Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Design of a self-paced brain-computer interface based on mental tasks Faradji, Farhad 2012

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2012_fall_faradji_farhad.pdf [ 3.26MB ]
Metadata
JSON: 24-1.0072796.json
JSON-LD: 24-1.0072796-ld.json
RDF/XML (Pretty): 24-1.0072796-rdf.xml
RDF/JSON: 24-1.0072796-rdf.json
Turtle: 24-1.0072796-turtle.txt
N-Triples: 24-1.0072796-rdf-ntriples.txt
Original Record: 24-1.0072796-source.json
Full Text
24-1.0072796-fulltext.txt
Citation
24-1.0072796.ris

Full Text

DESIGN OFA SELF-PACED BRAIN?COMPUTER INTERFACEBASED ON MENTAL TASKSbyFarhad FaradjiB.Sc., Amirkabir University of Technology, 2005B.Sc., Amirkabir University of Technology, 2007M.Sc., Amirkabir University of Technology, 2007A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDOCTOR OF PHILOSOPHYinThe Faculty of Graduate Studies(Electrical and Computer Engineering)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)May 2012c? Farhad Faradji, 2012AbstractPeople with very severe motor-disabilities have to accept a much-reduced quality of life. Unfortu-nately, these people cannot use assistive devices as present devices require motor activities.Brain?computer interfaces (BCIs) provide an alternative means of communication between thebrain and assistive devices. There are two types of BCIs, synchronous and self-paced. Self-pacedBCIs are more practical as they can be used at any time. The vast majority of existing self-pacedBCIs are activated by real, attempted, or imagined movements. Few are activated by mental tasks.The high false positive rates (FPRs) of existing self-paced BCIs render them impractical. Thereare no self-paced BCIs based on motor movements that have low FPRs. However, self-paced BCIswith low FPRs based on mental tasks have been proposed. Designing a self-paced mental task-based BCI with a zero or near zero FPR and a reasonable true positive rate (TPR) is the goal ofthis thesis.We investigated the feasibility of having a self-paced mental task-based BCI with a zero ornear zero FPR. The EEG signals from 6 electrodes of 4 subjects performing 4 mental tasks areused. Features were extracted using autoregressive modeling. Different classifiers were tested.The results were promising in that zero FPRs were obtained. The data used, however, had not beencollected in a self-paced paradigm.We then collected the EEG signals from 29 channels of 4 subjects performing 4 mental tasksin a self-paced paradigm. We evaluated the performance of our BCI using this dataset. It yieldediiAbstractan average TPR of 67.26% and zero FPR.To make the system practical, we decrease the number of channels from 29 to 7 and 5, using2 approaches that yield local optimal results. The average TPR is sufficiently high (54.60% and59.98% for systems with 5 and 7 channels) while the FPRs remain zero.We also study the effects on the performance, as the segment length is varied. For the 7-channelBCIs, the optimum length is between 1 and 2.5 sec. The average TPR is improved to 63.47%. TheFPRs are zero. We also show that our BCIs are robust to artifacts.iiiPrefaceI am the principal contributor of all chapters of this dissertation. I developed the methods, collectedand analyzed the data, interpreted the results, wrote the manuscripts, and acted as the correspond-ing author with the editors of journals and conferences the papers were submitted to. All chaptersare co-authored with Prof. Rabab K. Ward and Dr. Gary E. Birch, who supervised the developmentof the work and helped in interpreting the results, writing the manuscripts, and evaluations. Theexperimental protocol of this thesis has been reviewed by the Behavioural Research Ethics Board(BREB) of UBC. The number of the ethics certificate obtained is H08-00998.The following publications describe the work completed during this thesis.1. Farhad Faradji, Rabab K. Ward, and Gary E. Birch, ?Plausibility Assessment of a 2-StateSelf-Paced Mental Task-Based BCI Using the No-Control Performance Analysis,? Journalof Neuroscience Methods, vol. 180, no. 2, pp. 330?339, June 2009.2. Farhad Faradji, Rabab K. Ward, and Gary E. Birch, ?Toward Development of a Two-StateBrain?Computer Interface Based on Mental Tasks,? Journal of Neural Engineering, vol. 8,no. 4, p. 046014 (9pp), June 2011.3. Farhad Faradji, Rabab K. Ward, and Gary E. Birch, ?A Self-Paced 2-State Mental Task-Based Brain?Computer Interface Using Few EEG Channels,? Journal of Neural Engineer-ing, under revision.ivPreface4. Farhad Faradji, Rabab K. Ward, and Gary E. Birch, ?A Study on Finding the OptimumLength of the EEG Signal Segments When Used in Self-Paced 2-State Mental Task-BasedBCIs,? ready for submission.Conference Papers1. Farhad Faradji, Rabab K. Ward, and Gary E. Birch, ?A Simple Approach to Find the BestWavelet Basis in Classification Problems,? in Proc. of the 20th International Conferenceon Pattern Recognition (ICPR), pp. 641?644, Turkey, August 2010.2. Farhad Faradji, Rabab K. Ward, and Gary E. Birch, ?A Self-Paced BCI Using StationaryWavelet Packets,? in Proc. of the 31st Annual International Conference of the IEEE Engi-neering in Medicine and Biology Society (EMBC), pp. 962?965, USA, September 2009.3. Farhad Faradji, Rabab K. Ward, and Gary E. Birch, ?A Custom-Designed Mental Task-Based Brain?Computer Interface,? in Proc. of the 34th IEEE International Conference onAcoustics, Speech, and Signal Processing (ICASSP), pp. 529?532, Taiwan, April 2009.4. Farhad Faradji, Rabab K. Ward, and Gary E. Birch, ?Design of a Mental Task-Based Brain?Computer Interface with a Zero False Activation Rate Using Very Few EEG ElectrodeChannels,? in Proc. of the 4th International IEEE EMBS Conference on Neural Engineer-ing, pp. 403?406, Turkey, April 2009.5. Farhad Faradji, Rabab K. Ward, and Gary E. Birch, ?A Brain?Computer Interface Basedon Mental Tasks with a Zero False Activation Rate,? in Proc. of the 4th International IEEEEMBS Conference on Neural Engineering, pp. 355?358, Turkey, April 2009.6. Farhad Faradji, Rabab K. Ward, and Gary E. Birch, ?Self-Paced BCI Using Multiple SWT-Based Classifiers,? in Proc. of the 30th Annual International Conference of the IEEE Engi-neering in Medicine and Biology Society (EMBC), pp. 2095?2098, Canada, August 2008.vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiiList of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviiDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 What is a Brain?Computer Interface? . . . . . . . . . . . . . . . . . . . . . . . . . 11.1.1 Brain function monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . 21.1.2 System-paced versus self-paced . . . . . . . . . . . . . . . . . . . . . . . 31.1.3 System performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.1.4 Neurological phenomena . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.2 Mental Task-Based BCIs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6viTable of Contents1.2.1 Self-paced mental task-based BCI systems . . . . . . . . . . . . . . . . . . 101.2.2 The advantages of mental task-based BCIs . . . . . . . . . . . . . . . . . . 141.3 Contributions and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151.4 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCI . . . . . . . 182.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.3 Autoregressive Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242.4 Classification Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.4.1 Linear discriminant analysis . . . . . . . . . . . . . . . . . . . . . . . . . 262.4.2 Quadratic discriminant analysis . . . . . . . . . . . . . . . . . . . . . . . 292.4.3 Mahalanobis discriminant analysis . . . . . . . . . . . . . . . . . . . . . . 302.4.4 Support vector machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312.4.5 Radial basis function neural network . . . . . . . . . . . . . . . . . . . . . 312.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332.5.1 LDA classifier results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332.5.2 QDA classifier results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332.5.3 MDA classifier results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342.5.4 SVM classifier results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342.5.5 RBF NN classifier results . . . . . . . . . . . . . . . . . . . . . . . . . . . 342.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382.6.1 Comparing classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382.6.2 Selecting the classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392.6.3 Most discriminatory tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . 422.6.4 High TPRs versus low FPRs . . . . . . . . . . . . . . . . . . . . . . . . . 43viiTable of Contents2.6.5 Computational cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473 Towards Development of a Self-Paced 2-State Mental Task-Based BCI . . . . . . . . 483.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493.1.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493.1.2 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623.3.1 The least effective tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643.3.2 Optimum AR model orders . . . . . . . . . . . . . . . . . . . . . . . . . . 643.3.3 Subject dependency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643.3.4 Mental task dependency . . . . . . . . . . . . . . . . . . . . . . . . . . . 643.3.5 Comparison to our previous work . . . . . . . . . . . . . . . . . . . . . . 653.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684 A Self-Paced 2-State Mental Task-Based BCI Using Few EEG Channels . . . . . . . 694.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704.1.1 EEG channel selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704.1.2 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 764.3 Summary and Discussion of Results . . . . . . . . . . . . . . . . . . . . . . . . . 814.3.1 System performance of 5-channel BCIs . . . . . . . . . . . . . . . . . . . 814.3.2 System performance of 7-channel BCIs . . . . . . . . . . . . . . . . . . . 824.3.3 Discussion of results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 834.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85viiiTable of Contents5 A Study on Finding the Optimum Length of the EEG Signal Segments When Usedin Self-Paced 2-State Mental Task-Based BCIs . . . . . . . . . . . . . . . . . . . . . 875.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 875.1.1 EEG channel selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 885.1.2 Finding the optimum length of the EEG signal segments . . . . . . . . . . 895.1.3 Artifact rejection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 905.2 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 915.2.1 Optimum length of the segments . . . . . . . . . . . . . . . . . . . . . . . 915.2.2 System performance with artifact rejection . . . . . . . . . . . . . . . . . . 995.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1026 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1036.1 Research Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1036.2 Suggestions for Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108ixList of TablesTable 1.1 Mental Tasks Investigated in BCI Systems . . . . . . . . . . . . . . . . . . . . 7Table 1.2 Mental Tasks Investigated in BCI Systems . . . . . . . . . . . . . . . . . . . . 8Table 1.3 Mental Tasks Investigated in BCI Systems . . . . . . . . . . . . . . . . . . . . 9Table 1.4 Self-Paced Mental Task-Based BCI Systems . . . . . . . . . . . . . . . . . . . 11Table 1.5 Self-Paced Mental Task-Based BCI Systems . . . . . . . . . . . . . . . . . . . 12Table 1.6 Self-Paced Mental Task-Based BCI Systems . . . . . . . . . . . . . . . . . . . 13Table 2.1 The Number of Completed Trials for Each Subject . . . . . . . . . . . . . . . . 21Table 2.2 Cross-Validation Results for Different Subjects, Tasks, and Classifiers . . . . . . 36Table 2.3 Testing Results for Different Subjects, Tasks, and Classifiers with SelectedModelOrders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37Table 2.4 T-Test Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40Table 2.5 Average Performances of Classifiers for Different Subjects and Tasks . . . . . . 41Table 2.6 The Best Classifier for Each Subject and Each Mental Task . . . . . . . . . . . 43Table 2.7 Typical Processing Time Required in Different Parts of the BCI System . . . . . 46Table 3.1 Cross-Validation and Testing Results for Different Subjects and Mental Tasks . . 60Table 3.2 Average of Testing Results over Mental Tasks and Subjects . . . . . . . . . . . 61Table 3.3 System Performance of the Best BCI for Each Subject . . . . . . . . . . . . . . 61xList of TablesTable 3.4 p-Values Calculated Using Kolmogorov-Smirnov Normality Test for EachMen-tal Task of Each Subject . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61Table 3.5 p-Values Calculated Using Welch?s T-Test between Every Pair of the MentalTasks for Each Subject . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62Table 3.6 p-Values Calculated Using Welch?s T-Test between Every Pair of the Subjectsfor Each Mental Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63Table 3.7 System Performance [60] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67Table 4.1 Channels Selected for Different Subjects and Mental Tasks Using Channel Se-lection Method One (MDelete) . . . . . . . . . . . . . . . . . . . . . . . . . . 75Table 4.2 Channels Selected for Different Subjects and Mental Tasks Using Channel Se-lection Method Two (MForm) . . . . . . . . . . . . . . . . . . . . . . . . . . . 76Table 4.3 Cross-Validation and Testing Results of the Better Channel Selection Methodfor 5-Channel Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78Table 4.4 The Difference in the Performance between MDelete and MForm for the 5-Channel Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78Table 4.5 Cross-Validation and Testing Results of the Better Channel Selection Methodfor 7-Channel Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79Table 4.6 The Difference in the Performance between MDelete and MForm for the 7-Channel Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79Table 4.7 p-Values Calculated Using Welch?s T-Test between the Four Mental Tasks forEach Subject (5-Channel Systems) . . . . . . . . . . . . . . . . . . . . . . . . 80Table 4.8 p-Values Calculated Using Welch?s T-Test between the Four Mental Tasks forEach Subject (7-Channel Systems) . . . . . . . . . . . . . . . . . . . . . . . . 80Table 4.9 Average Performance of the BCI Systems . . . . . . . . . . . . . . . . . . . . . 84Table 4.10 System Performance of the Best BCI for Each Subject . . . . . . . . . . . . . . 84xiList of TablesTable 5.1 Channels Selected in Each BCI of Each Subject . . . . . . . . . . . . . . . . . 90Table 5.2 System Performance during Validation and Test without Artifact Rejection . . . 98Table 5.3 Average Performance of the BCI Systems . . . . . . . . . . . . . . . . . . . . . 99Table 5.4 System Performance of the Best BCI for Each Subject . . . . . . . . . . . . . . 99Table 5.5 System Performance during Validation and Test with Artifact Rejection . . . . . 101xiiList of FiguresFigure 1.1 System performance. The output of the BCI system can be a TP, a TN, a FP, ora FN, based on the user?s intent. . . . . . . . . . . . . . . . . . . . . . . . . . 4Figure 2.1 EEG signals were recorded from 6 electrodes according to the 10-20 system. . 22Figure 2.2 Block diagram of the BCI system. . . . . . . . . . . . . . . . . . . . . . . . . 23Figure 2.3 The architecture of an RBF neural network. . . . . . . . . . . . . . . . . . . . 32Figure 3.1 EEG signals were recorded from 29 electrodes distributed over the scalp ac-cording to the 10-10 system. . . . . . . . . . . . . . . . . . . . . . . . . . . . 50Figure 3.2 EOG electrodes are placed around the eyes as depicted. . . . . . . . . . . . . . 50Figure 3.3 Epoch timing: After a break of length of 15?2.5 sec, a ?Start Cue? was dis-played on the screen for 4 sec. The subject was told to wait about 1 sec afterthe cue disappeared before performing a mental task for about 10 sec. ?StopCue" was displayed on the screen for 2.5 sec informing the subject of the endof the 10-sec interval. The next epoch then started. . . . . . . . . . . . . . . . 51Figure 3.4 ROC curves for BCIs of Subject 1. The values shown on the plot are thecorresponding AR model orders. . . . . . . . . . . . . . . . . . . . . . . . . . 56Figure 3.5 ROC curves for BCIs of Subject 2. The values shown on the plot are thecorresponding AR model orders. . . . . . . . . . . . . . . . . . . . . . . . . . 57xiiiList of FiguresFigure 3.6 ROC curves for BCIs of Subject 3. The values shown on the plot are thecorresponding AR model orders. . . . . . . . . . . . . . . . . . . . . . . . . . 58Figure 3.7 ROC curves for BCIs of Subject 4. The values shown on the plot are thecorresponding AR model orders. . . . . . . . . . . . . . . . . . . . . . . . . . 59Figure 4.1 Channel selection method 1 (MDelete) . . . . . . . . . . . . . . . . . . . . . . 72Figure 4.2 Channel selection method 2 (MForm) . . . . . . . . . . . . . . . . . . . . . . 73Figure 5.1 TPRs and AR orders of different segment lengths for BCIs of Subject 1. . . . . 94Figure 5.2 TPRs and AR orders of different segment lengths for BCIs of Subject 2. . . . . 95Figure 5.3 TPRs and AR orders of different segment lengths for BCIs of Subject 3. . . . . 96Figure 5.4 TPRs and AR orders of different segment lengths for BCIs of Subject 4. . . . . 97xivList of AbbreviationsALS Amyotrophic lateral sclerosisANC Activity of neural cellsAR AutoregressiveBCI Brain?computer interfaceBREB Behavioural research ethics boardCP Cerebral palsyCSM Channel selection methodECOG Electro-corticographyEEG Electro-encephalographyFMRI Functional magnetic resonance imagingFP False positiveFPR False positive rateIC Intentional-controlLDA Linear discriminant analysisMDA Mahalanobis discriminant analysisMEG Magneto-encephalographyMRP Movement-related potentialsxvList of AbbreviationsMS Multiple sclerosisMT Mental taskNC No-controlNN neural networkPET Positron emission tomographyQDA Quadratic discriminant analysisRBF Radial basis functionRBF NN Radial basis function neural networkROC Receiver operating characteristicSCI Spinal cord injurySCP Slow cortical potentialsSSVEP Steady-state visual evoked potentialsSVM Support vector machineTP True positiveTPR True positive rateVEP Visual evoked potentialsxviAcknowledgmentsTO HIM BELONGS DOMINION AND TO HIM BELONGS ALL PRAISEI am most grateful to my supervisors, Prof. Rabab K. Ward and Dr. Gary E. Birch. Prof.Ward has spent much time helping me formulate my ideas and provided the opportunity to workin her research group. Dr. Birch?s constant support, guidance, and encouragement were invaluablethroughout my PhD program.I am also greatly thankful to all of my former teachers and professors.My sincere thanks to my friends for their prayers and help.I wish to thank those friends of mine who came and devoted their time to attend the EEGrecording sessions.I would like to express my deepest gratitude to my father, my mother, and my brother, whoseunconditional love has shaped my outlook of life, for their dedication, encouragement, advise, andthe many years of support.Last, but not least, I would like to thank my wife whose continued patience and encouragementwas in the end what made this dissertation possible. Without her, I would not have the drive tocomplete this thesis.This work was funded by the Natural Sciences and Engineering Research Council of Canadaand the Qatar National Research Fund.xviiDedicationTo my FamilyxviiiChapter 1Introduction1.1 What is a Brain?Computer Interface?Brain?computer interfaces (BCIs) aim at providing an alternative means of communication formotor-disabled people suffering from diseases such as brain injury, brainstem stroke, high-levelspinal cord injury (SCI), amyotrophic lateral sclerosis (ALS, also known as Maladie de Charcotor Lou Gehrig?s disease), muscular dystrophies, multiple sclerosis (MS), cerebral palsy (CP), orlocked-in syndrome (sometimes called ventral pontine syndrome, cerebromedullospinal discon-nection, pseudocoma, and de-efferented state)[171, 178, 179].People with very severe motor disabilities are forced to accept a much-reduced quality of lifedue to their inability to communicate orally or move their limbs including fingers, hands, and feet.Current assistive devices require some kind of motor activity; therefore, they cannot be used bypeople with very severe motor disabilities.Over the last two decades, the BCI field has emerged as a new frontier in assistive technology.This field of study aims at providing an alternative means of communication between the user?sbrain and the assistive device. A BCI is defined as ?a communication system in which messagesor commands that an individual sends to the external world do not pass through the brain?s normal1Chapter 1. Introductionoutput pathways of peripheral nerves and muscles? [178].A successful BCI design enables motor-disabled people to control their environment (such asopening and closing an automated door, turning on or off a light, a television, or other devices inthe room), actuate a neural prosthesis or even work with a computer. It can also be used by healthyindividuals for entertainment such as playing computer games.1.1.1 Brain function monitoringA BCI system monitors the state and activity of the brain. Some specific features pertaining to theuser?s intent are recognized and extracted from his or her brain state. The features from the brainsignals are translated into commands to control a target assistive device.The existing methods for monitoring brain activities include:? electro-encephalography (EEG),? electro-corticography (ECoG),? magneto-encephalography (MEG),? positron emission tomography (PET),? functional magnetic resonance imaging (fMRI), and? optical imaging.EEG and ECoG signals have short time constants, thus they can be used in real-time applica-tions [178, 179]. EEG signals are recorded from electrodes that are placed on the scalp. ECoGsignals are recorded using electrodes placed on the surface of the brain cortex, thus the use ofECoG forms an invasive procedure. Even though the use of MEG, PET, fMRI and optical imagingis less invasive than using ECoG, they require expensive and non-portable equipment. Therefore,EEG is most widely used in BCI systems because it is a non-invasive method, and also has theleast cost [178, 179].2Chapter 1. Introduction1.1.2 System-paced versus self-pacedExisting BCI systems, regardless of the monitoring methods used, are categorized into two majorclasses: system-paced (or synchronized) and self-paced (or asynchronous). In the system-pacedBCI, the earliest type of BCI, the user can only control the BCI in specific time intervals that arepredefined by the system and not by the user. The BCI is unavailable for control at other times. Aself-paced BCI, on the other hand, can be available for control at all times. It is clear that the secondclass is better and more efficient in terms of practicality and applicability to real-life applications.Two types of operational states (or modes) are usually defined as the outputs of a self-pacedBCI: the no-control (NC) state and the intentional control (IC) state. The BCI is in the NC modemost of the time. When the user issues a command, the system changes its state from the NC modeto the IC mode. After that, the BCI returns to the NC state. In other words, IC refers to the stateduring which the user controls the BCI, and NC is the state during which the user does not intendto activate the BCI.1.1.3 System performanceTo evaluate the performance of a system-paced BCI, the classification accuracy or informationtransfer rate is usually used. For self-paced systems, however, the true positive and the falsepositive rates form appropriate measures to evaluate the performance of these BCIs. Classifyingan IC input as an IC output state forms a true positive (TP), while misclassifying an NC input asan IC state is a false positive (FP) result. Refer to Figure 1.1.The true positive rate (TPR) is the number of TPs divided by the total number of IC classifi-cations. The false positive rate (FPR) is the ratio of the number of FPs to the total number of NCclassifications. Put differently, the TPR shows the rate of correctly classifying the IC states, whilethe FPR shows the rate of wrongly classifying the NC states.Due to the high false activation rates, self-paced BCI systems are deemed unsuccessful for use3Chapter 1. IntroductionFigure 1.1: System performance. The output of the BCI system can be a TP, a TN, a FP, or aFN, based on the user?s intent.in real-life applications. This is because false activations are a major cause of user frustration. Tofurther illustrate this point, suppose that the output rate of a self-paced BCI is 5 Hz (i.e., 5 outputsper sec, as in the BCIs designed in this thesis) and the FPR value is 1%. This FPR of 1% meansone false positive in every 100 outputs of the BCI. As the BCI generates 100 outputs in 20 sec,there would be three false activations in every minute, which is too high for practical purposes.Considering the fact that a self-paced BCI is in the no-control mode for most of the time, even alow FPR would frustrate any user. This is why for self-paced BCI systems, lowering the FPR is amajor concern.To better understand the significance of ?self-paced? BCIs that have ?very low FPRs? in reallife, let us consider two potential applications of BCI systems:? controlling different devices in a room (such as opening/closing an automated door, turningon/off the lights, a TV, etc) and? controlling a wheelchair.The ideal and practical BCI system for these applications is a self-paced BCI. In such applica-tions, the user wants to control the target device (through a BCI) whenever he/she wishes. He/shedoes not like to wait for the system to let him control the device in a predefined time interval.The importance of the BCI system being available all the time for control is more obvious in the4Chapter 1. Introductionwheelchair application since in some situations, the user needs to control the wheelchair immedi-ately.It is essential for a self-paced BCI system to have a very low (zero or near zero) FPR because:1. Each command resulting from a false positive needs to be neutralized by another command.Suppose the lights are turned off by an FP. This action needs to be counteracted by anothercommand from the user (to turn on the lights). Such a wrong command resulting form anFP makes the user easily frustrated, specially if it happens often.2. In some cases, FP commands may result in dangerous actions. An FP command that makesthe wheelchair move can cause a disaster for a user waiting at a traffic light.1.1.4 Neurological phenomenaBCIs exploit different kinds of neurological phenomena such as:? visual evoked potentials (VEPs),? steady-state visual evoked potentials (SSVEPs),? slow cortical potentials (SCPs),? brain rhythms such as the Mu, Beta and Gamma rhythms,? P300,? the activity of neural cells (ANCs),? mental tasks (MTs), and? movement-related potentials (MRPs).These phenomena are specific features in brain signals that are time-locked to brain activities.They can be categorized as exogenous (evoked by external stimuli) or endogenous phenomena (allothers). For a review of the field, refer to [20, 69, 113?116, 127, 170, 177, 179].5Chapter 1. Introduction1.2 Mental Task-Based BCIsMental (or cognitive) tasks are a class of neurological phenomena. They generally refer to inten-tional cognitive tasks that are done by the brain. Mental tasks have been used in the control of manyBCI systems proposed [10, 14, 15, 31, 33, 34, 37?40, 42, 43, 45, 46, 48?52, 55?57, 70, 73, 75, 77,78, 80?82, 85?88, 90?92, 96?99, 101?106, 109, 119, 120, 124?126, 128, 129, 134, 140, 141, 143?150, 153?156, 160?162, 166?169, 172?174, 181, 185].Mental tasks can be classified into two categories:? motor imagery tasks,? non-motor imagery tasks: mental mathematical calculations such as multiplication and count-ing, mental rotation, visualization, etc.Motor imagery has been so far investigated in many BCI studies [14, 15, 31, 33, 34, 39, 40,42, 43, 49, 50, 70, 75, 77, 80?82, 85?88, 90?92, 98, 101, 103?105, 109, 119, 120, 124?126, 128,141, 144?150, 153?155, 160, 166?169, 172?174, 181, 185]. The use of other types of mental tasksin BCI studies have received little or no attention in the literature. The papers that have studiednon-motor imagery mental tasks along with the motor imagery tasks include [3, 37, 38, 45, 48, 51,52, 55?58, 71, 73, 99, 102, 106, 129, 143, 156, 161, 162]. The studies that have only consideredthe non-motor imagery mental tasks are [5, 10, 46, 78, 96, 97, 134, 140, 165].The mental tasks investigated in the studies that have considered non-motor imagery mentaltasks along with the motor imagery tasks and in the studies that have only considered non-motorimagery mental tasks are summarized in Tables 1.1, 1.2, and 1.3.6Chapter 1. IntroductionTable 1.1: Mental Tasks Investigated in BCI SystemsStudy Mental Tasks[3]: ? imagined right hand clenching? imagined left foot moving up and down? counting backwards from 100 by 3? visualizing the computer screen tumbling in space[37, 73, 106, 161, 162]: ? the imagination of repetitive self-paced left or right hand movements*? the generation of words beginning with the same random letter*[38] and [102]: ? the imagination of repetitive self-paced left or right hand movements*? the generation of words beginning with the same random letter*? the imagination of the left or right hand movement?