prior Knowledge

Stimulants: Eye, Audio, Motion, Etc.

Brain-Wave Databank

Architecture and Intelligent Search

Conventional Approach

A New Approach

Threshhold Standardized User Brain Data Input and Retrieval

Sensor Interpreted

Headset

≠

FIGURE 3. A BCI technology road map—two approaches.

BCI headsets, injecting new break points in games and entertainment, deliver desirable special effects that can blend in our pursuit of wellness and rehabilitation.

38 IEEE ConsumEr ElECtronICs magazInE ^ july 2015

to a person’s brain condition varies from group to group. In gen- eral, the interpretation states whether a person is in attention or mediation. Exactly what attention and mediation means is left for one to judge.

Various digital processing algorithms were attempted for the purpose of making the brain signal interpretable via an increase of the signal-to-noise ratio, manipulation of sensor spatial and time domain parameters (TDPs), and fitting of a priori knowledge and various assumptions to yield some per- formance improvement. These digital processing algorithms are profiled as follows.

DIGITAL BRAIN SIGNAL PROCESSING ALGORITHMS Major research efforts have been spent developing digital processing algorithms to identify brain signals at various fre- quency bands. A plethora of inverse problem solving and pat- tern-matching analyses have been applied to signals measured from many sensors placed on a cap noninvasively covering a person’s head. Stable reconstructions of brain-sig- nal features could be achieved through the use of many tech- niques. Since the measurements and modeling techniques both contain noise and assumptions, the true solution could hardly be totally derived from the algorithms or totally deter- mined from the measurements. A comprehensive strategy for dealing with noisy data could also include data filtering and an optimal selection of geometric parameters, such as sensor positioning. One particularly notable technique suggested was regularization [13], which attempted to achieve a com- promise between a close fit to the data and stability of the algorithmic solution. By removing the high-frequency com- ponent from a derived solution, it believed that it effectively filtered a portion of the noise [13].

Various algorithms have demonstrated performances up to 90% accuracy in a controlled environment. They could be highly computationally intensive. They were primarily carried out offline in batch processing, with unproven real-time applica- tions. They also required stringent calibration and testing to facilitate performance dedicated to an individual. To appreciate the level of effort spent in the development of these brain signal processing algorithms, three major categories are elaborated: 1) band-power feature extraction, 2) common spatial patterns (CSP) analysis, and 3) statistical source separation [6]. The year cited, spanning from 1981 to 2014, refers to approximately the first introduction of the algorithm.

Band Power Feature extraction 1) band-pass filtering and power estimation taking tem-

poral average 2) periodogram (Fourier decomposition) 3) p owe r s p e c t r a l d e n s it y f r o m a u t o r e g r e s s ive

coefficients 4) wavelet scalogram (time-scale representation) 5) spectrogram (time-frequency decomposition with

averaged spectrums over time) [2008].

cSP analySiS 1) spatial filtering (SF)

i) bipolar ii) Laplacian

2) physical forward modeling: inverse solution methods i) Minimum current estimate [2008, 2009]

– Focal underdetermined system solver [1995] ii) Weighted minimum norm estimate

– LORETA, eLORETA, and sLORETA—all assuming smoothness [1987, 1994]

iii) Mixed norm estimate, combining sparsity and smoothness [2008] – S-FLEX Champagne—simple spatial structure

[2011] iv) minimum entropy v) source localization paradigms

– dipole modeling [1992] – multipole modeling – scanning

a) subspace methods (MUSIC/RAP-MUSIC) [1986, 1999]

b) beamformers [1997] i) LCMV beamformer ii) nulling beamformer

vi) depth compensation modeling [1987] c) CSP [1990]

i) supervised regulated spatial filters based on EEG and CSP

ii) filter bank CSP iii) discriminatory filter (DFCSP) iv) sparse CSP v) source power correlation analysis (SPoC)

[2014] vi) canonical spatial power comodulation [2014]

vii) steady-state auditory evoked potentials [1981] viii) steady-state visual evoked potentials

ix) spatiospectral decomposition.

StatiStical Backward Modeling: Blind Source SeParation

1) linear classifiers—linear discriminant analysis (LDA), support vector machine (SVM), and infinite impulse response [2010]

2) linear regression—OLS, ridge regression, LASSO [2005]

3) principal component analysis (PCA) [2005]

A comprehensive strategy for dealing with noisy data could also include data filtering and an optimal selection of geometric parameters, such as sensor positioning.

july 2015 ^ IEEE ConsumEr ElECtronICs magazInE 39

4) canonical correlation analysis (CCA)—hyperscanning ERP studies [2011]

5) independent component analysis (ICA) 6) Granger-causally interacting—SCSA, MVARICA—

brain connectivity studies [2008] 7) dimensionality reduction—stationary subspace analy-

sis (SSA) [2009]. Performance data ranging from 60.7%—with rudimentary

band-pass filtering—to 92.8%, with DFCSP, have been dem- onstrated in carefully calibrated and tested setup.

