0% found this document useful (0 votes)
65 views20 pagesCH 11
Biomedical Signal Processing chapter
Uploaded by
shivaspyCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
65 views20 pagesCH 11
Biomedical Signal Processing chapter
Uploaded by
shivaspyCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 20
ll
Other Time- and Frequency-Domain Techniques
Dorin Panescu
A biomedical signal is often corrupted by noise (¢.g., powerline interference, mus-
cle or motion artifacts, RF interference from electrosurgery or diathermy appara-
tus). Therefore, it is useful to know the frequency spectrum of the corrupting signal
in order to be able to design a filter to eliminate it. If we want to find out, for ex-
ample, how well the patient's cardiac output is correlated with the area of the QRS
complex, then we need to use proper correlation techniques. This chapter presents
time and frequency-domain techniques that might be useful for situations such as
those exemplified above.
11.1 THE FOURIER TRANSFORM
The digital computer algorithm for Fourier analysis called the fast Fourier trans-
form (FFT) serves as a basic tool for frequency-domain analysis of signals.
11.1.1 The Fourier transform of a discrete nonperiodic signal
Assuming that a discrete-time aperiodic signal exists as a sequence of data sampled
from an analog prototype with a sampling period of T,the angular sampling
frequency being as = 2x/T, we can write this signal in the time domain as a series
of weighted Dirac functions. Thus
x()= x(n) &t - nT) (1.1)
n
216 Other Time- and Frequency-Domain Techniques 217
The Fourier transform of this expression is
x@= J x eiorde (11.2)
or X(@) = J DL xm) &t-nT) eioar (11.3a)
a
‘The ordering of integration and summation can be changed to give
X@)= Y xn) J Ke-nTyeiora (11.3b)
Thus we obtain
DX xy) clon (11.3¢)
nwa
X(@)
And similarly, we can find that the inverse Fourier transform is
os
x(n ad X(@) &OrT deo a4)
One of the important properties of the Fourier transform, which is shown in
Figure 11.1(b), is its repetition at intervals of the sampling frequency in both posi-
tive and negative directions. Also it is remarkable that the components in the inter-
val 0< @ < @s/2 are the complex conjugates of the components in the interval
@s/2< @ < as. It is modem practice to use normalized frequencies, which means
that the sampling period T is taken to be 1. Therefore, the Fourier transform pair
for discrete signals, considering normalized frequencies, is
X(@= LY x(n) evden (11.5a)
i
©
x f X(@) el@nda (11.5b)
=
x(n) 218 Biomedical Digital Signal Processing
‘We observe that this kind of Fourier transform is continuous, and it repeats at in-
tervals of the sampling frequency.
Discrete
signal
Modulus:
20,
(a) (b)
Figure 11.1 (a) A discrete-time signal, and (b) the modulus of its Fourier transform. Symmetry
about @s/2, due to the sampling process, is illustrated.
11.1.2 The discrete Fourier transform for a periodic signal
The discrete Fourier transform (DFT) is the name given to the calculation of the
Fourier series coefficients for a discrete periodic signal. The operations are similar
to the calculation of Fourier coefficients for a periodic signal, but there are also
certain marked differences. The first is that the integrals become summations in the
discrete time domain. The second difference is that the transform evaluates only a
finite number of complex coefficients, the total being equal to the original number
of data points in one period of original signal. Because of this, each spectral line is
regarded as the k-th harmonic of the basic period in the data rather than identifying
with a particular frequency expressed in Hz or radian/s. Algebraically, the forward
and reverse transforms are expressed as
Nal kn 2m
X= LY xe (11.6a)
n=0
Nel ikn 2:
xm=h D Xwew (11.60)
k=0
Figure 11.2 shows a discrete periodic signal and the real and imaginary parts of
its DFT. The first spectral line (k = 0) gives the amplitude of the de component in
the signal, and the second line corresponds to that frequency which represents one Other Time- and Frequency-Domain Techniques 219
cycle in N data points. This frequency is 2n/N. The N-th line corresponds to the
sampling frequency of the discrete N-sample data sequence per period and the
k= N/2-th line corresponds to the Nyquist frequency. Using the symmetry of the
DFT, algorithms for fast computation have been developed. Also, the symmetry
has two important implications. The first is that the transformation will yield N
unique complex spectral lines. The second is that half of these are effectively re-
dundant because all of the information contained in a real time domain signal is
contained within the first N/2 complex spectral lines. These facts permitted the de-
velopment of the Fast Fourier ‘Transform (FFT), which is presented in the next
section.
Discrete signal
0123... NA n
(a) (b)
Figure 11.2 (a) A discrete periodic signal and (b) the real and imaginary parts of its DFT.