[45]: ? spatial navigation around a familiar environment? auditory imagery of a familiar tune? right motor imagery of opening and closing the hand? left motor imagery of opening and closing the hand[48, 52]: ? the imagination of right hand (or arm) movement? the imagination of left hand (or arm) movement? cube rotation? subtraction? word association[51]: ? the imagination of right hand movement? the imagination of left hand movement? cube rotation? subtraction[55]: ? the right hand extension motor imagery? the left hand extension motor imagery? subtraction? navigation imagery? auditory imagery? phone imagery? idle task* Data is provided by the IDIAP Research Institute in Switzerland [52].? Data is provided by the BCI laboratory of the Graz University of Technology in Austria [145].7Chapter 1. IntroductionTable 1.2: Mental Tasks Investigated in BCI SystemsStudy Mental Tasks[56]: ? the right hand flexion motor imagery? the left hand flexion motor imagery? subtraction? navigation imagery? auditory imagery? phone imagery? idle task.[57]: ? subtraction? navigation imagery? auditory recall? phone imagery? motor imageries of the right hand? motor imageries of the left hand[58]: ? imagined right hand movement? imagined left leg movement? counting backwards from 100 by 3? imagining a spinning computer[71]: ? imagined right hand clenching? counting backwards from 100 by 3[99]: ? the exact calculation of repetitive additions? imagination of left finger movement? mental rotation of a cube? evocation of a nonverbal audio signal[129]: ? left-hand movement motor imagery? right-hand movement motor imagery? foot movement motor imagery? tongue movement motor imagery? repeated subtraction of a constant number from a randomly chosen number8Chapter 1. IntroductionTable 1.3: Mental Tasks Investigated in BCI SystemsStudy Mental Tasks[143]: ? motor imagery of hand opening and closing? serially subtraction seven from a large number[156]: ? auditory recall? mental navigation? sensorimotor attention of the right hand? sensorimotor attention of the left hand? mental calculation? imaginary movement of the right hand? imaginary movement of the left hand[5, 10, 78, 96, 97], ? baseline?[134, 140, 165]: ? computing a nontrivial multiplication?? mentally composing a letter?? mentally rotating a 3D object?? visualizing a sequence of numbers being written on a blackboard?[46]: ? vowel speech imagery: imaginary speech of the two English vowels /a/ and /u/? Data is provided by Keirn and Aunon [96].9Chapter 1. Introduction1.2.1 Self-paced mental task-based BCI systemsFrom all systems developed in [10, 14, 15, 31, 33, 34, 37?40, 42, 43, 45, 46, 48?52, 55?57, 70,73, 75, 77, 78, 80?82, 85?88, 90?92, 96?99, 101?106, 109, 119, 120, 124?126, 128, 129, 134,140, 141, 143?150, 153?156, 160?162, 166?169, 172?174, 181, 185], only the BCI systems in[33, 48, 51, 52, 55?57, 73, 90, 120, 148, 155, 167?169] are self-paced. As shown in Tables 1.4,1.5, and 1.6, for these self-paced systems:? The FPR values are not given in [48, 52, 55?57, 73].? Even though the number of FPs and TPs are given in [168] and [169], the rates of FPs andTPs are not reported.? The FPR values are reported in [51, 90, 120, 148, 155, 167].? In [51], the average FPR is 5%.? In [90], the FPRs are between 2% and 82%.? In [120], the designed BCI was evaluated in terms of FPs during only one 3-minute pe-riod. No FPs were generated during this period; however, the BCI is generally consid-ered impractical for the real-life applications since it is too slow. The interval betweentwo subsequent active states of the system is at least 4 sec.? In [148], the false activation rate is 0?3.25 activations/minute.? In [155], the FPR values of the systems are in the range of 3.8?32.5%.? In [167], the reported FPRs are between 10% and 77%.? The specificity rates (i.e., 100?FPR%) are reported in [33]. Based on the specificity rates,the FPRs are in the range of 0.38?14.38%.In Tables 1.4, 1.5, and 1.6, the information about the methodology (features and classifiers) andthe number of channels used is given. TPR values and the classification accuracy (if reported inthe studies) are also provided. The features extracted are usually in the frequency domain. Lineardiscriminant analysis (LDA) is the mostly used classifier in these studies.10Chapter1.IntroductionTable 1.4: Self-Paced Mental Task-Based BCI SystemsMental TasksStudy (non-motor imagery + motor imagery) Features Classifier CH* FPR TPR CA?[73]: ? left/right hand motor imagery canonical variates analysis Gaussian 64 ? ? 15?89%? words generation[48]: ? right hand/arm motor imagery power spectrum local neural classifier 8 ? ? ?? left hand/arm motor imagery? cube rotation? subtraction? word association[51]: ? right hand motor imagery power spectrum local neural classifier 26 5% 70% ?? left hand motor imagery? cube rotation? subtraction[52]: ? right hand/arm motor imagery power spectrum Gaussian 8 ? 43?76.1% ?? left hand/arm motor imagery? cube rotation? subtraction? word association* number of EEG channels used? classification accuracy11Chapter1.IntroductionTable 1.5: Self-Paced Mental Task-Based BCI SystemsMental TasksStudy (non-motor imagery + motor imagery) Features Classifier CH* FPR TPR CA?[55]: ? right hand extension motor imagery band power LDA 64 ? ? 85.2?93.8%? left hand extension motor imagery reflection coefficients? subtraction? navigation imagery? auditory imagery? phone imagery? idle task[56]: ? right hand flexion motor imagery reflection coefficients adaptive classifier 64 ? ? 57?90%? left hand flexion motor imagery? subtraction? navigation imagery? auditory imagery? phone imagery? idle task.[57]: ? subtraction band power LDA 64 ? ? ?? navigation imagery reflection coefficients? auditory recall? phone imagery? right hand motor imagery? left hand motor imagery* number of EEG channels used? classification accuracy12Chapter1.IntroductionTable 1.6: Self-Paced Mental Task-Based BCI SystemsMental TasksStudy (motor imagery) Features Classifier CH* FPR TPR CA?[33]: ? right foot motor imagery principal component analysis k-NN 9 0.38?14.38% 45?96% ?locality preserving projections[90]: ? right/left arm motor imagery wavelet transform LDA 16 2?82% 1?83% ?? right/left hand motor imagery[120]: ? feet & left hand motor imagery frequency band LDA 6 ? (0 in 3 min) ? 71%[148]: ? brisk feet motor imagery frequency band LDA 5 0?3.25 /min 2.38?5.80 /min 85%[155]: ? left hand motor imagery band power LDA 22 3.8?32.5% 7.6?43.1% 60?88%? right hand motor imagery? foot motor imagery? tongue motor imagery[167]: ? right/left hand motor imagery normalized variance LDA 27 10?77% 45?94% ?[168]: ? left & right hand motor imagery band power LDA 10 ? ? ?[169]: ? left & right hand motor imagery band power LDA 10 ? ? ?* number of EEG channels used? classification accuracy13Chapter 1. Introduction1.2.2 The advantages of mental task-based BCIsMental task-based BCIs have some advantages over other types of BCIs.? There are some problems with the BCIs based on P300, VEPs, and SSVEPs:? Firstly, since an individual with severe motor disability can not move his/her bodyfor controlling the BCI, a panel with flashing lights needs to be placed in front of theindividual and in his/her sight line all the time. This highly restricts the use of the BCI.? Secondly, concentrating on a flashing panel continuously will typically make the userfrustrated and fatigue even in the short-term application. Therefore, this type of BCIscan not be used in long-term applications.In contrast, a mental task-based BCI does not have these limitations.? There is the risk with ALS and possibly other degenerative neurological diseases that move-ment related potentials (MRPs) become impaired and hence some other modality wouldneed to be explored. This is currently an open question in the research community. Thisis not a problem for non-motor imagery mental task-based BCIs since these BCIs are notbased on MRPs.? Non-motor imagery mental task-based BCIs have also an advantage over the motor imagery-based BCIs. Even though motor imagery-based BCIs are applicable for individuals sufferingfrom a spinal cord injury (SCI), who have movement experiences, they are not useful forpatients who were born with a disability. This group of people does not have any previousexperience of moving their limbs, and they are thus not able to perform motor imagerytasks. The inability of individuals who are born with motor disabilities has been overlookedin the present BCIs that are based on real or imaginary movements. However, this is not anissue for non-motor imagery mental task-based BCIs since mental tasks can be performedby motor-disabled individuals as long as they are cognitively healthy.14Chapter 1. Introduction1.3 Contributions and ResultsMany self-paced BCI systems have been designed by different research groups all over the world.These self-paced BCIs are mostly activated by real or attempted movements [16, 18, 19, 21?23, 26?28, 30, 32, 36, 41, 65?69, 114, 118, 175] and motor imageries [14, 15, 31, 33, 34, 39, 40, 42, 43, 49,50, 70, 75, 77, 80?82, 85?88, 90?92, 98, 101, 103?105, 109, 119, 120, 124?126, 128, 141, 144?150, 153?155, 160, 166?169, 172?174, 181, 185]. Very few self-paced BCI systems have usedmental tasks [33, 48, 51, 52, 55?57, 73, 90, 120, 148, 155, 167?169].To the best of our knowledge, no self-paced BCI system with a zero or near-zero false activationrate has been presented. Our main motivation for conducting this research is to design a self-paced 2-state mental task-based BCI system with a zero or near-zero false activation rate. Thecontributions of each chapter are as follows.? In Chapter 2, we assess the plausibility of having a self-paced 2-state mental task-based BCIwith a zero or near zero false activation rate using the dataset of Keirn and Aunon [96].The dataset contains the EEG signals of 6 channels. We use the scalar autoregressive (AR)model coefficients as the features. We study the performance of different classifiers suchas linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), Mahalanobisdiscriminant analysis (MDA), the support vector machine (SVM), and the radial basis func-tion neural network (RBF NN) when used in the design of the proposed BCI system. Weperform a 5?5 cross-validation for choosing the best AR model order for each of theseclassifiers. The results of this study are extremely encouraging and published in [60].? In Chapter 3, we carry out our own experiments to collect the EEG dataset which is muchlarger than the dataset of Keirn and Aunon [96]. The dataset contains the signals of 29channels from four subjects performing four different mental tasks and when they are in thebaseline state. We analyze this dataset to validate the findings of our previous study [60],which was based on the dataset of Keirn and Aunon [96]. The system performance obtained15Chapter 1. Introductionis very encouraging since the false positive rates (FPRs) reach zero, and the true positiverates (TPRs) are sufficiently high. The results of this study are published in [64].? In Chapter 4, we decrease the number of EEG channels used in our BCI system from 29to 5 or 7 using two different methods. We compare the performances of the system withthe channel sets obtained by these two methods and we select the better channel set for thefinal system design. This study is done since the proposed BCI in [64] with 29 channels isunfortunately impractical for use in real-life applications due to the large number of EEGchannels. The results are submitted for publication.? In Chapter 5, we study the change in the system performance of the 7-channel BCIs when theEEG signal segments have different lengths. The segment length was considered to be 1 sec(i.e., short in duration) in the BCIs developed in Chapters 2, 3, and 4 to have a fast responsesystem. In this chapter, we find the optimum length of the segments for the 7-channel BCIs.We also study the effect of artifact rejection on the overall system performance. The resultsare ready to be finalized and submitted for publication.1.4 Thesis OrganizationThe rest of the dissertation is organized as follows. In Chapter 2, we show that it is feasible tohave a self-paced 2-state mental task-based BCI with a zero or near zero FPR. In Chapter 3, wecollect and process our own dataset. The results confirm our previous findings about the feasibilityof self-paced 2-state BCIs based on mental tasks with zero or near-zero FPRs. In Chapter 4, wedevelop the design of the system and make the system applicable to real life by decreasing thenumber of EEG channels from 29 to 5 or 7. In Chapter 5, we find the optimum length of the EEGsignal segments when used in self-paced 2-state 7-channel BCIs. We also show that these systemsare robust to artifacts to a certain extent. Finally, Chapter 6 contains discussions of main results,conclusions, and suggestions of future research directions. Each main chapter of this dissertation16Chapter 1. Introductionis self-contained and included in a separate journal manuscript.17Chapter 2Plausibility Assessment of a Self-Paced2-State Mental Task-Based BCIDesigning a self-paced BCI that is activated by mental tasks and has a zero or near zero falseactivation rate is our main motivation for conducting this study. The dataset of Keirn and Aunon[96], which contains the EEG signals corresponding to five mental tasks, is used. Even though thisdataset has not been collected in a self-paced paradigm, it is used in the current study in a self-paced manner in the hope that mental tasks would pave the way for future research to designing aself-paced BCI based on mental tasks. Further details about this database are provided in the nextsection.A variety of studies such as [1, 2, 4?10, 17, 24, 47, 53, 54, 74, 76, 83, 84, 89, 94, 97, 107, 108,110?112, 121, 122, 132?140, 142, 151, 152, 157, 159, 163?165, 180, 182?184] have employedthis dataset. The aim was to classify the mental tasks. Most of them were successful in classifyingthe mental tasks to some extent, but only a few of them reported the resultant false positive rateor the confusion matrix [5, 10, 53, 54, 122, 132]. Three studies reported the classification error[112, 121, 151]. The remaining studies reported the correct classification rates for the differentmental tasks.As explained in Section 1.1.3, self-paced BCI systems are generally considered unusable in18Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIreal-life applications due to the high false activation rates. This is because false activations are amajor cause of user frustration. Since false activation is extremely important in the operation ofBCIs in real-life situations, we consider it in our work.In this study, we consider a 2-state self-paced BCI with a feature extractor based on the au-toregressive (AR) model coefficients and a general classifier. We then custom design the order ofthe AR model and the classifier for each of the five different mental tasks considered and for eachsubject.The output of each custom-designed BCI should be activated by only one of the 5 differentmental tasks. This task is denoted as the intentional-control (IC) task. The other four mental tasksare considered as no-control (NC) tasks. In other words, each custom-designed BCI has two states:the IC state and the NC state. For each subject, we have custom designed five BCIs, each wouldhave one mental task as the IC task and the other four tasks as NC tasks. We also consider thebaseline as one of the 5 tasks. Although it is meaningless to have a BCI in which the IC task is thebaseline, we consider this case for comparison purposes. We determine which mental task is thebest (most discriminatory) for each subject.The coefficients of the AR model are used as features. To select the optimal AR order, we varythe order from 2 to 20 during the cross-validation process and choose the best order based on theresulting TPR and FPR values. The AR order that yields the lowest FPR is selected. If there ismore than one order with the same lowest FPR, then the one with the maximum TPR is chosen.The performance of five different classifiers (i.e., linear discriminant analysis, quadratic dis-criminant analysis, Mahalanobis discriminant analysis, support vector machine, and radial basisfunction neural network) is studied. For each classifier, we perform its model order selection inde-pendently. At the end, for each mental task and for each subject, the classifier with the best overallperformance is selected.We have extracted different features from the EEG signals using the wavelets in the design19Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIof BCI systems. The results have been published [59, 61?63]. However, the features based onthe autoregressive modeling outperform those features in terms of the simplicity and the finalclassification accuracy.2.1 DatasetAs mentioned above, the data used were those collected by Keirn and Aunon [96]. This datasethas been made available at http://www.cs.colostate.edu/eeg/eegSoftware.html by Prof. Charles W.Anderson (Department of Computer Science, Colorado State University). The EEG recordings inthis dataset pertain to seven subjects, each performing five mental tasks:1. baseline,2. computing a nontrivial multiplication,3. mentally composing a letter (generating text for the letter),4. mentally rotating a 3D object, and5. visualizing a sequence of numbers being written on a blackboard.During the recordings of this database, the subjects should not have vocalized or gestured in anyway. Each recording session contains five trials of each of the five mental tasks (25 trials in total).In other words, there are five data runs per session. In a data run, each mental task is performedone time. Therefore, there are a total of 25 trials (5 of each task) in one session. Sessions wereperformed on different days. Each trial was 10 sec long.Subjects 2 and 7 completed only one session, while Subject 5 completed three sessions. Inthis study, we used the data of subjects who completed 10 or more trials. We did not use the EEGsignal dataset of Subject 4 because it was missing some data. For some trials, the missing dataoccurred in the first 2 sec of the recorded signal, or some parts of the signal were clipped. Trial 10also did not contain the data of the letter-composing mental task.20Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCITable 2.1: The Number of Completed Trials for Each SubjectSubject number Subject number Number of completedin original study in this study trials per mental task1 1 102 ? 53 2 104 ? 105 3 156 4 107 ? 5We assigned new numbers to the subjects used in our study (see Table 2.1). Table 2.1 showsthe number of completed trials for each subject in the original study.EEG signals were recorded from six electrodes (i.e., six channels), with the subjects seated ina sound-controlled room with dim lighting. The electrodes were placed on the scalp at C3, C4, P3,P4, O1, and O2, based on the International 10-20 System. They were referenced to two electricallylinked mastoids, A1 and A2. Fig. 2.1 shows the positions of the electrodes. The impedance of theelectrodes was kept below 5 k? during recordings.The sampling frequency was 250 Hz. A Lab Master 12-bit A/D converter was used, and aknown voltage was used to calibrate the system before each session. A bank of amplifiers (Grass7P511) with the band-pass filters set at 0.1-100 Hz was connected to the electrodes.Two electrodes were placed for detecting ocular artifacts, one of them at the outside cornerof the left eye and one below the left eye. In our study, the signals of these two electrodes werenot used, i.e., we did not reject or remove any part of the recorded EEG signals because of ocularartifacts. This is done to increase the realism of the test. Theoretically, it is also possible to havean automated system that generates a no-control output in the presence of ocular artifacts.This dataset was not collected in a self-paced paradigm; nevertheless, we are using this dataas an introductory exploration in laying the groundwork for future explorations of similar datacollected in a self-paced paradigm. Even though brain activities apparently do not change in self-21Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIFigure 2.1: EEG signals were recorded from 6 electrodes according to the 10-20 system.paced paradigms, the pacing information is not available. Therefore, training the BCI systembecomes more difficult since the exact start and end time of the mental tasks (pacing information)is not known.2.2 MethodologyAs mentioned earlier, the data of four subjects were used. For Subject 3, the first 10 trials (of 15available) were used, and all 10 trials were used for the other three subjects. Since the samplingfrequency was 250 Hz and the trial duration was 10 sec, each trial (of every mental task) has 2500samples.For feature extraction, we divide each trial into 46 overlapping segments. Each segment has250 samples and overlaps by 200 samples with the adjacent segment. Therefore, we have 460segments for each mental task of each subject. The 1-sec segments are sufficiently long to get agood estimate of the AR model [25].The AR coefficients are computed for each of these segments using the Burg algorithm, andare then used as features for classification. The AR model order is varied from 2 to 20. The perfor-mance of five different classifiers, i.e., linear discriminant analysis (LDA), quadratic discriminant22Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIFigure 2.2: Block diagram of the BCI system.analysis (QDA), Mahalanobis discriminant analysis (MDA), support vector machine (SVM), andradial basis function neural network (RBF NN) is investigated. Figure 2.2 illustrates the blockdiagram of the BCI system.For each subject, the 5?460 segments of the 5 mental tasks are divided into training, validationand test sets which are used to train the classifiers, select the best AR model order, and evaluatethe performance of the system, respectively. We use a 5-fold nested cross-validation (5?5 cross-validation), since the results of classification with a fixed split of data into training, validation andtest sets are not robust. The inner cross-validation is used for selecting the AR model order, whilethe outer cross-validation is used for estimating the performance of the system.The data are split into 5 outer folds. For each outer fold, 20% of the data is used for testing,and the remainder is for training and validation. The data assigned for training and validation arefurther divided into 5 inner folds. For each inner fold, 80% of the data is used for training and 20%is used for validation. Therefore, the cross-validation results report the average over 25 differentcases. The testing results are averaged over 5 cases.Subject-customized BCIs yield better results than general BCIs used by all subjects [18, 29].In this work, for each subject and each mental task, the customized BCI has a different AR orderand a different classifier.23Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIIn each of the five custom-designed BCIs for each subject, only one mental task is selected asthe intentional-control task and the other four mental tasks are treated as the no-control tasks. Weare including the data of no-control tasks for training, validation and testing purposes.2.3 Autoregressive ModelingThe autoregressive (AR) model of an order q is written as:y[n] =q?m=1amy[n?m]+u[n] (2.1)where y[n] is the one-dimensional signal at time n and am represents the AR coefficients. Theerror/noise u[n] is assumed to be a random process independent of previous values of the signaland with zero mean and a finite variance. The aim is to estimate the am coefficients from the finitesamples of the signal y.EEG signals are considered as random time series containing periodic information about rhythmsand stochastic noise. Although the EEG signals are not fully stationary, we assume that they aresufficiently stationary for each mental task considered to yield AR modeling that will discrimi-nate between them; since short segments with large overlaps are used, the stationary problem ismitigated. The EEG signal can therefore be treated as y[n] in Equation (2.1). In this study, eachsegment is 1 sec long with 80% overlap with the adjacent segment.There is no straightforward way for determining the correct model order. If the model order istoo low, only a small portion of the signal is captured and what remains is considered to be noise.If the order of the model is too high, not only is the signal captured but some portions of the noiseare also included in the model.Equation (2.1) represents a scalar autoregressive model. The generalization of Equation (2.1)24Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIfor a p-dimensional signal is the multivariate autoregressive model:Y [n] =q?m=1AmY [n?m]+U [n] (2.2)where Y [n] and U [n] are vectors with p elements, and Am is a p? p matrix. For EEG signals, p isthe number of channels.Existing techniques for choosing the order of multivariate AR models include the Final Pre-diction Error (FPE) criterion, the Information Theoretic criterion or Akaike Information Criterion(AIC), the Autoregressive Transfer Function criterion [72], and the Reflection Coefficient [137].In [72], AIC was used to obtain the model order. For the model of Equation (2.2), AIC is asfollows:AIC(q) = N ln(det(??q))+2qp2 (2.3)where N is the number of data points, ??q is the estimated covariance matrix of the prediction error,q is the model order, and p is the dimension of the data. For more details, see [72] or [9].AIC is a tradeoff between the prediction error (the first term) and model size (the second term).The prediction error decreases as the model order increases, but the second term increases with q.In [72], the values of AIC for 4-channel EEG data were calculated for different model orders.The region of the minimum AIC was flat and broad with model order 6 in the center. The powerspectra did not change significantly with the AR order. Therefore, the model order 6 was chosenin [72] for further analyses.In [96], features based on the Burg spectrum and Burg scalar AR coefficients were obtained.Classification accuracies mostly increased with the model order. The order 6 was selected since itwas high enough to yield good classification results.In [9], the best model order based on FPE criterion was 2, and the best order based on AIC wasbetween 1 and 3. But the authors decided to use a model of order 6 because of the result of studies25Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCI[96] and [72].As a result of the above studies, the AR model order has usually been set to 6 [4, 6, 8, 74,107, 139, 140], [47, 84, 89, 132, 135, 152, 159, 163, 164]. However, we believe that the modelorder should be adjusted in a way that meets the requirements of the application. AIC may not bea good criterion for BCI applications where FPR is as important as TPR, if not more important.Therefore, in this study, we determine the optimal AR order via a 5?5 cross-validation process.Using the validation set of the data, we calculate the TPR and FPR values of the system for differentAR orders, and select the order based on these values. The performance of the system with theselected AR order is evaluated using the test set, which is different from the validation set.The Burg algorithm [35] is probably the most popular method of estimating the AR coefficientssince it is computationally efficient and yields a stable AR model [123]. If the predictions of anAR model do not diverge over time, the AR model is stable. The Burg algorithm estimates thecoefficients at successive orders in forward and backward directions. We use the Burg algorithmfor computing the coefficients of the AR model.2.4 Classification MethodsIn this section, we briefly explain the classifiers used in this study.2.4.1 Linear discriminant analysisThe Linear discriminant analysis (LDA) classifier [12] assumes the classes have Normal (Gaussian)distributions and the same covariance matrix.Suppose we have k-dimensional vectors of x, each to be classified into one of the M classes.Each Class ?m has the Normal distribution:?m ? Nk(?m,?) (2.4)26Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIwhere m ? {1,2, ...,M}, ?m is the k-dimensional mean vector of Class ?m, and ? is the k?kcommon covariance matrix of all classes.The probability density function of Class ?m can be expressed as:fm,X(x) =1(2pi)k/2|?