04451421 (1)

Upload: debopam-datta

Post on 04-Apr-2018

215 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/31/2019 04451421 (1)

    1/5

    Abstract-- Electroencephalogram (EEG) signal has been found to

    be the most predictive and reliable indicator in wake-sleepresearch. It is a real time signal that reflects the brain states of asubject including the alertness. However the study of wake-sleepcondition using EEG signal is difficult due to the complexity ofthe EEG signal itself. The exact underlying dynamics of the EEGdata is still questionable. EEG signal varies from one individualto another and has an inter variability in the same physiological

    state. It is hard to compare the EEG to the specific pattern ofindividual or situation. This paper tries to investigate the use ofUMACE in distinguish between awake and sleep state of a

    subject. Normal EEG data from individual is used as an input inbuilding UMACE filter. From the result, we find UMACE hasthe capability to distinguish awake and sleep state of a subject.

    Index TermsElectroencephalogram, Uncostrained MovingAverage Correlation Energy (UMACE), Wake-sleep

    I. INTRODUCTION

    EG is one of the well known method applied in extracting

    signal from the brain. Researches in wake-sleep, emotion

    and epilepsy detection mostly utilize the information of the

    EEG signal. It is well known that EEG signal is a very

    complex signal and the exact underlying conditions that

    produce the signal is still uncertain. Two common drawbacksof EEG as inputs are inter-trial variability and inter-individual

    variability. Moreover, EEG has no specific pattern to indicate

    ages and conditions. It matures with the age of the person,

    therefore infant and adolescence will have different patterns of

    EEG signal. The signal is altered as the person grows older.

    Since maturation depends on age, it is difficult to make a

    specific comparison (distinct physiological state).

    In wake-sleep research, EEG is a known method that can

    depict the changes in brain states. The changes of the brain

    state for wake-sleep can be detected in the frontal area [1].

    The most prominent waves that showed a consistent pattern

    and occurrence are beta waves, alpha waves, theta waves and

    delta waves. In detection sleep onset, beta waves and alpha

    waves play an important rule in early stage of drowsiness and

    R. Ghafar is with Universiti Kebangsaan Malaysia, 43600 Bangi Selangor,Malaysia (e-mail: [email protected]).

    N. Md. Tahir is with Universiti Teknologi Mara, 40500 Shah Alam,Selangor, Malaysia (e-mail: [email protected])

    A. Hussain is with Universiti Kebangsaan Malaysia, 43600 Bangi

    Selangor, Malaysia (e-mail: a [email protected])S. Abdul Samad is with Universiti Kebangsaan Malaysia, 43600 Bangi

    Selangor, Malaysia (e-mail: sal [email protected])

    early stage of sleep. The description of EEG changes in sleep

    stage is as in Table 1 [2]. Some disadvantages of utilizing

    EEG as inputs for the sleep state are; in term of alpha wave,

    up to 10% of the population does not show alpha activity [3]

    and the maximum accuracy that can be achieved when using

    manual scoring by two or more neurologists are around 70 %

    to 80 % only [4]. Therefore, a computer assisted diagnosis is

    needed to increase the accuracy and further served as a

    guideline to diagnose wake sleep stage. The changes during

    transition of awake to sleep stage are individually differ from

    one to another person. Therefore it is hard to compare EEG toa specific pattern. In order to develop a technique that can be

    used for all EEG data, we try UMACE method using

    individual EEG data as an input. Three seconds segment is

    taken from the normal segment as a training set for each

    individual patient.

    TABLE1SCORING FOR WAKE-SLEEP USING EEG

    Stage EEG

    Awake Low-voltage, desynchonised, 10Hz alpha with eyes

    shut

    Stage 1 Low-voltage, desynchonised, no alpha. Some theta

    bursts or vertex spikes

    Stage 2 Low-voltage, desynchonised, with phasic 13-15Hz

    spindles and K complexes

    Stage 3 As stage 2, but increasing slow wave (50% of record

    Stage REM Low-voltage, desynchonised, sometimes preceded by

    3-7Hz saw-tooth waves.

    This paper investigates the used of UMACE filter todiscriminate two conditions namely, awake and sleep stateusing EEG as input data. Recently, UMACE filter have beenexplored and they have evolved into very effective algorithms

    for pattern recognition applications specifically biometrics

    verification [5]-[8]. For example, [8]-[9] proved that the

    advanced correlation filter such as MACE and UMACEoffered good matching performance in the presence of

    variability such as different facial expressions andillumination. These findings suggest that the UMACE could

    be explored to process the EEG signal with inter-trial and

    inter-individual variability.

