Proposal of a P300-based BCI Speller using a Predictive Text System

Ricardo Ron-Angevin and Leandro da Silva-Sauer

Departamento de Tecnolog

ıa Electr

onica, E.T.S.I. Telecomunicaci

University of M

alaga, Campus de Teatinos s/n, 29071 M

alaga, Spain

Keywords:

Brain-computer Interface (BCI), P300, Speller, T9 Interface.

Abstract:

This paper presents a P300-based BCI speller system that uses a virtual 4 x 3 keyboard based on the T9

interface developed on mobile phones in order to increase the writing speed. To validate the effectiveness of

the proposed BCI, we compared it with two adaptations of the classical Farwell and Donchin speller, which is

based on a 6 x 6 symbol matrix. Three healthy subjects took part in the experiment. The preliminary results

conﬁrm the effectiveness of T9-based speller, since the time needed to spell words and complete sentences

was considerably reduced.

1 INTRODUCTION

Brain-computer interface (BCI) systems (Wolpaw

et al., 2002; Birbaumer, 2006) are devices that trans-

form a user’s brain activity into commands that are

interpreted by a machine. Such systems offer a non-

muscular channel for users to interact with their en-

vironment. This is particularly useful for people

suffering from neurodegenerative disorders, such as

amyotrophic lateral sclerosis, as they can eventually

present severe motor disabilities, to the point of losing

control of the muscles that are responsible of volun-

tary body movements, including eye movement and

breathing itself. In those cases, BCIs turn out to be

the only way for patients to gain some degree of com-

munication and autonomy in their daily lives.

The most widely used BCI systems are those

based on electroencephalographic (EEG) signal

recording, due to its non-invasiveness, but also to its

good temporal resolution and ease of use. Three types

of EEG-based BCI systems have been used for com-

munication purposes, namely those based on: (a) slow

cortical potentials (SCPs), (b) P300 event-related po-

tentials (ERP), and (c) sensorimotor rhythms (SMR)

(Mak and Wolpaw, 2009). BCIs based on SCP and

SMR demand that users are extensively trained be-

fore they show sufﬁcient control of their brain ac-

tivity. In contrast, BCIs based on P300 rely on a

common, expected human response to infrequent tar-

get stimuli—usually visual—and thus require mini-

mal training. The P300 signal, recorded over the cen-

tral and parietal regions, is a positive deﬂection of

brain wave at a latency of about 300 ms after stim-

ulus presentation.

The main applications of P300-based BCI sys-

tems are aimed at communication purposes. They

are based on the P300 speller ﬁrst developed by Far-

well and Donchin (1988), which is still referenced and

intensely studied (Bianchi et al., 2010; Kleih et al.,

2010; Krusienski et al., 2008; Sellers et al., 2006). In

this BCI, a 6 x 6 matrix of letters, arranged in rows

and columns, is shown to the subject. The user fo-

cuses his/her attention on the matrix element he/she

wishes to select as each row and column is ﬂashed

(i.e., intensiﬁed) randomly, one after the other. Af-

ter a number of ﬂashes, the symbol that the user was

supposedly attending at is presented on screen.

The efectiveness of the P300-based BCI speller

system is guaranteed by a number of studies carried

out not only on healthy subjects (Donchin et al., 2000;

Wang et al., 2005) but also on subjects affected by

some motor disability (Sellers and Donchin, 2006).

Overall, these studies conclude that the P300 speller

processor is an effective communication tool for peo-

ple who have lost or are losing their ability to write

or speak. However, it is still needed to improve the

usability of these BCI speller systems. Some factors,

such as the mental fatigue induced by a long use (Ke-

ceci et al., 2006; Murata and Uetake, 2001), the sus-

tained attention at a symbol on screen (Mangun and

Buck, 1998), the user motivation (Kececi et al., 2006;

Kleih et al., 2010) or his/her frustration due to a mis-

take (Kleih et al., ress) can inﬂuence the amplitude

and latency of the P300 component (See Polich and

Ron Angevin R. and da Silva-Sauer L..