|1/2exp(?12(x??m)T??1(x??m))(2.5)Based on the Bayes discriminant rule, the input vector x is classified into Class ?i if:Cipii fi,X(x) = maxj(C jpi j f j,X(x)), j ? {1,2, ...,M} (2.6)where pii is the a-priori probability of Class ?i, andCi is the total cost of misclassifying a memberof Class ?i to the other classes. Note that:M?j=1pi j = 1. (2.7)When only two classes exist, the decision rule can be simplified as follows:x ?{?1 :C1pi1 f1,X(x)?C2pi2 f2,X(x)?2 :C1pi1 f1,X(x) <C2pi2 f2,X(x)(2.8)This is equivalent to:x ???????????1 : ln f1,X(x)? ln f2,X(x)? ln(pi2pi1.C2C1)?2 : ln f1,X(x)? ln f2,X(x) < ln(pi2pi1.C2C1)(2.9)27Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIUsing Equation (2.5) in Equation (2.9), the discriminant rule becomes:x ?{?1 : ld f (x)? 0?2 : ld f (x) < 0(2.10)where the linear discriminant function, ld f (x), is defined as:ld f (x) = (?1??2)T??1x?12(?1??2)T??1(?1 +?2)? ln(pi2pi1.C2C1)(2.11)The mean vectors (i.e., ?1 and ?2) and the common covariance matrice (i.e., ?) are estimatedfrom the data samples. The common estimated covariance matrix is calculated as:?= pi1?1 +pi2?2 (2.12)pi1 =n1n1 +n2(2.13)pi2 =n2n1 +n2(2.14)where n1 and n2 are the number of observations in class 1 and 2, and ?1 and ?2 are the estimatedcovariance matrices of the 2 classes.In this thesis, we used the same value for C1 and C2, the cost of false negative and the cost offalse positive classification, respectively, as in Equation (2.11).The a-priori probabilities for the two classes (IC and NC) were also assumed to be equal. If theaim is only to classify the whole dataset, the equality assumption about the a-priori probabilitiesmay not be correct, since the ratio of the number of IC tasks segments to the number of NC taskssegments is 1:4 for each of the BCIs. Therefore, the a-priori probabilities for the two classesshould be 0.2 and 0.8, respectively. In real-time experiments, the a-priori probabilities of theclasses are usually not known beforehand. One possible solution is to update these probabilities28Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIperiodically, based on the history. However, for the present offline evaluation, we have assumedequal probabilities, which represent the first and the simplest choice in this evaluation.2.4.2 Quadratic discriminant analysisQuadratic discriminant analysis (QDA) [12] is the quadratic version of LDA. Like LDA, Normaldistributions are assumed for the classes. The only difference between QDA and LDA relates tothe covariance matrices of the classes. In QDA, unlike LDA, the covariance matrices of the classesare not assumed to be the same. Therefore, QDA is more general than LDA.Suppose we have k-dimensional vectors of x to be classified into one of the M classes with theNormal distributions:?m ? Nk(?m,?m) (2.15)where m? {1,2, ...,M}, ?m is the k-dimensional mean vector, and ?m is the k?k covariance matrixof Class ?m.The probability density function of Class ?m can be expressed as:fm,X(x) =1(2pi)k/2|?m|1/2exp(?12(x??m)T??1m (x??m))(2.16)Using Equation (2.16) in Equation (2.9), the discriminant rule becomes:x ?{?1 : qd f (x)? 0?2 : qd f (x) < 0(2.17)where the quadratic discriminant function, qd f (x), is defined as:qd f (x) =?12xT (??11 ???12 )x+(?T1 ??11 ??T2 ??12 )x?12ln(|?1||?2|)29Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCI?12(?T1 ??11 ?1??T2 ??12 ?2)? ln(pi2pi1.C2C1)(2.18)The mean vectors (i.e., ?1 and ?2) and the covariance matrices (i.e., ?1 and ?2) are estimatedfrom the data samples. In this study, the same value for the a-priori probabilities pi1 and pi2, andthe same value for the cost parametersC1 andC2 are assumed.2.4.3 Mahalanobis discriminant analysisThe Mahalanobis discriminant analysis (MDA) classifier [93] is similar to QDA, with some excep-tions:1. The classes do not need to have Normal distributions.2. It is assumed that the a-priori probabilities are identical for all classes.3. The classification is based only on the squared Mahalanobis distances.Suppose we have k-dimensional vectors of x to be classified into one of the M classes. TheMahalanobis distance between the vector x and Class ?m is:dM(x,?m) =?(x??m)T??1m (x??m) (2.19)where m? {1,2, ...,M}, ?m is the k-dimensional mean vector, and ?m is the k?k covariance matrixof Class ?m.The input vector x is classified as belonging to Class ?i if:dM(x,?i) = minj(dM(x,? j)), j ? {1,2, ...,M} (2.20)The decision rule can be simplified as follows if only two classes exist:x ?{?1 : dM(x,?1)? dM(x,?2)?2 : dM(x,?1) > dM(x,?2)(2.21)30Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIUsing Equation (2.19) in Equation (2.21), the discriminant rule becomes:x ?{?1 : md f (x)? 0?2 : md f (x) < 0(2.22)where the Mahalanobis discriminant function, md f (x), is defined as:md f (x) =?xT (??11 ???12 )x+2(?T1 ??11 ??T2 ??12 )x? (?T1 ??11 ?1??T2 ??12 ?2) (2.23)The mean vectors (i.e., ?1 and ?2) and the covariance matrices (i.e., ?1 and ?2) are estimatedfrom the data samples.2.4.4 Support vector machineA support vector machine (SVM) [44] is a supervised learning method that is applicable to classi-fication and regression. The SVM algorithm is based on statistical learning theory.An SVM classifies the data by constructing an N-dimensional hyper-plane that optimally sep-arates the data into two classes.SVM is closely related to artificial neural networks. To train the multi-layer perceptron neuralnetwork, an unconstrained non-convex minimization problem should be solved. In SVM, however,a quadratic programming problem with linear constraints should be solved.Different kernel functions such as polynomial, Gaussian radial basis function, and sigmoid canbe used in an SVM.2.4.5 Radial basis function neural networkThe Radial Basis Function (RBF) neural network (NN) is one of the most widely used nets. TheRBF network has two layers:? The hidden layer of radial basis functions. This layer is usually referred to as the radial basis31Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIFigure 2.3: The architecture of an RBF neural network.layer.? The output layer.Fig. 2.3 presents the architecture of an RBF neural network. The input vector, X , is a p-dimensional vector, and the numbers of the nodes in the hidden layer and the output layer are qand r, respectively.Each neuron in the radial basis layer has two parameters: the center vector, and the radius(or the spread). These parameters are determined during the training process. At the ith neuronof the radial basis layer, the Euclidean distance between the input vector and the center vector iscalculated and divided by the radius of the neuron. The obtained distance is then applied to theradial basis function to form the output of the node, as follows:Zi = exp(?(X?Ci)T (X?Ci)r2i)(2.24)where X is the input vector,Ci is the center vector of the ith neuron in the radial basis layer, and riand Zi are the radius and the output of the ith neuron in the radial basis layer, respectively.32Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIThe values of the neurons in the output layer are calculated based on the following equation:y j =q?i=1ZiWi j (2.25)where y j is the value of the jth neuron in the output layer, and Wi j is the weight of the connectionbetween the ith node in the hidden layer and the jth node in the output layer.Different algorithms based on solving a set of linear equations or gradient descent exist fortraining the network. For further details about RBF neural networks and existing training algo-rithms, please refer to [79].2.5 Results2.5.1 LDA classifier resultsThe TPR values obtained using the LDA classifier were high, but unfortunately the FPR valueswere also high (average rate above 10%). The BCI system was thus deemed impractical, as highrates of false positives frustrate the user. It was thus concluded that the performance of the LDAclassifier is not acceptable; hence the results are not given here.2.5.2 QDA classifier resultsBy analyzing the results of the cross-validation of the QDA classifier, we found that when theAR model order increased, the TPR first increased and then decreased, while the FPR alwaysdecreased. Unlike the LDA classifier, the QDA classifier even reached zero FPR values at someAR orders. We selected the AR order with a zero FPR. If the FPR had the same minimum for morethan one AR order, then the one with the highest TPR was chosen. The selected orders and thecross-validation results are presented in Table 2.2. The classification test results are given in Table2.3. The optimal orders were chosen independently for different tasks and different subjects.33Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCI2.5.3 MDA classifier resultsGenerally, for this classifier, as the AR model order increased, the TPR first increased and thendecreased, while the FPR always decreased. In all cases, a zero FPR was obtained at lower ARorders than for the QDA classifier. To select the optimal order, the same approach was used as forthe QDA classifier. The classification results for cross-validation and testing are shown in Tables2.2 and 2.3, respectively.2.5.4 SVM classifier resultsThe kernel function used in SVM was the Gaussian function with a scaling factor 0.5. The 0.5scaling factor gave the best results among scaling factors 0.25, 0.75, 1, 2, and 4.As before, the TPR first increased with increasing AR model order and then decreased at highorders. The FPR decreased as the order increased. The TPRs of the SVM classifier were muchhigher than the TPRs of other classifiers. The drawback of the SVM classifier is that it did notyield zero FPRs except in two cases:? the letter-composing task of Subject 1, and? the multiplication task of Subject 3.The performances over cross-validation and testing data are reported in Tables 2.2 and 2.3,respectively.2.5.5 RBF NN classifier resultsWe used the algorithm based on solving a set of linear equations to train the network. Table 2.2also illustrates the cross-validation results of the RBF NN classifier. The number of RBF networkinputs was equal to the number of features. In this case, the TPR also increased as the order of theAR model increased and sometimes decreased at high orders. The FPR decreased with the order.The TPRs were less than the TPRs of SVM, but in many cases a zero FPR was reached. Results of34Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCItesting the system are shown in Table 2.3.35Chapter2.PlausibilityAssessmentofaSelf-Paced2-StateMentalTask-BasedBCITable 2.2: Cross-Validation Results for Different Subjects, Tasks, and ClassifiersBaseline Multiplication Letter Composing Rotation CountingTPR FPR TPR FPR TPR FPR TPR FPR TPR FPRSubject Classifier AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SD1 QDA 13 82.52 1.31 0.00 0.00 14 82.08 1.06 0.00 0.00 16 75.73 1.73 0.00 0.00 17 70.03 1.48 0.00 0.00 17 72.66 1.62 0.00 0.00MDA 8 62.19 1.21 0.00 0.00 10 46.08 1.58 0.00 0.00 11 53.32 1.79 0.00 0.00 12 37.26 2.96 0.00 0.00 12 45.26 1.09 0.00 0.00SVM 20 91.95 1.57 0.04 0.08 19 93.15 0.64 0.05 0.09 19 92.55 0.93 0.00 0.00 19 92.27 2.09 0.10 0.10 20 90.14 0.69 0.05 0.09RBF NN 16 77.04 2.19 0.00 0.00 13 80.88 1.85 0.00 0.00 18 73.10 1.92 0.00 0.00 16 72.00 3.71 0.03 0.08 16 70.85 1.61 0.00 0.002 QDA 20 67.78 1.43 0.07 0.10 18 68.49 2.11 0.00 0.00 18 72.22 2.07 0.00 0.00 18 67.62 1.17 0.00 0.00 19 67.95 2.46 0.00 0.00MDA 20 40.11 1.44 0.00 0.00 14 47.67 1.30 0.00 0.00 7 26.79 2.23 0.00 0.00 11 44.77 2.08 0.00 0.00 10 27.01 2.68 0.00 0.00SVM 17 86.41 1.61 0.01 0.07 20 77.86 0.93 0.05 0.13 19 88.44 1.31 0.08 0.02 19 80.93 1.43 0.03 0.08 20 81.10 1.54 0.26 0.13RBF NN 16 72.49 1.82 0.00 0.00 19 69.53 1.42 0.00 0.00 15 78.52 2.42 0.00 0.00 19 65.86 1.29 0.00 0.00 18 69.37 3.00 0.08 0.023 QDA 17 66.36 2.08 0.00 0.00 18 68.38 1.80 0.00 0.00 18 69.81 1.42 0.00 0.00 16 75.51 1.06 0.00 0.00 18 67.01 1.21 0.00 0.00MDA 10 40.93 2.31 0.00 0.00 11 39.84 1.19 0.00 0.00 13 42.79 1.50 0.00 0.00 10 52.00 1.83 0.00 0.00 10 26.68 1.59 0.00 0.00SVM 20 86.08 1.42 0.01 0.07 20 88.77 1.12 0.00 0.00 20 89.86 1.76 0.21 0.19 19 92.77 1.20 0.14 0.05 20 87.40 1.08 0.08 0.08RBF NN 16 66.25 1.42 0.00 0.00 17 71.89 2.51 0.01 0.07 18 70.63 2.84 0.03 0.14 17 76.49 0.98 0.00 0.00 20 67.45 2.82 0.01 0.074 QDA 18 71.73 2.10 0.00 0.00 17 76.99 1.45 0.00 0.00 16 68.27 2.51 0.00 0.00 19 69.37 0.44 0.00 0.00 17 75.89 1.68 0.00 0.00MDA 12 45.75 2.18 0.00 0.00 9 42.63 1.06 0.00 0.00 11 41.04 1.38 0.00 0.00 9 49.10 1.77 0.00 0.00 12 48.71 2.11 0.00 0.00SVM 20 84.66 1.30 0.05 0.09 20 87.45 1.85 0.04 0.08 19 78.58 2.74 0.01 0.07 20 86.03 1.62 0.01 0.07 20 86.47 1.28 0.14 0.13RBF NN 18 72.88 2.38 0.00 0.00 20 74.52 1.37 0.00 0.00 16 62.63 2.58 0.00 0.00 17 69.81 2.55 0.01 0.07 16 76.55 1.16 0.01 0.07The table shows the mean and standard deviation values of the TPR and FPR measures. The results of different classifiers are given in four separate rows for each subject and eachmental task. The optimum AR order for each case is also included in the table.36Chapter2.PlausibilityAssessmentofaSelf-Paced2-StateMentalTask-BasedBCITable 2.3: Testing Results for Different Subjects, Tasks, and Classifiers with Selected Model OrdersBaseline Multiplication Letter Composing Rotation CountingTPR FPR TPR FPR TPR FPR TPR FPR TPR FPRSubject Classifier AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SD1 QDA 13 84.78 3.84 0.00 0.00 14 85.65 5.72 0.00 0.00 16 75.00 6.20 0.00 0.00 17 75.22 7.02 0.00 0.00 17 72.61 5.35 0.00 0.00MDA 8 63.26 6.81 0.00 0.00 10 48.04 8.40 0.00 0.00 11 51.52 5.25 0.00 0.00 12 38.26 3.47 0.00 0.00 12 49.57 6.18 0.00 0.00SVM 20 92.61 0.91 0.00 0.00 19 94.35 3.02 0.11 0.15 19 93.04 3.22 0.00 0.00 19 94.57 2.03 0.05 0.12 20 91.09 3.02 0.00 0.00RBF NN 16 77.83 3.40 0.00 0.00 13 80.87 6.90 0.00 0.00 18 75.00 2.55 0.00 0.00 16 74.78 4.76 0.00 0.00 16 72.17 5.36 0.00 0.002 QDA 20 69.35 6.54 0.05 0.12 18 68.70 5.40 0.00 0.00 18 75.00 2.43 0.00 0.00 18 68.04 2.73 0.05 0.12 19 66.96 7.32 0.00 0.00MDA 20 41.30 5.60 0.00 0.00 14 49.57 5.83 0.00 0.00 7 26.96 4.57 0.00 0.00 11 45.22 2.25 0.00 0.00 10 30.43 6.29 0.00 0.00SVM 17 86.96 6.29 0.00 0.00 20 80.43 3.92 0.00 0.00 19 88.70 5.52 0.05 0.12 19 80.65 2.09 0.05 0.12 20 81.96 2.94 0.11 0.15RBF NN 16 76.30 3.72 0.00 0.00 19 70.43 3.21 0.00 0.00 15 79.57 4.44 0.05 0.12 19 66.09 4.02 0.00 0.00 18 71.09 5.78 0.00 0.003 QDA 17 69.57 5.60 0.00 0.00 18 68.91 4.25 0.00 0.00 18 72.39 1.82 0.00 0.00 16 74.57 6.55 0.00 0.00 18 69.35 6.31 0.00 0.00MDA 10 43.48 6.70 0.00 0.00 11 37.17 4.31 0.00 0.00 13 42.17 8.04 0.00 0.00 10 51.96 4.02 0.00 0.00 10 26.30 4.17 0.00 0.00SVM 20 87.39 1.65 0.00 0.00 20 89.13 1.54 0.05 0.12 20 90.65 2.25 0.11 0.15 19 93.70 1.42 0.16 0.15 20 87.83 2.09 0.11 0.15RBF NN 16 66.96 4.52 0.00 0.00 17 73.70 3.56 0.00 0.00 18 73.48 7.15 0.00 0.00 17 76.74 4.11 0.00 0.00 20 66.30 4.68 0.00 0.004 QDA 18 71.74 3.84 0.00 0.00 17 78.70 4.52 0.00 0.00 16 71.09 4.18 0.00 0.00 19 72.83 5.10 0.00 0.00 17 76.09 3.77 0.00 0.00MDA 12 46.74 8.83 0.00 0.00 9 44.78 5.56 0.00 0.00 11 40.87 7.48 0.00 0.00 9 52.17 3.26 0.00 0.00 12 48.70 5.40 0.00 0.00SVM 20 85.22 3.96 0.05 0.12 20 89.57 5.88 0.11 0.15 19 77.39 4.02 0.00 0.00 20 85.87 3.17 0.05 0.12 20 86.09 6.36 0.16 0.15RBF NN 18 71.52 4.95 0.05 0.12 20 74.35 4.71 0.00 0.00 16 61.30 4.46 0.00 0.00 17 71.52 5.87 0.00 0.00 16 73.91 3.99 0.00 0.00The table shows the mean and standard deviation values of the TPR and FPR measures. The results of different classifiers are given in four separate rows for each subject and eachmental task. The optimum AR order for each case is also included in the table.37Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCI2.6 DiscussionAs seen from Table 2.3, the results of testing the system are robust and in line with the cross-validation results. In this table, the cases with non-zero FPRs are shown in bold. In most cases, ifthe FPR is zero for the cross-validation, then the FPR is also zero in the testing process. For thefour rare cases in which the results of testing are worse than those of cross-validation, the numbersare underlined in the table.In Table 2.4, the results of the t-test (on TPR values) for measuring the differences between theclassifiers are given. Table 2.5 presents the average performances of the classifiers for each subject(over tasks) and for each task (over subjects).2.6.1 Comparing classifiersThe AR orders selected for QDA are higher than those for MDA. But QDA outperforms MDA,since it has higher TPR values. The results of the t-test for these two classifiers show the significantperformance differences. The FPRs of MDA for all subjects and tasks are zero. In two cases, theobtained FPR of QDA is not zero (Subject 2, the baseline and rotation tasks, see Table 2.3). In thecase of Subject 2, and the rotation task, we tested the system with the next higher order (19) andobtained a TPR with mean 65.43% and standard deviation 1.94%, and an FPR with mean 0 andstandard deviation of 0. That is, FPR is zero and TPR is still higher than the 45.22% of MDA. Itseems that by increasing the order in the other case (Subject 2 and the baseline task), we obtain azero FPR while the TPR would remain still higher than for the MDA. Therefore, we can be surethat in terms of TPR and FPR, the QDA classifier outperforms MDA for higher AR orders.For the QDA and the RBF neural network classifiers, very close values were obtained for theAR orders, TPRs and FPRs. Their performances are not significantly different in most cases,as shown in Table 2.4. The RBF network is, however, more complex; therefore, using QDA ispreferable.38Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIThe SVM classifier has the highest AR orders. The TPR of SVM is also the highest. But thedrawback is its FPR, which did not reach zero. The t-test shows significant differences betweenSVM and the other classifiers. One point that should be mentioned here is that the orders selectedfor SVM are mostly either 19 or 20. Therefore, it may be possible that we can get a zero FPR inorders higher than 20, but we cannot be sure whether the TPR would be still high or not.2.6.2 Selecting the classifierTo choose the best classifier for each subject and each task, the most important emphasis consideredis on whether or not a zero FPR is reached. The classifier should have a zero FPR not only duringthe cross-validation, but also during the testing process.? For Subject 1:? For the baseline task, the best classifier is SVM according to Table 2.3, since it has thehighest TPR and a zero FPR. But if we consider the results of cross-validation (Table2.2), SVM does not have a zero FPR; therefore, the best classifier would be QDA,which has zero FPRs in both cross-validation and testing.? For the multiplication task, the best classifier is also QDA.? For the letter-composing task, SVM is the best classifier, since it shows zero FPRsduring cross-validation and testing, and has the highest TPR.? For the rotation task, QDA is the best.? For the counting task, QDA and the RBF network have almost the same performance,but QDA is preferable because of its simplicity. For this mental task, SVM has a zeroFPR on the testing data, but since its FPR during cross-validation is not zero, it is notselected as the best classifier.39Chapter2.PlausibilityAssessmentofaSelf-Paced2-StateMentalTask-BasedBCITable 2.4: T-Test ResultsBaseline Multiplication Letter Composing Rotation CountingSubject Classifier MDA SVM RBF NN MDA SVM RBF NN MDA SVM RBF NN MDA SVM RBF NN MDA SVM RBF NN1 QDA < 0.0005 < 0.005 < 0.01 < 0.0001 < 0.01 > 0.05 < 0.0001 < 0.0005 > 0.05 < 0.0001 < 0.0005 > 0.05 < 0.0005 < 0.0001 > 0.05MDA < 0.0001 < 0.005 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0005SVM < 0.0001 < 0.005 < 0.0001 < 0.0001 < 0.00012 QDA < 0.0001 < 0.005 < 0.05 < 0.0005 < 0.005 > 0.05 < 0.0001 < 0.0005 < 0.05 < 0.0001 < 0.0001 > 0.05 < 0.0001 < 0.005 > 0.05MDA < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001SVM < 0.01 < 0.005 < 0.05 < 0.0001 < 0.0053 QDA < 0.0001 < 0.0001 > 0.05 < 0.0001 < 0.0001 < 0.05 < 0.0001 < 0.0001 > 0.05 < 0.0001 < 0.0005 > 0.05 < 0.0001 < 0.0005 > 0.05MDA < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001SVM < 0.0001 < 0.0001 < 0.0005 < 0.0001 < 0.00014 QDA < 0.0005 < 0.0005 > 0.05 < 0.0001 < 0.01 > 0.05 < 0.0001 < 0.05 < 0.005 < 0.0001 < 0.001 > 0.05 < 0.0001 < 0.01 > 0.05MDA < 0.0001 < 0.0005 < 0.0001 < 0.0001 < 0.0001 < 0.0005 < 0.0001 < 0.0001 < 0.0001 < 0.0001SVM < 0.001 < 0.001 < 0.0005 < 0.001 < 0.005p?0.05 means there is no significant difference.40Chapter2.PlausibilityAssessmentofaSelf-Paced2-StateMentalTask-BasedBCITable 2.5: Average Performances of Classifiers for Different Subjects and TasksQDA MDA SVM RBF NNTPR FPR TPR FPR TPR FPR TPR FPRAR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SDAverage over tasksSubject 1 15.40 78.65 5.63 0.00 0.00 10.60 50.13 6.02 0.00 0.00 19.40 93.13 2.44 0.03 0.05 15.80 76.13 4.59 0.00 0.00Subject 2 18.60 69.61 4.88 0.02 0.05 12.40 38.70 4.91 0.00 0.00 19.00 83.74 4.15 0.04 0.08 17.40 72.70 4.23 0.01 0.02Subject 3 17.40 70.96 4.91 0.00 0.00 10.80 40.22 5.45 0.00 0.00 19.80 89.74 1.79 0.09 0.11 17.60 71.44 4.80 0.00 0.00Subject 4 17.40 74.09 4.28 0.00 0.00 10.60 46.65 6.11 0.00 0.00 19.80 84.83 4.68 0.07 0.11 17.40 70.52 4.80 0.01 0.02Average over subjectsBaseline 17.00 73.86 4.96 0.01 0.03 12.50 48.70 6.99 0.00 0.00 19.25 88.05 3.20 0.01 0.03 16.50 73.15 4.15 0.01 0.03Multiplication 16.75 75.49 4.97 0.00 0.00 11.00 44.89 6.03 0.00 0.00 19.75 88.37 3.59 0.07 0.11 17.25 74.84 4.60 0.00 0.00Letter Composing 17.00 73.37 3.66 0.00 0.00 10.50 40.38 6.34 0.00 0.00 19.25 87.45 3.75 0.04 0.07 16.75 72.34 4.65 0.01 0.03Rotation 17.50 72.67 5.35 0.01 0.03 10.50 46.90 3.25 0.00 0.00 19.25 88.70 2.18 0.08 0.13 17.25 72.28 4.69 0.00 0.00Counting 17.75 71.25 5.69 0.00 0.00 11.00 38.75 5.51 0.00 0.00 20.00 86.74 3.60 0.10 0.11 17.50 70.87 4.95 0.00 0.00Total Average 17.20 73.33 4.92 0.01 0.01 11.10 43.92 5.62 0.00 0.00 19.50 87.86 3.27 0.06 0.09 17.05 72.70 4.61 0.01 0.0141Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCI? For Subject 2:? For the baseline, the multiplication, and the rotation tasks, the RBF net is the bestclassifier.? QDA is the selected classifier for the remaining two tasks.? The RBF network is not selected for the counting task, because its FPR during cross-validation is not zero.? For Subject 3:? QDA is selected for all tasks.? For the rotation task, one can think of the RBF neural network as the best classifier, butsince it is not significantly different from QDA (Table 2.4), QDA is selected instead.? For Subject 4:? The best classifier for all tasks is also QDA.As mentioned before, the main factor in selecting the best classifier is having a zero FPR. TPRand simplicity are the next factors. The best classifier should be selected based on the individualsneeds for the application. For each subject, the best classifier once determined will be used in thefinal design of the BCI system.2.6.3 Most discriminatory tasksBased on the best classifier selected for each case, the most discriminatory task for each subject isthen found:? For Subject 1, the letter-composing task is the most discriminatory task. The TPR reachedhas a mean of 93.04% and a standard deviation of 3.22%.? For Subject 2, the baseline task is the best choice, followed by the letter-composing task. ABCI based on the baseline task however, does not make sense and is not practical. This isbecause the output of the BCI would be activated when the subject wants to relax and think42Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCITable 2.6: The Best Classifier for Each Subject and Each Mental TaskBaseline Multiplication Letter Composing Rotation CountingTPR TPR TPR TPR TPRSubject Classifier Mean SD Classifier Mean SD Classifier Mean SD Classifier Mean SD Classifier Mean SD1 QDA 84.78 3.84 QDA 85.65 5.72 SVM 93.04 3.22 QDA 75.22 7.02 QDA 72.61 5.352 RBF 76.30 3.72 RBF 70.43 3.21 QDA 75.00 2.43 RBF 66.09 4.02 QDA 66.96 7.323 QDA 69.57 5.60 QDA 68.91 4.25 QDA 72.39 1.82 QDA 74.57 6.55 QDA 69.35 6.314 QDA 71.74 3.84 QDA 78.70 4.52 QDA 71.09 4.18 QDA 72.83 5.10 QDA 76.09 3.77For each subject, the information about the most discriminatory task is in bold. The FPR value is zero for every case.of nothing. Therefore, in this case, the letter-composing task is selected, with TPR meanand standard deviation of 75.00% and 2.43%, respectively.? For Subject 3, the rotation task is the most discriminatory task, since QDA (which wasselected as the best classifier) has the best results for this task. The obtained TPR value hasa mean of 74.57% and a standard deviation of 6.55%.? For the same reason, the multiplication task is selected for Subject 4 as the most discrimina-tory task. In this case, the TPR mean is 78.70% and the TPR standard deviation is 4.52%.For all of the most discriminatory tasks, the FPR value reaches zero. These results are summa-rized in Table 2.6.2.6.4 High TPRs versus low FPRsA low TPR and a high FPR will make the user frustrated. Suppose the proposed system is im-plemented in a real time experiment and it outputs at the same frequency as the sampling, whichis 250 Hz. Due to the EEG segment length considered, the first output of the system would beavailable after the first second has passed (i.e., there exists a 1-sec delay). After the 1-sec delay,the system produces an output for each obtained sample. Hence, the number of generated outputsin a 1-minute period is 250?60?1=14999?15000. Since an FPR value of 0.05% equals 5 falsepositives in every 10000 outputs, the system would have a false activation rate of 7.5 activations43Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIper minute. This rate of false activation will likely frustrate the user. Suppose another design for areal time implementation of the system produces outputs at 5 Hz. Once the first second has passedand the first output is generated, the system yields an output every 50 samples, i.e., the systemconsiders 250-sample segments overlapping by 200 samples with adjacent segments. In this case,the FPR value of 0.05% would yield 1.5 false activations every 10 minutes. This false activationrate may be acceptable depending on the application. The usefulness and acceptability of an FPRvalue of 0.05% depends on the frequency at which the BCI system outputs a control signal, and onthe application.The ideal value for a TPR is 100%; nevertheless, a TPR value of 75% means that 3 out ofevery 4 attempts yield true activations. This may work for many applications with a reasonableuser satisfaction. What we are trying to emphasize is that reaching a zero or a near-zero falsepositive rate is much more important than getting a TPR value of 100%. However, the degree ofimportance of each rate depends on the application for which the BCI system is being used. Atradeoff between achieving a high TPR and a low FPR is inevitable.2.6.5 Computational costThe computational cost increases as the AR model order increases. Table 2.7 shows the typicalprocessing time required for estimating the AR coefficients, training, and testing the classifiersfor different AR orders. For the QDA and MDA classifiers, the training time is negligible andis included in the testing time. The system was implemented in MATLAB 7.4.0 on a personalcomputer with a dual-core 2.13 GHz CPU and 1 GB RAM. As shown in Table 2.7, the processingtime increases with the AR model order. The time required to estimate the AR coefficients variesfrom 48.14 msec to 61.81 msec. The QDA and MDA classifiers are almost the same in terms ofcomputational cost, and they both need the least processing time, on the order of 10-6 sec. TheSVM classifier requires the most time to be trained, from 5.63 sec for the AR order 2, to 172.9444Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCIsec for the AR order 20. The testing of SVM takes little time (on the order of 10-6 sec), but longerthan QDA and MDA. Even though the RBF network requires less time for training than SVM, itneeds more time (on the order of msec) for testing.45Chapter2.PlausibilityAssessmentofaSelf-Paced2-StateMentalTask-BasedBCITable 2.7: Typical Processing Time Required in Different Parts of the BCI SystemAR model order2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20AR model estimation (msec) 48.14 48.50 49.29 49.36 50.47 51.05 51.46 52.95 53.16 54.36 54.76 55.77 56.48 58.04 58.27 59.66 59.93 61.10 61.81QDA (?sec) 12 13 16 21 27 33 38 47 55 64 73 84 91 103 116 132 142 159 168MDA (?sec) 12 13 17 21 26 33 39 47 54 63 73 82 91 103 116 128 139 157 172SVMTraining (sec) 5.63 4.43 7.27 11.61 20.65 40.97 64.33 99.38 131.08 150.27 160.43 167.58 170.77 171.68 171.24 170.52 171.39 170.25 172.94Testing (?sec) 41 55 79 109 137 173 268 270 296 365 323 405 462 477 521 560 556 596 654RBF NNTraining (sec) 