An efficient word typing P300-BCI system using a modified T9 interface and random forest classifier

7
An efcient word typing P300-BCI system using a modied T9 interface and random forest classier Faraz Akram, Seung Moo Han, Tae-Seong Kim n Department of Biomedical Engineering, Kyung Hee University, Yongin-si, Republic of Korea article info Article history: Received 31 July 2014 Accepted 25 October 2014 Keywords: Brain computer interface BCI EEG P300 speller Word typing paradigm Dictionary Random forest Humancomputer interaction abstract Background: A typical P300-based spelling brain computer interface (BCI) system types a single character with a character presentation paradigm and a P300 classication system. Lately, a few attempts have been made to type a whole word with the help of a smart dictionary that suggests some candidate words with the input of a few initial characters. Methods: In this paper, we propose a novel paradigm utilizing initial character typing with word suggestions and a novel P300 classier to increase word typing speed and accuracy. The novel paradigm involves modifying the Text on 9 keys (T9) interface, which is similar to the keypad of a mobile phone used for text messaging. Users can type the initial characters using a 3 3 matrix interface and an integrated custom-built dictionary that suggests candidate words as the user types the initials. Then the user can select one of the given suggestions to complete word typing. We have adopted a random forest classier, which signicantly improves P300 classication accuracy by combining multiple decision trees. Results and discussion: We conducted experiments with 10 subjects using the proposed BCI system. Our proposed paradigms signicantly reduced word typing time and made word typing more convenient by outputting complete words with only a few initial character inputs. The conventional spelling system required an average time of 3.47 min per word while typing 10 random words, whereas our proposed system took an average time of 1.67min per word, a 51.87% improvement, for the same words under the same conditions. & 2014 Elsevier Ltd. All rights reserved. 1. Introduction Brain computer interfaces (BCIs) are systems that provide direct communication with a computer through signals generated by the brain. P300 event related potentials are a commonly used EEG signal for BCI known as a P300-BCI. A typical P300-based character-spelling BCI system consists of a stimulus-presenting paradigm and a classier. The most widely adopted paradigm for character spelling was rst introduced by Farwell and Donchin [1]. Their speller paradigm involved a 6 6 matrix of characters and numbers in which each row or column is randomly intensied. A user looks at a target character and the resulting intensication of either the row or column containing the target character evokes P300. Classication of the row and column containing the evoked P300 leads to the recognition of the target character. Most later research followed the same scheme and focused on improving classication accuracy and speed. In recent years, there is a growing interest in designing more efcient paradigms. Multiple attempts have been made to modify the Farwell and Donchin (FD) paradigm. These modications include using different matrix sizes [2], changing visual aspects such as color, symbol dimensions, and distance between the symbols [3,4], single character ipping instead of row and column intensication [5], and using a checkerboard paradigm [6]. Some works suggested the potential of region-based [7] or submatrix- based [8] paradigms. A common drawback of all these conventional P300 spelling systems is the requirement to spell out each character of a word one at a time. This spelling process is slow and can take several minutes to nish a single word. A new paradigm is needed so a user can type a whole word with less typing time. Ryan et al. [9] proposed a predictive spelling scheme to improve the word-typing speed. Their proposed predictive spelling increased the typing speed, but decreased the classication accuracy because of higher workload in their displays. Kaufmann et al. [10] also used a predictive speller with some commonly used German words. Their proposed paradigm showed six suggested words in the rst column of the matrix. Their results also indicated that predictive spelling could signicantly decrease spelling time. Ahi et al. [11] used a dictionary-driven P300 speller. They integrated a custom Contents lists available at ScienceDirect journal homepage: www.elsevier.com/locate/cbm Computers in Biology and Medicine http://dx.doi.org/10.1016/j.compbiomed.2014.10.021 0010-4825/& 2014 Elsevier Ltd. All rights reserved. n Correspondence to: Dept. of Biomedical Engineering, College of Electronics and Information, Kyung Hee University, 1 Seocheon-dong, Giheung-gu, Yongin-si 446-701, Gyeonggi-do, Republic of Korea. Tel.: þ82 31 201 3731; fax: þ82 31 201 3666. E-mail address: [email protected] (T.-S. Kim). Computers in Biology and Medicine 56 (2015) 3036