Proposal of a P300-based BCI Speller using a Predictive Text System.

DOI: 10.5220/0004612300350040

In Proceedings of the International Congress on Neurotechnology, Electronics and Informatics (NEUROTECHNIX-2013), pages 35-40

ISBN: 978-989-8565-80-8

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

Kok, 1995, for a review). In this regard, the inﬂuence

on performance of the temporal and spatial aspects

of the user interfaces of these systems is increasingly

drawing the attention of researchers (Lu et al., 2013;

McFarland et al., 2011; Shih et al., 2013)

This paper presents a study aimed at contributing

to this line of research. We propose a P300-based

BCI speller system that uses the T9 interface devel-

oped for mobile phones (Grover et al., 1998). Specif-

ically, the virtual keyboard of this interface consisted

of a 4 x 3 matrix that was based on that of the T9

interface. Compared to the 6 x 6 matrix used by con-

ventional P300-based spellers (Donchin et al., 2000),

using a smaller matrix was expected not only to lead

to a higher accuracy, but also to reduce the time re-

quired to select a character, since the number of rows

and columns that had to be ﬂashed to detect it was

reduced. On the other hand, we provided this speller

with a text predictive system because selecting some

characters in a 4 x 3 matrix would demand more than

one choice from the user. Such a system would render

those extra choices hardly unnecessary and thus in-

crease the user’s writing speed (as discussed by Dun-

lop and Crossan (2000)). To validate the effectiveness

of the proposed BCI, we compared it with two adap-

tations of the classical Farwell and Donchin speller in

terms of the time needed to write a target sentence.

2 METHOD

2.1 Participants, Procedure and Data

Acquisition

Three healthy subjects (one man, aged 36, and two

women, aged 22 and 25) took part in the experiment.

None of them had previous experience with BCI sys-

tems. The experiment was divided into three sessions.

The objective of each session was to evaluate one of

the proposed interfaces, namely the two adaptations

of the classical speller (i.e., the Spellermod and the

SpellermodPred) and a speller based on the T9 inter-

face (i.e., the SpellerT9). A description of these three

interfaces will be done in the following sections.

Each participant carried out the three sessions in

different days. To avoid effects due to learning, the

order in which participants completed the three ses-

sions was randomized.

Each session was divided into two phases: a ﬁrst

one for calibration purposes and a second one to eval-

uate the interface. Participants were instructed to

silently count how many times the selected symbol

was intensiﬁed (i.e., ﬂashed). In the calibration phase,

they were asked to copy ﬁve sequences of three to four

characters. In the evaluation phase, they were asked

to write the sentence “Experiencia BCI en la Universi-

dad de M

alaga” (i.e., BCI experience at the University

of M

alaga).

As we wanted to compare the three interfaces in

terms of the writing speed their users could achieve,

the values of all the temporary parameters related to

the selection of a symbol (i.e., a letter) were equal

across them. These values were based on those used

by Donchin et al. (2000). Speciﬁcally, each row and

column was randomly ﬂashed 10 times. Therefore,

each character was randomly intensiﬁed 20 times.

The duration of each ﬂash was 125 ms and the inter-

stimulus interval (ISI) between ﬂashes was also 125

ms. There was a pause of 2 s after each sequence of

ﬂashes (i.e., after a character had been selected) and

also in the beginning of each trial. It is important to

notice that the duration of a sequence of ﬂashes de-

pended on the size matrix, so that bigger matrices en-

tailed longer sequences.

Eight-channel EEG data were recorded at the elec-

trodes Fz, Cz, Pz, Oz, P3, P4, PO7 and PO8—

according to the 10/20 international system—, with

FPz as reference and the left mastoid as ground. Data

were acquired by a biosignal ampliﬁer (g.BSamp,

Guger Technologies) with a sampling rate of 256 Hz

and a 12-bit resolution data acquisition NI-USB-6210

(National Instruments) card.