4.79 4.58 4.86 5.13 5.42 7.70 5.96 6.32 6.61 6.87 7.24 7.58 7.88 8.28 8.55 8.92 9.24 9.60 9.97Testing (msec) 0.91 0.99 1.15 1.34 1.17 1.48 1.93 2.15 2.37 2.61 2.85 3.10 3.35 3.59 3.85 4.07 4.32 4.55 4.8946Chapter 2. Plausibility Assessment of a Self-Paced 2-State Mental Task-Based BCI2.7 SummaryIn this study, we present different designs for a 2-state self-paced mental task-based BCI. Our goalwas to minimize the false activation rate in a BCI, in which the output is controlled by one of fivemental tasks. In other words, whenever the subject performs a certain mental task, the output ofthe BCI should be activated.The scalar autoregressive model coefficients of EEG signals were used as features. The per-formance of the different classifiers such as LDA, QDA, MDA, SVM, and RBF NN was studied.We performed a 5?5 cross-validation for choosing the best autoregressive model order for each ofthese classifiers. We also tested the classifiers with the selected orders.The performance of the LDA classifier was not acceptable. The QDA classifier had betterresults than MDA. RBF NN performed almost as well as QDA. The only two classifiers that werenot significantly different were QDA and RBF NN. SVM showed a higher TPR, but its non-zeroFPR was a disadvantage.The MDA classifier needed the lowest AR model orders and the SVM classifier needed thehighest. The AR orders that yielded the optimal performance for the QDA and RBF NN classifierswere not very different. It may be possible to obtain a zero FPR with SVM if we increase the ARmodel order. We found that the best discriminatory task varied from one subject to another, whichconfirms the findings of previous studies using the same dataset. For example, see [133, 134, 140,163].The results of this study are extremely encouraging, since they show that it is feasible to havea self-paced BCI that would be activated by mental tasks and yield a false activation rate thatapproaches zero. However, since the data used were not collected in a self-paced paradigm, wecannot yet claim that the designed 2-state self-paced BCIs would work as well in real life. The nextimportant step is to collect our own data in a self-paced paradigm in order to study the feasibilityof self-paced mental task-based BCIs in real-life applications.47Chapter 3Towards Development of a Self-Paced2-State Mental Task-Based BCIIn the previous chapter, we have investigated the plausibility of having a mental task-based BCIwith a zero or near zero false activation rate. However, since the dataset used was old and small,we could not make a strong hypothesis based on the obtained results. We have thus carried outour own experiments to collect an appropriate EEG dataset to test the self-paced BCI design. Thedetails of our data collection process are provided in the next section.In this study, similar to our previous work, we develop a 2-state mental task-activated BCI. Theaim is to design a self-paced BCI system with a zero or near-zero false activation rate. The EEGsignals obtained from four subjects performing four different mental tasks and when they are inthe baseline state are analyzed.The design of the proposed BCI is simple. The coefficients obtained from the autoregressive(AR) modeling of the EEG signals are used as the features and quadratic discriminant analysis(QDA) serves as the classifier.The BCI is custom designed for each subject and each mental task. In other words, the optimumAR model order is obtained via five-fold cross-validation for each task performed by each subjectseparately. Therefore, the AR orders of the different cases are not necessarily the same.48Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCIIn each BCI, one of the mental tasks acts as the intentional-control (IC) task that activates theBCI. The BCI should be in the no-control (NC) state during the other three mental tasks as well asduring the baseline condition. For each subject, we thus have four BCIs based on the four mentaltasks. Based on the results of the five-fold cross-validation process, we are able to determine theBCI with the best performance for each subject.3.1 Methodology3.1.1 DatasetThe EEG signals were collected from four subjects. All subjects were male, right-handed, andmedication-free. They were aged 24 to 31, with no previous head injury, and with general goodhealth and good vision (i.e., the ability to easily recognize experimental symbols on the display).Our experimental protocol had been reviewed by the Behavioural Research Ethics Board (BREB)of the University of British Columbia and the procedures had been found to be acceptable on ethi-cal grounds for research involving human subjects. All subjects signed the consent form requiredby the BREB.The subjects were seated comfortably in a chair in a 4?4 m2 room, approximately 75 cm infront of a computer monitor. An electrode cap was placed on the subject?s head to capture the EEGsignals from 29 channels. The electrodes were distributed over the scalp at Fpz, AF3, AF4, F7, F3,Fz, F4, F8, FC5, FC1, FC2, FC6, T7, C3, Cz, C4, T8, CP5, CP1, CP2, CP6, P7, P3, Pz, P4, P8,PO3, PO4, and Oz, according to the 10-10 (or 10%) system [95, 131] to study the signals of thedifferent brain regions. Fig. 3.1 shows the electrode positions.As shown in Fig. 3.2, five electrodes were placed around the eyes to record the EOG signals.Two electrodes were also positioned on the right forearm to monitor the EMG activities related tothe hand extension.One reference electrode was attached to each of the subject?s earlobes and linked together.49Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCIFigure 3.1: EEG signals were recorded from 29 electrodes distributed over the scalp accord-ing to the 10-10 system.Figure 3.2: EOG electrodes are placed around the eyes as depicted.The EEG signals were amplified using L64 Dr. Sagura amplifier and sampled at 500 Hz by a PCequipped with a 12-bit analog-to-digital converter. The signals were bandpassed at 0.03?150 Hz.The EEG signals of subjects which were asked to perform four different mental tasks wererecorded in three sessions (each session on a different day). In each session, after the preparationand setup, the EEG recording was done in six runs.Each run consisted of approximately 12 minutes of recording. The subjects were instructed totake breaks as long as they wanted between the runs. During each run, five epochs of EEG signalsof four different mental tasks (20 epochs in total) were recorded. Each epoch took 32.5?2.5 sec.50Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCIFigure 3.3: Epoch timing: After a break of length of 15?2.5 sec, a ?Start Cue? was displayedon the screen for 4 sec. The subject was told to wait about 1 sec after the cue disap-peared before performing a mental task for about 10 sec. ?Stop Cue" was displayed onthe screen for 2.5 sec informing the subject of the end of the 10-sec interval. The nextepoch then started.Please see Fig. 3.3. The subjects performed four different mental tasks in a random order so as toavoid possible adaptation.The data were recorded when the subjects were performing guided mental tasks. The eyes wereopen during the experiment.Each epoch started with a break of a length 15?2.5 sec. The break lengths varied within therange of ?2.5 sec in different epochs so as to avoid possible adaptation. After a break, a cue,called ?Start Cue?, was displayed on the screen for 4 sec prompting the subject to perform aspecific mental task. The subject was told to start to perform the mental task about 1 sec after thecue disappears. He or she should keep performing the mental task for 10 sec. Another cue, called?Stop Cue", was displayed on the screen for 2.5 sec informing the subject of the end of the 10-secinterval. The next epoch then started. Refer to Fig. 3.3 to see the timing of each epoch.The screen background was black all the time. The Start Cue prompting the subject to performa mental task had the name of that mental task written in white and in a size which could be easilyread from a distance of 75 cm. The Stop Cue was the word ?STOP" written in a green circle.During the break, ?Break? was written on the screen in white and having the same size as the StartCue.The mental tasks included:1. Visualizing some words being written on a board: the subject should imagine a board on51Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCIwhich he was writing his full name.2. Non-trivial multiplication: the subject mentally multiplied a 2-digit number by another 2-digit number. The numbers were given to the subject as the Start Cue.3. Mentally rotating a 3D object: the subject imagined that he was rotating a laptop mentally.4. Motor imagery: the subject imagined a right hand extension.During the break interval of each epoch, the subject was asked to be in the baseline state. Inthis state, the subject should not perform any of the four mental tasks of the experiment and wassupposed to remain looking at the screen without moving (i.e., attain the same physical conditionas that assumed during the mental tasks states).Each session contains 300 sec (30 10-sec epochs) of EEG signals for each of the four mentaltasks, and about 1800 sec (120 epochs) of EEG signals for the baseline. Therefore, at the end ofthe last session, we had 90 epochs of each mental task and 360 epochs of the baseline for eachsubject.During every session, the subjects were informed and frequently reminded that they would notbe evaluated on their performance and that there were absolutely no expectations. This was tominimize the concern for adverse psychological effects.3.1.2 ProcedureFor each subject, 90 epochs of each mental task and 360 epochs of the baseline are obtained.Each mental task epoch is 10 sec long; however, the length of the baseline epochs is variable (i.e.,15?2.5 sec). The sampling rate is 500 Hz; therefore, there are 5000 samples in each mental taskepoch. The baseline epochs have between 6250 and 8750 samples.We divide the epochs into 500-sample overlapping segments. Each segment overlaps with theadjacent segment by 400 samples. The features used for classification are the set of coefficients ofthe AR model of the segments. Based on the studies [60] and [25], 1 sec of the signal is sufficient52Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCIto estimate the AR model.The AR coefficients of the 29 EEG channels are concatenated into a single vector to form thefinal feature vector. The classifier is based on quadratic discriminant analysis. Brief descriptionsof the feature extraction method and the classifier are given in the next sections.To find the optimum AR model order, a 5-fold nested cross-validation process is performed.The whole set of segments is divided into five sections of equal size (i.e., the outer fold). For eachouter fold, the system is trained and validated with four of the data sections and tested with theremaining section. The four sections assigned to training and validation is further divided into fiveequal-sized data partitions that form the inner fold. For each inner fold, the training is performedwith four partitions and the validation is completed with the remaining partition. Thus, for eachouter fold, there are five inner folds. The cross-validation results are averaged over 25 differentcases. The testing results report the average over 5 cases. Performing cross-validation makes theresults more robust.For every subject and every mental task, the initial AR order is equal to 2 in the cross-validationprocess. If the FPR value for an AR order reaches zero, that order is selected as the optimum ARorder; otherwise, the order is increased by one, and the FPR of the new order is then calculated.Increasing the AR order is continued until the FPR value is zero or a maximum AR order of 40 isreached. In the latter case, the AR order corresponding to the minimum FPR value is chosen.It has been proven that customizing the BCI for each subject improves the overall performanceof the system [18, 29]. Here, we also custom design the BCI system by allowing different ARorders to be selected for different subjects and mental tasks during the cross-validation process.3.2 ResultsWe used WestGrid computing resources for processing. WestGrid is one of seven partner consortiathat make up Compute Canada, a national platform that integrates high performance computing53Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCIresources across the country to create a dynamic computational resource.The receiver operating characteristic (ROC) curves of our BCI system for the four subjectsand four mental tasks are shown in Fig. 3.4?3.7. The performance evaluation results of our BCIsystems are summarized in Table 3.1. This table shows the mean and standard deviation values ofthe TPR and FPR measures. The results of the cross-validation and testing processes are given intwo separate rows for each subject and each mental task. The optimum AR order for each case isalso included. In this table, the values related to the mental tasks with the lowest and the highestperformance are shown in italics and in bold, respectively. Table 3.2 illustrates the average systemperformance over the mental tasks and subjects. Table 3.3 shows the system performance of thebest BCI for each subject.To investigate whether the TPR values (for each mental task of each subject) meet the Normaldistribution condition, the one-sample Kolmogorov-Smirnov test [117] is used. The significancelevel is set at ? = 0.05. Based on the results shown in Table 3.4, it is concluded that the TPRs ofeach case have a Normal distribution.The Welch?s t-test [176], as a statistical significance test, is performed on the TPR values formeasuring the performance differences between every pair of the four mental tasks and betweenevery pair of the four subjects.The null hypothesis is considered to be that there is no difference between the TPR values ofthe 2 groups (that can be 2 mental tasks or 2 subjects). Assuming a 5% significance level, the nullhypothesis is rejected if the resultant p-value is less than 0.05. This implies that the TPR valuesof the 2 test groups are significantly different. If p ? 0.05, the null hypothesis cannot be rejectedat the 5% significance level. This means that there is no significant difference between the TPRvalues of the test groups.A significance level of ? = 0.05 is used for the analyses. For each subject, the results of thet-test between any 2 of the four mental tasks are given in Table 3.5. Table 3.6 shows the p-values54Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCIbetween the four subjects for each mental task.55Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCI(a) Visualization (b) Multiplication(c) Object rotation (d) Motor imageryFigure 3.4: ROC curves for BCIs of Subject 1. The values shown on the plot are the corre-sponding AR model orders.56Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCI(a) Visualization (b) Multiplication(c) Object rotation (d) Motor imageryFigure 3.5: ROC curves for BCIs of Subject 2. The values shown on the plot are the corre-sponding AR model orders.57Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCI(a) Visualization (b) Multiplication(c) Object rotation (d) Motor imageryFigure 3.6: ROC curves for BCIs of Subject 3. The values shown on the plot are the corre-sponding AR model orders.58Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCI(a) Visualization (b) Multiplication(c) Object rotation (d) Motor imageryFigure 3.7: ROC curves for BCIs of Subject 4. The values shown on the plot are the corre-sponding AR model orders.59Chapter3.TowardsDevelopmentofaSelf-Paced2-StateMentalTask-BasedBCITable 3.1: Cross-Validation and Testing Results for Different Subjects and Mental TasksSentence Visualization Multiplication Object Rotation Motor ImageryTPR FPR TPR FPR TPR FPR TPR FPRSubject Process AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SD1 Validation 23 67.28 0.27 0.00 0.00 26 68.33 0.39 0.00 0.00 21 70.53 1.19 0.00 0.00 26 65.51 0.94 0.00 0.00Test 67.45 1.56 0.00 0.00 69.18 1.91 0.00 0.00 70.86 2.89 0.00 0.00 66.67 1.04 0.00 0.002 Validation 17 71.47 0.31 0.00 0.00 20 70.92 0.76 0.00 0.00 19 69.76 0.64 0.00 0.00 19 66.78 0.35 0.00 0.00Test 72.65 1.52 0.00 0.00 70.84 1.16 0.00 0.00 70.79 1.72 0.00 0.00 67.25 2.55 0.00 0.003 Validation 20 70.76 0.42 0.00 0.00 21 71.78 0.44 0.00 0.00 21 69.80 1.07 0.00 0.00 19 71.82 0.16 0.00 0.00Test 70.63 1.46 0.00 0.00 71.05 1.63 0.00 0.00 69.67 0.67 0.00 0.00 71.85 1.19 0.00 0.004 Validation 26 64.46 0.73 0.00 0.00 29 59.98 0.67 0.00 0.00 34 51.43 0.45 0.00 0.00 27 63.13 1.53 0.00 0.00Test 65.06 1.57 0.00 0.00 58.44 1.91 0.00 0.00 51.50 3.06 0.00 0.00 62.24 2.08 0.00 0.00The table shows the mean and standard deviation values of the TPR and FPR measures. The results of the cross-validation and testing processes aregiven in two separate rows for each subject and each mental task. The optimum AR order for each case is also included in the table.60Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCITable 3.2: Average of Testing Results over Mental Tasks and SubjectsTPR FPRAR Mean SD Mean SDAverage over TasksSubject 1 24 68.54 2.55 0.00 0.00Subject 2 19* 70.38 2.67 0.00 0.00Subject 3 20* 70.80 1.51 0.00 0.00Subject 4 29 59.31 5.55 0.00 0.00Average over SubjectsSentence Visualization 21* 68.95 3.29 0.00 0.00Multiplication 24 67.38 5.48 0.00 0.00Object Rotation 24* 65.71 8.53 0.00 0.00Motor Imagery 23* 67.00 3.86 0.00 0.00Total Average 23 67.26 5.78 0.00 0.00* The value is rounded to the nearest integer.Table 3.3: System Performance of the Best BCI for Each SubjectTPR FPRAR Mean SD Mean SDSubject 1 21 70.86 2.89 0.00 0.00Subject 2 17 72.65 1.52 0.00 0.00Subject 3 19 71.85 1.19 0.00 0.00Subject 4 26 65.06 1.57 0.00 0.00Average 21* 70.11 1.79 0.00 0.00* The value is rounded to the nearest integer.Table 3.4: p-Values Calculated Using Kolmogorov-Smirnov Normality Test for Each MentalTask of Each SubjectMental TaskSubject Sentence Visualization Multiplication Object Rotation Motor Imagery1 0.2382 0.6946 0.9434 0.98252 0.9986 0.5738 0.9659 0.99963 0.9001 0.9468 0.9475 0.98624 0.2396 0.9820 0.9658 0.9961p? 0.05 means ?the TPR values have a normal distribution".61Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCITable 3.5: p-Values Calculated Using Welch?s T-Test between Every Pair of the Mental Tasksfor Each SubjectMental TaskSubject Mental Task Sentence Visualization Multiplication Object Rotation Motor Imagery1 Sentence Visualization 0.1569 0.0583 0.3833Multiplication 0.1569 0.3144 0.0406Object Rotation 0.0583 0.3144 0.0283Motor Imagery 0.3833 0.0406 0.02832 Sentence Visualization 0.0696 0.1081 0.0055Multiplication 0.0696 0.9585 0.0310Object Rotation 0.1081 0.9585 0.0368Motor Imagery 0.0055 0.0310 0.03683 Sentence Visualization 0.6793 0.2331 0.1870Multiplication 0.6793 0.1369 0.4036Object Rotation 0.2331 0.1369 0.0108Motor Imagery 0.1870 0.4036 0.01084 Sentence Visualization 0.0004 0.0001 0.0441Multiplication 0.0004 0.0039 0.0170Object Rotation 0.0001 0.0039 0.0003Motor Imagery 0.0441 0.0170 0.0003p? 0.05 means ?there is no significant difference". These cases are shown in bold.3.3 DiscussionAs can be seen from Table 3.1, the FPR value is zero for all subjects and mental tasks. Themaximum TPR, which is 72.65%, is reached by the BCI based on the sentence visualization taskof Subject 2. The rotation task-based BCI of Subject 4 has a minimum TPR value equal to 51.50%.From Tables 3.1 and 3.5, we deduce the following:? For Subject 1, the significant differences in the performance only exist between the motorimagery-based BCI and those based on multiplication and rotation tasks. The BCIs basedon the sentence visualization, multiplication, and rotation tasks all have similar performancewith TPRs in the range of 67.45%?70.86%.? For Subject 2, the significant differences only exist between the motor imagery and each ofthe other mental tasks. As in Subject 1, the performance of the BCIs based on the sentence62Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCITable 3.6: p-Values Calculated Using Welch?s T-Test between Every Pair of the Subjects forEach Mental TaskSubjectMental Task Subject 1 2 3 4Sentence Visualization 1 0.0007 0.0105 0.04222 0.0007 0.0645 0.00013 0.0105 0.0645 0.00044 0.0422 0.0001 0.0004Multiplication 1 0.1433 0.1354 0.00002 0.1433 0.8209 0.00003 0.1354 0.8209 0.00004 0.0000 0.0000 0.0000Object Rotation 1 0.9643 0.4158 0.00002 0.9643 0.2309 0.00003 0.4158 0.2309 0.00014 0.0000 0.0000 0.0001Motor Imagery 1 0.6565 0.0001 0.00562 0.6565 0.0118 0.00993 0.0001 0.0118 0.00014 0.0056 0.0099 0.0001p? 0.05 means ?there is no significant difference".These cases are shown in bold.visualization, multiplication, and rotation tasks are not significantly different from each otherand the TPRs are between 70.79% and 72.65%.? For Subject 3, the motor imagery task is significantly different from the rotation task. Allother cases are statistically similar to each other. The TPR varies in the range 69.67%?71.85% for different mental tasks.? For Subject 4, all mental tasks are significantly different from each other (Table 3.5). TheBCI based on the sentence visualization task has the best performance with a TPR of 65.06%and the motor imagery task-based BCI with 62.24% TPR is the second best BCI.63Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCI3.3.1 The least effective tasksFor Subjects 1 and 2, the task that yielded the least effective performance is motor imagery. ForSubject 1, this task is significantly different from other tasks except for sentence visualization. ForSubject 2, the motor imagery task is significantly different in performance from all other tasks.For Subject 3, the rotation task has the lowest TPR; however, it is not significantly differentfrom the sentence visualization and rotation tasks.For Subject 4, the rotation task has the poorest performance and is significantly different fromall other tasks.3.3.2 Optimum AR model ordersAccording to Tables 3.1 and 3.2, the optimum AR orders are in the range 17 to 34, with an overallmean of 23. The average AR orders for Subject 4 is 29 which is the highest amongst the subjects,while the lowest AR order average is 19 which belongs to Subject 2.3.3.3 Subject dependencyThe performance of the system is completely subject-dependent. This can be easily recognized bylooking at the average of the performance over the mental tasks in Table 3.2. Among the subjects,Subject 4 has the BCIs with the poorest performance with an average TPR of 59.31%. Accordingto Table 3.6, the results obtained for Subject 4 are significantly different from the ones obtainedfor other subjects. The BCIs belonging to Subject 3 have an average TPR of 70.80% which is thehighest. Please note that these are not significantly different from the BCIs of Subject 2.3.3.4 Mental task dependencyThe average performances of the BCIs over the subjects are also given in Table 3.2. The BCIsbased on the sentence visualization task have the highest average TPR of 68.95% and the rotationtask-based BCIs have the lowest TPR of 65.71%. However, no significant difference was found64Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCIbetween the average performances of the mental tasks over the subjects.3.3.5 Comparison to our previous workIn the previous chapter, we used 5 different classifiers and also customized the classificationmethod for every subject and mental task. The classifiers included LDA, QDA, MDA, SVM,and the RBF neural network. Our findings showed that QDA outperforms the others in terms ofaccuracy and simplicity. This is why the QDA classifier has been selected as the classifier in thepresent study.The dataset used in the previous chapter is much smaller than the dataset employed in thecurrent work. The previous dataset contains only ten 10-sec epochs of EEG signals for each of thefour mental tasks and for the baseline. Our new dataset contains ninety 10-sec epochs of each ofthe four mental tasks and three hundred and sixty 15-sec epochs of the baseline.Moreover, the subjects, some of the mental tasks, and the experimental conditions such asrecording protocols, sampling frequencies and the number of channels are completely different.In spite of the above-mentioned dissimilarities, the performance of the system that uses theQDA classifier in the previous work is shown in Table 3.7. The purpose is to give a rough measureof comparison.Two of the mental tasks (i.e., the non-trivial multiplication task and the 3D object rotation task)of our previous study are almost similar to the corresponding tasks in the present study. For theletter composing task, the subjects were asked to mentally compose a letter to a friend. During thecounting task, the subjects were visualizing a sequence of numbers being written on a blackboard.Comparing Table 3.1 to Table 3.7, we see that the new results are generally consistent with ourprevious work. Specifically, for the BCIs based on the mental tasks that are similar for both studies(i.e., the rotation and the multiplication tasks), the results are of the same order of magnitude withthe exception of the multiplication task-based BCI of Subject A of the previous study which had a65Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCIhigh TPR of 85.65%. The FPR of the rotation task-based BCI of Subject B was 0.05%. It shouldreach zero if the AR order increases by one or two.As mentioned earlier, the results of the previous study is based on a small and old dataset ofEEG signals. The newly obtained results are hence much more reliable than our previous results.66Chapter3.TowardsDevelopmentofaSelf-Paced2-StateMentalTask-BasedBCITable 3.7: System Performance [60]Counting Multiplication Object Rotation Letter ComposingTPR FPR TPR FPR TPR FPR TPR FPRSubject Process AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SD AR Mean SD Mean SDA Validation 17 72.66 1.62 0.00 0.00 14 82.08 1.06 0.00 0.00 17 70.03 1.48 0.00 0.00 16 75.73 1.73 0.00 0.00Test 72.61 5.35 0.00 0.00 85.65 5.72 0.00 0.00 75.22 7.02 0.00 0.00 75.00 6.20 0.00 0.00B Validation 19 67.95 2.46 0.00 0.00 18 68.49 2.11 0.00 0.00 18 67.62 1.17 0.00 0.00 18 72.22 2.07 0.00 0.00Test 66.96 7.32 0.00 0.00 68.70 5.40 0.00 0.00 68.04 2.73 0.05 0.12 75.00 2.43 0.00 0.00C Validation 18 67.01 1.21 0.00 0.00 18 68.38 1.80 0.00 0.00 16 75.51 1.06 0.00 0.00 18 69.81 1.42 0.00 0.00Test 69.35 6.31 0.00 0.00 68.91 4.25 0.00 0.00 74.57 6.55 0.00 0.00 72.39 1.82 0.00 0.00D Validation 17 75.89 1.68 0.00 0.00 17 76.99 1.45 0.00 0.00 19 69.37 0.44 0.00 0.00 16 68.27 2.51 0.00 0.00Test 76.09 3.77 0.00 0.00 78.70 4.52 0.00 0.00 72.83 5.10 0.00 0.00 71.09 4.18 0.00 0.00The table shows the mean and standard deviation values of the TPR and FPR measures. The results of the cross-validation and testing processes are givenin two separate rows for each subject and each mental task. The optimum AR order for each case is also included in the table.67Chapter 3. Towards Development of a Self-Paced 2-State Mental Task-Based BCI3.4 SummaryIn this study, we evaluate the performance of a previously designed mental task-based BCI usinga dataset which we have collected. The system performance is very encouraging since the FPRmeasures are zeros, and the TPRs are sufficiently high for real-life applications.As mentioned before, the main problem with current BCI systems is their high rate of falseactivations which prevents their use in real-life applications. Since a self-paced BCI is in the NCmode most of the time, even a low rate of FP can easily frustrate the subject and hence make theBCI undesirable to use. Therefore, the FPR is much more important than the TPR when operatinga self-paced BCI system. It is more tolerable for subjects if they can activate the BCI with even arate of 50% (i.e., one out of two attempts) if they are sure to some extent that the system is undertheir control and does not get activated by itself due to a false positive result.In this study, the BCI with the highest performance has a TPR above 70% for three of thefour subjects. For Subject 4, this rate is around 65%, which is also acceptable based on the abovediscussion since the FPR is zero.This study shows that it is feasible to have a BCI that is based on mental tasks to yield a zero ornear-zero false activation rate. We have processed and used all the signals from 29 EEG channelsin our design. The next step is to find the best channels so as to decrease the number of channelsfrom 29. This is a necessary step to make our BCI applicable to real-life applications.Another direction for future work is to find a minimum performing length for each mental taskto obtain a reasonable performance. This should result in a BCI system with a faster response.68Chapter 4A Self-Paced 2-State Mental Task-BasedBCI Using Few EEG ChannelsIn the previous chapter, we have shown that it is possible to have a two-state self-paced BCI witha zero (or near zero) false activation rate. Such a BCI is activated by a mental task using thesignals of 