Transcript of An efficient word typing P300-BCI system using a modified T9 interface and random forest classifier

An efficient word typing P300-BCI system using a modified T9interface and random forest classifier

Faraz Akram, Seung Moo Han, Tae-Seong Kim n

Department of Biomedical Engineering, Kyung Hee University, Yongin-si, Republic of Korea

a r t i c l e i n f o

Article history:Received 31 July 2014Accepted 25 October 2014

Keywords:Brain computer interfaceBCIEEGP300 spellerWord typing paradigmDictionaryRandom forestHuman–computer interaction

a b s t r a c t

Background: A typical P300-based spelling brain computer interface (BCI) system types a single characterwith a character presentation paradigm and a P300 classification system. Lately, a few attempts havebeen made to type a whole word with the help of a smart dictionary that suggests some candidate wordswith the input of a few initial characters.Methods: In this paper, we propose a novel paradigm utilizing initial character typing with wordsuggestions and a novel P300 classifier to increase word typing speed and accuracy. The novel paradigminvolves modifying the Text on 9 keys (T9) interface, which is similar to the keypad of a mobile phoneused for text messaging. Users can type the initial characters using a 3�3 matrix interface and anintegrated custom-built dictionary that suggests candidate words as the user types the initials. Then theuser can select one of the given suggestions to complete word typing. We have adopted a random forestclassifier, which significantly improves P300 classification accuracy by combining multiple decision trees.Results and discussion: We conducted experiments with 10 subjects using the proposed BCI system. Ourproposed paradigms significantly reduced word typing time and made word typing more convenient byoutputting complete words with only a few initial character inputs. The conventional spelling systemrequired an average time of 3.47 min per word while typing 10 random words, whereas our proposedsystem took an average time of 1.67 min per word, a 51.87% improvement, for the same words under thesame conditions.

& 2014 Elsevier Ltd. All rights reserved.

1. Introduction

Brain computer interfaces (BCIs) are systems that providedirect communication with a computer through signals generatedby the brain. P300 event related potentials are a commonly usedEEG signal for BCI known as a P300-BCI. A typical P300-basedcharacter-spelling BCI system consists of a stimulus-presentingparadigm and a classifier. The most widely adopted paradigm forcharacter spelling was first introduced by Farwell and Donchin [1].Their speller paradigm involved a 6�6 matrix of characters andnumbers in which each row or column is randomly intensified. Auser looks at a target character and the resulting intensification ofeither the row or column containing the target character evokesP300. Classification of the row and column containing the evokedP300 leads to the recognition of the target character. Most laterresearch followed the same scheme and focused on improvingclassification accuracy and speed.

In recent years, there is a growing interest in designing moreefficient paradigms. Multiple attempts have been made to modifythe Farwell and Donchin (FD) paradigm. These modificationsinclude using different matrix sizes [2], changing visual aspectssuch as color, symbol dimensions, and distance between thesymbols [3,4], single character flipping instead of row and columnintensification [5], and using a checkerboard paradigm [6]. Someworks suggested the potential of region-based [7] or submatrix-based [8] paradigms.

A common drawback of all these conventional P300 spellingsystems is the requirement to spell out each character of a wordone at a time. This spelling process is slow and can take severalminutes to finish a single word. A new paradigm is needed so auser can type a whole word with less typing time. Ryan et al. [9]proposed a predictive spelling scheme to improve the word-typingspeed. Their proposed predictive spelling increased the typingspeed, but decreased the classification accuracy because of higherworkload in their displays. Kaufmann et al. [10] also used apredictive speller with some commonly used German words. Theirproposed paradigm showed six suggested words in the firstcolumn of the matrix. Their results also indicated that predictivespelling could significantly decrease spelling time. Ahi et al. [11]used a dictionary-driven P300 speller. They integrated a custom

Contents lists available at ScienceDirect

journal homepage: www.elsevier.com/locate/cbm

Computers in Biology and Medicine

http://dx.doi.org/10.1016/j.compbiomed.2014.10.0210010-4825/& 2014 Elsevier Ltd. All rights reserved.

n Correspondence to: Dept. of Biomedical Engineering, College of Electronics andInformation, Kyung Hee University, 1 Seocheon-dong, Giheung-gu, Yongin-si446-701, Gyeonggi-do, Republic of Korea.Tel.: þ82 31 201 3731; fax: þ82 31 201 3666.