    Furthermore, UMACE is chosen in this study due to its

    capability to function in a limited memory device, ease of

    installation and produces fast results [9]. In real time, EEG

    EEG Analysis of Wake-sleep Data using

    UMACE filterR. Ghafar, N. Md. Tahir, A. Hussain, Member, IEEE& S. A. Samad, Sr. Member, IEEE.

    E

    The 5h

    Student Conference on Research and Development SCOReD 200711-12 December 2007, Malaysia

    1-4244-1470-9/07/$25.00 2007 IEEE.

  • 7/31/2019 04451421 (1)

    2/5

    signal changes rapidly over time, therefore an efficient

    processing technique is needed for first-rate results.

    II. MACE AND UMACE FILTERS

    Minimum Average Correlation Energy (MACE) filter is an

    evolution from matched filtered that is design to produce

    optimal detection from a known reference images in the

    presence of Gaussian white noise. Matched filter performanceis easily deteriorates by the scale changes, rotation and pose of

    the image. Synthetic discriminant function filter and equal

    correlation peak filter is created to overcome the problem arise

    in using matched filter. This filtered used several training

    images to create a single correlation filter. Synthetic

    Discriminant Function (SDF) produced a pre-specified peak

    known as peak constrain. The peak constrains lead to

    miscalculation when the sidelobes are larger than the

    controlled origin. Thus, to reduce large sidelobes the MACE

    and UMACE filters are designed. MACE filter minimizes the

    average correlation energy of the training image while

    constraining the correlation output at the origin to a specific

    value for each training images. On the other hand, UMACE

    minimizes the average correlation output while maximizing

    the correlation output at the origin..

    The equations for MACE and UMACE are as given in (1)

    and (2) respectively:

    h=D-1X(X+D-1X)-1u (1)

    h=D-1m (2)

    whereD is a diagonal matrix with the average power spectrum

    of the training image placed along diagonal elements; Xconsists of the Fourier transform of the training imageslexicographically re-ordered and placed along each column; u

    is a column vector containing the desired correlation output at

    the origin for each training images and m is a column vector

    containing the mean of the Fourier Transforms of the training

    images. Detail derivation of both MACE and UMACE

    equations can be found in [8]. Studies have shown that theMACE filter has no built-in immunity to noise and it is often

    excessively sensitive to intra-class variations as compared to

    UMACE filter [7]-[9]. Therefore only the latter will be

    adopted in this study.

    III. TECHNICAL WORKPREPARATION

    In this study, data was recorded using the Biopac systems

    and the AcqKnowlegde software (Model MP 100 Data

    Acquisition Systems; Biopac Systems Inc, Goleta, CA; using

    AcqKnowledge 881 Version 3.1.2 software). The recording

    was done at sampling rate of 256 Hz. The subjects involved

    aged between 21 to 35 years. Several steps were taken in order

    to determine better result such as the avoidance of applying

    any type of hair gel or conditioners for effective contact

    between the scalp and electrodes and all subjects are advised

    from taking any form of medication prior to recording.

    Besides, they are also recommended to reduce their coffee

    intake and deprive their sleep the night before in order to get a

    quick sleep data during recording. Subjects were seated in a

    sound-controlled air conditioning room with dim lighting to

    provide comfortable sleeping condition for the recording

    purpose. During recording, subjects were also requested to

    perform brain related activities namely reading and some

    simple mental calculation.

    Eighteen surface electrodes were placed at specific location

    all over the scalp based on the 10-20 systems. The electrodes

    are connected by wires to an EEG machines. Data taken using

    Biopac system was converted into text file prior to processing

    in the MATLAB environment. The EEG data is filtered using

    the low-pass filter that allows only frequency below 50Hz to

    be processed; this helps to discard the 50Hz noise that was

    generated during the recording. Since the component that

    triggered certain situation is still unknown, we prefer to use

    only filtered data as an input to the UMACE filter. The further

    processed data are rarely used because certain information

    might be missing and unavailable [10].

    Typically, UMACE filters utilized images as their inputs.

    Image is a 2D data where as EEG signal is a one dimensional

    (1D) data. Therefore, embedding method has been applied to

    the EEG data in order to increase the EEG data

    dimensionality. Next, UMACE is applied to the filtered and

    embedded EEG data. Three consecutive seconds are taken

    from the sleep portion data and used as the training set. Fig. 1

    depicts the overall system using UMACE filter for wake sleep

    detection.