29 EEG channels. The proposed BCI system is however impractical for use in real-lifeapplications due to the use of a large number of channels (29). This is mainly because:? Setting up a system with too many channels consumes a considerable amount of time.? The impedance of the EEG channels must be kept low. If the number of channels is high,this task would be overly cumbersome.? The large number of channels potentially makes the system less portable and generally morecomplicated to set-up and use.In this study, we address this problem and decrease the number of channels to 5 or 7 using twodifferent methods. The performance of the systems with the channel sets obtained by these twomethods are compared with each other and the better channel set is selected for the final systemdesign.As before, the classifier used is based on the quadratic discriminant analysis (QDA) due to itssimplicity and accuracy. The features to be classified are the scalar autoregressive (AR) coefficients69Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG Channelsof the EEG signals. The feature extraction and the classification methods employed are efficient interms of computational complexity. The cross-validation process is performed so as to obtain theoptimal order of AR coefficients as well as the best EEG channels for every mental task of everysubject.As in the previous chapter, four different BCI systems are developed for each subject. EachBCI is activated by one of the four mental tasks, i.e., in each BCI, one mental task is selected as theIC task. The other three mental tasks are considered as NC tasks. The BCI system should thereforeremain in the NC mode during the NC tasks and the baseline.Even though the system performance is evaluated offline, the EEG signals are analyzed in aself-paced manner. A signal trial is divided into overlapping segments. Each segment is labeled aseither IC or NC, depending on whether or not it belongs to the IC task. The performance of thesystem is then evaluated in terms of its true positive rate (TPR) and its false positive rate (FPR).4.1 Methodology4.1.1 EEG channel selectionThe dataset collected is formed of the signals from 29 electrodes. However, as stated above, a BCIsystem with 29 channels is impractical for use in real-life applications. For practical applications,the number of channels should be as small as possible. We thus select a subset of channels fromthe 29 channels that would yield the best performance for the final design of our system.Suppose that we have a BCI system with n channels and we need to select the BCI that hasm channels (m<n) and yields the best performance. The ideal but not always computationallypractical way, is to consider all the possible m-channel combinations of the n channels and selectthe combination that yields the best system performance. For instance, if we want to decrease thenumber of channels from 29 to 7, the system performance needs to be evaluated for all 1,560,780different 7-channel combinations and then compared. Moreover, in order to make the results more70Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG Channelsrobust, the performance evaluation of the system is usually carried out for different training andtesting sets via a cross-validation process. If we assume that the number of evaluation runs in cross-validation is 5, the number above (1,560,780) should be multiplied by 5. This forms a prohibitivelylarge amount of computations as the processing time will take several days. It is thus impracticalfor implementation.In this study, two approaches are performed for selecting the best system that has m channels:backward elimination and forward selection. The first approach, backward elimination, deletes thechannels (from the 29-channel system) that results in the least reduction in the performance of theremaining system. The channels are deleted one by one. The second approach, forward selection,builds a new system by adding channels one by one to the newly built system. A channel added tothe new system is selected from the 29-channel system so that the performance of the new systemis maximal. These 2 approaches result in 2 methods that we denote by MDelete and MForm. Eventhough these methods are not optimal as the method mentioned above (i.e., considering all possiblecombinations), they still reach the goal to a certain extent.Channel selection method oneIn channel selection method one (MDelete), all 29 channels are first considered. The resultantFPR and TPR values after deleting each channel are obtained. The BCI system with 28 channelsthat yields the best performance is selected. That is the channel whose removal results in the best28-channel system is detected and deleted from the list. This task is repeated on the remaining list(i.e., on 28 channels) and the best 27-channel system is found. This is repeated again until all butm channels are omitted. The flowchart of this method is given in Fig. 4.1.Channel selection method twoChannel selection method two (MForm) is similar to the first one except that it is carried in thereverse direction and the selection of the channels differs as explained below. Fig. 4.2 shows the71Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG ChannelsFigure 4.1: Channel selection method 1 (MDelete)flowchart of the method.In the first iteration, the BCI system is assumed to have one channel only. The channel withthe best performance among the 29 existing channels is thus detected. This will form the best BCIthat has one channel only. In the second iteration, the BCI system is assumed to have 2 channelsonly. One of these 2 channels is the one already selected in iteration 1. Thus the performance ofeach channel in the remaining 28 channels, together with the channel selected in the first iterationis obtained. Amongst these 28 possibilities, the channel (that together with the already selectedone) yields the best performance is selected and added to the list of the best channels. In the thirditeration, three channels are considered. These are formed of each channel from the remaining list(i.e., 27 channels) and the two channels already selected in the previous iterations. The channelthat together with the two already selected channels yields the 3-channel BCI system with the bestperformance is added to the list. This procedure is repeated until m channels are added to the listof the best channels.72Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG ChannelsFigure 4.2: Channel selection method 2 (MForm)4.1.2 ProcedureThe length of each mental task epoch is 10 sec. The baseline epochs have a variable length in therange of 15?2.5 sec. Since the sampling frequency is 500 samples/sec, each mental task epochconsists of 5000 samples and the number of samples in a baseline epoch varies between 6250 and8750.To process the data, every epoch is divided into overlapping segments. Each segment is oflength 1 sec (i.e., 500 samples) and overlaps with the previous segment by 400 samples. In otherwords, the BCI system generates an output every 100 samples using the last 500 samples of thesignals. Since 100 samples is equivalent to 0.2 sec, the output rate of the BCI is 5 Hz.73Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG ChannelsFeature selection and classificationAutoregressive (AR) modeling is used to obtain the features from the segments. Based on theresults in [60] and [25], 1 sec of the EEG signal is sufficiently long for the AR model estimation.The feature vector is formed by concatenating the AR coefficients (estimated from the segments ofthe selected channels) into a single vector. This vector is then fed to the classifier for classificationpurposes. Classification is performed using the QDA classifier.Custom designingCustom designing the system for every subject yields improvements in the overall BCI perfor-mance [18, 29]. In this study, the BCI system is customized for each subject and for each mentaltask by selecting the channels and AR orders during cross-validation.Cross-validationTo perform the cross-validation, we randomly divide the whole set of segments into five equal-sized sections. Four of the data sections are used to train and validate the system. Testing is carriedon the remaining section. The four data sections assigned to training and validation are furtherdivided randomly into five data partitions of equal size. Four partitions are used for training andone is used for validation.Selecting the best five and the best seven channelsWe select the top 5 and also the top 7 best performing channels for future processing usingMDeleteand MForm with the AR model order of 40. We then compare the results of the 5-channel caseswith those of the 7-channel cases to figure out the final design for the BCI system. Channelselection is accomplished separately for different subjects and different mental tasks. Tables 4.1and 4.2 list the best channels selected using MDelete and MForm, respectively. For each subjectand each mental task, each channel selected by both MDelete and MForm is shown in bold in the74Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG ChannelsTable 4.1: Channels Selected for Different Subjects and Mental Tasks Using Channel Selec-tion Method One (MDelete)Selected Channels (from best to worst)1 2 3 4 5 6 7Subject 1Sentence Visualization T8 F8 F3 AF3 Oz CP6 P7Non-Trivial Multiplication FC5 FC6 PO3 T7 C4 CP6 C33D Object Rotation FC6 T8 F3 AF3 CP6 AF4 F8Motor Imagery CP6 T8 P8 AF3 F7 FC2 T7Subject 2Sentence Visualization P8 FC6 T7 F3 PO4 P7 P3Non-Trivial Multiplication Oz Fpz Fz T7 FC6 P3 C33D Object Rotation Oz T7 P3 FC5 Fz C4 CzMotor Imagery CP2 Oz F7 Fpz P7 CP6 F4Subject 3Sentence Visualization PO3 P8 P4 AF4 T7 P7 C4Non-Trivial Multiplication Oz Fpz Pz T7 CP1 PO4 CP53D Object Rotation FC2 CP1 PO4 P8 Cz F4 FC6Motor Imagery T7 P7 CP2 Fpz CP6 F7 FC6Subject 4Sentence Visualization P8 Cz PO3 CP5 T7 PO4 AF3Non-Trivial Multiplication F4 CP5 T8 P7 P8 Oz F33D Object Rotation AF3 FC5 AF4 P7 C3 T8 PO4Motor Imagery Cz P8 P7 FC6 T7 F3 CP5tables.MDelete and MForm give BCI systems with different channel combinations. This is becauseeach of these methods obtain a channel set which is locally optimum. The globally optimumset can be obtained by the exhaustive method mentioned earlier in which the best combinationis selected by considering all possible combinations of channels. Finding the global optimumis computationally impossible for our application. We hence have to be satisfied with the localoptima.Finding the optimum AR model orderAfter selecting the best 5 and the best 7 channels for each subject and each mental task, we findthe optimum AR model order for each of these cases in the cross-validation process. The initial75Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG ChannelsTable 4.2: Channels Selected for Different Subjects and Mental Tasks Using Channel Selec-tion Method Two (MForm)Selected Channels (from best to worst)1 2 3 4 5 6 7Subject 1Sentence Visualization T8 FC5 F8 P4 T7 P7 OzNon-Trivial Multiplication F7 Oz FC6 T8 F8 P7 CP13D Object Rotation T7 FC6 P8 CP2 P7 Oz F4Motor Imagery Oz P7 C3 CP6 T8 P8 FC6Subject 2Sentence Visualization P8 F7 FC6 T7 FC5 AF3 T8Non-Trivial Multiplication CP6 P4 T8 Fpz C4 F7 FC63D Object Rotation P7 P4 C4 Oz FC6 AF4 F7Motor Imagery CP5 Fz FC6 T7 FC1 Oz C4Subject 3Sentence Visualization PO3 Oz Fpz P8 P7 Pz CP1Non-Trivial Multiplication Oz P8 Fpz P7 Cz T8 AF33D Object Rotation CP5 FC2 P7 P4 T7 Oz FzMotor Imagery T8 Oz P3 PO3 P7 AF4 AF3Subject 4Sentence Visualization CP6 P8 FC5 AF4 T7 CP5 OzNon-Trivial Multiplication AF4 F7 T8 P8 P7 T7 Oz3D Object Rotation F3 F7 AF4 PO3 FC5 Oz C3Motor Imagery CP2 P3 AF4 F8 P8 F3 F7AR model order is equal to 41. If the FPR for an order reaches zero, that order is selected asthe optimum order; if not, the order is increased by one, and the FPR of the new order is thencalculated. Increasing the AR order is terminated once the FPR is zero or a maximum order of 136is reached. The order corresponding to the minimum FPR is chosen as the optimum AR order forthe latter case.4.2 ResultsFor each subject and each mental task, there are two sets of 5 best channels (one is obtained usingMDelete and the other is obtained using MForm). The performance of the 5-channel set thatyielded the better performance is summarized in Table 4.3. For each subject and each mental task,76Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG Channelsthe table shows whether the channels obtained using MDelete (1) or MForm (2) are selected. Thisis indicated under the CSM (channel selection method) column. It also shows the mean values ofthe TPR and FPR obtained from the cross-validation and testing processes in two separate rows.The optimum AR model order is also included in the table. In Table 4.4, the difference in theperformance between the 5-channel BCIs using MDelete and MForm is given.The performance of the 7-channel BCIs using the two channel selection methods are comparedwith each other and the performance of the better method is given in Table 4.5 for each subjectand each mental task. The performance difference between the two channel selection methods isshown in Table 4.6.In Tables 4.3 and 4.5, the values related to the highest performance are shown in bold, whilethose related to the lowest performance are underlined.The Welch?s t-test [176], as a statistical significance test, is performed on the TPR values formeasuring the performance difference between MDelete and MForm (for each subject and eachmental task) and between every pair of the four mental tasks (for each subject).The null hypothesis is that there is no difference between the TPR values of 2 groups (these 2groups can be the 2 channel selection methods or any 2 mental tasks). We assume a 5% significancelevel. The null hypothesis is rejected if the resultant p-value is less than 0.05. This means that theTPR values of the 2 test groups are significantly different. If p? 0.05, the null hypothesis cannotbe rejected at the 5% significance level. This implies that there is no significant difference betweenthe TPR values of the test groups.The p-value between MDelete and MForm for each mental task of each subject is given inTables 4.4 and 4.6 under the p-value column. Tables 4.7 and 4.8 show the results of the t-testbetween any 2 of the four mental tasks for each subject in the 5-channel and the 7-channel BCIs,respectively.77Chapter4.ASelf-Paced2-StateMentalTask-BasedBCIUsingFewEEGChannelsTable 4.3: Cross-Validation and Testing Results of the Better Channel Selection Method for 5-Channel SystemsVisualization Multiplication Object Rotation Motor ImagerySubject Process CSM AR TPR FPR CSM AR TPR FPR CSM AR TPR FPR CSM AR TPR FPR1 Validation 1 87 59.82 0.00 2 98 61.52 0.00 2 89 59.82 0.00 2 93 56.56 0.00Testing 57.78 0.00 61.24 0.00 59.71 0.00 55.30 0.002 Validation 1 91 57.37 0.00 1 90 58.29 0.00 2 84 57.23 0.00 1 90 54.68 0.00Testing 58.56 0.00 57.93 0.00 56.00 0.00 54.38 0.003 Validation 2 81 61.98 0.00 2 82 63.25 0.00 2 95 57.23 0.00 1 90 57.25 0.00Testing 62.68 0.00 63.72 0.00 55.37 0.00 58.51 0.004 Validation 1 101 55.21 0.00 1 129 42.89 0.00 1 136 33.03 0.20 1 134 41.62 0.01Testing 53.89 0.00 42.61 0.01 33.50 0.18 42.37 0.00The table shows the mean values of the TPR and FPR measures. The results of the cross-validation and testing processes are given intwo separate rows for each subject and each mental task. The selected CSM and AR order for each case are also included in the table.Table 4.4: The Difference in the Performance between MDelete and MForm for the 5-Channel SystemsVisualization Multiplication Object Rotation Motor ImagerySubject Process dAR? dTPR? dFPR? p-value? dAR dTPR dFPR p-value dAR dTPR dFPR p-value dAR dTPR dFPR p-value1 Validation ?8 1.31 0.00 0.2063 7 ?5.61 0.00 0.0004 5 ?2.04 0.00 0.0593 16 ?5.11 0.00 0.0029Testing 0.29 0.00 0.7615 ?5.24 0.00 0.0005 ?1.78 0.00 0.2160 ?4.57 0.00 0.02012 Validation ?2 1.59 0.00 0.0121 4 2.40 0.00 0.1009 5 ?0.19 0.00 0.8667 ?5 4.82 0.00 0.0000Testing 1.63 0.00 0.3012 2.88 0.00 0.0654 ?0.43 0.00 0.7867 4.16 0.00 0.00953 Validation 6 ?1.37 0.00 0.3534 9 ?4.96 0.00 0.0005 6 ?3.07 0.00 0.1215 ?3 0.91 0.00 0.1889Testing ?2.19 0.00 0.0620 ?6.10 0.00 0.0001 ?3.19 0.00 0.1863 1.51 0.00 0.10764 Validation ?26 6.36 0.00 0.0004 ?7 2.43 0.00 0.1970 0 0.67 0.01 0.5523 ?2 2.90 ?0.01 0.0878Testing 6.41 0.00 0.0021 1.98 0.00 0.0146 1.83 ?0.01 0.0756 2.27 ?0.01 0.0777? dV= VMDelete?VMForm, V?{AR,TPR,FPR}.? p< 0.05 means ?there is a significant difference". These cases are shown in bold.78Chapter4.ASelf-Paced2-StateMentalTask-BasedBCIUsingFewEEGChannelsTable 4.5: Cross-Validation and Testing Results of the Better Channel Selection Method for 7-Channel SystemsVisualization Multiplication Object Rotation Motor ImagerySubject Process CSM AR TPR FPR CSM AR TPR FPR CSM AR TPR FPR CSM AR TPR FPR1 Validation 2 63 63.23 0.00 2 73 65.91 0.00 1 64 65.45 0.00 1 69 63.35 0.00Testing 63.70 0.00 64.51 0.00 66.11 0.00 62.51 0.002 Validation 2 62 64.59 0.00 2 61 63.07 0.00 2 60 61.55 0.00 1 64 59.39 0.00Testing 64.77 0.00 62.73 0.00 61.97 0.00 60.68 0.003 Validation 1 64 63.32 0.00 2 56 69.94 0.00 1 65 66.33 0.00 1 65 64.57 0.00Testing 64.90 0.00 70.51 0.00 64.73 0.00 65.32 0.004 Validation 2 79 61.56 0.00 2 95 51.90 0.00 2 129 27.58 0.01 1 97 49.30 0.00Testing 59.71 0.00 50.21 0.00 27.36 0.01 50.00 0.00The table shows the mean values of the TPR and FPR measures. The results of the cross-validation and testing processes are given intwo separate rows for each subject and each mental task. The selected CSM and AR order for each case are also included in the table.Table 4.6: The Difference in the Performance between MDelete and MForm for the 7-Channel SystemsVisualization Multiplication Object Rotation Motor ImagerySubject Process dAR? dTPR? dFPR? p-value? dAR dTPR dFPR p-value dAR dTPR dFPR p-value dAR dTPR dFPR p-value1 Validation 2 ?1.77 0.00 0.1354 7 ?6.70 0.00 0.0017 3 0.58 0.00 0.5591 1 1.71 0.00 0.3435Testing ?2.56 0.00 0.0084 ?5.12 0.00 0.0047 0.80 0.00 0.5357 2.36 0.00 0.18762 Validation 2 ?1.58 0.00 0.0954 4 ?2.25 0.00 0.1739 6 ?0.70 0.00 0.6486 4 1.25 0.00 0.4317Testing ?2.24 0.00 0.0960 ?0.44 0.00 0.7226 ?2.48 0.00 0.2445 2.58 0.00 0.00453 Validation ?2 1.00 0.00 0.4971 6 ?4.86 0.00 0.0007 4 0.73 0.00 0.7233 4 ?0.06 0.00 0.9542Testing 2.22 0.00 0.0455 ?5.42 0.00 0.0009 0.91 0.00 0.6633 1.52 0.00 0.11104 Validation ?3 ?1.59 0.00 0.2486 ?5 1.62 0.00 0.1907 ?3 4.26 0.01 0.0173 5 0.00 0.00 1.0000Testing ?0.37 0.00 0.7741 2.26 0.01 0.0037 4.02 0.00 0.0012 0.54 0.00 0.7501? dV= VMDelete?VMForm, V?{AR,TPR,FPR}.? p< 0.05 means ?there is a significant difference". These cases are shown in bold.79Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG ChannelsTable 4.7: p-Values Calculated UsingWelch?s T-Test between the FourMental Tasks for EachSubject (5-Channel Systems)Mental TaskSubject Mental Task Multiplication Object Rotation Motor Imagery1 Visualization 0.0146 0.1759 0.1427Multiplication 0.2472 0.0053Object Rotation 0.02512 Visualization 0.7203 0.1695 0.0330Multiplication 0.2499 0.0408Object Rotation 0.29493 Visualization 0.3236 0.0049 0.0040Multiplication 0.0034 0.0001Object Rotation 0.10624 Visualization 0.0000 0.0000 0.0000Multiplication 0.0000 0.8060Object Rotation 0.0000p < 0.05 means ?there is a significant difference". These cases are shown in bold.Table 4.8: p-Values Calculated UsingWelch?s T-Test between the FourMental Tasks for EachSubject (7-Channel Systems)Mental TaskSubject Mental Task Multiplication Object Rotation Motor Imagery1 Visualization 0.4889 0.0519 0.3890Multiplication 0.2553 0.2243Object Rotation 0.04052 Visualization 0.0809 0.1582 0.0035Multiplication 0.6683 0.0328Object Rotation 0.46013 Visualization 0.0007 0.9199 0.6681Multiplication 0.0135 0.0013Object Rotation 0.73214 Visualization 0.0001 0.0000 0.0002Multiplication 0.0000 0.8619Object Rotation 0.0000p < 0.05 means ?there is a significant difference". These cases are shown in bold.80Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG Channels4.3 Summary and Discussion of ResultsTables 4.4 shows that the system performance of the 5-channel BCIs obtained using MDelete andMForm are close to each other for the majority of the cases. The same is true for the 7-channelsystems (Table 4.6). From Tables 4.3 and 4.5, it is shown that MDelete yields better results in nineout of the sixteen cases of the 5-channel BCIs and in seven out of the sixteen cases of the 7-channelBCIs.4.3.1 System performance of 5-channel BCIsFrom Tables 4.3 and 4.7, we find the following:1. The FPR values are zero for all 5-channel BCIs of Subjects 1, 2, and 3 during the cross-validation and testing processes irrespective of the task type. For Subject 4, this is also truefor the sentence visualization task-based BCI; the BCI based on the multiplication task has0.01% FPR for the testing process; the BCI based on the motor imagery task has an FPR of0.01% for the cross-validation process; and the BCI based on the object rotation task hasFPR values of 0.20% and 0.18% for cross-validation and testing, respectively.2. Amongst all 5-channel BCIs, the highest performance (TPR= 63.72%) is reached by themultiplication task of Subject 3. The object rotation task-based BCI of Subject 4 has thelowest performance with TPR and FPR values of 33.50% and 0.18%, respectively.3. For Subject 1: Multiplication is the best in performance (TPR= 61.24%) although it is notsignificantly different from object rotation. Motor imagery is the poorest in performance(TPR= 55.30%) although it is not significantly different from sentence visualization.4. For Subject 2: Sentence visualization, multiplication, and object rotation have statisticallysimilar performances with TPRs in the range of 56.00?58.56%. Motor imagery has thepoorest performance (TPR= 54.38%) although it is not significantly different from objectrotation.81Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG Channels5. For Subject 3: Multiplication and sentence visualization have similar and the highest per-formances (TPRs= 63.72% and 62.68%). Object rotation and motor imagery have similarand the lowest performances (TPRs= 55.37% and 58.51%).6. For Subject 4: Sentence visualization has the best performance (TPR= 53.89%) and objectrotation has the poorest performance (TPR= 33.50% and FPR= 0.18%).4.3.2 System performance of 7-channel BCIsFrom Tables 4.5 and 4.8, the following are found:1. The FPR values reach zero for 15 out of the 16 cases. That is for all cases except for theobject rotation task of Subject 4 which has 0.01% FPR for each of the cross-validation andtesting processes.2. The results as to which BCIs yield the best and worst performance are exactly as thosein the 5-channel systems. Amongst all the 7-channel BCIs, the best performance with aTPR of 70.51% is reached by the BCI based on the multiplication task of Subject 3 (whichis better than the TPR= 63.72% of the corresponding 5-channel BCI). The object rotationtask-based BCI of Subject 4 has the worst performance with TPR and FPR values of 27.36%and 0.01%, respectively (which is also better than the FPR= 0.18% of the corresponding5-channel BCI).3. For Subject 1: Motor imagery is significantly different from object rotation. All othercases are statistically similar to each other. The TPR varies in the range 62.51?66.11% fordifferent mental tasks.4. For Subject 2: Sentence visualization, multiplication, and object rotation have statisticallysimilar performances with TPRs in the range of 61.97?64.77%. Motor imagery has thepoorest performance (TPR= 60.68%) although it is not significantly different from objectrotation.82Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG Channels5. For Subject 3: Multiplication has the highest performance (TPR= 70.51%). Sentence vi-sualization, object rotation and motor imagery have similar performances with TPRs in therange of 64.73?65.32%.6. For Subject 4: Sentence visualization has the best performance (TPR= 59.71%) and objectrotation has the poorest performance (TPR= 27.36% and FPR= 0.01%).4.3.3 Discussion of resultsComparing Tables 4.3 and 4.5, it can be noticed that increasing the number of channels from 5 to 7enhances the performance of every BCI by an increase of 5.38% in TPR on average (see Table 4.9).Therefore, there is a trade-off between using fewer channels and having better system performance.The choice should be made depending on the applications, the situations in which the BCI systemis used, and the computational power available. Table 4.10 shows the system performance of thebest BCI for each subject.In terms of the AR model orders, the 5-channel BCIs need higher orders than the 7-channelBCIs. According to Table 4.9, the overall mean of the AR model order is 98 and 73 for the 5-channel and 7-channel BCI systems, respectively.Studies [11, 13, 100, 130, 158] have also used high AR model orders in their analyses. In [11],the sampling frequency is 1200 Hz. The top AR model orders are up to 70. In [13], the samplingrate is 173.61 Hz. The AR model orders selected are between 12 and 49. In [100], the samplingfrequency is 160 Hz and the performance exhibits an increase as the AR model order increases upto 26. In [130], the AR order 100 at a sampling rate of 10 Hz is found to be the most suitable order.In [158], the AR model order is found to be in the range of 4?133 using a sampling rate of 128 Hz.The AR model order depends on the sampling frequency [100]. Since our sampling frequencyis 500 Hz (which is high), then the AR order can also be high. In other words, if the AR orderbelonging to sampling frequency 100 Hz is 25, the AR order belonging to sampling frequency 50083Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG ChannelsTable 4.9: Average Performance of the BCI Systems5-Channel Design 7-Channel Design 29-Channel DesignAR TPR FPR AR TPR FPR AR TPR FPRAverage over TasksSubject 1 92* 58.51 0.00 67* 64.21 0.00 24 68.54 0.00Subject 2 89* 56.72 0.00 62* 62.54 0.00 19* 70.38 0.00Subject 3 87 60.07 0.00 63* 66.37 0.00 20* 70.80 0.00Subject 4 125 43.09 0.05 100 46.82 0.00 29 59.31 0.00Average over SubjectsVisualization 90 58.23 0.00 67 63.27 0.00 21* 68.95 0.00Multiplication 100* 56.38 0.00 71* 61.99 0.00 24 67.38 0.00Object Rotation 101 51.15 0.05 80* 55.04 0.00 24* 65.71 0.00Motor Imagery 102* 52.64 0.00 74* 59.63 0.00 23* 67.00 0.00Total Average 98* 54.60 0.01 73* 59.98 0.00 23 67.26 0.00* The value is rounded to the nearest integer.Table 4.10: System Performance of the Best BCI for Each Subject5-Channel Design 7-Channel Design 29-Channel DesignAR TPR FPR AR TPR FPR AR TPR FPRSubject 1 98 61.24 0.00 64 66.11 0.00 21 70.86 0.00Subject 2 91 58.56 0.00 62 64.77 0.00 17 72.65 0.00Subject 3 82 63.72 0.00 56 70.51 0.00 19 71.85 0.00Subject 4 101 53.89 0.00 79 59.71 0.00 26 65.06 0.00Average 93 59.35 0.00 65* 65.28 0.00 21* 70.11 0.00* The value is rounded to the nearest integer.Hz is 125. This is because:1. The AR order is the number of previous samples of the signal which make the currentsample of the signal. Please refer to Equation 2.1.2. AR order= 25 at freq= 100 Hz means that we need the last 0.25 sec of the signal to makethe current sample. Now, if freq= 500 Hz, since we need the last 0.25 sec of the signal, theAR order should be 0.25?500= 125.From Table 4.9, it can be easily recognized that the system performance depends on the sub-ject and the mental task. Among all subjects, Subject 3 has the BCIs with the highest averageperformance with average TPRs of 60.07% and 66.37% for the 5-channel and 7-channel systems,84Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG Channelsrespectively. Subject 4 has the poorest average performance with an average TPR of 43.09% forthe 5-channel systems and 46.82% for the 7-channel systems.Among the mental tasks, sentence visualization and multiplication are the best tasks overall.Object rotation and motor imagery yield the least performances on average.Table 4.9 also shows the average performances of the 29-channel systems over the subjectsand mental tasks for comparison purposes. It can be seen that by decreasing the number of EEGchannels from 29 to 7 and 5, the system performance degrades by a decrease of 12.66% and7.28%, respectively, in TPR on average . This degradation in the system performance is the trade-off to have a simpler system that can be easily setup and requires less computational powers.Decreasing the number of channels more would result in an average TPR less than 54.60% whichis not acceptable in real-life applications.According to Table 4.9, the AR model order increases as the number of channels decreasesfrom 29 to 7 and 5. In the 29-channel systems, since the number of channels are high, AR modelswith low orders are sufficient for classification purposes. However, as the number of channelsdecreases to 7 and 5, AR models with higher orders are required. This is because each channel hasa certain amount of information. By increasing the AR order, the more percentage of the availableinformation in each channel