E-mail address: [email protected] (T.-S. Kim).

Computers in Biology and Medicine 56 (2015) 30–36

dictionary into the classification system of their P300 speller forautomatic correction of misspellings. However, the dictionary wasused only for word correction and the user still had to spell everycharacter of a target word. Akram et al. [12] integrated a customdictionary with the conventional FD paradigm to give wordsuggestions to the users, speeding up the typing speed. Höhneet al. [13] used a German-language Text on Nine keys (T9) systemwith an auditory event related potential-based speller. The userspelled on a 3�3-scheme with audio stimuli and suggestionswere shown after the user spelled a complete word. Word typingusing their system was slow because when typing a five-letterword, the user must go through five character inputs and at leasttwo additional inputs: one for switching to the control mode andthe other for the selection of a target word.

To improve classification accuracy, many techniques for P300extraction and classification have been used, including supportvector machine (SVM) [14,15], linear discriminant analysis (LDA)[16], neural network [17], independent component analysis (ICA)[18], and, lately, constrained independent component analysis(cICA) [19], and adaptive neural networks [20].

In this study, we propose a modified T9 interface with adictionary to give word suggestions as the user types. In the mobiletext messaging system, the T9 is used to compose text messages athigh speeds using a keypad with twelve keys. Each of the eight keysis associated with several unique characters and the dictionary isused to suggest words according to the sequence of keys a userpresses. Four other keys are used for other functions. In this study,we have modified it to use only nine keys: eight keys are associatedwith characters and one key for a delete function for errors. Wordtyping is done in two steps. A user spells a few characters of a targetword using the 3�3 matrix of the modified T9 keys, and then thedictionary module suggests some words starting with the spelledcombination of keys. Finally, the user can select one out of the givensuggestions using the same 3�3 matrix. The suggested paradigmreduces the typing time significantly.

Based on the analysis of the previously used classifiers, wepropose a random forest (RF) classifier for P300 detection. RF hasbeen well developed in many fields, but is relatively unknown inthe field of BCI. RF is a machine learning technique developed byBreiman [21]. RF consists of a number of decision trees where eachtree is grown independently and it casts a vote in favor of one classand the class with a majority vote wins.

We have tested our P300 words typing BCI with 10 healthysubjects and each subject was instructed to type 10 randomlyselected words. Our results show that the proposed P300 wordtyping paradigm takes an average of 1.67 min to type a word,whereas the conventional paradigm required 3.47 min for thesame words. The proposed scheme reduces the word-typing timeby 51.87%, thereby increasing the information transfer rateconsiderably.

2. Methodology

2.1. The conventional P300 speller

The conventional Farwell and Donchin (FD) speller consists of a6�6 matrix of characters and numbers in which each row orcolumn flashes randomly. A user is asked to focus on a targetcharacter and silently count the number of times of the targetcharacter flashes. P300 is elicited when the row or columncontaining the target character is flashed. A classifier detects onerow and one column with P300, thereby identifying the targetcharacter. Classification results are shown to the user as feedback.The user completes word typing by spelling every character of the

target word. SVM is the most commonly used technique for P300classification.

2.2. Overview of our proposed P300-based word typing

In our word typing BCI system, we used a modified T9paradigm with a word suggestion mechanism to reduce typingtime. Word typing is done in two steps: spelling initial charactersusing a modified T9 interface and selecting from the suggestedwords. When a session starts, a user starts to spell the initialcharacters of a target word using a T9 paradigm that consists of a3�3 matrix of keys in which each key is associated with severalcharacters and each key is randomly intensified. The user focuseson the key containing the target character. RF is used for P300classification. The classifier identifies the spelled key and theclassification results are fed into the dictionary module. Thedictionary module searches for the words starting with prefixesmade by the classified combinations of the keys. If the number ofwords is less than a threshold (nine, in our case), a suggestionscreen is shown to the user and the user is asked to select onefrom those suggestions. If the number of suggestions is greaterthan the threshold, the user continues to type the next character ofthe target word. Fig. 1 shows the flow diagram of the proposedsystem.

