Lecture Notes in Computer Science
Download 12.42 Mb. Pdf ko'rish
|
- Bu sahifa navigatsiya:
- Decomposing EEG Data into Space-Time-Frequency Components Using Parallel Factor Analysis and Its Relation with Cerebral Blood Flow
- Keywords
Fig. 4. The Hinton diagram of the first and the second projection matrices obtained by PTA, as shown in the left and the right sub-figures, respectively with the following requirements: 1 2 3 6 , , 10 l l l ≤ ≤ , 1 2 3 2 , , 5 l l l ′ ′ ′
≤ ≤ , 1 1
l ′ <
, 2 2 l l ′ <
, and 3 3 l l ′ <
. The total model selection errors are also 0. In every experiment, the value of the reconstruction error is very small. 4 Conclusion Vector data are normally used for probabilistic graphical models with probabilistic inference. However,
data, i.e., multidimensional arrays, are actually natural representations of a large amount of data, in data mining, computer vision, and many other applications. Aiming at breaking the huge gap between vectors and
tensors in
conventional statistical tasks, e.g., model selection, this paper proposes a decoupled probabilistic algorithm, named probabilistic tensor analysis (PTA) with Akaike information criterion (AIC) and Bayesian information criterion (BIC). PTA associated AIC and BIC can select suitable models for
data, as demonstrated by empirical studies.
We authors would like to thank Professor Andrew Blake at the Microsoft Research Cambridge for encouragement of developing the tensors probabilistic graphic model. This research was supported by the Competitive Research Grants at the Hong Kong Polytechnic University (under project number A-PH42 and A-PC0A) and the National Natural Science Foundation of China (under grant number 60703037). References [1] Bader, B.W., Kolda, T.G.: Efficient MATLAB Computations with Sparse and Factored Tensors. Technical Report SAND2006-7592, Sandia National Laboratories, Albuquerque, NM and Livermore, CA (2006) [2] Bader, B.W., Kolda, T.G.: MATLAB Tensor Classes for Fast Algorithm Prototyping. ACM Transactions on Mathematical Software 32(4) (2006) Probabilistic Tensor Analysis with Akaike and Bayesian Information Criteria 801 [3] Bishop, C.M.: Neural Networks for Pattern Recognition. Oxford University Press, Oxford (1995) [4] Kroonenberg, P., Leeuw, J.D.: Principal Component Analysis of Three-Mode Data by Means of Alternating Least Square Algorithms. Psychometrika, 45 (1980) [5] Lathauwer, L.D.: Signal Processing Based on Multilinear Algebra, Ph.D. Thesis. Katholike Universiteit Leuven (1997) [6] Sun, J., Tao, D., Faloutsos, C.: Beyond Streams and Graphs: Dynamic Tensor Analysis. In: The Twelfth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Philadelphia, PA, USA, pp. 374–383 (2006) [7] Tao, D., Li, X., Wu, X., Maybank, S.J.: General Tensor Discriminant Analysis and Gabor Features for Gait Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(10) (2007) [8] Tipping, M.E., Bishop, C.M.: Probabilistic Principal Component Analysis. Journal of the Royal Statistical Society, Series B 21(3), 611–622 (1999) [9] Tucker, L.R.: Some Mathematical Notes on Three-mode Factor Analysis. Psychometrika 31(3) (1966) [10] Vasilescu, M.A.O., Terzopoulos, D.: Multilinear Subspace Analysis of Image Ensembles. In: IEEE Proc. International Conference on Computer Vision and Pattern Recognition, Madison, Wisconsin, USA, vol. 2, pp. 93–99 (2003) [11] Ye, J., Janardan, R., Li, Q.: Two-Dimensional Linear Discriminant Analysis. In: Schölkopf, B., Platt, J., Hofmann, T. (eds.) Advances in Neural Information Processing Systems, Vancouver, British Columbia, Canada, vol. 17 (2004) [12] Ye, J.: Generalized Low Rank Approximations of Matrices. Machine Learning 61(1-3), 167–191 (2005)
M. Ishikawa et al. (Eds.): ICONIP 2007, Part I, LNCS 4984, pp. 802–810, 2008. © Springer-Verlag Berlin Heidelberg 2008 Decomposing EEG Data into Space-Time-Frequency Components Using Parallel Factor Analysis and Its Relation with Cerebral Blood Flow Fumikazu Miwakeichi 1 , Pedro A. Valdes-Sosa 3 , Eduardo Aubert-Vazquez 3 ,
3 , Jobu Watanabe 4 , Hiroaki Mizuhara 2 , and Yoko Yamaguchi 2
Medical System Course, Graduate School of Engineering, Chiba University 1-33, Yayoi-cho, Inage-ku, Chiba-shi, Chiba, 263-8522 Japan miwake1@faculty.chiba-u.jp 2 Laboratory for Dynamics of Emergent Intelligence, RIKEN Brain Science Institute, Japan 3 Cuban Neuroscience Center, Cuba 4 Waseda Institute for Advance Study, Waseda University, Japan