2.2 System Description

The three P300-based spellers were implemented

through the BCI2000 platform (Schalk et al., 2004).

The ﬁrst speller, called Spellermod, was an adapta-

tion of the classical Farwell and Donchin speller. Un-

like this, which consisted of a 6 x 6 symbol matrix,

the Spellermod presented a 7 x 6 virtual keyboard, in

which additional symbols were included. We will re-

fer to the Spellermod as the reference speller. The

second speller, called SpellermodPred, was identical

to the previous one, save for the fact that the Speller-

modPred included a text predictive system in order

to reduce the time needed to write. Finally, the third

speller, called SpellerT9, consisted of a small 4 x 3

symbol matrix and included the same text predictive

system used in the SpellermodPred. Next, we give a

better description of these interfaces.

2.2.1 Spellermod

The Spellermod used a 7 x 6 matrix (See Figure 1).

The last row contained only two special characters,

BORRAR and ESPACIO, for deleting a character and

introducing a blank space, respectively. Therefore,

NEUROTECHNIX2013-InternationalCongressonNeurotechnology,ElectronicsandInformatics

in case of a spelling error, the user was able to erase

the last letter by selecting the special character BOR-

RAR. Besides, the digit zero was also included in the

sixth row.

Figure 1: Interface of the Spellermod.

The characters that the user had selected were dis-

played inside a text box situated at the bottom of the

screen. According to the temporary parameters de-

scribed in section 2.1, the time needed to select a sym-

bol through this interface was 34.5 s—the initial 2 s

of pause plus the duration of ﬂashing seven rows and

six columns 10 times each during 125 ms with an ISI

of 125 ms—.

2.2.2 SpellermodPred

Although the main objective of this study was to com-

pare the SpellerT9 with the reference speller (i.e., the

Spellermod), we considered very interesting to as-

sess the performance of a classical speller when it in-

cluded a text predictive system. So, we developed the

SpellermodPred speller (See Figure 2). Its interface

was like the Spellermod’s 7 x 6 matrix, only now it in-

cluded the special character VALIDAR. Since the ma-

trix dimensions were the same as in the Spellermod,

the time needed to select a symbol using the Speller-

modPred was also 34.5 s.

The characters (i.e., letters) that the user selected

were displayed in a text box below the symbol ma-

trix. These characters were fed into the text predic-

tor, which in turn displayed the most frequently used

Spanish word starting with those letters in an addi-

tional text box, just above the former. For example, if

the subject had selected the letter G and then the let-

ter O, the predictor would suggest the word gobierno

(i.e., the Spanish word for government).

In case the suggested word were the one the user

wanted to write, he/she could conﬁrm it by selecting

the special character VALIDAR. As a consequence,

the rest of the word was added to the letters he/she

had already spelled, allowing the user to write a new

Figure 2: Interface of the SpellermodPred.

word. Importantly, validating a word also added a

blank space at the end of it.

The list of the most frequently used Spanish words

was obtained from the Royal Spanish Academy. In

case of wanting to write in a different language, it

would only be necessary to change the dictionary used

by the text predictor.

2.2.3 SpellerT9

The SpellerT9 used the T9 predictive text system de-

veloped for mobile phones. Its interface (See Figure

3) consisted of a 4 x 3 matrix of elements or keys,

in which only eight keys—the ones corresponding to

the numbers 2 to 9—were used for spelling. Each of

those keys corresponded to three to four letters. Im-

portantly, as the matrix size was 4 x 3, the time needed

to select a symbol in the SpellerT9 was just 19.5 s—

the initial 2 s of pause plus the duration of ﬂashing

four rows and three columns 10 times each during 125

ms with an ISI of 125 ms—, less than that of the other

spellers.

Figure 3: Interface of the SpellerT9.