2.3. Word typing paradigm

T9 predictive spelling is used in mobile phones for fastercomposition of text messages using a keypad of 12 keys [22].Our modified T9 word typing paradigm consists of a 3�3 matrixof nine keys. Eight keys contain three or four characters and theninth key is for a ‘delete’ function used to delete the last spelledcharacter in case of a spelling mistake, as shown in Fig. 2(a). Theinterface is similar to the keypad of a mobile phone, where textmessages are written on a keypad with 12 keys and a T9 predictivedictionary is used for faster typing. We used only nine keys andeach of nine keys flashes randomly with a 100-ms intensificationtime and a 75-ms blank time between intensifications. A userfocuses on the key containing the target character and silently

Fig. 1. A flow diagram for the proposed word typing system.

F. Akram et al. / Computers in Biology and Medicine 56 (2015) 30–36 31

counts the number of times the target key flashes. Detection ofone key-containing P300 identifies the target key. Each keycontains three or four characters. The dictionary module generatesa list of words starting with the characters in the spelled key. Theuser continues to write next characters of the word until thenumber of words in the suggestion list becomes less than thethreshold, at which point the word suggestion screen is shown tothe user (Fig. 2(b)) and the user is asked to select one of thesuggested words. The user inputs the number corresponding tothe target word as shown in Fig. 2(c) to complete typing the word.

The final typed word is shown to the user (Fig. 2(d)) and thenthe user is returned to the main paradigm to type the next word. Ifa spelling mistake occurs during the word selection step, the usercan select the ‘delete’ key and start inputting again from the wordselection step.

Table 1 shows the time specification of the conventional para-digm as used in the BCI competition III dataset II [23] and theproposed T9 paradigm.

2.4. Dictionary module

We used a custom-built dictionary of the 1000 most commonlyused English words. The dictionary is implemented in the form ofa ternary search tree (TST). TST is a kind of prefix tree datastructure that can be used to find words having a given prefix. Thedictionary gives suggestions based on the combinations of prefixesthat are formed by the user typed keys. For example, if a userselects ‘Key-5’ while typing the first character of a target word,‘Key-5’ contains three characters (M, N, and O): the dictionary willsuggest all words starting with these three characters. Afterwards,

if the user selects ‘Key-5’ again for the second character, there canbe a total of nine prefixes (MM, MN, MO, NM, NN, NO, OM, ON, andOO). Then the dictionary module keeps only the words that startwith these prefixes. The number of words decreases in thesuggestion list after the user types each character and, once thenumber falls below the threshold, a suggestion screen is shown tothe user. The user chooses one of the suggestions to completeword typing.

2.5. Classification

Decision tree classification techniques have been used success-fully for a wide range of classification problems. Decision treesoffer substantial advantages for classification problems because oftheir flexibility, simplicity, and computational efficiency. The majordrawback of decision tree classification is its low accuracy. Therehave been many efforts to improve this classification accuracy. Oneof the best ways to improve the accuracy of decision tree-based

Fig. 2. Our word typing paradigm. (a) The first step for spelling the initial characters of a target word (in this case ‘mountain’), (b) a word suggestion screen with the typedkeys (in this case 557), (c) the step of word selection from nine suggested words, and (d) a display showing the final word.

Table 1Timing information for the paradigms.

Conventionalparadigm

Proposedparadigm

Intensification time 100 ms 100 msBlank time betweenintensifications

75 ms 75 ms

Total stimuli 12 9Character repeat 15 15Blank time between characters 2.5 s 2.5 s

F. Akram et al. / Computers in Biology and Medicine 56 (2015) 30–3632

algorithms is to use an ensemble of trees. The theory of ensembleclassification allows us to combine a group of weak classifiers toform a strong classifier.