data has been a long-standing problem in electrophysiology. Our previous works showed that Parallel Factor Analysis (PARAFAC) can effectively per- form atomic decomposition of the time-varying EEG spectrum in space/ frequency/time domain. In this study, we propose to use PARAFAC for extract- ing significant activities in EEG data that is concurrently recorded with func- tional Magnetic Resonance Imaging (fMRI), and employ the temporal signature of the atom for investigating the relation between brain electrical activity and the changing of BOLD signal that reflects cerebral blood flow. We evaluated the statistical significance of dynamical effect of BOLD respect to EEG based on the modeling of BOLD signal by plain autoregressive model (AR), its AR with exogenous EEG input (ARX) and ARX with nonlinear term (ARNX). Keywords: Parallel Factor Analysis, EEG space/frequency/time decomposi- tion, Nonlinear time series analysis, Concurrent fMRI/EEG data. 1 Introduction The electroencephalogram (EEG) is recorded as multi-channel time-varying data. In the history of EEG study, there are so many types of oscillatory phenomena in spon- taneous and evoked EEG have been observed and reported. A statistical description of the oscillatory phenomena of the EEG was carried out first in the frequency domain by estimation of the power spectrum for quasi-stationary segments of data [1]. More recent characterizations of transient oscillations are carried out by estimation of the time-varying (or evolutionary) spectrum in the frequency/time domain [2]. These evolutionary spectra of EEG oscillations will have a topographic distribution on the sensors that is contingent on the spatial configuration of the neural sources that gener- ate them as well as the properties of the head as a volume conductor [3].
Decomposing EEG Data into Space-Time-Frequency Components Using PARAFAC 803
There is a long history of atomic decompositions for the EEG. However, to date, atoms have not been defined by the triplet spatial, spectral and temporal signatures but rather pairwise combinations of these components. Space/time atoms are the basis of both Principal Component Analysis (PCA) and Independent Component Analysis (ICA) as applied to multi-channel EEG. PCA has been used for artifact removal and to extract significant activities in the EEG [4,5]. A basic problem is that atoms defined by only two signatures (space and time) are not determined uniquely. In PCA orthogonality is therefore imposed between the corre- sponding signatures of different atoms. And there is the well-known non-uniqueness of PCA that allows the arbitrary choice of rotation of axes (e.g. Varimax and Quarti- max rotations). More recently, ICA has become a popular tool for space/time atomic decomposition [6,7]. It avoids the arbitrary choice of rotation (Jung et al. 2001). Uniqueness, however, is achieved at the price of imposing a constraint even stronger than orthogonality, namely, statistical independence. In both PCA and ICA the fre- quency information may be obtained from the temporal signature of the extracted atoms in a separate step. For the purpose of decomposing of single channel EEG into frequency/time atoms the Fast Fourier Transformation (FFT) with sliding window [8] or the wavelet transformation [9,10] have been employed. In fact, any of the frequency/time atomic decompositions currently available [11] could, in principle, be used for the EEG. How- ever, these methods do not address the topographic aspects of the EEG time/frequency analysis. It has long been known, especially in the chemometrics literature, that unique mul- tilinear decompositions of multi-way arrays of data (more than 2 dimensions) are possible under very weak conditions [12]. In fact, this is the basic argument for Paral- lel Factor Analysis (PARAFAC). This technique recently has been improved by Bro [13]. The important difference between PARAFAC and techniques such as PCA or ICA, is that the decomposition of multi-way data is unique even without additional orthogonality or independence constraints. Thus, PARAFAC can be employed for a space/frequency/time atomic decomposition of the EEG. This makes use of the fact that multi-channel evolutionary spectra are multi-way arrays, indexed by electrode, frequency and time. The inherent uniqueness of the PARAFAC solution leads to a topographic time/frequency decomposition with a minimum of a priori assumptions. It has been shown that PARAFAC can effectively perform a time/frequency/spatial (T/F/S) atomic decomposition which is suitable for identification of fundamental modes of oscillatory activity in the EEG [14,15]. In this paper, the theory of PARAFAC and its applications to EEG analysis will be showed. Moreover the possibility of analysis for investigating the relation between extracted EEG atom and cerebral blood flow will be discussed.