As in the T9 interface of mobile phones, the user

of the SpellerT9 had to select a sequence of the men-

ProposalofaP300-basedBCISpellerusingaPredictiveTextSystem

tioned eight keys in order to write a word. As the user

selected a sequence of those keys, the most favorable

sequence of letters corresponding to those keys was

displayed in the text box at the bottom of the inter-

face. To increase writing speed, the SpellerT9 used

a text predictive system, exactly like the one used in

the SpellermodPred. This system identiﬁed the four

more frequently used Spanish words starting with the

suggested sequence of letters. The most frequent of

them was displayed in the text box beside the key-

board and the other three were kept in memory. For

example, after selecting the sequence of keys 2272,

the four predicted words would be casa, cara, capa,

and basa, being casa—the most frequently used—the

one displayed beside the keyboard. In case that were

the word that the user wanted to write, he/she could

conﬁrm it by selecting the key 1 (i.e., validar). This,

as in the SpellermodPred, added a blank space at the

end of the word. Otherwise, the user could select

the key c (i.e., cambiar) to switch to a new interface

that displayed the four predicted words (See Figure

4). The user could then select one of them or alterna-

tively go back to the virtual keyboard by selecting the

left arrow key (i.e., volver).

Figure 4: 4 x 3 matrix of the SpellerT9 showing the four

most frequently used Spanish words associated with the se-

quence of keys 2272.

Like in the other two spellers, users could delete

characters—by selecting the key x (i.e., borrar)—or

include blank spaces—by selecting the key 0 (i.e., es-

pacio)—. The SpellerT9 made it also possible for

users to write digits. To do so, he/she had to select

the key c prior to selecting any letter of a new word.

In this way, the user could select the keys 0 to 9 to

write the corresponding digits. Once the desired dig-

its had been written, the user had to select the key c

again to continue writing words.

3 RESULTS AND DISCUSSION

The spelling times needed by each of the three partic-

ipants to write the target sentence (i.e., “Experiencia

BCI en la Universidad de M

alaga”) through each of

the three compared spellers are shown in Table 1. We

did not discount the times associated with errors. The

table also displays the minimum required times that

a perfect user (i.e., a user that selected each charac-

ter at the ﬁrst attempt) would need to write the target

sentence in each case.

As we can see, the three participants wrote nearly

all words faster with the SpellerT9 than with any of

the other two spellers. Even participant 2, who did

not gain too much control of any of the three inter-

faces, achieved the shortest spelling time for the full

target sentence when using the SpellerT9. As for

the other two participants, they made very few mis-

takes, as their spelling times were slightly longer than

the minimum required, regardless of the speller they

used.

The average spelling time for the full target sen-

tence was 650.5 s, 1115.5 s and 1849.5 s when using

the SpellerT9, the SpellermodPred and the Speller-

mod, respectively. The average spelling times for

each word when using each of the three interfaces are

shown in Figure 5. These results strongly suggest that

SpellerT9 system was the fastest writing interface of

all three.

Furthermore, our study supports the usefulness of

a text predictive system irrespective of matrix size, as

participants achieved shorter spelling times when us-

ing the SpellermodPred than when using to conven-

tional interface of the Spellermod.

4 CONCLUSIONS

The objective of this study was to propose and to val-

idate a new P300-based BCI speller system aimed at

increasing the user’s writing speed. The proposed

speller (i.e., SpellerT9) was based on the T9 interface

developed for mobile phones. It presented a 4 x 3 vir-

tual keyboard and incorporated the T9 text predictive

system.

Although the study has been carried out with a

very small sample, our preliminary results suggest

that the SpellerT9 leads to lower spelling times than

those of an adaptation of the classical Farwell and

Donchin speller (i.e., Spellermod), even when the lat-

ter interface is improved by adding a text predictor

(i.e., SpellermodPred). The three participants in our

experiment achieved the lowest spelling times when

they wrote a target sequence through the SpellerT9

NEUROTECHNIX2013-InternationalCongressonNeurotechnology,ElectronicsandInformatics

Table 1: Spelling times needed by the three participants to write each word of the target sentence as a function of the speller

they used.