To detect P300s, we have adopted a RF classifier. RF, anensemble learning technique introduced by Breiman [21], is apowerful classifier and has been used effectively in many fields,but is relatively unknown in the field of BCI. The underlyingconcept of RF is the combination of multiple independent decisiontrees with the resulting decisions casting votes for a class. RFcreates multiple distinct decision trees and the class that ispredicted by majority of the trees becomes the final output. Forexample, if there are 10 trees and 7 of them vote ‘yes’ and 3 vote‘no’ in the binary classification of the question “is there P300 ornot,” then the final classification result of the forest will be ‘yes’based on the majority vote.

Each tree in the forest is trained with a different subset of thetraining data, specifically, a random sample drawn with replace-ment from the training set. Therefore, some data may be usedmore than once in the training of classifiers while others mightnever be used. This increases classifier stability and robustnesswhen facing slight variations in data. Some randomness is addedat the node level in each tree. At every node, a random subset ofvariables is selected and only those variables are used to find thebest split for the node. This randomness removes the correlationbetween the trees of the forest. RF can improve the classificationaccuracy and can be quite robust. We have compared the perfor-mance of RF against the widely used SVM [14].

In the training session, users were asked to spell 10 randomlyselected characters. This data was used to train both classifiers. Weapplied a band-pass filter with cutoff frequencies of 0.1 Hz and25 Hz. Then, epochs of 600 ms after the stimulus onset wereextracted. A total of eight channels were used for classification.The epochs from the eight channels were concatenated to createfeature vectors used to train classifiers.

3. Results

3.1. Experimental setup

This study was conducted using 10 healthy male subjects.Subjects did not have a history of any neurological brain diseasesand had normal or corrected vision. All participants in our studyprovided written informed consent in accordance with the Insti-tutional Review Board of Kyung Hee University.

The EEG data was acquired through a 32-channel BrainAmp MRamplifier with a sampling frequency 250 Hz. Electrodes were placedaccording to the 10–20 international standard. Each subject partici-pated in training and testing sessions. In the training session, eachparticipant was instructed to spell 10 randomly selected characters;the resulting data was used to train the classifier. In the test session,

each subject typed 10 randomly selected words. An example set oftarget words is as follows: “window,” “wings,” “beautiful,” “state-ment,” “understand,” “poem,” “stone,” “piece,” “smell,” and “answer.”Table 2 shows the words typed by 10 subjects during the test session.

3.2. Waveform morphologies

Event-related potentials for target and non-target stimuliaveraged across all subjects are shown in Fig. 3. Averaged wave-forms of 800 ms after the stimulus onset are shown from the eightchannels. We can clearly see P300 for the target cases in all of thechannels used.

3.3. Theoretical timing comparison

Using the information shown in Table 1, we computed the timerequired to spell one character using both the conventional FD andthe proposed paradigms. For the conventional paradigm, we had180 intensifications in total (12 stimuli�15 repetitions) and175 ms (intensification timeþblank time between intensifications)per intensification. Therefore, the time required to spell a singlecharacter using the conventional paradigm is 34 s ([180�175 ms]þ2.5 s blank time between characters). The proposed paradigm,on the other hand, uses 26.125 s (i.e., [135�175 ms]þ2.5 s) percharacter, as the total number of intensifications per character is135 (9 stimuli�15 repetitions). With these numbers, we com-puted the time required by the conventional paradigm to spell thetarget words and compared it to the time required by theproposed paradigm for the same words. Table 3 shows the timerequired to spell these words using the conventional and proposedparadigm considering no error.

The conventional spelling required an average time of 3.47 minper word while the proposed word typing scheme required onlyan average typing time of 1.67 min per word for the same targetwords, decreasing the typing time by 51.87%.

3.4. Experimental results

In practical and experimental settings, spelling mistakes canoccur and the subjects have to correct the spelling mistakes usingthe ‘delete’ key and then typing the correct letter or word again.Correcting misspellings takes extra time. Table 4 shows the actualtime taken by the proposed method to type the target words for all10 subjects. Note that the lapsed time for the proposed schemeincludes the error correction time.

3.5. Classification accuracy using RF

We also compared the offline classification results of RF againstSVM for different numbers of stimulus repetitions. Our resultsshow that the RF classifier can better classify P300s even with

Table 2Target words.