For the purpose of EEG analysis, we define the d f t N N N × × data matrix S as the three-way time-varying EEG spectrum array obtained by applying a wavelet trans- formation, where ,
and
t N are the numbers of channels, frequency steps and 804 F. Miwakeichi et al. time points, respectively. For the wavelet transformation a complex Morlet mother function was used. The energy d f t S of channel d at frequency f and time t is
given by the squared norm of the convolution of a Morlet wavelet with the EEG sig- nal
( , ) v d t ,
2 ( , )
( , ) d f t S w t f v d t = ∗ ,
(1) where the complex Morlet wavelet ( , )
w t f
is defined by 2 2 ( , ) b t i f t b w t f e e σ π πσ ⎛ ⎞ − ⎜ ⎟ ⎝ ⎠ = with
b σ denoting the bandwidth parameter. The width of the wavelet, 2
m f πσ = , is set to 7 in this study. The basic structural model for a PARAFAC decomposition of the data matrix ( )
× × S with elements
S is defined by
1 ˆ k N d f t d k f k t k d f t d f t d f t k S a b c e S e = = + = + ∑ .
(2) The problem is to find the loading matrices A ,
and
, the elements of which are denoted by
, d k f k a b and
t k c in Eq. (2). Here we will refer to components k as
“atoms”, and the corresponding loading vectors { }
{ } { }
, ,
dk k fk k tk a b c = = = a b c will be said to represent the spatial, spectral and temporal signatures of the atoms (Fig. 1). The uniqueness of the decomposition (2) is guaranteed if rank( )+rank( )+rank( ) 2( 1) k N ≥ + A B C . As can be seen, this is a less stringent condition than either orthogonality or statistical independence [12].
spectrum
is obtained from a channel by channel wavelet transform. S is a three-way data array indicated by channel, frequency and time. PARAFAC decomposes this array into the sum of “atoms”. The k-th atom is the trilinear product of loading vectors representing spatial ( k a ),
spectral ( k b ) and temporal ( k c ) “signatures”. Under these condition PARAFAC can be sum- marized as finding the matrices { }
k =
a , { } k =
b and
{ } k =
c that explain S with
minimal residual error. Decomposing EEG Data into Space-Time-Frequency Components Using PARAFAC 805
The decomposition (2) can be obtained by evaluating
2
ˆ min d k f k t k d f t d f t a b c S S − . Since the
d f t S can be regarded as representing spectra, this minimization should be carried out under a non-negativity constraint for the loading vectors. The result of the PARAFAC decomposition is given by the ( 1)
× vector k a , representing the topog- raphical map of the
-th component, the ( 1)
N × vector k b representing the fre- quency spectrum of the
-th component, and the ( 1)
× vector k c representing the temporal signature of the
-th component. 3 Results 3.1 Extracting Relevant Components from EEG The experiment consisted of two conditions, eyes-closed resting condition and mental arithmetic condition; for each condition five epochs, each lasting 30 s per condition, were recorded. During the mental arithmetic epochs, the subjects were asked to count backwards from 1000 in steps of a constant single-digit natural number, while keep- ing their eyes closed. At the beginning of each mental arithmetic epoch this number was randomly chosen by a computer and presented to the subjects through head- phones. The end of each mental arithmetic epoch was announced by a beeping sound (4 kHz, 20 ms). Mental arithmetic and resting epochs were occurring five times within each trial and arranged in alternating order; each subject was examined in two trials. During the experiment, fMRI and EEG were recorded simultaneously. The elec- trode set consisted of 61 EEG channels, two ECG channels and one EOG channel. The reference electrode was at FCz. Raw EEG signals were sampled at a sampling frequency of 5 kHz, using a 1Hz high-pass filter provided by the Brain Vision Re- corder software (Brain Products, Munich, Germany). The fMRI was acquired as