Time (s) for each word

Speller “Experiencia” “BCI” “en” “la” “Universidad” “de” “M

alaga” Total time (s)

Participant 1

Spellermod 414 138 103.5 103.5 621 103.5 276 1759.5

SpellermodPred 172.5 138 103.5 69 207 103.5 207 1000.5

SpellerT9 78 78 58.5 39 97.5 39 117 507

Participant 2

Spellermod 483 345 310.5 172.5 483 103.5 276 2173.5

SpellermodPred 138 103.5 103.5 517.5 172.5 69 345 1449

SpellerT9 78 156 58.5 78 214.5 78 177 840

Participant 3

Spellermod 414 138 103.5 103.5 414 172.5 270 1615.5

SpellermodPred 138 138 103.5 69 172.5 69 207 897

SpellerT9 78 117 78 39 97.5 39 156 604.5

A perfect user

Spellermod 414 138 103.5 103.5 414 103.5 270 1546.5

SpellermodPred 138 103.5 103.5 69 172.5 69 207 862.5

SpellerT9 78 78 58.5 39 97.5 39 117 507

Figure 5: Average spelling times needed to write each word of the target sentence using each of the three spellers.

and the highest when they used the Spellermod. Be-

sides, the writing times of each subject indicate that

the SpellerT9 is not more difﬁcult to use than any of

the other two spellers, as participants performed just

slightly worse than a perfect user.

Nevertheless, these conclusions should be sup-

ported by extending this experiment to a greater sam-

ple of participants, so that statistical tests of signiﬁ-

cance could be performed, and also by obtaining ad-

ditional direct measures of user performance. More-

over, sentences of different complexity could be used

to compare the usability of these interfaces.

ACKNOWLEDGEMENTS

This work was partially supported by the Innovation,

Science and Enterprise Council of the Junta de An-

daluc

ıa (Spain), project P07-TIC-03310, the Span-

ish Ministry of Science and Innovation, project TEC

2011-26395 and by the European fund ERDF.

ProposalofaP300-basedBCISpellerusingaPredictiveTextSystem

REFERENCES

Bianchi, L., Sami, S., Hillebrand, A., Fawcett, I.,

Quitadamo, L., and Seri, S. (2010). Which phys-

iological components are more suitable for visual

ERP based brain-computer interface? A preliminary

MEG/EEG study. Brain Topography, 23(2):180–185.

Birbaumer, N. (2006). Breaking the silence: Brain-

computer interfaces (BCI) for communication and

motor control. Psychophysiology, 43(6):517–532.

Donchin, E., Spencer, K., and Wijesinghe, R. (2000). The

mental prosthesis: Assessing the speed of a P300-

based brain-computer interface. IEEE Transactions

on Rehabilitation Engineering, 8(2):174–179.

Dunlop, M. D. and Crossan, A. (2000). Predictive text entry

methods for mobile phones. Personal Technologies,

4(2-3):134–143.

Farwell, L. and Donchin, E. (1988). Talking off the top of

your head: toward a mental prosthesis utilizing event-

related brain potentials. Electroencephalography and

Clinical Neurophysiology, 70(6):510 – 523.

Grover, D. L., King, M. T., and Kushler, C. A. (1998). Re-

duced keyboard disambiguating computer. U.S. Patent

No. 5,818,437. Washington, DC: U.S. Patent and

Trademark Ofﬁce.

Kececi, H., Degirmenci, Y., and Atakay, S. (2006). Ha-

bituation and dishabituation of P300. Cognitive and

Behavioral Neurology, 19(3):130–134.

Kleih, S., Nijboer, F., Halder, S., and K

ubler, A. (2010).