Word number S1 S2 S3 S4 S5 S6 S7 S8 S9 S10

1 Window Winter Complete Young Water Watch Between After Snow Correct2 Wings Fruit Movement Happy Song Fruit Hair Experience Study Beautiful3 Beautiful Mountain Huge Water Signal Little Read Second Mother Experience4 Statement King Notice Scientist North Equal Temperature Express Reason Against5 Understand Information Uncle Summer Radio Hard Position Equation King Underline6 Poem Africa Whole Story Reason Record Second Little Better Wind7 Stone List Garden Hundred School Hair After Mountain Human Exercise8 Piece Explain Stone Human Thousand Between Teacher Position Paragraph Explain9 Smell Corner Piece Example Garden Light Example Game Symbol Notice10 Answer Hard Table Today Hand Warm Science Soldier Opposite Stood

F. Akram et al. / Computers in Biology and Medicine 56 (2015) 30–36 33

fewer repetitions. Fig. 4 shows the average classification accuracyfor both SVM and RF for the first 5 subjects and the overall averageof all 10 subjects. For higher number of repetitions, the accuracy ofSVM and RF are comparable, but as we reduce the number ofrepetitions, the accuracy of SVM decreases more rapidly than thatof RF. For fewer repetitions, the accuracy of RF is significantlyhigher than that of SVM.

4. Discussion and conclusion

The proposed method has a considerable advantage in accuracyover the conventional spelling method because in the proposedparadigm, the prior probability of the target is lower and accuracyis inversely proportional to the prior probability of the target [2].The conventional paradigm uses a 6�6 matrix with flashing rows

0 200 400 600 800

-15

-10

-5

0

5

10

15

Am

plitu

de Cz

0 200 400 600 800

-15

-10

-5

0

5

10

15

Pz

0 200 400 600 800

-15

-10

-5

0

5

10

15

P3

0 200 400 600 800

-15

-10

-5

0

5

10

15

P4

0 200 400 600 800

-15

-10

-5

0

5

10

15

O1

0 200 400 600 800

-15

-10

-5

0

5

10

15

Time (msecs)

O2

0 200 400 600 800

-15

-10

-5

0

5

10

15

P7

0 200 400 600 800

-15

-10

-5

0

5

10

15

P8

Fig. 3. Averaged amplitude of target (solid) and non-target (dashed) waveforms of the eight electrodes used.

Table 3Comparison of the time required by the conventional FD speller and our proposedword typing paradigm.

Word set Time required to spell the target words (min)

Conventional paradigm Proposed paradigm

S1 3.6371.2 1.7470.00S2 3.4671.24 1.6570.18S3 3.2370.76 1.6170.21S4 3.3470.78 1.6570.18S5 3.1270.67 1.7470.00S6 2.8970.56 1.6170.21S7 3.7471.17 1.6570.18S8 3.9170.98 1.7070.14S9 3.3470.9 1.6170.21S10 4.0871.06 1.7070.14Mean 3.4770.98 1.6770.16

Table 4Word typing time using the proposed paradigm.

Word number Word typing time using the proposed scheme (min)

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10

1 1.74 1.74 1.74 1.74 1.74 1.31 1.74 1.74 1.74 2.612 1.74 2.61 2.61 1.74 1.74 1.74 1.74 1.74 2.61 2.613 1.74 1.74 1.31 1.74 2.61 2.18 1.74 1.74 1.74 1.744 1.74 1.31 1.74 2.18 1.74 1.74 2.18 2.61 1.74 1.745 1.74 1.74 1.74 2.61 1.74 2.61 1.74 1.74 1.31 2.616 1.74 1.74 1.74 1.74 2.61 1.74 2.61 1.31 1.74 1.747 1.74 1.31 1.74 1.31 2.61 1.74 1.74 1.74 1.31 1.748 1.74 2.61 1.31 2.18 1.74 1.74 1.31 1.74 2.61 2.619 1.74 1.74 1.74 1.74 1.74 1.31 1.74 1.74 2.18 1.7410 1.74 1.74 1.31 1.74 1.74 1.31 1.74 2.61 1.31 1.74Mean 1.74 1.83 1.7 1.87 2 1.74 1.83 1.87 1.83 2.09Grand mean 1.8570.12

F. Akram et al. / Computers in Biology and Medicine 56 (2015) 30–3634

or columns. With two targets and 12 intensifications, the prob-ability of landing on the target is 1/6, whereas the proposedparadigm has a prior probability of landing on the target is 1/9.