    Fig. 1. Application of UMACE filter for wake sleep detection

    Fourier transform of the embedded training data is used to

    build a synthesized correlation filter; the UMACE filter. The

    Fourier transform of the embedded test image is then cross-

    Rearrange into 2D

    FFT

    CorrelationFilter

    Filter

    Design

    IFFT

    Training data

    Test data

    FFT

    FFT

    FFT

    Correlation Plane

    ROI

    Sum of ROI

  • 7/31/2019 04451421 (1)

    3/5

    correlated with the synthesized UMACE filter and produced

    the correlation output. Most of the previous research on

    advance correlation filters used the Peak to side lobe ratio

    (PSR) measurement from correlation output as an indicator for

    discrimination or classification. However, PSR value was

    found to be unreliable for the case of EEG as input. These are

    due to the fluctuations and inconsistency of the EEG data.

    However, there are consistent changes in the correlation plot

    of UMACE filtered. As depicted in Fig. 2, the correlation plotfor sleep portion is almost black with pixel values is mostly

    near zero. In contrast, the wake portion correlation output

    produces a tiny gray pixel in the center of the plot as

    illustrated in Fig. 2b.

    The difference which is located in the center of the plot is

    difficult to visualize. Fig. 3 is used to reveal the variation in

    the plot between sleep and wake stage. The differences

    between sleep and wake states can be seen more prominently

    in the inversion plot as shown in Fig. 3. The changes in the

    middle of the correlation output plot are consistent for three

    subjects. Instead of using PSR, the sum of the pixel value is

    calculated for the middle area in order to producedistinguishable result. The middle area is called ROI. Then,

    the sum of ROI is plotted as indicated in Fig. 3, Fig. 4 and Fig.

    5 respectively. Preliminary results demonstrated the existence

    of very small differences in distinguishing the awake and sleep

    states. The result is further smoothed using moving average

    filter to improve the result and enhance the differences

    between the two states.

    (a) sleep (b) awake

    Fig. 2. Correlation output

    (a) sleep (b) awake

    Fig. 3. Correlation output (inversion)

    IV. RESULTS

    In this section, the results of the UMACE filter as depicted

    in Fig. 4, Fig. 5 and Fig. 6 are examined. The first plot (top) is

    the normalized sum of ROI (pre-filtered sum of ROI). The

    second plot (bottom) is the smoothen sum of ROI plot using

    moving average filter.

    Fig. 4. Subject 1

    Fig. 5. Subject 2

    .

    Awake AwakeSleep

    Awake AwakeSleep

  • 7/31/2019 04451421 (1)

    4/5

    Fig. 6. Subject 3

    It can be seen that the pre-filtered sum of ROI results

    showed extreme fluctuation. After filtering, the processedEEG data (i.e. the sum of ROI) has less fluctuation and thesleep region for all subjects can be determined easily since the

    fluctuation has reduced to values between 0 to 0.3. The sleep

    region has normalized magnitude of 0.3 or less. The plot

    showed consistent results in the sleep region and enormous

    fluctuation in the wake region. As for the wake region, the

    fluctuation is greater and the magnitude varies from 0 to 1.The preliminary findings of the use of sum of ROI obtained

    from UMACE method can be applied as guidelines orreferences in building an application for wake-sleep detection.

    Fluctuations occurred in the wake region are due to

    movement, conversation and mental task activities performed

    during recording and therefore are uncontrollable. In thisstudy, subjects were not restraint as they are allowed to move

    slightly in order to provide greater comfort in finding the best

    position for them to get a good sleep. Besides, with or without

    movement, it is natural for the EEG signal to have fluctuations

    presence.

    V. CONCLUSIONSIn conclusion, the UMACE method has been proven

    applicable to detect the wake sleep condition in EEG data. The

    changes of the correlation plot are vital and therefore are

    considered instead of the normal PSR ratio. From the result, it

    can be concluded that the ROI plot can be used to classify the

    wake and sleep stage of a subject. Sleep stage show smaller

    fluctuations in the form of ripples as contrast to the wake state.

    These characteristics can be utilized as an aid to interpret the

    EEG signal. Further work has to be done to improve the

    detection and automate the process in order to facilitate in the

    wake-sleep scoring process.

    VI. ACKNOWLEDGMENT

    The authors expressed their gratitude to Miss Fatimah

    Abdul Hamid for the data used in this study.