blood-oxygenation-sensitive (T2*-weighted) echoplanar images, using a 1.5 T MR scanner (Staratis II, Hitachi Medico, Japan) and a standard head coil. Thirty slices (4mm in thickness, gapless), covering almost the entire cerebrum, were measured, under the following conditions: TR 5 s, TA 3.3 s, TE 47.2 ms, FA 901, FoV 240mm, matrix size 64×64. In order to minimize head motion, heads of subjects were immobi- lized by a vacuum pad. SPM99 was used for preprocessing of fMRI images (motion correction, slice timing correction). Using bilinear interpolation, images were normal- ized for each subject to a standard brain defined by Montreal Neurological Institute; the normalized images were subsequently smoothed using a Gaussian kernel (full- width half-maximum: 10 mm). For each voxel, the BOLD signal time series were high-pass filtered (>0.01Hz), in order to remove fluctuations with frequencies lower than the frequency defined by the switching between resting and task. The EEG signal ( , )
v d t was subsampled to 500Hz, and, in order to construct a three-way (channel/frequency/time) time-varying EEG
× × data matrix S , a
806 F. Miwakeichi et al.
(a) Spectral signatures of atoms of Parallel Factor Analysis (PARAFAC) for a typical subject. Note the recurrent appearance of frequency peaks in the theta and alpha bands. The horizontal axis represents frequency in Hz, the vertical axis represents the normalized ampli- tude. (b) Spatial signatures of atoms displayed as a topographic map, for the theta, alpha and high alpha atoms (from above) of Parallel Factor Analysis (PARAFAC) for the same subject. (c) Temporal signatures of atoms, same order as in (a). The horizontal axis represents time in units of scans (time between scans is five seconds), the vertical axis represents normalized intensity. The black and red colored lines are corresponding to resting and task stages, respec- tively. (See colored figure in CD-R). wavelet transform was applied in the frequency range from 0.5 to 40.0 Hz with 0.5 Hz steps, using the complex Morlet wavelet as mother function. For adjusting the temporal resolution to the resolution of the fMRI data, the wavelet-transformed EEG was aver- aged over consecutive 5 seconds intervals. Then, by applying PARAFAC to this three- way data set, major signal components in the frequency range from 3.0 to 30.0 Hz were
Decomposing EEG Data into Space-Time-Frequency Components Using PARAFAC 807
extracted. The reason for choosing this frequency band was that at lower frequency EEG data typically contains eye movement artifacts, while at higher frequency there is no relevant activity in the data set analyzed in this study. In Fig.2(a) spectral signatures of the identified atoms are shown: theta (around 6Hz), alpha (around 9Hz), and high alpha (around 12Hz) atoms were found; in this study we denote the frequency of the last atom as “high alpha”. We find that if the order of the decomposition is increased beyond three, two atoms with overlapping frequency peaks appear. Since this indicates overfit- ting, we select three as the order of the decomposition. In Fig. 2(b) spatial signatures of atoms are shown. It can be seen that the main power of theta and alpha atoms is focused in frontal and occipital areas, respectively. In Fig.2(c) temporal signatures of atoms are shown; red color refers to task condition, black color to resting condition. By comparing the amplitudes of temporal signatures during both conditions by a standard T-test, we find that for the theta atom the amplitudes of the temporal signature are higher during the task condition, while for the alpha atom they are higher during the resting condition. For the high alpha atom no significant result is found with this test, therefore this atom will be omitted from further study. Download 12.42 Mb. Do'stlaringiz bilan baham: |
ma'muriyatiga murojaat qiling