is used in modeling. By decreasing the number of channels, the morepercentage of the available information in the channels is needed for classification. Therefore, theAR model order increases.4.4 SummaryIn this study, we have shown that it is feasible to design self-paced mental task-based BCIs with azero false activation rate using very few (i.e., five or seven) EEG channels. The system performancewas evaluated on a dataset which we collected from four subjects.Although the evaluation was carried off-line, the methodology can be used in real-time self-85Chapter 4. A Self-Paced 2-State Mental Task-Based BCI Using Few EEG Channelspaced systems after slight modifications. This is because the feature extraction and the classifica-tion processes are not computationally demanding and there is no need for the timing informationof the EEG signals after training the classifier with the training data.In this study, the best performance BCI systems have zero FPRs and sufficiently high TPRs(i.e., 53.89?63.72% for the 5-channel systems, and 59.71?70.51% for the 7-channel systems).Hence, in terms of the system performance, they are acceptable for use in the real-life applications.As reported in Chapter 3, the best performance BCI systems with 29 channels have zero FPRs and65.06?72.65% TPRs.In this study, the EEG trials are divided into 1-sec segments for classification, and the BCIsystem gives an output every 0.2 sec. Finding the optimum values for the segment length and thesystem output rate are other directions for future work. The use of these values might result in amore accurate system.86Chapter 5A Study on Finding the Optimum Length ofthe EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIsIn Chapter 4, we demonstrated that it is possible to have a self-paced 2-state BCI that is activatedby mental tasks and yields a zero (or near zero) false activation rate using the signals of 7 or 5 EEGchannels. The EEG signal segments used for classification in the proposed BCI have a length of 1sec so as to have a fast response system.In this chapter, we study the impact of the length of the EEG signal segments on the perfor-mance of our self-paced BCIs. The optimum segment length is found for each subject and eachmental task. The effect of artifact rejection on the overall performance of the system is also studiedin this chapter.5.1 MethodologyAs in the previous chapters, four different BCIs are designed for each subject. In each BCI, onemental task is considered as the IC task and the other 3 mental tasks are considered as NC tasks.During the NC tasks and the baseline, the system should remain in the NC mode. In other words,the BCI should only be activated by one mental task selected for each subject.87Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIsIn the dataset we have collected, each epoch of a mental task is 10 sec long. That is the subjectwas asked to continue performing the mental task for 10 sec. The length of each baseline epochis 15?2.5 sec. Every epoch is divided into segments. Each segment overlaps with the previoussegment by 80%. In order to study the effect of the length of the signal segments on the systemperformance, the segment length is varied between 1 and 8 sec with steps of 0.5 sec. The segmentlength shorter than 1 sec is not tested since the subjects may not be able to switch between mentaltasks faster than 1 sec.Feature selection is based on the scalar AR modeling of the segments. It has been shownthat 1 sec of the EEG signal is sufficiently long for estimating the coefficients of its AR model[25, 60]. The feature vector is formed by concatenating the AR coefficients of the segments of thedifferent channels into a single vector. The classification stage is based on quadratic discriminantanalysis (QDA). Explanations of the AR modeling and QDA were given in Sections 2.3 and 2.4.2of Chapter 2.Since custom designing the system for each person improves the overall performance [18, 29],each BCI system is customized by selecting its own AR model order and segment length via cross-validation.To perform cross-validation, the whole set of segments (of a certain length) is randomly dividedinto five equal-sized sections. Four of these sections are used for training and validating the system.The remaining section is used for testing. The data sections assigned to training and validation arefurther divided randomly into five equal-sized data parts. Four of these parts are used for trainingand one part is used for validation.5.1.1 EEG channel selectionIn this study, we use the 7-channel sets that have been selected for each subject and each mentaltask based on the methods discussed in Chapter 4.88Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIsIn Chapter 4, we decreased the number of channels from 29 to 7 using two different methods(i.e., MDelete and MForm). In MDelete, at each iteration, the channel with the least effect on theperformance (i.e., the FPR and TPR values) was found and then deleted. This task was repeateduntil all but 7 channels were omitted. In MForm, at each iteration, the channel whose additionto the channels (that we selected in the previous iterations) resulted in the system with the bestperformance was selected. This procedure was repeated until 7 channels were selected.In Chapter 4, for each BCI of each subject, the two methods for channel selection were per-formed on 1-sec EEG signal segments with the AR order of 40. Two 7-channel sets were obtained.For each set, the optimum AR model order was found as follows: the initial AR model order wasset to 41. If the FPR obtained for an AR model order was zero, that order was considered as theoptimum order; if not, the order was increased by one. Increasing the AR order continued until theFPR obtained was zero or a maximum order of 136 was reached. The AR order corresponding tothe minimum FPR was considered as the optimum AR order for the latter case (i.e., a maximumorder of 136 was reached).The performance of the system with the two 7-channel sets selected by MDelete and MForm(when the optimum AR orders were considered) were compared with each other and the set thatyielded the better results was chosen for the final system design. For each BCI of each subject,Table 5.1 shows the channel selection method with the better results under the CSM column. The7 selected channels are also shown in the table.5.1.2 Finding the optimum length of the EEG signal segmentsUsing the 7-channel set listed in Table 5.1 for each BCI of each subject, we vary the length of thesegments of the epochs of the EEG signals. For each segment length of 1.5, 2, 2.5, 3, 3.5, 4, 4.5,5, 5.5, 6, 6.5, 7, 7.5, and 8 sec, the new optimum AR order is selected via cross-validation. InSection 5.2.1, the results are discussed and the optimum length of the segments for each BCI of89Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIsTable 5.1: Channels Selected in Each BCI of Each SubjectSelected Channels (from best to worst)CSM* 1 2 3 4 5 6 7Subject 1Sentence Visualization 2 T8 FC5 F8 P4 T7 P7 OzNon-Trivial Multiplication 2 F7 Oz FC6 T8 F8 P7 CP13D Object Rotation 1 FC6 T8 F3 AF3 CP6 AF4 F8Motor Imagery 1 CP6 T8 P8 AF3 F7 FC2 T7Subject 2Sentence Visualization 2 P8 F7 FC6 T7 FC5 AF3 T8Non-Trivial Multiplication 2 CP6 P4 T8 Fpz C4 F7 FC63D Object Rotation 2 P7 P4 C4 Oz FC6 AF4 F7Motor Imagery 1 CP2 Oz F7 Fpz P7 CP6 F4Subject 3Sentence Visualization 1 PO3 P8 P4 AF4 T7 P7 C4Non-Trivial Multiplication 2 Oz P8 Fpz P7 Cz T8 AF33D Object Rotation 1 FC2 CP1 PO4 P8 Cz F4 FC6Motor Imagery 1 T7 P7 CP2 Fpz CP6 F7 FC6Subject 4Sentence Visualization 2 CP6 P8 FC5 AF4 T7 CP5 OzNon-Trivial Multiplication 2 AF4 F7 T8 P8 P7 T7 Oz3D Object Rotation 2 F3 F7 AF4 PO3 FC5 Oz C3Motor Imagery 1 Cz P8 P7 FC6 T7 F3 CP5* channel selection method (1= MDelete, 2= MForm)each subject is found.5.1.3 Artifact rejectionWe study the effects of the EOG and EMG artifacts on the system performance. All segments con-taminated by artifacts are deleted and the system is then re-trained and re-tested with the artifact-free segments of the dataset. The performance of the BCIs with and without artifact rejection arethen compared with each other.A segment is considered to be contaminated by artifacts if one of the following conditionshappens:1. The difference between the max and min values of any selected EEG channels exceeds 100?V . This case can be caused by any external noise sources.90Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIs2. The difference between the max and min values of each bipolar channels of RU-RD, RR-C, and C-LL exceeds 100 ?V . This could indicate artifact contamination due to EOGactivities.3. The difference between the max and min values of the channel Fpz or the bipolar channelFpz-C exceeds 100 ?V . These channels are checked to reject any probable EMG artifactsmade by the forehead muscles.4. The difference between the max and min values of the bipolar EMG channel placed onthe right forearm exceeds 100 ?V . This would reject the unexpected movement-relatedpotentials (MRPs) of the right forearm muscles.Using the cleaned data, the optimum AR order is re-selected via cross-validation for the opti-mum length of the segments for each BCI of each subject. The results are given in Section 5.2.2.5.2 Results and DiscussionWe first report the results we obtained on finding the optimum segment length in Section 5.2.1. Wethen discuss the effects of artifact rejection on the system performance in Section 5.2.2.5.2.1 Optimum length of the segmentsFor each subject and each mental task, the system performance of the 7-channel BCI is evaluatedfor different segment lengths via cross-validation. The FPR reaches zero for all cases. The TPRvalues and the optimum AR orders are given in Fig. 5.1?5.4. The following can be seen from thefigures:1. Generally, as the segment length increases, the TPR first increases and then decreases.2. The optimum length is in the range of 1 to 4.5 sec.3. The TPR value usually goes below 50% for lengths greater than 5 sec.4. The AR orders generally decrease as the segment length increases.91Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIsTo decide on the optimum length of the segments for each BCI, we proceed as follows:1. We first consider 1-sec segments.2. If the 1-sec segments yield the highest performance amongst all segments with differentlengths, the optimum length is selected to be 1 sec and the task is done. Otherwise, twocases can happen:(a) If there is no significant difference between the performance with 1-sec segments anda higher performance with a longer segment, the optimum length is again selected tobe 1 sec and the task is done.(b) If the higher performance with a longer segment is significantly different from theperformance with 1-sec segments, the length of 1 sec is not the optimum length. Wecontinue to the next step.3. We consider 1.5-sec segments and redo the previous step with 1.5-sec segments instead of1-sec segments. If the length of 1.5 sec is not selected as the optimum length, we consider2-sec segments. We keep increasing the segment length until the optimum length is chosen.The Welch?s t-test [176], as a statistical significance test, is performed on the TPR values formeasuring the differences in the performance. A significance level of ? = 0.05 is used for theanalyses.Table 5.2 shows the selected lengths of the segments for the BCIs of each subject. The systemperformance (in terms of TPR and FPR during validation and test) and the optimum AR order arealso reported in the table for each case.1. For Subject 1, the optimum length is selected to be 1 sec for the BCIs based on the visu-alization, object rotation, and motor imagery tasks. The multiplication-based BCI has theoptimum length of 1.5 sec.2. For Subject 2, all BCIs have the optimum length of 1 sec.3. For Subject 3, the segment length is found to be 1.5 sec for the BCIs based on the visual-92Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIsization and motor imagery tasks. For the other two BCIs, the optimum length is found tobe 1 sec.4. For Subject 4, the BCIs based on the visualization and multiplication tasks have the opti-mum length of 1.5 sec. The lengths for the object rotation-based and motor imagery-basedBCIs are 2 and 2.5 sec, respectively.The results are promising in terms of the FPR and TPR values. FPR is zero for all BCIs. TPRis in the range of 60.68%?70.51% during the test for all except one BCI (i.e., the object rotation-based BCI of Subject 4 with 41.56% TPR).Table 5.3 shows the average performance of 7-channel BCI systems when the optimum lengthis selected for the segments and when 1-sec segments are used (Chapter 4). The average perfor-mance of 29-channel systems with the 1-sec segments (Chapter 3) is also given in the table forcomparison purposes. It can be seen from the table that for 7-channel BCIs, using the segmentswith the optimum length improves the overall TPR by 3.49% (from 59.98% to 63.47%). For the29-channel BCIs with the 1-sec segments, the average TPR is 67.26%. Table 5.4 shows the systemperformance of the best BCI for each subject.93Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIs(a) Visualization (b) Multiplication(c) Object rotation (d) Motor imageryFigure 5.1: TPRs and AR orders of different segment lengths for BCIs of Subject 1.94Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIs(a) Visualization (b) Multiplication(c) Object rotation (d) Motor imageryFigure 5.2: TPRs and AR orders of different segment lengths for BCIs of Subject 2.95Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIs(a) Visualization (b) Multiplication(c) Object rotation (d) Motor imageryFigure 5.3: TPRs and AR orders of different segment lengths for BCIs of Subject 3.96Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIs(a) Visualization (b) Multiplication(c) Object rotation (d) Motor imageryFigure 5.4: TPRs and AR orders of different segment lengths for BCIs of Subject 4.97Chapter5.AStudyonFindingtheOptimumLengthoftheEEGSignalSegmentsWhenUsedinSelf-Paced2-StateMentalTask-BasedBCIsTable 5.2: System Performance during Validation and Test without Artifact RejectionVisualization-based BCI Multiplication-based BCI Object rotation-based BCI Motor imagery-based BCISubject Process Length AR TPR FPR Length AR TPR FPR Length AR TPR FPR Length AR TPR FPR1 Validation 1 63 63.23 0.00 1.5 53 68.39 0.00 1 64 65.45 0.00 1 69 63.35 0.00Test 63.70 0.00 69.04 0.00 66.11 0.00 62.51 0.002 Validation 1 62 64.59 0.00 1 61 63.07 0.00 1 60 61.55 0.00 1 64 59.39 0.00Test 64.77 0.00 62.73 0.00 61.97 0.00 60.68 0.003 Validation 1.5 50 69.30 0.00 1 56 69.94 0.00 1 65 66.33 0.00 1.5 48 70.74 0.00Test 67.26 0.00 70.51 0.00 64.73 0.00 68.29 0.004 Validation 1.5 53 67.00 0.00 1.5 59 61.92 0.00 2 68 40.67 0.00 2.5 39 63.39 0.00Test 67.91 0.00 62.57 0.00 41.56 0.00 61.21 0.00The table shows the mean values of the TPR and FPR measures. The results of the validation and test processes are given in two separate rowsfor each BCI of each subject. The optimum AR order and the optimum segment length for each system are also included in the table.98Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIsTable 5.3: Average Performance of the BCI Systems7-Channel Design 7-Channel Design 29-Channel Design(Optimum Length) (1-sec Length) (1-sec Length)AR TPR FPR AR TPR FPR AR TPR FPRAverage over TasksSubject 1 62* 65.34 0.00 67* 64.21 0.00 24 68.54 0.00Subject 2 62* 62.54 0.00 62* 62.54 0.00 19* 70.38 0.00Subject 3 55* 67.70 0.00 63* 66.37 0.00 20* 70.80 0.00Subject 4 55* 58.31 0.00 100 46.82 0.00 29 59.31 0.00Average over SubjectsVisualization 57 65.91 0.00 67 63.27 0.00 21* 68.95 0.00Multiplication 57* 66.21 0.00 71* 61.99 0.00 24 67.38 0.00Object Rotation 64* 58.59 0.00 80* 55.04 0.00 24* 65.71 0.00Motor Imagery 55 63.17 0.00 74* 59.63 0.00 23* 67.00 0.00Total Average 58* 63.47 0.00 73* 59.98 0.00 23 67.26 0.00* The value is rounded to the nearest integer.Table 5.4: System Performance of the Best BCI for Each Subject7-Channel Design 7-Channel Design 29-Channel Design(Optimum Length) (1-sec Length) (1-sec Length)AR TPR FPR AR TPR FPR AR TPR FPRSubject 1 53 69.04 0.00 64 66.11 0.00 21 70.86 0.00Subject 2 62 64.77 0.00 62 64.77 0.00 17 72.65 0.00Subject 3 56 70.51 0.00 56 70.51 0.00 19 71.85 0.00Subject 4 53 67.91 0.00 79 59.71 0.00 26 65.06 0.00Average 56 68.06 0.00 65* 65.28 0.00 21* 70.11 0.00* The value is rounded to the nearest integer.5.2.2 System performance with artifact rejectionThe performance of BCI systems on the cleaned data is given in Table 5.5. The portion of theEEG signal segments that are rejected due to artifact contamination is also reported in the table.Comparing this table with Table 5.2, we see that:1. The system performances with and without artifact rejection are generally consistent witheach other.2. For some BCIs (i.e., object rotation-based BCI of Subject 1; visualization-based BCI ofSubject 2; visualization-based, multiplication-based, and motor imagery-based BCIs of99Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIsSubject 3; visualization-based and motor imagery-based BCIs of Subject 4), the perfor-mance with artifact rejection is even poorer than the performance using the original datathat contained artifacts. This performance degradation in TPR is because of using a fixedthreshold to reject the segments. Some artifact-free segments might have been wronglyrejected from the training set and this have caused a decrease in TPR.3. The performances of the multiplication-based and object rotation-based BCIs of Subject 4are improved with artifact rejection.4. The optimum AR order is lower when the artifacts are rejected.The reason why we evaluated the system performance with cleaned data was to make surethat our BCIs are not activated by artifacts instead of mental tasks. By comparing the systemperformance with and without rejecting the segments contaminated by artifacts, we deduce that:1. The results without artifact rejection are valid and the source of system activation is thedesignated mental task.2. There is no need anymore to collect the EOG and EMG signals and to reject artifacts,since the performance without artifact rejection is generally similar to the performancewith artifact rejection.3. The proposed 7-channel BCIs based on mental tasks are robust to the artifacts.100Chapter5.AStudyonFindingtheOptimumLengthoftheEEGSignalSegmentsWhenUsedinSelf-Paced2-StateMentalTask-BasedBCIsTable 5.5: System Performance during Validation and Test with Artifact RejectionVisualization-based BCI Multiplication-based BCI Object rotation-based BCI Motor imagery-based BCISubject Process Length AR TPR FPR ATF* Length AR TPR FPR ATF Length AR TPR FPR ATF Length AR TPR FPR ATF1 Validation 1 46 61.54 0.00 36% 1.5 47 66.53 0.00 18% 1 61 62.88 0.00 10% 1 48 63.12 0.00 34%Test 61.35 0.00 68.41 0.00 63.04 0.00 63.95 0.002 Validation 1 51 62.64 0.00 23% 1 52 64.67 0.00 22% 1 52 61.71 0.00 23% 1 53 58.51 0.00 23%Test 61.85 0.00 62.57 0.00 61.74 0.00 59.12 0.003 Validation 1.5 39 62.20 0.00 25% 1 51 62.77 0.00 17% 1 54 65.51 0.00 17% 1.5 39 65.84 0.00 25%Test 63.08 0.00 64.58 0.00 65.44 0.00 65.65 0.004 Validation 1.5 48 62.55 0.00 23% 1.5 36 65.17 0.00 22% 2 28 45.37 0.00 44% 2.5 26 51.86 0.00 34%Test 64.79 0.00 66.69 0.00 50.02 0.00 53.68 0.00* The portion of the segments that are rejected due to artifact contamination.The table shows the mean values of the TPR and FPR measures. The results of the validation and test processes are given in two separate rowsfor each BCI of each subject. The optimum AR order and the optimum segment length for each system are also included in the table.101Chapter 5. A Study on Finding the Optimum Length of the EEG Signal Segments When Used inSelf-Paced 2-State Mental Task-Based BCIs5.3 SummaryIn this study, we investigate the effect of the EEG signal segment length on the performance of ourself-paced mental task-based BCIs. The optimum lengths obtained are 1 and 1.5 sec for all excepttwo systems (which have optimum lengths of 2 and 2.5 sec).The false activation rate based on the off-line evaluation reach zero. True positive rates aresufficiently high (i.e., 60?70% for all except one system). The system performance is evaluated ona dataset which we collected from four subjects.The methodology can be used in real-time systems after slight modifications since featureextraction and classification are not computationally demanding.We also show that our 7-channel mental task-based BCI systems are robust to the EOG andEMG artifacts.102Chapter 6Conclusions and Future WorkIn this chapter, we conclude the dissertation by summarizing our results and highlighting the con-tributions of the thesis. We also suggest some directions for further research.6.1 Research Contributions? In this thesis, we designed and developed a self-paced 2-state mental task-based brain?computer interface with a zero (or near zero) false activation rate.? In Chapter 2, we assessed the plausibility of a self-paced 2-state mental task-based BCIwith a zero or near zero false activation rate. We presented different designs for thesystem with the aim of minimizing the false activation rate. The scalar autoregressivemodeling was used as the feature extraction method. We studied the performance ofdifferent classifiers such as LDA, QDA, MDA, SVM, and RBF NN. The results wereextremely encouraging, since they showed that it is feasible to have a self-paced BCIbased on mental tasks with a false activation rate that approaches zero.? In Chapter 3, we evaluated the performance of our BCI using the dataset that we havecollected from four subjects. The FPRs reached zero and TPRs were sufficiently highfor real-life applications. For three subjects, the BCI with the highest performance had103Chapter 6. Conclusions and Future Worka TPR above 70%. For the forth subject, this rate was around 65%.? In Chapter 4, we decreased the number of EEG channels used in our BCI system from29 to 5 and 7. We showed that it is feasible to have a self-paced mental task-based BCIswith a zero false activation rate using very few channels. The system performance wasevaluated off-line on our dataset. The BCIs with the highest performance had zeroFPRs and sufficiently high TPRs (i.e., 53.89?63.72% for the 5-channel systems, and59.71?70.51% for the 7-channel systems). Therefore, they are acceptable for use inthe real-life applications in terms of the system performance. The methodology can beused in real time after slight modifications since feature extraction and classificationare not computationally demanding and there is no need for the timing information ofthe signals after training the classifier.? In Chapter 5, the change in the performance of the 7-channel BCI systems was studiedwhen the length of the EEG signal segments were varied. The optimum length obtainedwas 1 and 1.5 sec for all except two systems (which had the optimum lengths of 2 and2.5 sec). The false activation rate reached zero. True positive rates were in the rangeof 60?70% for all except one system. We also demonstrated that the 7-channel mentaltask-based BCIs are robust to the artifacts to a certain extent.? The performance of different classifiers such as LDA, QDA, MDA, SVM, and RBF NN wasstudied in Chapter 2. The results showed that:? The QDA classifier outperformed other classifiers in terms of TPRs, FPRs, and com-putational costs.? The performance of the classifier based on LDA was not acceptable.? The QDA classifier had better performance than the MDA classifier.? The RBF NN classifier performed almost as well as the QDA classifier, however, theQDA classifier is preferable because of its simplicity.104Chapter 6. Conclusions and Future Work? The only two classifiers that were not significantly different from each other in theperformance were the QDA classifier and the RBF NN classifier.? The SVM classifier showed a higher TPR than the QDA and the RBF NN classifiers,but its non-zero FPR was a disadvantage.? We extracted different features from the EEG signals using the wavelets in the design ofBCI systems. The results were published [59, 61?63]. However, the features based on theautoregressive modeling outperformed those features in terms of the simplicity and the finalclassification accuracy.? Even though many BCI systems have been designed based on motor imagery as mentionedin Chapter 1, our findings showed that the motor imagery task yields the BCI systems withthe poorest performance for 2 out of 4 subjects. This implies that to develop a self-pacedBCI, mental tasks that are not motor imagery also need to be considered. The final selectionof mental tasks should be made after comparing the performance of the system based onmotor imagery and non-motor imagery mental tasks.6.2 Suggestions for Future WorkThe following are some directions for future research:? Assigning different costs for misclassifying IC and NC statesIn this thesis, we used the same value for C1 and C2, the cost of false negative and the costof false positive classification, respectively, as in Equation (2.18). One can assign differentcosts to false negatives and false positives and investigate how non-equal costs would affectthe results.? Leave-one-out cross-validationWe performed 5-fold cross-validation to evaluate the performance of our BCI systems. Onecan think of leave-one-out cross-validation instead. To do this, one complete epoch is held105Chapter 6. Conclusions and Future Workout for testing. Training and validation are performed using the other epochs.? Considering different brain states and activitiesIn the BCI systems developed in this dissertation, one mental task was considered as the ICstate and the baseline state and the other three mental tasks were considered as the NC state.During actual use, all states and activities other than the IC task will be placed in the NCstate. In other words, the NC state is not limited to the baseline state and the other threemental tasks considered in this study. Considering different brain states and activities as theNC state is one important direction for future work.? Frequency domain analysisWe can analyze the signals in the frequency domain to find out the frequency bands with thehighest information rate. This might allow us to decrease the sampling rate from its currentvalue of 500 Hz and make the signal processing simpler.? Selecting different features and classifiersSelecting different methods for feature extraction and classification can improve the systemperformance in terms of the TPR and FPR values. One can also vary the SVM classifierparameters to investigate whether it outperforms the QDA classifier.? Subject trainingWe believe that with further subject training during sessions with feedback, the results wouldlikely be improved. This is something that can be investigated in future work.? Running on-line experimentsRunning on-line experiments is another direction for future work in order to evaluate theperformance of our BCI system in real time.? Making a portable BCI systemMaking the system portable and easy to setup is the final step in developing the BCI systems.The signal processing (i.e., the preprocessing, the feature extraction, and the classification106Chapter 6. Conclusions and Future Workstages) can be implemented on a digital signal processor. This would make the BCI systemportable. Also, to make the system setup easy, one can think of easy to use electrodes (e.g.,dry electrodes) instead of regular electrodes.107Bibliography[1] F. Abdollahi and A. Motie-Nasrabadi. Combination of frequency bands in EEG for featurereduction in mental task classification. In Proceedings of the 28th Annual InternationalConference of the IEEE Engineering in Medicine and Biology Society, pages 1146?1149,September 2006. doi:10.1109/IEMBS.2006.260229. URLhttp://dx.doi.org/10.1109/IEMBS.2006.260229. ? pages 18[2] F. Abdollahi, S. K. Setarehdan, and A. M. Nasrabadi. Locating information maximizationtime in EEG signals recorded during mental tasks. In Proceedings of the 5th InternationalSymposium on Image and Signal Processing and Analysis, pages 238?241, September2007. doi:10.1109/ISPA.2007.4383697. URLhttp://dx.doi.org/10.1109/ISPA.2007.4383697. ? pages 18[3] C. Anderson, E. Forney, D. Hains, and A. Natarajan. Reliable identification of mental tasksusing time-embedded EEG and sequential evidence accumulation. Journal of NeuralEngineering, 8(2):025023, 2011. ISSN 1741-2560, 1741-2552.doi:10.1088/1741-2560/8/2/025023. URLhttp://dx.doi.org/10.1088/1741-2560/8/2/025023. ? pages 6, 7[4] C. W. Anderson. Effects of variations in neural network topology and output averaging onthe discrimination of mental tasks from spontaneous electroencephalogram. Journal ofIntelligent Systems, 7:165?190, 1997. ? pages 18, 26[5] C. W. Anderson and J. A. Bratman. Translating thoughts into actions by finding patterns inbrainwaves. In 14th Yale Workshop on Adaptive and Learning Systems, pages 1?6, 2008.? pages 6, 9, 18[6] C. W. Anderson and Z. Sijerc?ic?. Classification of EEG