It should be noted that every method performs well if its specific assumptions are met. Unfortunately, no method can perform well in all real cases. It is anticipated that multiple methods may be combined to lead to better solutions. It is also possible that multiple methods exist for the same solution. With a common goal to characterize brain activity of interest, there is still no assurance that one method is better than others in all circumstances.

Assumptions play a major role in the derivative and con- vergence of these methods: 1) Assumptions often made with inverse analysis and blind

source separation: Brain activity is assumed to be: – correlated with behavior or stimulus variables [OLS,

ridge regression, LASSO] – reflected in the strongest components of EEG [PCA] – correlated across subjects/stimulus repetitions [CCA] – stationary, major signaling stays local, unaffected by

neurons away from the measuring point [SSA] – different from experimental conditions [LDA, SVM].

Brain components are assumed to be: – mutually independent [ICA] – Granger-causally interacting [SCSA, MVARICA].

2) Not all EEG phenomena are phase-locked to certain events. There are rhythms depending on the mental state. Most rhythms are idle, attenuated during activation (e.g., eyes open/close, arm at rest/moves).

3) Sensor-spatial analysis assumes smoothness and sparsity where neighboring voxels (discrete volume elements) show similar activity, and only a small part of the brain is active for a single task. A limited evaluation has provided some insight as to how

to improve the BCI performance. The performance of the digital brain signal algorithmic processing has reached nearly 93% under a controlled environment. For example, the basic band-power approach demonstrated 60.7% average accuracy in a BCI competition [6]. Applying the Laplacian SF tech- nique, the average accuracy reached 68%. Using the bipolar SF technique, the average accuracy was improved to 70.5%, slightly better than Laplacian. Combining supervised regulat- ed spatial filters on EEG with CSP and TDP, the average accuracy was demonstrated between 78.7 and 88.9%, a range too large to lend enough confidence. Another combination of TDP, SF, and CSP provided 80.1% accuracy. Using filter bank CSP, the average accuracy achieved was between 81.1 and 90.9%. The best accuracy of 92.8% was accomplished by

adopting DFCSP [6]. Thus, the performance has not demon- strated enough robustness for all known algorithms, to say the least.

One reaches a diminishing return if one continues searching for better algorithmic processing of brain activities. Training and testing input have been employed to augment these algo- rithms. Recently, another notable approach by means of virtual reality and gaming has opened up new frontiers for BCI, in lieu of the imprecise algorithms.

BCI TOOLS FOR DEVELOPMENT ACCELERATION Two major research tools have been developed extensively: BCI2000 [15] from the Wadsworth Center in New York and BCILAB [8] from the University of California, San Diego, the Swartz Center of Computational Neuroscience (SCCN). These pioneering platforms of the BCI were initiated for the disabled to operate wheelchairs/computers. Later, they were extended to support more efforts for rehabilitation, education, and enter- tainment purposes. Recently, commercial tools offered by Neu- rosky [5], Interaxon [4], and eMotiv [3], bundled with products, have also been made available with various degrees of open utility and readiness. Most of the commercial software development kits (SDKs) present various degrees of stability and maturity. There is enough momentum from the develop- ment community to trial and inject further “kickstarter” initia- tives. Benefiting from more than two decades of neuroscience research and development, these tools have become widely available within the last several years. Table 1 summarizes the status of these tools.

It is clear that both a vigorous research tool and a product realization kit can be chosen for development. It should also be noted that the number of sensors gathering brain signals has purposefully decreased significantly in commercial product realization. Some of the earlier algorithms of broader spatial coverage might lose their effectiveness as sensor placements are minimized for user comfort.

With all due diligence exploring the intricate brain, it becomes evident that collective efforts among engineering, medical, and user disciplines have to come together for optimal use of the brain signal data. Thus, forming a stan- dardized databank should be an obvious next step. The naming preference, “databank” instead of “data base,” is to acknowledge the amount and the dynamics of the data (big data) associated with various algorithms, training, testing, and feedback that can eventually realize brain activities. It

Collective efforts among engineering, medical, and user disciplines have to come together for optimal use of the brain signal data.

40 IEEE ConsumEr ElECtronICs magazInE ^ july 2015

is necessary to recognize user brain reactions eventually across the board: gender, age, environment, stimulus, intent, processing algorithm, intelligent search, and pattern associ- ation, not to omit involvement from big data, cloud comput- ing, and sensor networking.

