On feature selection for brain computer interfaces

H. Lakany

    Research output: Contribution to journalArticle

    Abstract

    Currently, Brain Computer Interfaces (BCI) are of limited practical use requiring prolonged user training. BCI usability can improve by developing methods for EEG feature selection that allow for a more intuitive link between user intention and device control and which can provide multiple degrees of freedom in command selection. EEG was recorded from 28 electrodes forming an array over the left sensorimotor cortex in normal subjects with approval of a Local Ethics Committee. Right wrist muscle EMG was also recorded. Subjects viewed a PC monitor randomly presenting a set of five spatially distinct target positions corresponding to neutral, flexed, extended, ulnar and radial deviations of the wrist. When presented with a new target subjects were required to move (or imagine move) their wrist to the target as fast as possible creating datasets of repeats of 20 different movement (or imagined movement) combinations. Continuous wavelet transform was used to extract spatio-temporal characteristics of the EEG signals relative to time of movement initiation or target appearance in case of imagined motion. Using PCA, 10 principal components generated 95% correct classification of pre-movement EEG with actual direction of motion. During imagined movement classification success reduced to 67%. In contrast, SVM classification of the imagined movement data outperformed PCA with up to 87% correct classification for intended direction of movement. During imagined movement SVM outperforms PCA. However, SVM is computationally demanding. We conclude that a hybrid method combining merits of both PCA and SVM may have potential for use in BCI intention detection.
    LanguageEnglish
    JournalClinical Neurophysiology
    Volume117
    Issue numberS71
    DOIs
    Publication statusPublished - 2006

    Fingerprint

    Brain-Computer Interfaces
    Passive Cutaneous Anaphylaxis
    Electroencephalography
    Wrist
    Wavelet Analysis
    Ethics Committees
    Electrodes
    Equipment and Supplies
    Muscles
    Direction compound

    Keywords

    • feature selection
    • brain computer interfaces
    • bioengineering
    • neurophysiology

    Cite this

    @article{abcc5d1621484201a5f642aa558dcbf8,
    title = "On feature selection for brain computer interfaces",
    abstract = "Currently, Brain Computer Interfaces (BCI) are of limited practical use requiring prolonged user training. BCI usability can improve by developing methods for EEG feature selection that allow for a more intuitive link between user intention and device control and which can provide multiple degrees of freedom in command selection. EEG was recorded from 28 electrodes forming an array over the left sensorimotor cortex in normal subjects with approval of a Local Ethics Committee. Right wrist muscle EMG was also recorded. Subjects viewed a PC monitor randomly presenting a set of five spatially distinct target positions corresponding to neutral, flexed, extended, ulnar and radial deviations of the wrist. When presented with a new target subjects were required to move (or imagine move) their wrist to the target as fast as possible creating datasets of repeats of 20 different movement (or imagined movement) combinations. Continuous wavelet transform was used to extract spatio-temporal characteristics of the EEG signals relative to time of movement initiation or target appearance in case of imagined motion. Using PCA, 10 principal components generated 95{\%} correct classification of pre-movement EEG with actual direction of motion. During imagined movement classification success reduced to 67{\%}. In contrast, SVM classification of the imagined movement data outperformed PCA with up to 87{\%} correct classification for intended direction of movement. During imagined movement SVM outperforms PCA. However, SVM is computationally demanding. We conclude that a hybrid method combining merits of both PCA and SVM may have potential for use in BCI intention detection.",
    keywords = "feature selection, brain computer interfaces, bioengineering, neurophysiology",
    author = "H. Lakany",
    note = "Also published in Proceedings of the Third International Conference on Advances in Medical, Signal and Information Processing (IET/IEEE) MEDSIP2006",
    year = "2006",
    doi = "10.1016/j.clinph.2006.06.059",
    language = "English",
    volume = "117",
    journal = "Clinical Neurophysiology",
    issn = "1388-2457",
    number = "S71",

    }

    On feature selection for brain computer interfaces. / Lakany, H.

    In: Clinical Neurophysiology, Vol. 117, No. S71, 2006.

    Research output: Contribution to journalArticle

    TY - JOUR

    T1 - On feature selection for brain computer interfaces

    AU - Lakany, H.

    N1 - Also published in Proceedings of the Third International Conference on Advances in Medical, Signal and Information Processing (IET/IEEE) MEDSIP2006

    PY - 2006

    Y1 - 2006

    N2 - Currently, Brain Computer Interfaces (BCI) are of limited practical use requiring prolonged user training. BCI usability can improve by developing methods for EEG feature selection that allow for a more intuitive link between user intention and device control and which can provide multiple degrees of freedom in command selection. EEG was recorded from 28 electrodes forming an array over the left sensorimotor cortex in normal subjects with approval of a Local Ethics Committee. Right wrist muscle EMG was also recorded. Subjects viewed a PC monitor randomly presenting a set of five spatially distinct target positions corresponding to neutral, flexed, extended, ulnar and radial deviations of the wrist. When presented with a new target subjects were required to move (or imagine move) their wrist to the target as fast as possible creating datasets of repeats of 20 different movement (or imagined movement) combinations. Continuous wavelet transform was used to extract spatio-temporal characteristics of the EEG signals relative to time of movement initiation or target appearance in case of imagined motion. Using PCA, 10 principal components generated 95% correct classification of pre-movement EEG with actual direction of motion. During imagined movement classification success reduced to 67%. In contrast, SVM classification of the imagined movement data outperformed PCA with up to 87% correct classification for intended direction of movement. During imagined movement SVM outperforms PCA. However, SVM is computationally demanding. We conclude that a hybrid method combining merits of both PCA and SVM may have potential for use in BCI intention detection.

    AB - Currently, Brain Computer Interfaces (BCI) are of limited practical use requiring prolonged user training. BCI usability can improve by developing methods for EEG feature selection that allow for a more intuitive link between user intention and device control and which can provide multiple degrees of freedom in command selection. EEG was recorded from 28 electrodes forming an array over the left sensorimotor cortex in normal subjects with approval of a Local Ethics Committee. Right wrist muscle EMG was also recorded. Subjects viewed a PC monitor randomly presenting a set of five spatially distinct target positions corresponding to neutral, flexed, extended, ulnar and radial deviations of the wrist. When presented with a new target subjects were required to move (or imagine move) their wrist to the target as fast as possible creating datasets of repeats of 20 different movement (or imagined movement) combinations. Continuous wavelet transform was used to extract spatio-temporal characteristics of the EEG signals relative to time of movement initiation or target appearance in case of imagined motion. Using PCA, 10 principal components generated 95% correct classification of pre-movement EEG with actual direction of motion. During imagined movement classification success reduced to 67%. In contrast, SVM classification of the imagined movement data outperformed PCA with up to 87% correct classification for intended direction of movement. During imagined movement SVM outperforms PCA. However, SVM is computationally demanding. We conclude that a hybrid method combining merits of both PCA and SVM may have potential for use in BCI intention detection.

    KW - feature selection

    KW - brain computer interfaces

    KW - bioengineering

    KW - neurophysiology

    UR - http://dx.doi.org/10.1016/j.clinph.2006.06.059

    U2 - 10.1016/j.clinph.2006.06.059

    DO - 10.1016/j.clinph.2006.06.059

    M3 - Article

    VL - 117

    JO - Clinical Neurophysiology

    T2 - Clinical Neurophysiology

    JF - Clinical Neurophysiology

    SN - 1388-2457

    IS - S71

    ER -