signals from four subjects duringfive mental tasks. In Proceedings of the Conference on Engineering Applications in NeuralNetworks, pages 407?414, June 1996. ? pages 26[7] C. W. Anderson, S. V. Devulapalli, and E. A. Stolz. Determining mental state from EEGsignals using parallel implementations of neural networks. Scientific Programming, 4(3):171?183, 1995. ISSN 1058-9244. URLhttp://iospress.metapress.com/content/10KK81326751964R. ? pages108Bibliography[8] C. W. Anderson, E. A. Stolz, and S. Shamsunder. Discriminating mental tasks using EEGrepresented by AR models. In Proceedings of the 17th Annual International Conference ofthe IEEE Engineering in Medicine and Biology Society, volume 2, pages 875?876,September 1995. doi:10.1109/IEMBS.1995.579248. URLhttp://dx.doi.org/10.1109/IEMBS.1995.579248. ? pages 26[9] C. W. Anderson, E. A. Stolz, and S. Shamsunder. Multivariate autoregressive models forclassification of spontaneous electroencephalographic signals during mental tasks. IEEETransactions on Biomedical Engineering,, 45(3):277?286, March 1998. ISSN 0018-9294.doi:10.1109/10.661153. URL http://dx.doi.org/10.1109/10.661153. ? pages 25[10] C. W. Anderson, J. N. Knight, T. O?Connor, M. J. Kirby, and A. Sokolov. Geometricsubspace methods and time-delay embedding for EEG artifact removal and classification.IEEE Transactions on Neural Systems and Rehabilitation Engineering, 14(2):142?146,June 2006. ISSN 1534-4320. doi:10.1109/TNSRE.2006.875527. URLhttp://dx.doi.org/10.1109/TNSRE.2006.875527. ? pages 6, 9, 10, 18[11] N. R. Anderson, K. Wisneski, L. Eisenman, D. W. Moran, E. C. Leuthardt, and D. J.Krusienski. An offline evaluation of the autoregressive spectrum for electrocorticography.IEEE Transactions on Biomedical Engineering, 56(3):913?916, 2009. ISSN 0018-9294.doi:10.1109/TBME.2009.2009767. URL http://dx.doi.org/10.1109/TBME.2009.2009767.? pages 83[12] A. C. Atkinson, M. Riani, and A. Cerioli. Exploring Multivariate Data with the ForwardSearch. Springer, New York, 2004. ? pages 26, 29[13] S. Aydin. Determination of autoregressive model orders for seizure detection. TurkishJournal of Electrical Engineering & Computer Sciences, 18:23?30, 2010.doi:10.3906/elk-0906-83. URL http://dx.doi.org/10.3906/elk-0906-83. ? pages 83[14] F. Babiloni, F. Cincotti, L. Lazzarini, J. Mill?n, J. Mouri?o, M. Varsta, J. Heikkonen,L. Bianchi, and M. G. Marciani. Linear classification of low-resolution EEG patternsproduced by imagined hand movements. IEEE Transactions on RehabilitationEngineering, 8(2):186?188, 2000. ISSN 1063-6528. doi:10.1109/86.847810. URLhttp://dx.doi.org/10.1109/86.847810. ? pages 6, 10, 15[15] F. Babiloni, F. Cincotti, L. Bianchi, G. Pirri, J. del R. Mill?n, J. Mouri?o, S. Salinari, andM. G. Marciani. Recognition of imagined hand movements with low resolution surfaceLaplacian and linear classifiers. Medical Engineering & Physics, 23(5):323?328, 2001.ISSN 1350-4533. doi:10.1016/S1350-4533(01)00049-2. URLhttp://dx.doi.org/10.1016/S1350-4533(01)00049-2. ? pages 6, 10, 15[16] O. Bai, P. Lin, S. Vorbach, J. Li, S. Furlani, and M. Hallett. Exploration of computationalmethods for classification of movement intention during human voluntary movement from109Bibliographysingle trial EEG. Clinical Neurophysiology, 118(12):2637?2655, 2007. ISSN 1388-2457.doi:10.1016/j.clinph.2007.08.025. URL http://dx.doi.org/10.1016/j.clinph.2007.08.025. ?pages 15[17] G. A. Barreto, R. A. Frota, and F. N. S. de Medeiros. On the classification of mental tasks:A performance comparison of neural and statistical approaches. In Proceedings of the 14thIEEE Signal Processing Society Workshop on Machine Learning for Signal Processing,pages 529?538, October 2004. doi:10.1109/MLSP.2004.1423016. URLhttp://dx.doi.org/10.1109/MLSP.2004.1423016. ? pages 18[18] A. Bashashati, M. Fatourechi, R. K. Ward, and G. E. Birch. User customization of thefeature generator of an asynchronous brain interface. Annals of Biomedical Engineering,34(6):1051?1060, June 2006. doi:10.1007/s10439-006-9097-5. URLhttp://dx.doi.org/10.1007/s10439-006-9097-5. ? pages 15, 23, 53, 74, 88[19] A. Bashashati, S. G. Mason, R. K. Ward, and G. E. Birch. An improved asynchronousbrain interface: making use of the temporal history of the LF-ASD feature vectors. Journalof Neural Engineering, 3(2):87?94, 2006. doi:10.1088/1741-2560/3/2/002. URLhttp://dx.doi.org/10.1088/1741-2560/3/2/002. ? pages 15[20] A. Bashashati, M. Fatourechi, R. K. Ward, and G. E. Birch. A survey of signal processingalgorithms in brain-computer interfaces based on electrical brain signals. Journal of NeuralEngineering, 4(2):R32?R57, June 2007. doi:10.1088/1741-2560/4/2/R03. URLhttp://dx.doi.org/10.1088/1741-2560/4/2/R03. ? pages 5[21] A. Bashashati, S. G. Mason, J. F. Borisoff, R. K. Ward, and G. E. Birch. A comparativestudy on generating training-data for self-paced brain interfaces. IEEE Transactions onNeural Systems and Rehabilitation Engineering, 15(1):59?66, 2007. ISSN 1534-4320.doi:10.1109/TNSRE.2007.891382. URL http://dx.doi.org/10.1109/TNSRE.2007.891382.? pages 15[22] A. Bashashati, B. Nouredin, R. K. Ward, P. Lawrence, and G. E. Birch. Effect ofeye-blinks on a self-paced brain interface design. Clinical Neurophysiology, 118(7):1639?1647, 2007. ISSN 1388-2457. doi:10.1016/j.clinph.2007.03.020. URLhttp://dx.doi.org/10.1016/j.clinph.2007.03.020. ? pages[23] A. Bashashati, R. K. Ward, and G. E. Birch. Towards development of a 3-state self-pacedbrain?computer interface. Computational Intelligence and Neuroscience, 2007:1?8, 2007.ISSN 1687-5265. doi:10.1155/2007/84386. URL http://dx.doi.org/10.1155/2007/84386.? pages 15[24] M. I. Bhatti, A. Pervaiz, and M. H. Baig. EEG signal decomposition and improved spectralanalysis using wavelet transform. In Proceedings of the 23rd Annual InternationalConference of the IEEE Engineering in Medicine and Biology Society, volume 2, pages110Bibliography1862?1864, October 2001. doi:10.1109/IEMBS.2001.1020587. URLhttp://dx.doi.org/10.1109/IEMBS.2001.1020587. ? pages 18[25] G. E. Birch, P. D. Lawrence, J. C. Lind, and R. D. Hare. Application of prewhitening toAR spectral estimation of EEG. IEEE Transactions on Biomedical Engineering, 35(8):640?645, August 1988. ISSN 0018-9294. doi:10.1109/10.4597. URLhttp://dx.doi.org/10.1109/10.4597. ? pages 22, 52, 74, 88[26] G. E. Birch, P. D. Lawrence, and R. D. Hare. Single-trial processing of event-relatedpotentials using outlier information. IEEE Transactions on Biomedical Engineering, 40(1):59?73, 1993. ISSN 0018-9294. doi:10.1109/10.204772. URLhttp://dx.doi.org/10.1109/10.204772. ? pages 15[27] G. E. Birch, Z. Bozorgzadeh, and S. G. Mason. Initial on-line evaluations of the LF-ASDbrain?computer interface with able-bodied and spinal-cord subjects using imaginedvoluntary motor potentials. IEEE Transactions on Neural Systems and RehabilitationEngineering, 10(4):219?224, 2002. ISSN 1534-4320. doi:10.1109/TNSRE.2002.806839.URL http://dx.doi.org/10.1109/TNSRE.2002.806839. ? pages[28] G. E. Birch, S. G. Mason, and J. F. Borisoff. Current trends in brain?computer interfaceresearch at the Neil Squire foundation. IEEE Transactions on Neural Systems andRehabilitation Engineering, 11(2):123?126, 2003. ISSN 1534-4320.doi:10.1109/TNSRE.2003.814450. URL http://dx.doi.org/10.1109/TNSRE.2003.814450.? pages 15[29] G. Blanchard and B. Blankertz. BCI competition 2003-data set IIa: spatial patterns ofself-controlled brain rhythm modulations. IEEE Transactions on Biomedical Engineering,51(6):1062?1066, June 2004. ISSN 0018-9294. doi:10.1109/TBME.2004.826691. URLhttp://dx.doi.org/10.1109/TBME.2004.826691. ? pages 23, 53, 74, 88[30] B. Blankertz, G. Dornhege, M. Krauledat, K. R. Muller, V. Kunzmann, F. Losch, andG. Curio. The Berlin brain?computer interface: EEG-based communication withoutsubject training. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 14(2):147 ?152, 2006. ISSN 1534-4320. doi:10.1109/TNSRE.2006.875557. URLhttp://dx.doi.org/10.1109/TNSRE.2006.875557. ? pages 15[31] B. Blankertz, G. Dornhege, M. Krauledat, K.-R. M?ller, and G. Curio. The non-invasiveBerlin Brain?Computer Interface: Fast acquisition of effective performance in untrainedsubjects. NeuroImage, 37(2):539?550, 2007. ISSN 1053-8119.doi:10.1016/j.neuroimage.2007.01.051. URLhttp://dx.doi.org/10.1016/j.neuroimage.2007.01.051. ? pages 6, 10, 15[32] J. F. Borisoff, S. G. Mason, A. Bashashati, and G. E. Birch. Brain?computer interfacedesign for asynchronous control applications: improvements to the LF-ASD asynchronous111Bibliographybrain switch. IEEE Transactions on Biomedical Engineering, 51(6):985?992, 2004. ISSN0018-9294. doi:10.1109/TBME.2004.827078. URLhttp://dx.doi.org/10.1109/TBME.2004.827078. ? pages 15[33] A. T. Boye, U. Q. Kristiansen, M. Billinger, O. F. do Nascimento, and D. Farina.Identification of movement-related cortical potentials with optimized spatial filtering andprincipal component analysis. Biomedical Signal Processing and Control, 3(4):300?304,2008. ISSN 1746-8094. doi:10.1016/j.bspc.2008.05.001. URLhttp://dx.doi.org/10.1016/j.bspc.2008.05.001. ? pages 6, 10, 13, 15[34] C. Brunner, B. Z. Allison, D. J. Krusienski, V. Kaiser, G. R. M?ller-Putz, G. Pfurtscheller,and C. Neuper. Improved signal processing approaches in an offline simulation of a hybridbrain?computer interface. Journal of Neuroscience Methods, 188(1):165?173, 2010. ISSN0165-0270. doi:10.1016/j.jneumeth.2010.02.002. URLhttp://dx.doi.org/10.1016/j.jneumeth.2010.02.002. ? pages 6, 10, 15[35] J. P. Burg. A new analysis technique for time series data. NATO Adv. Study Inst. on SignalProcessing with Emphasis on Underwater Acoustics, Enschede, The Netherlands, Aug.1968, reprinted in Modern Spectrum Analysis, D.G. Childers, ed., IEEE Press, pages42?48, 1978. ? pages 26[36] D. P. Burke, S. P. Kelly, P. de Chazal, R. B. Reilly, and C. Finucane. A parametric featureextraction and classification strategy for brain?computer interfacing. IEEE Transactions onNeural Systems and Rehabilitation Engineering, 13(1):12?17, 2005. ISSN 1534-4320.doi:10.1109/TNSRE.2004.841881. URL http://dx.doi.org/10.1109/TNSRE.2004.841881.? pages 15[37] S. Chiappa and D. Barber. EEG classification using generative independent componentanalysis. Neurocomputing, 69(7-9):769?777, 2006. ISSN 0925-2312.doi:10.1016/j.neucom.2005.12.028. URLhttp://dx.doi.org/10.1016/j.neucom.2005.12.028. ? pages 6, 7, 10[38] A. Cichocki, H. Lee, Y.-D. Kim, and S. Choi. Non-negative matrix factorization with?-divergence. Pattern Recognition Letters, 29(9):1433?1440, 2008. ISSN 0167-8655.doi:10.1016/j.patrec.2008.02.016. URL http://dx.doi.org/10.1016/j.patrec.2008.02.016. ?pages 6, 7[39] F. Cincotti, D. Mattia, C. Babiloni, F. Carducci, S. Salinari, L. Bianchi, M. G. Marciani,and F. Babiloni. The use of EEG modifications due to motor imagery for brain?computerinterfaces. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 11(2):131?133, 2003. ISSN 1534-4320. doi:10.1109/TNSRE.2003.814455. URLhttp://dx.doi.org/10.1109/TNSRE.2003.814455. ? pages 6, 15[40] F. Cincotti, D. Mattia, F. Aloise, S. Bufalari, L. Astolfi, F. D. V. Fallani, A. Tocci,L. Bianchi, M. G. Marciani, S. Gao, J. Millan, and F. Babiloni. High-resolution EEG112Bibliographytechniques for brain?computer interface applications. Journal of Neuroscience Methods,167(1):31?42, 2008. ISSN 0165-0270. doi:10.1016/j.jneumeth.2007.06.031. URLhttp://dx.doi.org/10.1016/j.jneumeth.2007.06.031. ? pages 6, 10, 15[41] M. Congedo, F. Lotte, and A. L?cuyer. Classification of movement intention by spatiallyfiltered electromagnetic inverse solutions. Physics in Medicine and Biology, 51(8):1971,2006. doi:10.1088/0031-9155/51/8/002. URLhttp://dx.doi.org/10.1088/0031-9155/51/8/002. ? pages 15[42] E. J. X. Costa and E. F. Cabral Jr. EEG-based discrimination between imagination of leftand right hand movements using adaptive gaussian representation. Medical Engineering &Physics, 22(5):345?348, 2000. ISSN 1350-4533. doi:10.1016/S1350-4533(00)00051-5.URL http://dx.doi.org/10.1016/S1350-4533(00)00051-5. ? pages 6, 10, 15[43] D. Coyle, T. M. McGinnity, and G. Prasad. Improving the separability of multiple EEGfeatures for a BCI by neural-time-series-prediction-preprocessing. Biomedical SignalProcessing and Control, 5(3):196?204, 2010. ISSN 1746-8094.doi:10.1016/j.bspc.2010.03.004. URL http://dx.doi.org/10.1016/j.bspc.2010.03.004. ?pages 6, 10, 15[44] N. Cristianini and J. Shawe-Taylor. An Introduction to Support Vector Machines and OtherKernel-based Learning Methods. Cambridge University Press, 1 edition, 2000. ISBN0521780195. ? pages 31[45] E. Curran, P. Sykacek, M. Stokes, S. J. Roberts, W. Penny, I. Johnsrude, and A. M. Owen.Cognitive tasks for driving a brain-computer interfacing system: a pilot study. IEEETransactions on Neural Systems and Rehabilitation Engineering, 12(1):48?54, 2003. ISSN1534-4320. doi:10.1109/TNSRE.2003.821372. URLhttp://dx.doi.org/10.1109/TNSRE.2003.821372. ? pages 6, 7, 10[46] C. S. DaSalla, H. Kambara, M. Sato, and Y. Koike. Single-trial classification of vowelspeech imagery using common spatial patterns. Neural Networks, 22(9):1334?1339, 2009.ISSN 0893-6080. doi:10.1016/j.neunet.2009.05.008. URLhttp://dx.doi.org/10.1016/j.neunet.2009.05.008. ? pages 6, 9, 10[47] M. Daud and J. Yunus. Classification of mental tasks using de-noised EEG signals. InProceedings of the 7th International Conference on Signal Processing, volume 3, pages2206?2209, August 2004. doi:10.1109/ICOSP.2004.1442216. URLhttp://dx.doi.org/10.1109/ICOSP.2004.1442216. ? pages 18, 26[48] J. del R. Mill?n and J. Mouri?o. Asynchronous BCI and local neural classifiers: Anoverview of the adaptive brain interface project. IEEE Transactions on Neural Systems andRehabilitation Engineering, 11(2):159?161, 2003. ISSN 1534-4320.doi:10.1109/TNSRE.2003.814435. URL http://dx.doi.org/10.1109/TNSRE.2003.814435.? pages 6, 7, 10, 11, 15113Bibliography[49] J. del R. Mill?n, J. Mouri?o, F. Cincotti, M. Varsta, J. Heikkonen, F. Topani, M. G.Marciani, K. Kaski, and F. Babiloni. EEG patterns associated to spontaneous execution ofmental tasks. NeuroImage, 11(5, Supplement 1):S78?S78, 2000. ISSN 1053-8119.doi:10.1016/S1053-8119(00)91012-7. URLhttp://dx.doi.org/10.1016/S1053-8119(00)91012-7. ? pages 6, 15[50] J. del R. Mill?n, M. Franz?, J. Mouri?o, F. Cincotti, and F. Babiloni. Relevant EEGfeatures for the classification of spontaneous motor-related tasks. Biological Cybernetics,86(2):89?95, 2002. ISSN 0340-1200. doi:10.1007/s004220100282. URLhttp://dx.doi.org/10.1007/s004220100282. ? pages 6, 15[51] J. del R. Mill?n, J. Mouri?o, M. Franz?, F. Cincotti, M. Varsta, J. Heikkonen, andF. Babiloni. A local neural classifier for the recognition of EEG patterns associated tomental tasks. IEEE Transactions on Neural Networks, 13(3):678?686, 2002. ISSN1045-9227. doi:10.1109/TNN.2002.1000132. URLhttp://dx.doi.org/10.1109/TNN.2002.1000132. ? pages 6, 7, 10, 11, 15[52] J. del R. Mill?n, F. Renkens, J. Mouri?o, and W. Gerstner. Brain-actuated interaction.Artificial Intelligence, 159(1-2):241?259, 2004. ISSN 0004-3702.doi:10.1016/j.artint.2004.05.008. URL http://dx.doi.org/10.1016/j.artint.2004.05.008. ?pages 6, 7, 10, 11, 15[53] D. M. Dobrea and M. C. Dobrea. An EEG (bio) technological system for assisting thedisabled people. In Proceedings of the 5th IEEE International Conference onComputational Cybernetics, pages 191?196, October 2007.doi:10.1109/ICCCYB.2007.4402033. URLhttp://dx.doi.org/10.1109/ICCCYB.2007.4402033. ? pages 18[54] D. M. Dobrea, M. C. Dobrea, and M. Costin. An EEG coherence based method used formental tasks classification. In Proceedings of the 5th IEEE International Conference onComputational Cybernetics, pages 185?190, October 2007.doi:10.1109/ICCCYB.2007.4402032. URLhttp://dx.doi.org/10.1109/ICCCYB.2007.4402032. ? pages 18[55] M. Dyson, F. Sepulveda, and J. Q. Gan. Mental task classification against the idle state: Apreliminary investigation. In Proceedings of the 30th Annual International Conference ofthe IEEE Engineering in Medicine and Biology Society, pages 4473?4477, 2008. ISBN1557-170X. doi:10.1109/IEMBS.2008.4650206. URLhttp://dx.doi.org/10.1109/IEMBS.2008.4650206. ? pages 6, 7, 10, 12, 15[56] M. Dyson, F. Sepulveda, J. Q. Gan, and S. J. Roberts. Sequential classification of mentaltasks vs. idle state for EEG based BCIs. In Proceedings of the 4th International IEEEEMBS Conference on Neural Engineering, pages 351?354, 2009.114Bibliographydoi:10.1109/NER.2009.5109305. URL http://dx.doi.org/10.1109/NER.2009.5109305. ?pages 8, 12[57] M. Dyson, F. Sepulveda, and J. Q. Gan. Localisation of cognitive tasks used in EEG-basedBCIs. Clinical Neurophysiology, 121(9):1481?1493, 2010. ISSN 1388-2457.doi:10.1016/j.clinph.2010.03.011. URL http://dx.doi.org/10.1016/j.clinph.2010.03.011. ?pages 6, 8, 10, 12, 15[58] K. Ericson, S. Pallickara, and C. W. Anderson. Analyzing electroencephalograms usingcloud computing techniques. In IEEE Second International Conference on CloudComputing Technology and Science (CloudCom), pages 185?192, 2010.doi:10.1109/CloudCom.2010.80. URL http://dx.doi.org/10.1109/CloudCom.2010.80. ?pages 6, 8[59] F. Faradji, R. K. Ward, and G. E. Birch. Self-paced BCI using multiple SWT-basedclassifiers. In 30th Annual International Conference of the IEEE Engineering in Medicineand Biology Society (EMBS), pages 2095?2098, 2008. doi:10.1109/IEMBS.2008.4649606.URL http://dx.doi.org/10.1109/IEMBS.2008.4649606. ? pages 20, 105[60] F. Faradji, R. K. Ward, and G. E. Birch. Plausibility assessment of a 2-state self-pacedmental task-based BCI using the no-control performance analysis. Journal of NeuroscienceMethods, 180(2):330?339, June 2009. ISSN 0165-0270.doi:10.1016/j.jneumeth.2009.03.011. URLhttp://dx.doi.org/10.1016/j.jneumeth.2009.03.011. ? pages xi, 15, 52, 67, 74, 88[61] F. Faradji, R. K. Ward, and G. E. Birch. A custom-designed mental task-basedbrain?computer interface. In IEEE International Conference on Acoustics, Speech andSignal Processing (ICASSP), pages 529?532, 2009. doi:10.1109/ICASSP.2009.4959637.URL http://dx.doi.org/10.1109/ICASSP.2009.4959637. ? pages 20, 105[62] F. Faradji, R. K. Ward, and G. E. Birch. A self-paced BCI using stationary wavelet packets.In Annual International Conference of the IEEE Engineering in Medicine and BiologySociety (EMBC), pages 962?965, 2009. doi:10.1109/IEMBS.2009.5332802. URLhttp://dx.doi.org/10.1109/IEMBS.2009.5332802. ? pages[63] F. Faradji, R. K. Ward, and G. E. Birch. A simple approach to find the best wavelet basis inclassification problems. In 20th International Conference on Pattern Recognition (ICPR),pages 641?644, 2010. doi:10.1109/ICPR.2010.162. URLhttp://dx.doi.org/10.1109/ICPR.2010.162. ? pages 20, 105[64] F. Faradji, R. K. Ward, and G. E. Birch. Toward development of a two-statebrain?computer interface based on mental tasks. Journal of Neural Engineering, 8(4):046014, 2011. ISSN 1741-2560. doi:10.1088/1741-2560/8/4/046014. URLhttp://dx.doi.org/10.1088/1741-2560/8/4/046014. ? pages 16115Bibliography[65] M. Fatourechi, A. Bashashati, G. E. Birch, and R. K. Ward. Automatic user customizationfor improving the performance of a self-paced brain interface system. Medical &Biological Engineering & Computing, 44(12):1093?1104, 2006. ISSN 0140-0118,1741-0444. doi:10.1007/s11517-006-0125-2. URLhttp://dx.doi.org/10.1007/s11517-006-0125-2. ? pages 15[66] M. Fatourechi, G. E. Birch, and R. K. Ward. Application of a hybrid wavelet featureselection method in the design of a self-paced brain interface system. Journal ofNeuroengineering and Rehabilitation, 4:11, 2007. ISSN 1743-0003.doi:10.1186/1743-0003-4-11. URL http://dx.doi.org/10.1186/1743-0003-4-11. PMID:17470288. ? pages[67] M. Fatourechi, G. E. Birch, and R. K. Ward. A self-paced brain interface system that usesmovement related potentials and changes in the power of brain rhythms. Journal ofComputational Neuroscience, 23(1):21?37, 2007. ISSN 0929-5313, 1573-6873.doi:10.1007/s10827-006-0017-3. URL http://dx.doi.org/10.1007/s10827-006-0017-3. ?pages[68] M. Fatourechi, R. K. Ward, and G. E. Birch. Performance of a self-paced brain computerinterface on data contaminated with eye-movement artifacts and on data recorded in asubsequent session. Computational Intelligence and Neuroscience, page 749204, 2008.ISSN 1687-5265. doi:10.1155/2008/749204. URL http://dx.doi.org/10.1155/2008/749204.PMID: 18497872. ? pages[69] M. Fatourechi, R. K. Ward, and G. E. Birch. A self-paced brain?computer interface systemwith a low false positive rate. Journal of Neural Engineering, 5(1):9?23, 2008.doi:10.1088/1741-2560/5/1/002. URL http://dx.doi.org/10.1088/1741-2560/5/1/002. ?pages 5, 15[70] S. Fazli, F. Popescu, M. Dan?czy, B. Blankertz, K.-R. M?ller, and C. Grozea.Subject-independent mental state classification in single trials. Neural Networks, 22(9):1305?1312, 2009. ISSN 0893-6080. doi:10.1016/j.neunet.2009.06.003. URLhttp://dx.doi.org/10.1016/j.neunet.2009.06.003. ? pages 6, 10, 15[71] E. M. Forney and C. W. Anderson. Classification of EEG during imagined mental tasks byforecasting with Elman Recurrent Neural Networks. In The 2011 International JointConference on Neural Networks (IJCNN), pages 2749?2755, 2011.doi:10.1109/IJCNN.2011.6033579. URL http://dx.doi.org/10.1109/IJCNN.2011.6033579.? pages 6, 8[72] P. J. Franaszczuk, K. J. Blinowska, and M. Kowalczyk. The application of parametricmultichannel spectral estimates in the study of electrical brain activity. BiologicalCybernetics, 51(4):239?247, January 1985. doi:10.1007/BF00337149. URLhttp://dx.doi.org/10.1007/BF00337149. ? pages 25, 26116Bibliography[73] F. Gal?n, M. Nuttin, E. Lew, P. Ferrez, G. Vanacker, J. Philips, and J. del R. Mill?n. Abrain-actuated wheelchair: Asynchronous and non-invasive brain?computer interfaces forcontinuous control of robots. Clinical Neurophysiology, 119(9):2159?2169, 2008. ISSN1388-2457. doi:10.1016/j.clinph.2008.06.001. URLhttp://dx.doi.org/10.1016/j.clinph.2008.06.001. ? pages 6, 7, 10, 11, 15[74] D. Garrett, D. A. Peterson, C. W. Anderson, and M. H. Thaut. Comparison of linear,nonlinear, and feature selection methods for EEG signal classification. IEEE Transactionson Neural Systems and Rehabilitation Engineering, 11(2):141?144, June 2003. ISSN1534-4320. doi:10.1109/TNSRE.2003.814441. URLhttp://dx.doi.org/10.1109/TNSRE.2003.814441. ? pages 18, 26[75] T. Geng, J. Q. Gan, M. Dyson, C. S. Tsui, and F. Sepulveda. A novel design of 4-class BCIusing two binary classifiers and parallel mental tasks. Computational Intelligence andNeuroscience, 2008, 2008. doi:10.1155/2008/437306. URLhttp://dx.doi.org/10.1155/2008/437306. Article ID 437306, 5 pages, 2008. ? pages 6, 10,15[76] C. Gope, N. Kehtarnavaz, and D. Nair. Neural network classification of EEG signals usingtime-frequency representation. In Proceedings of the IEEE International Joint Conferenceon Neural Networks, volume 4, pages 2502?2507, August 2005.doi:10.1109/IJCNN.2005.1556296. URL http://dx.doi.org/10.1109/IJCNN.2005.1556296.? pages 18[77] C. Guger, A. Schl?gl, C. Neuper, D. Walterspacher, T. Strein, and G. Pfurtscheller. Rapidprototyping of an EEG-based brain?computer interface (BCI). IEEE Transactions onNeural Systems and Rehabilitation Engineering, 9(1):49?58, 2001. ISSN 1534-4320.doi:10.1109/7333.918276. URL http://dx.doi.org/10.1109/7333.918276. ? pages 6, 10,15[78] L. Guo, Y. Wu, L. Zhao, T. Cao, W. Yan, and X. Shen. Classification of mental task fromEEG signals using immune feature weighted support vector machines. IEEE Transactionson Magnetics, 47(5):866?869, 2011. ISSN 0018-9464. doi:10.1109/TMAG.2010.2072775.URL http://dx.doi.org/10.1109/TMAG.2010.2072775. ? pages 6, 9, 10[79] M. M. Gupta, L. Jin, and N. Homma. Static and Dynamic Neural Networks: FromFundamentals to Advanced Theory. Wiley-IEEE Press, 1 edition, 2003. ISBN0471219487. ? pages 33[80] E. Gysels and P. Celka. Phase synchronization for the recognition of mental tasks in abrain?computer interface. IEEE Transactions on Neural Systems and RehabilitationEngineering, 12(4):406?415, 2004. ISSN 1534-4320. doi:10.1109/TNSRE.2004.838443.URL http://dx.doi.org/10.1109/TNSRE.2004.838443. ? pages 6, 10, 15117Bibliography[81] E. Gysels, P. Renevey, and P. Celka. SVM-based recursive feature elimination to comparephase synchronization computed from broadband and narrowband EEG signals inBrain?Computer Interfaces. Signal Processing, 85(11):2178?2189, 2005. ISSN0165-1684. doi:10.1016/j.sigpro.2005.07.008. URLhttp://dx.doi.org/10.1016/j.sigpro.2005.07.008. ? pages[82] M. K. Hazrati and A. Erfanian. An online EEG-based brain?computer interface forcontrolling hand grasp using an adaptive probabilistic neural network. MedicalEngineering & Physics, 32(7):730?739, 2010. ISSN 1350-4533.doi:10.1016/j.medengphy.2010.04.016. URLhttp://dx.doi.org/10.1016/j.medengphy.2010.04.016. ? pages 6, 10, 15[83] C. R. Hema, M. P. Paulraj, R. Nagarajan, S. Yaacob, and A. H. Adorn. Fuzzy basedclassification of EEG mental tasks for a brain machine interface. In Proceedings of the 3rdInternational Conference on Intelligent Information Hiding and Multimedia SignalProcessing, volume 1, pages 53?56, November 2007. doi:10.1109/IIHMSP.2007.4457491.URL http://dx.doi.org/10.1109/IIHMSP.2007.4457491. ? pages 18[84] S. M. Hosni, M. E. Gadallah, S. F. Bahgat, and M. S. AbdelWahab. Classification of EEGsignals using different feature extraction techniques for mental-task BCI. In Proceedings ofthe International Conference on Computer Engineering Systems, pages 220?226,November 2007. doi:10.1109/ICCES.2007.4447052. URLhttp://dx.doi.org/10.1109/ICCES.2007.4447052. ? pages 18, 26[85] W.-Y. Hsu. EEG-based motor imagery classification using neuro-fuzzy prediction andwavelet fractal features. Journal of Neuroscience Methods, 189(2):295?302, 2010. ISSN0165-0270. doi:10.1016/j.jneumeth.2010.03.030. URLhttp://dx.doi.org/10.1016/j.jneumeth.2010.03.030. ? pages 6, 10, 15[86] W.-Y. Hsu and Y.-N. Sun. EEG-based motor imagery analysis using weighted wavelettransform features. Journal of Neuroscience Methods, 176(2):310?318, 2009. ISSN0165-0270. doi:10.1016/j.jneumeth.2008.09.014. URLhttp://dx.doi.org/10.1016/j.jneumeth.2008.09.014. ? pages[87] B. hua Yang, G. zheng Yan, R. guo Yan, and T. Wu. Adaptive subject-based featureextraction in brain?computer interfaces using wavelet packet best basis decomposition.Medical Engineering & Physics, 29(1):48?53, 2007. ISSN 1350-4533.doi:10.1016/j.medengphy.2006.01.009. URLhttp://dx.doi.org/10.1016/j.medengphy.2006.01.009. ? pages[88] B. hua Yang, G. zheng Yan, T. Wu, and R. guo Yan. Subject-based feature extraction usingfuzzy wavelet packet in brain?computer interfaces. Signal Processing, 87(7):1569?1574,2007. ISSN 0165-1684. doi:10.1016/j.sigpro.2006.12.018. URLhttp://dx.doi.org/10.1016/j.sigpro.2006.12.018. ? pages 6, 10, 15118Bibliography[89] N. J. Huan and R. Palaniappan. Classification of mental tasks using fixed and adaptiveautoregressive models of EEG signals. In Proceedings of the 2nd International IEEEEMBS Conference on Neural Engineering, pages 633?636, March 2005.doi:10.1109/CNE.2005.1419704. URL http://dx.doi.org/10.1109/CNE.2005.1419704. ?pages 18, 26[90] E. I??ez, J. M. Azor?n, A. ?beda, J. M. Ferr?ndez, and E. Fern?ndez. Mental tasks-basedbrain?robot interface. Robotics and Autonomous Systems, 58(12):1238?1245, 2010. ISSN0921-8890. doi:10.1016/j.robot.2010.08.007. URLhttp://dx.doi.org/10.1016/j.robot.2010.08.007. ? pages 6, 10, 13, 15[91] N. F. Ince, A. H. Tewfik, and S. Arica. Extraction subject-specific motor imagerytime-frequency patterns for single trial EEG classification. Computers in Biology andMedicine, 37(4):499?508, 2007. ISSN 0010-4825.doi:10.1016/j.compbiomed.2006.08.014. URLhttp://dx.doi.org/10.1016/j.compbiomed.2006.08.014. ? pages[92] N. F. Ince, F. Goksu, A. H. Tewfik, and S. Arica. Adapting subject specific motor imageryEEG patterns in space-time-frequency for a brain computer interface. Biomedical SignalProcessing and Control, 4(3):236?246, 2009. ISSN 1746-8094.doi:10.1016/j.bspc.2009.03.005. URL http://dx.doi.org/10.1016/j.bspc.2009.03.005. ?pages 6, 10, 15[93] Y. Ito, C. Srinivasan, and H. Izumi. Discriminant analysis by a neural network withMahalanobis distance. In S. Kollias, A. Stafylopatis, W. Duch, and E. Oja, editors,Artificial Neural Networks - ICANN 2006, volume 4132 of Lecture Notes in ComputerScience, pages 350?360. Springer Berlin / Heidelberg, 2006. ISBN 978-3-540-38871-5.doi:10.1007/11840930_36. URL http://dx.doi.org/10.1007/11840930_36. ? pages 30[94] Z. Jiang, Y. Ning, B. An, A. Li, and H. Feng. Detecting mental EEG properties usingdetrended fluctuation analysis. In Proceedings of the 27th Annual InternationalConference of the IEEE Engineering in Medicine and Biology Society, pages 2017?2020,September 2005. doi:10.1109/IEMBS.2005.1616852. URLhttp://dx.doi.org/10.1109/IEMBS.2005.1616852. ? pages 18[95] V. Jurcak, D. Tsuzuki, and I. Dan. 10/20, 10/10, and 10/5 systems revisited: Their validityas relative head-surface-based positioning systems. NeuroImage, 34(4):1600?1611, 2007.ISSN 1053-8119. doi:10.1016/j.neuroimage.2006.09.024. URLhttp://dx.doi.org/10.1016/j.neuroimage.2006.09.024. ? pages 49[96] Z. A. Keirn and J. I. Aunon. A new mode of communication between man and hissurroundings. IEEE Transactions on Biomedical Engineering, 37(12):1209?1214,December 1990. ISSN 0018-9294. doi:10.1109/10.64464. URLhttp://dx.doi.org/10.1109/10.64464. ? pages 6, 9, 10, 15, 18, 20, 25, 26119Bibliography[97] Z. A. Keirn and J. I. Aunon. Man-machine communications through brain-waveprocessing. IEEE Engineering in Medicine and Biology Magazine, 9(1):55?57, March1990. ISSN 0739-5175. doi:10.1109/51.62907. URL http://dx.doi.org/10.1109/51.62907.? pages 6, 9, 18[98] R. Krepki, G. Curio, B. Blankertz, and K.