This databank, built upon a standard brain signal profile format plus intelligent linking factors can be destined to make the retrieval and triggering function much more timely and meaningful. Once standardized, new applications and benefits can be then accelerated beyond imagination. Attributes that

Rehabilitation

Computer and Wheelchair Operation

Attention Oblivion Defocused Neglect Random Hyped

Marketing Applications

Like Dislike Maybe Trial Committed Etc.

Signal Characteristics + Algorithmic Processing

Feature Vector

Gamma 25~100 Hz

Beta 12~30 Hz

Alpha 8~12 Hz

Mu 8~13 Hz

Theta 4~7 Hz

Delta 1~4 Hz

Empirical Composite

Learning Factor from Training and Testing

Band Power

Spacial Filter Temporal Average

Periodogram Fourier Decomposition

Power Spectral Density + AR Coefficients

Wavelet Scalogram Time-Scale Representation

Spectrogram Time- Frequency Decomposition

Analysis Type

ERP: Inverse Solution and CSP

BSS, Statistic Modeling

CCA ICA LDA PCA SPoC SSA

Entertainment Directives

Happy Sad Excited Scary Disgusted Don’t Care Etc.

Database Classifier

Gender Age Eye Blinking

Audio Effect

Facial Muscle Movement: Chewing, Clenching Teeth, Grinding Jaw, Etc.

Hand/Foot Imagery Motion

Chemical, Light, Color, Etc. Stimulant

FIGURE 4. Attributes in a brain signal databank.

Table 1. The BCI SDK/platform [3]–[5], [7], [14]–[16].

Company/ university Product/Platform Year

sensor/ Channel tools/Platforms apps

NeuroSky MindWave Mobile 2013 1 MWM SDK Rehab for ADD, stroke; education, entertainment

Interaxon MUSE 2014 7/4 Basic SDK for con- nection

Entertainment device control

eMotiv EPOC 2014 9/14 SDK Lite Entertainment neurotherapy

Insight 2015 9/5

SCCN/ UCSD

BCILAB/ LSL

2012 Many Open/ MATLAB

Focus: Comparative evaluation of BCI methods

Wadsworth Center, New York BCI 2000 2010 Many Open Rehab + general purpose

Various Developers Pyff in Python 2010 – Open and free Standardization of feedback and stimulus

july 2015 ^ IEEE ConsumEr ElECtronICs magazInE 41

require consideration in such a standardization effort are intro- duced in Figure 4 as a starting point. These attributes are not meant to be exclusive.

CHALLENGES IN TIME One would wonder about the time frame for the BCI to become fully developed from the first patent/prototype to mass produc- tion for general consumption. Some historical timelines of simi- lar technology can be referenced. Considering the following track records:

▼ Wired telephone: patent established in 1876 to mass pro- duction in 1970s

▼ Mobile phone: 1946–1994 (iPhone was introduced in 2007) ▼ Brain-wired cap: from 1981 onward ▼ Brain headset: since 2012, limited products have been introduced. One can thus anticipate the acceleration point on BCI production

in the next two decades or sooner. A few crucial questions are still waiting to be tackled by the research community [6], [11], [12].

▼ What are the fundamental accuracy limits imposed by the current EEG sensors?

▼ What assumptions are widely agreeable, and what empiri- cal data are required to improve the accuracy of the avail- able mathematical models?

▼ How can hierarchical models be constructed to include data from multiple people, environments, and applications?

▼ As the brain-signal performance improves, how will sensor convergence be handled?

▼ How are auxiliary data included (e.g., muscle movement, eye contact, and chemical change)?

▼ How can designing methods directly target real-world applications with robustness?

▼ Will there be standardization of the brain databank? ▼ What are the privacy issues?

SUMMARY Major thrusts combining neuroscience, sensor chip design, and software development have already shown remarkable advance- ment, regardless of the many uncertainties and challenges. Entrepreneurs have started to capture the low-hanging fruit from the BCI technology evolutionary “branches.” The BCI headsets, injecting new break points in games and entertainment, deliver desirable special effects that can blend in our pursuit of wellness and rehabilitation. To foster these promises, brainwave databank standardization can play a major role in converging the collec- tion and utilization of users’ essential, private brain information, following the example of DNA and fingerprints.

ABOUT THE AUTHOR Narisa N.Y. Chu ([email protected]) is the cofounder of CWLab International and is currently focusing on BCI research. She is also a member of the IEEE Consumer Elec- tronics Society Board of Directors. It is with the latter role that she contributed to the writing of this article.