Motivation modulates the P300 amplitude during

brain-computer interface use. Clinical Neurophysiol-

ogy, 121(7):1023 – 1031.

Kleih, S. C., Kaufmann, T., Hammer, E., Pisotta, I., Pi-

chiorri, F., Riccio, A., Mattia, D., and K

ubler, A. (in

press). Motivation and SMR-BCI: Fear of failure af-

fects BCI performance. In Proceedings of the Fifth

International BCI Meeting.

Krusienski, D., Sellers, E., McFarland, D., Vaughan, T.,

and Wolpaw, J. (2008). Toward enhanced P300

speller performance. Journal of Neuroscience Meth-

ods, 167(1):15 – 21.

Lu, J., Speier, W., Hu, X., and Pouratian, N. (2013). The

effects of stimulus timing features on P300 speller

performance. Clinical Neurophysiology, 124(2):306

– 314.

Mak, J. and Wolpaw, J. (2009). Clinical applications of

brain-computer interfaces: Current state and future

prospects. IEEE Reviews in Biomedical Engineering,

2:187–199.

Mangun, G. R. and Buck, L. A. (1998). Sustained visual-

spatial attention produces costs and beneﬁts in re-

sponse time and evoked neural activity. Neuropsy-

chologia, 36(3):189 – 200.

McFarland, D. J., Sarnacki, W. A., Townsend, G., Vaughan,

T., and Wolpaw, J. R. (2011). The P300-based brain-

computer interface (BCI): Effects of stimulus rate.

Clinical Neurophysiology, 122(4):731 – 737.

Murata, A. and Uetake, A. (2001). Evaluation of mental fa-

tigue in human-computer interaction - Analysis using

feature parameters extracted from event-related poten-

tial. In 10th IEEE International Workshop on Robot

and Human Interactive Communication, pages 630–

635.

Polich, J. and Kok, A. (1995). Cognitive and biological de-

terminants of P300: An integrative review. Biological

Psychology, 41(2):103 – 146.

Schalk, G., McFarland, D., Hinterberger, T., Birbaumer, N.,

and Wolpaw, J. (2004). Bci2000: A general-purpose

brain-computer interface (BCI) system. IEEE Trans-

actions on Biomedical Engineering, 51(6):1034–

1043.

Sellers, E. W. and Donchin, E. (2006). A P300-based

brain–computer interface: Initial tests by ALS pa-

tients. Clinical Neurophysiology, 117(3):538 – 548.

Sellers, E. W., Krusienski, D. J., McFarland, D. J., Vaughan,

T. M., and Wolpaw, J. R. (2006). A P300 event-related

potential brain–computer interface (BCI): The effects

of matrix size and inter stimulus interval on perfor-

mance. Biological Psychology, 73(3):242 – 252.

Shih, J. J., Townsend, G., Krusienski, D. J., Shih,

K. D., Shih, R. M., Heggeli, K., Paris, T., and

Meschia, J. F. (2013). Comparison of checker-

board P300 speller vs. row-column speller in

normal elderly and aphasic stroke population.

Paper presented at the Fifth International Brain-

Computer Interface Meeting, Asilomar Confer-

ence Grounds, Paciﬁc Grove, CA. Retrieved from

http://castor.tugraz.at/doku/BCIMeeting2013/020.pdf.

Wang, C., Guan, C., and Zhang, H. (2005). P300 brain-

computer interface design for communication and

control applications. In 27th Annual International

Conference of the Engineering in Medicine and Bi-

ology Society IEEE-EMBS 2005, pages 5400–5403.

Wolpaw, J. R., Birbaumer, N., McFarland, D. J.,

Pfurtscheller, G., and Vaughan, T. M. (2002). Brain-

computer interfaces for communication and control.

Clinical Neurophysiology, 113(6):767–791.

NEUROTECHNIX2013-InternationalCongressonNeurotechnology,ElectronicsandInformatics