-R. M?ller. Berlin Brain?ComputerInterface?The HCI communication channel for discovery. International Journal ofHuman-Computer Studies, 65(5):460?477, 2007. ISSN 1071-5819.doi:10.1016/j.ijhcs.2006.11.010. URL http://dx.doi.org/10.1016/j.ijhcs.2006.11.010. ?pages 6, 15[99] J. Kronegg, G. Chanel, S. Voloshynovskiy, and T. Pun. EEG-based synchronizedbrain?computer interfaces: A model for optimizing the number of mental tasks. IEEETransactions on Neural Systems and Rehabilitation Engineering, 15(1):50?58, 2007. ISSN1534-4320. doi:10.1109/TNSRE.2007.891389. URLhttp://dx.doi.org/10.1109/TNSRE.2007.891389. ? pages 6, 8, 10[100] D. J. Krusienski, D. J. McFarland, and J. R. Wolpaw. An evaluation of autoregressivespectral estimation model order for brain?computer interface applications. In 28th AnnualInternational Conference of the IEEE Engineering in Medicine and Biology Society(EMBS), pages 1323?1326, 2006. doi:10.1109/IEMBS.2006.259822. URLhttp://dx.doi.org/10.1109/IEMBS.2006.259822. ? pages 83[101] T. N. Lal, M. Schr?der, T. Hinterberger, J. Weston, M. Bogdan, N. Birbaumer, andB. Sch?lkopf. Support vector channel selection in BCI. IEEE Transactions on BiomedicalEngineering, 51(6):1003?1010, 2004. ISSN 0018-9294. doi:10.1109/TBME.2004.827827.URL http://dx.doi.org/10.1109/TBME.2004.827827. ? pages 6, 10, 15[102] H. Lee, A. Cichocki, and S. Choi. Kernel nonnegative matrix factorization for spectralEEG feature extraction. Neurocomputing, 72(13-15):3182?3190, 2009. ISSN 0925-2312.doi:10.1016/j.neucom.2009.03.005. URLhttp://dx.doi.org/10.1016/j.neucom.2009.03.005. ? pages 6, 7[103] R. Leeb, F. Lee, C. Keinrath, R. Scherer, H. Bischof, and G. Pfurtscheller. Brain?computercommunication: Motivation, aim, and impact of exploring a virtual apartment. IEEETransactions on Neural Systems and Rehabilitation Engineering, 15(4):473?482, 2007.ISSN 1534-4320. doi:10.1109/TNSRE.2007.906956. URLhttp://dx.doi.org/10.1109/TNSRE.2007.906956. ? pages 6, 15[104] X. Lei, P. Yang, and D. Yao. An empirical Bayesian framework for brain?computerinterfaces. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 17(6):521?529, December 2009. ISSN 1534-4320. doi:10.1109/TNSRE.2009.2027705. URLhttp://dx.doi.org/10.1109/TNSRE.2009.2027705. ? pages120Bibliography[105] Y. Li, J. Long, T. Yu, Z. Yu, C. Wang, H. Zhang, and C. Guan. An EEG-based BCI systemfor 2-D cursor control by combining Mu/Beta rhythm and P300 potential. IEEETransactions on Biomedical Engineering, 57(10):2495?2505, October 2010. ISSN0018-9294. doi:10.1109/TBME.2010.2055564. URLhttp://dx.doi.org/10.1109/TBME.2010.2055564. ? pages 6, 15[106] C.-J. Lin and M.-H. Hsieh. Classification of mental task from EEG data using neuralnetworks based on particle swarm optimization. Neurocomputing, 72(4-6):1121?1130,2009. ISSN 0925-2312. doi:10.1016/j.neucom.2008.02.017. URLhttp://dx.doi.org/10.1016/j.neucom.2008.02.017. ? pages 6, 7, 10[107] D. Liu, Z. Jian, and H. Feng. Separating the different components of spontaneous EEG byoptimized ICA. In Proceedings of the International Conference on Neural Networks andSignal Processing, volume 2, pages 1334?1337, December 2003.doi:10.1109/ICNNSP.2003.1281118. URLhttp://dx.doi.org/10.1109/ICNNSP.2003.1281118. ? pages 18, 26[108] D. Liu, Z. Jiang, W. Cong, and H. Feng. Detect determinism of spontaneous EEG with amulti-channel reconstruction method. In Proceedings of the International Conference onNeural Networks and Signal Processing, volume 1, pages 708?711, December 2003.doi:10.1109/ICNNSP.2003.1279373. URLhttp://dx.doi.org/10.1109/ICNNSP.2003.1279373. ? pages 18[109] G. Liu, G. Huang, J. Meng, and X. Zhu. A frequency-weighted method combined withCommon Spatial Patterns for electroencephalogram classification in brain?computerinterface. Biomedical Signal Processing and Control, 5(2):174?180, 2010. ISSN1746-8094. doi:10.1016/j.bspc.2010.02.004. URLhttp://dx.doi.org/10.1016/j.bspc.2010.02.004. ? pages 6, 10, 15[110] H. Liu, J. Wang, and C. Zheng. Mental tasks classification and their eeg structures analysisby using the growing hierarchical self-organizing map. In Proceedings of the 1stInternational Conference on Neural Interface and Control, pages 115?118, May 2005.doi:10.1109/ICNIC.2005.1499856. URL http://dx.doi.org/10.1109/ICNIC.2005.1499856.? pages 18[111] H. Liu, J. Wang, C. Zheng, and P. He. Study on the effect of different frequency bands ofEEG signals on mental tasks classification. In Proceedings of the 27th AnnualInternational Conference of the IEEE Engineering in Medicine and Biology Society, pages5369?5372, September 2005. doi:10.1109/IEMBS.2005.1615695. URLhttp://dx.doi.org/10.1109/IEMBS.2005.1615695. ? pages[112] V. A. Maiorescu, M. Serban, and A. M. Lazar. Classification of EEG signals representedby AR models for cognitive tasks - a neural network based method. In Proceedings of theInternational Symposium on Signals, Circuits and Systems, volume 2, pages 441?444, July121Bibliography2003. doi:10.1109/SCS.2003.1227084. URLhttp://dx.doi.org/10.1109/SCS.2003.1227084. ? pages 18[113] J. N. Mak and J. R. Wolpaw. Clinical applications of brain?computer interfaces: Currentstate and future prospects. IEEE Reviews in Biomedical Engineering, 2:187?199, 2009.ISSN 1937-3333. doi:10.1109/RBME.2009.2035356. URLhttp://dx.doi.org/10.1109/RBME.2009.2035356. ? pages 5[114] S. G. Mason and G. E. Birch. A brain-controlled switch for asynchronous controlapplications. IEEE Transactions on Biomedical Engineering, 47(10):1297?1307, 2000.ISSN 0018-9294. doi:10.1109/10.871402. URL http://dx.doi.org/10.1109/10.871402. ?pages 15[115] S. G. Mason and G. E. Birch. A general framework for brain?computer interface design.IEEE Transactions on Neural Systems and Rehabilitation Engineering, 11(1):70?85, 2003.ISSN 1534-4320. doi:10.1109/TNSRE.2003.810426. URLhttp://dx.doi.org/10.1109/TNSRE.2003.810426. ? pages[116] S. G. Mason, A. Bashashati, M. Fatourechi, K. F. Navarro, and G. E. Birch. Acomprehensive survey of brain interface technology designs. Annals of BiomedicalEngineering, 35(2):137?169, February 2007. doi:10.1007/s10439-006-9170-0. URLhttp://dx.doi.org/10.1007/s10439-006-9170-0. ? pages 5[117] F. J. Massey. The Kolmogorov-Smirnov test for goodness of fit. Journal of the AmericanStatistical Association, 46(253):68?78, 1951. ISSN 0162-1459. doi:10.2307/2280095.URL http://www.jstor.org/stable/2280095. ? pages 54[118] R. G. d. P. Menendez, Q. Noirhomme, F. Cincotti, D. Mattia, F. Aloise, andS. Gonz?lez Andino. Modern electrophysiological methods for brain?computer interfaces.Computational Intelligence and Neuroscience, 2007:1?8, 2007. ISSN 1687-5265.doi:10.1155/2007/56986. URL http://dx.doi.org/10.1155/2007/56986. ? pages 15[119] K.-R. M?ller, M. Tangermann, G. Dornhege, M. Krauledat, G. Curio, and B. Blankertz.Machine learning for real-time single-trial EEG-analysis: From brain?computer interfacingto mental state monitoring. Journal of Neuroscience Methods, 167(1):82?90, 2008. ISSN0165-0270. doi:10.1016/j.jneumeth.2007.09.022. URLhttp://dx.doi.org/10.1016/j.jneumeth.2007.09.022. ? pages 6, 10, 15[120] G. R. M?ller-Putz, R. Scherer, G. Pfurtscheller, and R. Rupp. EEG-based neuroprosthesiscontrol: A step towards clinical practice. Neuroscience Letters, 382(1-2):169?174, 2005.ISSN 0304-3940. doi:10.1016/j.neulet.2005.03.021. URLhttp://dx.doi.org/10.1016/j.neulet.2005.03.021. ? pages 6, 10, 13, 15[121] K. Nakayama, Y. Kaneda, and A. Hirano. A brain computer interface based on FFT andmultilayer neural network - feature extraction and generalization. In Proceedings of the122BibliographyInternational Symposium on Intelligent Signal Processing and Communication Systems,pages 826?829, November 2007. doi:10.1109/ISPACS.2007.4446015. URLhttp://dx.doi.org/10.1109/ISPACS.2007.4446015. ? pages 18[122] K. Nakayaman and K. Inagaki. A brain computer interface based on neural network withefficient pre-processing. In Proceedings of the International Symposium on IntelligentSignal Processing and Communications, pages 673?676, December 2006.doi:10.1109/ISPACS.2006.364745. URL http://dx.doi.org/10.1109/ISPACS.2006.364745.? pages 18[123] S. V. Narasimhan and S. Veena. Signal processing: principles and implementation. AlphaScience International, 2005. ISBN 9781842651995. ? pages 26[124] B. Nasihatkon, R. Boostani, and M. Z. Jahromi. An efficient hybrid linear and kernel CSPapproach for EEG feature extraction. Neurocomputing, 73(1-3):432?437, 2009. ISSN0925-2312. doi:10.1016/j.neucom.2009.07.012. URLhttp://dx.doi.org/10.1016/j.neucom.2009.07.012. ? pages 6, 10, 15[125] K. Nazarpour, P. Praamstra, R. C. Miall, and S. Sanei. Steady-state movement relatedpotentials for brain?computer interfacing. IEEE Transactions on Biomedical Engineering,56(8):2104?2113, August 2009. ISSN 0018-9294. doi:10.1109/TBME.2009.2021529.URL http://dx.doi.org/10.1109/TBME.2009.2021529. ? pages[126] C. Neuper, G. R. M?ller, A. K?bler, N. Birbaumer, and G. Pfurtscheller. Clinicalapplication of an EEG-based brain?computer interface: a case study in a patient withsevere motor impairment. Clinical Neurophysiology, 114(3):399?409, 2003. ISSN1388-2457. doi:10.1016/S1388-2457(02)00387-5. URLhttp://dx.doi.org/10.1016/S1388-2457(02)00387-5. ? pages 6, 10, 15[127] M. A. L. Nicolelis. Brain?machine interfaces to restore motor function and probe neuralcircuits. Nature Reviews Neuroscience, 4(5):417?422, 2003. ISSN 1471-003X.doi:10.1038/nrn1105. URL http://dx.doi.org/10.1038/nrn1105. ? pages 5[128] B. Obermaier, C. Guger, C. Neuper, and G. Pfurtscheller. Hidden Markov models foronline classification of single trial EEG data. Pattern Recognition Letters, 22(12):1299?1309, 2001. ISSN 0167-8655. doi:10.1016/S0167-8655(01)00075-7. URLhttp://dx.doi.org/10.1016/S0167-8655(01)00075-7. ? pages 6, 10, 15[129] B. Obermaier, C. Neuper, C. Guger, and G. Pfurtscheller. Information transfer rate in afive-classes brain?computer interface. IEEE Transactions on Neural Systems andRehabilitation Engineering, 9(3):283?288, 2001. ISSN 1534-4320.doi:10.1109/7333.948456. URL http://dx.doi.org/10.1109/7333.948456. ? pages 6, 8, 10[130] L. F. Oliveira, D. M. Simpson, and J. Nadal. Autoregressive spectral analysis ofstabilometric signals. In 16th Annual International Conference of the IEEE Engineering in123BibliographyMedicine and Biology Society, Engineering Advances: New Opportunities for BiomedicalEngineers, pages 1300?1301, 1994. doi:10.1109/IEMBS.1994.415442. URLhttp://dx.doi.org/10.1109/IEMBS.1994.415442. ? pages 83[131] R. Oostenveld and P. Praamstra. The five percent electrode system for high-resolution EEGand ERP measurements. Clinical Neurophysiology, 112(4):713?719, 2001. ISSN1388-2457. doi:10.1016/S1388-2457(00)00527-7. URLhttp://dx.doi.org/10.1016/S1388-2457(00)00527-7. ? pages 49[132] R. Palaniappan. Identifying individuality using mental task based brain computer interface.In Proceedings of the 3rd International Conference on Intelligent Sensing and InformationProcessing, pages 239?242, December 2005. doi:10.1109/ICISIP.2005.1619442. URLhttp://dx.doi.org/10.1109/ICISIP.2005.1619442. ? pages 18, 26[133] R. Palaniappan. Brain computer interface design using band powers extracted duringmental tasks. In Proceedings of the 2nd International IEEE EMBS Conference on NeuralEngineering, pages 321?324, March 2005. doi:10.1109/CNE.2005.1419622. URLhttp://dx.doi.org/10.1109/CNE.2005.1419622. ? pages 47[134] R. Palaniappan. Utilizing Gamma band to improve mental task based brain-computerinterface design. IEEE Transactions on Neural Systems and Rehabilitation Engineering,14(3):299?303, September 2006. ISSN 1534-4320. doi:10.1109/TNSRE.2006.881539.URL http://dx.doi.org/10.1109/TNSRE.2006.881539. ? pages 6, 9, 10, 47[135] R. Palaniappan and N. J. Huan. Improving the performance of two-state mental taskbrain-computer interface design using linear discriminant classifier. In Proceedings of theInternational Conference on Computer as a Tool, volume 1, pages 409?412, November2005. doi:10.1109/EURCON.2005.1629949. URLhttp://dx.doi.org/10.1109/EURCON.2005.1629949. ? pages 26[136] R. Palaniappan and P. Raveendran. A new mode of EEG based communication. InProceedings of the International Joint Conference on Neural Networks, volume 4, pages2679?2682, July 2001. doi:10.1109/IJCNN.2001.938793. URLhttp://dx.doi.org/10.1109/IJCNN.2001.938793. ? pages[137] R. Palaniappan, P. Raveendran, S. Nishida, and N. Saiwaki. Autoregressive spectralanalysis and model order selection criteria for EEG signals. In Proceedings of the IEEERegion 10 Technical Conference (TENCON 2000), volume 2, pages 126?129, September2000. doi:10.1109/TENCON.2000.888403. URLhttp://dx.doi.org/10.1109/TENCON.2000.888403. ? pages 25[138] R. Palaniappan, P. Raveendran, S. Nishida, and N. Saiwaki. Evolutionary fuzzy ARTMAPfor autoregressive model order selection and classification of EEG signals. In Proceedingsof the IEEE International Conference on Systems, Man, and Cybernetics, volume 5, pages124Bibliography3682?3686, October 2000. doi:10.1109/ICSMC.2000.886582. URLhttp://dx.doi.org/10.1109/ICSMC.2000.886582. ? pages[139] R. Palaniappan, P. Raveendran, S. Nishida, and N. Saiwaki. Fuzzy ATRMAP classificationof mental tasks using segmented and overlapped EEG signals. In Proceedings of the IEEERegion 10 Technical Conference (TENCON 2000), volume 2, pages 388?391, September2000. doi:10.1109/TENCON.2000.888768. URLhttp://dx.doi.org/10.1109/TENCON.2000.888768. ? pages 26[140] R. Palaniappan, R. Paramesran, S. Nishida, and N. Saiwaki. A new brain?computerinterface design using fuzzy ARTMAP. IEEE Transactions on Neural Systems andRehabilitation Engineering, 10(3):140?148, September 2002. ISSN 1534-4320.doi:10.1109/TNSRE.2002.802854. URL http://dx.doi.org/10.1109/TNSRE.2002.802854.? pages 6, 9, 10, 18, 26, 47[141] L. Parra, C. Alvino, A. Tang, B. P. N. Yeung, A. Osman, and P. Sajda. Linear spatialintegration for single-trial detection in encephalography. NeuroImage, 17(1):223?230,2002. ISSN 1053-8119. doi:10.1006/nimg.2002.1212. URLhttp://dx.doi.org/10.1006/nimg.2002.1212. ? pages 6, 10, 15[142] M. P. Paulraj, C. R. Hema, R. Nagarajan, S. Yaacob, and A. H. Adorn. EEG classificationusing radial basis PSO neural network for brain machine interfaces. In Proceedings of the5th Student Conference on Research and Development, pages 1?5, December 2007.doi:10.1109/SCORED.2007.4451377. URLhttp://dx.doi.org/10.1109/SCORED.2007.4451377. ? pages 18[143] W. D. Penny, S. J. Roberts, E. A. Curran, and M. J. Stokes. EEG-based communication: Apattern recognition approach. IEEE Transactions on Rehabilitation Engineering, 8(2):214?215, 2000. ISSN 1063-6528. doi:10.1109/86.847820. URLhttp://dx.doi.org/10.1109/86.847820. ? pages 6, 9, 10[144] G. Pfurtscheller and C. Neuper. Motor imagery and direct brain?computer communication.Proceedings of the IEEE, 89(7):1123?1134, 2001. ISSN 0018-9219.doi:10.1109/5.939829. URL http://dx.doi.org/10.1109/5.939829. ? pages 6, 15[145] G. Pfurtscheller, C. Neuper, G. R. M?ller, B. Obermaier, G. Krausz, A. Schl?gl, R. Scherer,B. Graimann, C. Keinrath, D. Skliris, M. W?rtz, G. Supp, and C. Schrank. Graz-BCI: stateof the art and clinical applications. IEEE Transactions on Neural Systems andRehabilitation Engineering, 11(2):177?180, 2003. ISSN 1534-4320.doi:10.1109/TNSRE.2003.814454. URL http://dx.doi.org/10.1109/TNSRE.2003.814454.? pages 7[146] G. Pfurtscheller, C. Brunner, A. Schl?gl, and F. H. L. da Silva. Mu rhythm(de)synchronization and EEG single-trial classification of different motor imagery tasks.125BibliographyNeuroImage, 31(1):153?159, 2006. ISSN 1053-8119.doi:10.1016/j.neuroimage.2005.12.003. URLhttp://dx.doi.org/10.1016/j.neuroimage.2005.12.003. ? pages[147] G. Pfurtscheller, R. Leeb, C. Keinrath, D. Friedman, C. Neuper, C. Guger, and M. Slater.Walking from thought. Brain Research, 1071(1):145?152, 2006. ISSN 0006-8993.doi:10.1016/j.brainres.2005.11.083. URLhttp://dx.doi.org/10.1016/j.brainres.2005.11.083. ? pages[148] G. Pfurtscheller, T. Solis-Escalante, R. Ortner, P. Linortner, and G. M?ller-Putz. Self-pacedoperation of an SSVEP-based orthosis with and without an imagery-based ?brain switch:"A feasibility study towards a hybrid BCI. IEEE Transactions on Neural Systems andRehabilitation Engineering, 18(4):409?414, August 2010. ISSN 1534-4320.doi:10.1109/TNSRE.2010.2040837. URLhttp://dx.doi.org/10.1109/TNSRE.2010.2040837. ? pages 10, 13, 15[149] K. Qian, P. Nikolov, D. Huang, D.-Y. Fei, X. Chen, and O. Bai. A motor imagery-basedonline interactive brain-controlled switch: Paradigm development and preliminary test.Clinical Neurophysiology, 121(8):1304?1313, 2010. ISSN 1388-2457.doi:10.1016/j.clinph.2010.03.001. URL http://dx.doi.org/10.1016/j.clinph.2010.03.001. ?pages[150] H. Ramoser, J. M?ller-Gerking, and G. Pfurtscheller. Optimal spatial filtering of singletrial EEG during imagined hand movement. IEEE Transactions on RehabilitationEngineering, 8(4):441?446, 2000. ISSN 1063-6528. doi:10.1109/86.895946. URLhttp://dx.doi.org/10.1109/86.895946. ? pages 6, 10, 15[151] R. Rao and R. Derakhshani. A comparison of EEG preprocessing methods using timedelay neural networks. In Proceedings of the 2nd International IEEE EMBS Conference onNeural Engineering, pages 262?264, March 2005. doi:10.1109/CNE.2005.1419607. URLhttp://dx.doi.org/10.1109/CNE.2005.1419607. ? pages 18[152] S. Rezaei, K. Tavakolian, and K. Naziripour. Comparison of five different classifiers forclassification of mental tasks. In Proceedings of the 27th Annual International Conferenceof the IEEE Engineering in Medicine and Biology Society, pages 6007?6010, September2005. doi:10.1109/IEMBS.2005.1615860. URLhttp://dx.doi.org/10.1109/IEMBS.2005.1615860. ? pages 18, 26[153] R. Ron-Angevin and A. D?az-Estrella. Brain?computer interface: Changes in performanceusing virtual reality techniques. Neuroscience Letters, 449(2):123?127, 2009. ISSN0304-3940. doi:10.1016/j.neulet.2008.10.099. URLhttp://dx.doi.org/10.1016/j.neulet.2008.10.099. ? pages 6, 10, 15[154] C. Sannelli, M. Braun, and K.-R. M?ller. Improving BCI performance by task-related trialpruning. Neural Networks, 22(9):1295?1304, 2009. ISSN 0893-6080.126Bibliographydoi:10.1016/j.neunet.2009.08.006. URL http://dx.doi.org/10.1016/j.neunet.2009.08.006.? pages[155] R. Scherer, F. Lee, A. Schl?gl, R. Leeb, H. Bischof, and G. Pfurtscheller. Towardself-paced brain?computer communication: Navigation through virtual worlds. IEEETransactions on Biomedical Engineering, 55(2):675?682, 2008. ISSN 0018-9294.doi:10.1109/TBME.2007.903709. URL http://dx.doi.org/10.1109/TBME.2007.903709. ?pages 6, 10, 13, 15[156] F. Sepulveda, M. Dyson, J. Q. Gan, and C. S. L. Tsui. A comparison of mental taskcombinations for asynchronous EEG-based BCIs. In Proceedings of the 29th AnnualInternational Conference of the IEEE Engineering in Medicine and Biology Society, pages5055?5058, 2007. ISBN 1557-170X. doi:10.1109/IEMBS.2007.4353476. URLhttp://dx.doi.org/10.1109/IEMBS.2007.4353476. ? pages 6, 9, 10[157] M. C. Serban and D. M. Dobrea. Discrimination between cognitive tasks - a comparativestudy. In Proceedings of the International Symposium on Signals, Circuits and Systems,volume 2, pages 805?808, July 2005. doi:10.1109/ISSCS.2005.1511363. URLhttp://dx.doi.org/10.1109/ISSCS.2005.1511363. ? pages 18[158] D. M. Simpson, A. F. C. Infantosi, J. F. C. Junior, A. J. Peixoto, and L. M. d. S. Abrantes.On the selection of autoregressive order for electroencephalographic (EEG) signals. In38th Midwest Symposium on Circuits and Systems, volume 2, pages 1353?1356, 1995.doi:10.1109/MWSCAS.1995.510349. URLhttp://dx.doi.org/10.1109/MWSCAS.1995.510349. ? pages 83[159] B. T. Skinner, H. T. Nguyen, and D. K. Liu. Classification of EEG signals using agenetic-based machine learning classifier. In Proceedings of the 29th Annual InternationalConference of the IEEE Engineering in Medicine and Biology Society, pages 3120?3123,August 2007. doi:10.1109/IEMBS.2007.4352990. URLhttp://dx.doi.org/10.1109/IEMBS.2007.4352990. ? pages 18, 26[160] A. Spiegler, B. Graimann, and G. Pfurtscheller. Phase coupling between different motorareas during tongue-movement imagery. Neuroscience Letters, 369(1):50?54, 2004. ISSN0304-3940. doi:10.1016/j.neulet.2004.07.054. URLhttp://dx.doi.org/10.1016/j.neulet.2004.07.054. ? pages 6, 10, 15[161] S. Sun, C. Zhang, and D. Zhang. An experimental evaluation of ensemble methods forEEG signal classification. Pattern Recognition Letters, 28(15):2157?2163, 2007. ISSN0167-8655. doi:10.1016/j.patrec.2007.06.018. URLhttp://dx.doi.org/10.1016/j.patrec.2007.06.018. ? pages 6, 7[162] S. Sun, C. Zhang, and Y. Lu. The random electrode selection ensemble for EEG signalclassification. Pattern Recognition, 41(5):1663?1675, 2008. ISSN 0031-3203.127Bibliographydoi:10.1016/j.patcog.2007.10.023. URL http://dx.doi.org/10.1016/j.patcog.2007.10.023.? pages 6, 7, 10[163] K. Tavakolian and S. Rezaei. Classification of mental tasks using Gaussian mixtureBayesian network classifiers. In Proceedings of the IEEE International Workshop onBiomedical Circuits and Systems, pages S3.6?9?11, December 2004.doi:10.1109/BIOCAS.2004.1454169. URLhttp://dx.doi.org/10.1109/BIOCAS.2004.1454169. ? pages 18, 26, 47[164] K. Tavakolian, S. Rezaei, and S. K. Setarehdan. Choosing optimal mental tasks forclassification in brain computer interfaces. In Proceedings of the 7th IASTED InternationalConference on Artificial Intelligence and Applications, pages 396?399, February 2004.URL http://www.actapress.com/Abstract.aspx?paperId=15817. ? pages 26[165] M. N. Teli and C. Anderson. Nonlinear dimensionality reduction of electroencephalogram(EEG) for brain computer interfaces. In Annual International Conference of the IEEEEngineering in Medicine and Biology Society (EMBC), pages 2486?2489, 2009.doi:10.1109/IEMBS.2009.5334802. URL http://dx.doi.org/10.1109/IEMBS.2009.5334802.? pages 6, 9, 18[166] G. Townsend and Y. Feng. Using phase information to reveal the nature of event-relateddesynchronization. Biomedical Signal Processing and Control, 3(3):192?202, 2008. ISSN1746-8094. doi:10.1016/j.bspc.2008.01.003. URLhttp://dx.doi.org/10.1016/j.bspc.2008.01.003. ? pages 6, 10, 15[167] G. Townsend, B. Graimann, and G. Pfurtscheller. Continuous EEG classification duringmotor imagery-simulation of an asynchronous BCI. IEEE Transactions on Neural Systemsand Rehabilitation Engineering, 12(2):258?265, 2004. ISSN 1534-4320.doi:10.1109/TNSRE.2004.827220. URL http://dx.doi.org/10.1109/TNSRE.2004.827220.? pages 10, 13, 15[168] C. S. L. Tsui and J. Q. Gan. Asynchronous BCI control of a robot simulator withsupervised online training. In H. Yin, P. Tino, E. Corchado, W. Byrne, and X. Yao, editors,Intelligent Data Engineering and Automated Learning - IDEAL 2007, volume 4881 ofLecture Notes in Computer Science, pages 125?134. Springer Berlin / Heidelberg, 2007.doi:10.1007/978-3-540-77226-2\_14. URLhttp://dx.doi.org/10.1007/978-3-540-77226-2_14. ? pages 10, 13[169] C. S. L. Tsui, J. Q. Gan, and S. J. Roberts. A self-paced brain?computer interface forcontrolling a robot simulator: An online event labelling paradigm and an extended Kalmanfilter based algorithm for online training. Medical and Biological Engineering andComputing, 47:257?265, 2009. ISSN 0140-0118. doi:10.1007/s11517-009-0459-7. URLhttp://dx.doi.org/10.1007/s11517-009-0459-7. ? pages 6, 10, 13, 15128Bibliography[170] T. M. Vaughan. Guest editorial brain?computer interface technology: A review of thesecond international meeting. IEEE Transactions on Neural Systems and RehabilitationEngineering, 11(2):94?109, June 2003. ISSN 1534-4320.doi:10.1109/TNSRE.2003.814799. URL http://dx.doi.org/10.1109/TNSRE.2003.814799.? pages 5[171] T. M. Vaughan, J. R. Wolpaw, and E. Donchin. EEG-based communication: prospects andproblems. IEEE Transactions on Rehabilitation Engineering, 4(4):425?430, 1996. ISSN1063-6528. doi:10.1109/86.547945. URL http://dx.doi.org/10.1109/86.547945. ? pages 1[172] C. Vidaurre, N. Kr?mer, B. Blankertz, and A. Schl?gl. Time domain parameters as afeature for EEG-based brain?computer interfaces. Neural Networks, 22(9):1313?1319,2009. ISSN 0893-6080. doi:10.1016/j.neunet.2009.07.020. URLhttp://dx.doi.org/10.1016/j.neunet.2009.07.020. ? pages 6, 10, 15[173] A. Vuckovic and F. Sepulveda. Quantification and visualisation of differences between twomotor tasks based on energy density maps for brain?computer interface applications.Clinical Neurophysiology, 119(2):446?458, 2008. ISSN 1388-2457.doi:10.1016/j.clinph.2007.10.015. URL http://dx.doi.org/10.1016/j.clinph.2007.10.015. ?pages[174] T. Wang, J. Deng, and B. He. Classifying EEG-based motor imagery tasks by means oftime-frequency synthesized spatial patterns. Clinical Neurophysiology, 115(12):2744?2753, 2004. ISSN 1388-2457. doi:10.1016/j.clinph.2004.06.022. URLhttp://dx.doi.org/10.1016/j.clinph.2004.06.022. ? pages 6, 10, 15[175] Y. Wang, Z. Zhang, Y. Li, X. Gao, S. Gao, and F. Yang. BCI competition 2003-data set IV:An algorithm based on CSSD and FDA for classifying single-trial EEG. IEEETransactions on Biomedical Engineering, 51(6):1081?1086, 2004. ISSN 0018-9294.doi:10.1109/TBME.2004.826697. URL http://dx.doi.org/10.1109/TBME.2004.826697. ?pages 15[176] B. L. Welch. The generalization of ?student?s? problem when several different populationvariances are involved. Biometrika, 34(1-2):28?35, 1947. doi:10.1093/biomet/34.1-2.28.URL http://dx.doi.org/10.1093/biomet/34.1-2.28. ? pages 54, 77, 92[177] J. R. Wolpaw. Brain?computer interfaces (BCIs) for communication and control: amini-review. Supplements to Clinical Neurophysiology, 57:607?613, 2004. ISSN1567-424X. doi:10.1016/S1567-424X(09)70400-3. URLhttp://dx.doi.org/10.1016/S1567-424X(09)70400-3. ? pages 5[178] J. R. Wolpaw, N. Birbaumer, W. J. Heetderks, D. J. McFarland, P. H. Peckham, G. Schalk,E. Donchin, L. A. Quatrano, C. J. Robinson, and T. M. Vaughan. Brain?computer interfacetechnology: a review of the first international meeting. IEEE Transactions on129BibliographyRehabilitation Engineering, 8(2):164?173, 2000. ISSN 1063-6528.doi:10.1109/TRE.2000.847807. URL http://dx.doi.org/10.1109/TRE.2000.847807. ?pages 1, 2[179] J. R. Wolpaw, N. Birbaumer, D. J. McFarland, G. Pfurtscheller, and T. M. Vaughan.Brain?computer interfaces for communication and control. Clinical Neurophysiology, 113(6):767?791, 2002. ISSN 1388-2457. doi:10.1016/S1388-2457(02)00057-3. URLhttp://dx.doi.org/10.1016/S1388-2457(02)00057-3. ? pages 1, 2, 5[180] X. Wu and X. Guo. Mental EEG analysis based on independent component analysis. InProceedings of the 3rd International Symposium on Image and Signal Processing andAnalysis, volume 1, pages 327?331, September 2003. doi:10.1109/ISPA.2003.1296917.URL http://dx.doi.org/10.1109/ISPA.2003.1296917. ? pages 18[181] Q. Xu, H. Zhou, Y. Wang, and J. Huang. Fuzzy support vector machine for classification ofEEG signals using wavelet-based features. Medical Engineering & Physics, 31(7):858?865, 2009. ISSN 1350-4533. doi:10.1016/j.medengphy.2009.04.005. URLhttp://dx.doi.org/10.1016/j.medengphy.2009.04.005. ? pages 6, 10, 15[182] J. Z. Xue, H. Zhang, C. X. Zheng, and X. G. Yan. Wavelet packet transform for featureextraction of EEG during mental tasks. In Proceedings of the International Conference onMachine Learning and Cybernetics, volume 1, pages 360?363, November 2003.doi:10.1109/ICMLC.2003.1264502. URL http://dx.doi.org/10.1109/ICMLC.2003.1264502.? pages 18[183] G. Yan, X. Guo, R. Yan, and B. Yang. Nonlinear quadratic phase coupling on EEG basedon 11/2-dimension spectrum. In Proceedings of the IET 3rd International Conference onAdvances in Medical, Signal and Information Processing, pages 1?4, July 2006. URLhttp://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4225248. ? pages[184] L. Zhiwei and S. Minfen. Classification of mental task EEG signals using wavelet packetentropy and SVM. In Proceedings of the 8th International Conference on ElectronicMeasurement and Instruments, pages 3?906?909, July 2007.doi:10.1109/ICEMI.2007.4351064. URL http://dx.doi.org/10.1109/ICEMI.2007.4351064.? pages 18[185] S.-M. Zhou, J. Q. Gan, and F. Sepulveda. Classifying mental tasks based on features ofhigher-order statistics from EEG signals in brain?computer interface. InformationSciences, 178(6):1629?1640, 2008. ISSN 0020-0255. doi:10.1016/j.ins.2007.11.012. URLhttp://dx.doi.org/10.1016/j.ins.2007.11.012. ? pages 6, 10, 15130

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0072796/manifest

Comment

Related Items