The use of a modified T9 interface offers faster spelling than theconventional FD method, as the proposed method requires fewerintensifications per character. In a single trial, the proposedparadigm required nine intensifications, whereas the conventionalparadigm required twelve. The use of a dictionary to give wordsuggestions further reduces the overall word typing time as theusers can type a whole word by inputting a few initial charactersand then selecting the target word from the suggestions.

Also, in the proposed paradigm, there are fewer chances to makemistakes because a user only needs to spell a few characters to type aword. In the case of a spelling error, the user can spell the ‘delete’ keyand then input the correct character again. This error correction timewas also reduced in the proposed paradigm compared with theconventional paradigm. The conventional paradigm took 34 s to usethe ‘delete’ key and another 34 s to input the correct character. Thecost of one mistake in the conventional paradigm was 68 s. Theproposed paradigm reduced this cost to 52.25 s.

We used the RF classifier instead of the conventional classifica-tion method, SVM. One of the important advantages of RF is that itis less sensitive to outliers in the training data, which makes RFsuitable for cases where some of the features are mislabeled. Thiscan be one of the reasons for better accuracy compared to SVMwhen there are fewer repetitions (Fig. 4).

In this study, we proposed an efficient dictionary-based P300word typing BCI paradigm using the modified T9 interface, adictionary, and RF. Our results showed that the proposed methodof P300 word typing system not only increased the speed of typingwith higher accuracy, but also made the spelling task easier forusers by reducing the amount of work required. These could beimportant advantages for a practical BCI, especially for use byindividuals with disabilities.

Conflict of interest statement

None declared.

Acknowledgments

This work was supported by the National Research Foundationof Korea (NRF) grant funded by the Korean government (MSIP)(no. 2008-0061908).

References

[1] L.A. Farwell, E. Donchin, Talking off the top of your head: toward a mentalprosthesis utilizing event-related brain potentials, Electroencephalogr. Clin.Neurophysiol. 70 (1988) 510–528.

[2] B.Z. Allison, J.A. Pineda, ERPs evoked by different matrix sizes: Implications fora brain computer interface (BCI) system, IEEE Trans. Neural Syst. Rehabil. Eng.11 (2003) 110–113.

[3] M. Salvaris, F. Sepulveda, Visual modifications on the P300 speller BCIparadigm, J. Neural Eng. 6 (2009) 046011.

[4] K. Takano, T. Komatsu, N. Hata, Y. Nakajima, K. Kansaku, Visual stimuli for theP300 brain–computer interface: a comparison of white/gray and green/blueflicker matrices, Clin. Neurophysiol. 120 (2009) 1562–1566.

[5] Guan, C., Thulasidas, M., Wu, J., (2004). High performance p300 speller forbrain–computer interface, in: IEEE International Workshop on BiomedicalCircuits and Systems, Singapore, 13–16.

[6] G. Townsend, B.K. LaPallo, C.B. Boulay, D.J. Krusienski, G.E. Frye, C.K. Hauser,N.E. Schwartz, T.M. Vaughan, J.R. Wolpaw, E.W. Sellers, A novel P300-basedbrain–computer interface stimulus presentation paradigm: moving beyondrows and columns, Clin. Neurophysiol. 121 (2010) 1109–1120.

[7] R. Fazel-Rezai, k. Abhari, A region-based P300 speller for brain–computerinterface, Can. J. Electr. Comput. Eng. 34 (2009) 81–85.

[8] J.H. Shi, J.Z. Shen, Y. Ji, F.L. Du, A submatrix-based P300 brain–computerinterface stimulus presentation paradigm, J. Zhejiang Univ. Sci. C: Comput.Electron. 13 (2012) 452–459.