    VII. REFERENCES

    [1] C. Cajochen, R. Foy, D. Dijk, (1999) Frontal Predominance of a

    Relative Increase in Sleep Delta and Theta EEG Activity after SleepLoss in Humans, Sleep Research Online, 2(3), pp:65-69. Availabel:http://www.sro.org/1999/Cajochen/65/65

    [2] J. Empson, Human Brainwaves. The Psychological Significance of theElectroencephalogram, London, Macmillan Press Ltd, 1986, p. 70.

    [3] U. Svenson, Blink behaviour based on drowsiness detection method

    development and validation, Linkopings Universitet, Tecg. Rep. LiU-IMT-EX--04/369--,Sept. 2004.

    [4] V. Gerla, L. Lhotska, V. Krajca, and K. Paul, Multichannel Analysis of

    the Newborn EEG Data.In Proc. The Int. Special Topic Conference onInformation Technology in Biomedicine (ITAB 2006). Available at :

    http://medlab.cs.uoi.gr/itab2006/proceedings/EEG%20Analysis/52.pdf

    [5] S. A. Samad, D. Athiar Ramli & A. Hussain., Person IdentificationUsing Lip Motion Sequance, B. Apolloni et at (Eds). Pp 839-846, 2007.Springer-Verlaq Berlin Heidelberg

    [6] S. A. Samad, D. Athiar Ramli & A. Hussain., Lower Face VerificationCentered on Lips using Correlation Filters, Information TechnologyJournal, 2007 (in press)

    [7] N. Md. Tahir, A. Hussain & S. Abdul Samad, On The Use ofAdvanced Correlation Filters for Human Posture Recognition, Journalof Applied Science (in press)

    [8] M. Savvides, B. V. K. Vijaya Kumar and P. K. Khosla, FaceVerification using Correlation Filters, in Proc. of the Third IEEE

    Automatic Identification Advanced Technologies, pp. 56-61, 2002

    [9] M. Savvides and B. V. K. Vijaya Kumar, Effificient Design of

    Advanced Correlation Filters For Robust Distortion-Tolerant FaceIdentification, inProc. IEEE Int. Conf. On Advanced Video and Signal

    Based Surveillance(AVVS), pp. 45-52, 2003[10] C. K. Ng, M. Savvides, and P. K. Khosla, Real-Time Face Verification

    Systems on a Cell-Phone Using Advanced Correlation Filters, in Fourth

    IEEE Workshop on Autimatic Identification Advanced Technologies(AutoID05), pp. 57-62.

    [11] D. Sommer, M. Chen, M. Golz, U. Trutschel, and D. Mandic, Fusion ofState Space and Frequency-Domain Features For Improved MicrosleepDetection (W. Duch et al, Eds.): ICANN 2005, LNCS 3697, pp. 753-759.

    VIII. BIOGRAPHIES

    Rosniwati Ghafar obtained her Bachelor of Engineering and Mathematics

    from Vanderbilt University, USA in 1997. In 2004, she received her Masterdegree from Universiti Sains Malaysia. Now workingtowards Ph.D degree in signal processing field at

    Faculty of Engineering, Universiti KebangsaanMalaysia.

    Nooritawati Md Tahir is a lecturer at UniversitiTeknologi Mara (UiTM) since December 1991. She

    obtained her BEng in Electrical Engineering from

    UiTM in 1988 and MSc (Electronics &Communication) in 1991 from the University ofLiverpool, UK. Currently, she is working toward her

    PhD degree in the field of image processing, patternrecognition and computational intelligence at theFaculty of Engineering, Universiti Kebangsaan

    Malaysia.

    Awake AwakeSleep

  • 7/31/2019 04451421 (1)

    5/5

    Aini Hussain (M97) received the B. Sc. inElectrical Engineering, M.Sc. in Control Engineering

    and Ph.D from Louisiana State University, UMIST,and Universiti Kebangsaan Malaysia, respectively.She is a professor at the Department of Electrical,

    Electronic & Systems Engineering, UniversitiKebangsaan Malaysia. Her research interests includesignal processing, pattern recognition and

    computational intelligence.

    Salina Abdul Samad (M95-SM07) received theBSEE and Ph.D in electrical engineering from

    University of Tennessee and University ofNottingham, respectively. She is a professor at theDepartment of Electrical, Electronic & Systems

    Engineering, Universiti Kebangsaan Malaysia. Herresearch interests include digital signal processing,filter design and multimodal biometric system.