Biomedical Engineering Reference
In-Depth Information
locations in a given direction. The PCA finds a set of orthogonal axes, a base, such
that each consecutive axis spans the directions with consecutively decreasing vari-
ance. The projections of the points onto these axes constitute the components. Each
component can be visualized as a time series. On the other hand the original time
series can be recovered as a linear combination of these components. Components
corresponding to the smallest variance can be neglected and in this way a reduction
of data dimensionality can be achieved.
3.6.1.2 Computation
PCA can be performed by means of the singular value decomposition (SVD) algo-
rithm. An epoch of
k
channels data of length
m
can be represented as a
m
×
k
matrix
x
. It is always possible to find three matrices
P
,
A
,and
M
such that:
PAM
T
x
=
(3.50)
where
P
is the
m
×
k
matrix containing
k
normalized principal component waveforms
(
P
T
P
k
matrix mapping components to original data such that
M
ij
is the contribution of
j
th
=
1),
A
is
k
×
k
, diagonal matrix of components amplitudes,
M
is a
k
×
component to
i
th
1. The SVD algorithm is implemented in
MATLAB as a function
svd
;tofind the decomposition of
x
into
P
,
A
,and
M
execute:
[P,A,M] = svd(x);
When data concerning multiple conditions or subjects are to be analyzed by PCA
the matrix
x
is composed by concatenating the matrixes of individual conditions or
subjects along the time dimension. Thus for example for
N
c
conditions we have to
decompose matrix
x
which has dimension
m
channel;
M
T
M
=
k
.The
n
th
row of resulting matrix
P
represents
N
c
concatenated time courses obtained for each condition for the
n
th
component. The corresponding topographical map is contained in the
n
th
column.
·
N
c
×
3.6.1.3
Possible applications
PCA can be used as a tool for decomposition of multichannel data into a linear
combination of components characterized by different spatial, temporal, and ampli-
tude distribution. These components can be used as a preprocessing step in source lo-
calization techniques. However, there is no direct correspondence between the prin-
cipal components and individual sources. A single source may produce a signal that
decomposes into a number of components, while on the other hand multiple sources
can contribute to a single component. In general, the individual components should
not be ascribed directly to any physiologically meaningful phenomenon [Lamothe
and Stroink, 1991].
A similar method, which leads to the reduction of dimensionality of the data is fac-
tor analysis (FA). FA is related to PCA but not identical. PCA performs a variance-
maximizing rotation (varimax) of the variable space, thus it takes into account all
variability in the data. In contrast, factor analysis estimates how much of the vari-
ability is due to common factors. The number of common factors is usually less than
the dimension of the original variable space [Bryant and Yarnold, 1994]. FA can de-
scribe a large set of recorded wave shapes from multiple sensors in terms of a small
Search WWH ::
Custom Search