[9] D.B. Ryan, G.E. Frye, G. Townsend, D.R. Berry, G. Mesa, N.A. Gates, E.W. Sellers,Predictive spelling with a P300-based brain–computer interface: increasingthe rate of communication, Int. J. Hum. Comput. Interact. 27 (2011) 69–84.

[10] T. Kaufmann, S. Völker, L. Gunesch, A. Kübler, Spelling is just a click away—auser-centered brain–computer interface including auto-calibration and pre-dictive text entry, Front. Neurosci. 6 (2012) 72.

[11] S.T. Ahi, H. Kambara, Y. Koike, A dictionary-driven P300 speller with amodified interface, IEEE Trans. Neural Syst. Rehabil. Eng. 19 (2011) 6–14.

[12] F. Akram, H.S. Han, T.S. Kim, A P300-based brain computer interface system forwords typing, Comput. Biol. Med. 45 (2014) 118–125.

[13] J. Höhne, M. Schreuder, B. Blankertz, M. Tangermann, A novel 9-class auditoryERP paradigm driving a predictive text entry system, Front. Neurosci (2011) 5.

[14] M. Kaper, P. Meinicke, U. Grossekathoefer, T. Lingner, H. Ritter, BCI competition2003—data set IIb: support vector machines for the P300 speller paradigm,IEEE Trans. Biomed. Eng. 51 (2004) 1073–1076.

[15] M. Thulasidas, C. Guan, J.K. Wu, Robust classification of EEG signal for brain–computer interface, IEEE Trans. Neural Syst. Rehabil. Eng. 14 (2006) 24–29.

0 5 10 1520

40

60

80

100 S1

0 5 10 1520

40

60

80

100 S2

0 5 10 1520

40

60

80

100 S3

0 5 10 1520

40

60

80

100

Acc

urac

y (%

)

S4

0 5 10 1520

40

60

80

100

Number of Repetitions

S5

0 5 10 1520

40

60

80

100Average

RFSVM

Fig. 4. Average classification accuracy for SVM (dotted) vs. RF (solid) relative to the number of repetitions.

F. Akram et al. / Computers in Biology and Medicine 56 (2015) 30–36 35

[16] U. Hoffmann, J.M. Vesin, T. Ebrahimi, K. Diserens, An efficient P300-basedbrain–computer interface for disabled subjects, J. Neurosci. Methods 167(2008) 115–125.

[17] H. Cecotti, A. Graser, Convolutional neural networks for P300 detection withapplication to brain–computer interfaces, IEEE Trans. Pattern Anal. Mach.Intell. 33 (2011) 433–445.

[18] H. Serby, E. Yom-Tov, G.F. Inbar, An improved P300-based brain–computerinterface, IEEE Trans. Neural Syst. Rehabil. Eng. 13 (2005) 89–98.

[19] O.I. Khan, F. Farooq, F. Akram, M.T. Choi, S.M. Han, T.S. Kim, Robust extractionof P300 using constrained ICA for BCI applications, Med. Biol. Eng. Comput. 50(2012) 231–241.

[20] D. Soetraprawata, A. Turnip, Autoregressive integrated adaptive neural net-works classifier for EEG-P300 classification, Mechatron. Electr. Power Veh.Technol. 4 (1) (2013) 1–8.

[21] L. Breiman, Random forests, Mach. Learn. 45 (1) (2001) 5–32.[22] Silfverberg, M., MacKenzie, I.S., Korhonen, P. (2000). Predicting text entry

speed on mobile phones, in: Proceedings of the SIGCHI Conference on HumanFactors in Computing Systems. ACM. pp. 9–16.

[23] B. Blankertz, K.R. Muller, D.J. Krusienski, G. Schalk, J.R. Wolpaw, A. Schlogl,G. Pfurtscheller, J.D.R. Millan, M. Schroder, N. Birbaumer, The BCI competitionIII: Validating alternative approaches to actual BCI problems, IEEE Trans.Neural Syst. Rehabil. Eng. 14 (2006) 153–159.

F. Akram et al. / Computers in Biology and Medicine 56 (2015) 30–3636