REFERENCES [1] P. Herman, G. Prasad, T. M. McGinnity, and D. Coyle, “Comparative analysis of spectral approaches to feature extraction for EEG-based motor imagery classification,” IEEE Trans. Neural Syst. Rehab. Eng., vol. 16, no. 4, pp. 317–326, Aug. 2008. [2] N. Brodu, F. Lotte, and A. Lécuyer, “Exploring two novel features for EEG-based brain-computer interfaces: Multifractal cumulants and pre- dictive complexity,” Neurocomputing, vol. 79, no. 1, pp. 87–94, 2012. [3] eMotiv. (2015). EPOC & INSIGHT. [Online]. Available: http:// emotiv.com/store/compare/ [4] Interaxon. (2015). MUSE. [Online]. Available: https://sites.google. com/a/interaxon.ca/muse-developer-site/data-files [5] NeuroSky. NeuroSky MindWave Mobile [Online]. Available: http:// www.braincorner.com.tw/en/?product=mindwave_mobile_myndplay [6] S. Haufe. (2014). Berlin brain-computer interface, Winter School Lecture. [Online]. Available: bbci2014_haufe_statistical_models_01. pdf [7] C. Kothe. (2012). Lectures, SCCN/UCSD. [Online]. Available: http://sccn. ucsd.edu/wiki/Introduction_To_Modern_Brain-Computer_Interface_Design [8] C. Kothe. (2013). BCILAB, SCCN/UCSD. [Online]. Available: ftp:// sccn.ucsd.edu/pub/bcilab [9] S. Haufe, V. V. Nikulin, A. Ziehe, K.-R. Muller, and G. Nolte, “Com- bining sparsity and rotational invariance in EEG/MEG source reconstruc- tion,” NeuroImage, vol. 42, no. 2, pp. 726–738, June 2008. [10] S. Haufe, V. V. Nikulin, A. Ziehe, K.-R. Muller, and G. Nolte, “Esti- mating vector fields using sparse basis field expansions,” in Advances in Neural Information Processing Systems, vol. 21, Cambridge, MA: MIT Press, 2008, pp. 617–624. [11] Y. Zhang and Y. Ding, “CTL model update for system modifica- tions,” J. Artificial Intell. Res., vol. 31, pp. 113–155, 2008. [12] K.-R. Muller, S. Fazli, P. von Bunau, F. Meinecke, W. Samek, G. Curio, and B. Blankertz. (2014). BCI and nonstationarity. BBCI Winter School, Berlin. [Online]. Available: bbci2014_muller_non_stationarity_01.pdf [13] F. Lotte. (2014). Oscillatory EEG-based BCI design: Signal process- ing and more. InriaBordeaux Sud-Ouest, France, Potiocteam. [Online]. Available: bbci2014_lotte_signal_processomg_01.pdf [14] I. F. Gorodnitsky, J. S. George, and B. D. Rao, “Neuromagnetic source imaging with FOCUSS: A recursive weighted minimum norm algorithm,” Electroencephalogr. Clin. Neurophysiol., vol. 95, no. 4, pp. 231–251, Oct. 1995. [15] C. Brunner, G. Andreoni, L. Bianchi, B. Blankertz, C. Breitwieser, S. Kanoh, C. A. Kothe, A. Lecuyer, S. Makeig, J. Mellinger, P. Perego, Y. Renard, G. Schalk, I. P. Susila, B. Venthur, and G. R. Muller-Putz. (2012). BCI software platforms. [Online]. Available: http://people. rennes.inria.fr/Anatole.Lecuyer/BOOKBCI_bciplatforms12.pdf [16] T. M. Vaughan, D. J. McFarland, G. Schalk, W. A. Sarnacki, D. J. Krusienski, E. W. Sellers, and J. R. Wolpaw, “The Wadsworth BCI research and development program: At home with BCI,” IEEE Trans. Neural Syst. Rehab. Eng., vol. 14, no. 2, pp. 229–233, June 2006. [17] B. Venthur, S. Scholler, J. Williamson, S. Dahne, M. S. Treder, M. T. Kramarek, K. R. Muller, and B. Blankertz, “Pyff—A pythonic framework for feedback applications and stimulus presentation in neuro- science,” Front. Neurosci., vol. 4:179, Dec. 2010. [18] EEG Resources. (2014, Jan. 5). Biomed research, Middle East Med- ical Information Center and Directory. Epilepsy Awareness Program. [Online]. Available: http://www.biomedresearches.com/root/pages/ researches/epilepsy/eeg_resources.html

Some of the earlier algorithms of broader spatial coverage might lose their effectiveness as sensor placements are minimized for user comfort.