11.1.3 The fast Fourier transform
For a complete discussion of this subject, see Oppenheim and Schafer (1975). The
term FFT applies to any computational algorithm by which the discrete Fourier
transform can be evaluated for a signal consisting of N equally spaced samples,
with N usually being a power of two. To increase the computation efficiency, we
must divide the DFT into successively smaller DFTs. In this process we will use
the symmetry and the periodicity properties of the complex exponential
Wii = eJOnmNykn
where Wy substitutes for e2/), Algorithms in which the decomposition is
based on splitting the sequence x(n) into smaller sequences are called decimation in
time algorithms. The principle of decimation in time is presented below for N equal
to an integer power of 2. We can consider, in this case, X(k) to be formed by two 220 Biomedical Digital Signal Processing
N/2-point sequences consisting of the even-numbered points in x(n) and odd-
numbered points in x(n), respectively. Thus we obtain
x@= YD sew + Y xywy a7)
n= pel n&%p
which can also be written as
N/2-1 N2=1
x= DY xOpywr+ YD xQp+ wees ve (11.8)
p=0 p=0
But Wy = Wyo and consequently Eq. (11.8) can be written as
N/2-1 N2-1
xo= ¥ xp) Whi + Wy La@r+ NWA =Xelk) + Wy Xo) (11.9)
p= p=
Each of the sums in Eq. (11.9) is an N/2-point DFT of the even- and odd-numbered
points of the original sequence, respectively.
‘After the two DFTs are computed, they are combined to give the DFT for the
original N-point sequence. We can proceed further by decomposing each of the two
N/2-point DFTs into two N/4-point DFTs and each of the four N/4-point DFTs into
two N/8-point DFTs, and so forth. Finally we reduce the computation of the N-
point DFT to the computation of the 2-point DFTs and the necessary additions and
multiplications.
Figure 11.3 shows the computations involved in computing X(k) for an 8-point
original sequence. Oppenheim and Schafer (1975) show that the total number of
complex additions and multiplications involved is Nlog>N. The original N-point
DFT requires N2 complex multiplications and additions; thus it turns out that the
FFT algorithm saves us considerable computing time.
Figure 11.4 shows the computation time for the FFT and the original DFT versus
N. The FFT requires at least an order of magnitude fewer computations than the
DFT. As an example, some modem microcomputers equipped with a math copro-
cessor are able to perform an FFT for a 1024-point sequence in much less than 1 s.
In the case when NV is not an integer power of 2, the common procedure is to aug-
ment the finite-length sequence with zeros until the total number of points reaches
the closest power of 2, or the power for which the FFT algorithm is written. This
technique is called zero padding. In order to make the error as low as possible,
sometimes the signal is multiplied with a finite-length window function.
Windowing is also applied when N is an integer power of 2 but the FFT-analyzed
signal does not contain an integer number of periods within the N points. In such
cases, the error introduced by the unfinished period of the signal may be reduced
by a proper choice of the window type. Other Time- and Frequency-Domain Techniques 221
Figure 11.3 The flow graph of the decimation-in-time of an 8-point DFT.
80000
—2— DFT computations
——t——-_ FFT computations
SL
100 200
Number of points (NV)
60000
40000
Number of computations
300
Figure 11.4 Computational savings with the FFT. 222 Biomedical Digital Signal Processing
11.1.4 C-language FFT function
‘The computational flow graph presented in Figure 11.3 describes an algorithm for
the computation of the FFT of finite-length signal applicable when the number of
points is an integer power of 2. When implementing the computations depicted in
this figure, we can imagine the use of two arrays of storage registers, one for the
array being computed and one for the data being used in computation. For
example, in computing the first array, one set of storage registers would contain the
input data and the second set of storage registers would contain the computed
results for the first stage.
In order to perform the computation based on the “butterfly” graph, the input
data must be stored in a nonsequential order. In fact, the order in which the input
data must be stored is the bit-reversed order.
To show what is meant by this, we write the index of the output data and the
index of the corresponding input data using three binary digits.
(000) — x(000)
X(001) — x(100)
X(010) — x(010)
X(11) — x(110)
X(100) — x(001)
X(101) — x(101)
X(110) — x11)
X(111) — x(111)
If (xg nq ng) is the binary representation of the index of the sequence x(n), which is
the input data, then (x(n)} must be rearranged such that the new position of
x(ng ny] ng) is x(n9 m1 2).
Figure 11.5 shows a C-language fragment for the FFT computation, which
reorders the input array, x[nn]. The FFT function in UW DigiScope allows the user
to zero-pad the signal or to window it with the different windows.
11.2: CORRELATION
We now investigate the concept of correlation between groups of data or between
signals. Correlation between groups of data implies that they move or change with
respect to each other in a structured way. To study the correlation between signals,
we will consider signals that have been digitized and that therefore form groups of
data. Other Time- and Frequency-Domain Techniques 223
¥define RORD(a,b) tempr=(a); (a)=(b); (b) =tempr
float tempr,x(512];
int i,3,m,n,nnz
Inn=512;
In=nn<4) {
RORD(x(3],x[i])7/*this is the bit-reversal section of*/
RORD (x(}+1],x[i+1]);/*a FFT computation routine*/
)
men >>1;
while (m>=2 && j>m){
j
im >>1;
)
jtsm;
Figure 11.5 The C-language program for bit-reversal computations,
11.2.1 Correlation in the time domain
For N pairs of data (x(m),y(n)}, the correlation coefficient is defined as
N _
LY (e@)-x} {y@)-y)
a eae
4) © (x) -x}2_Y (9) -y)2
a
If finite-length signals are to be analyzed, then we must define the crosscorrela-
tion function of the two signals.
(11.10)
N = =
© (x(n) -x) +0 -y}
) =
rxy(k) W a :
> (x) - x}? Y (Oy -y}?
n=l nel
(11.11) 224 Biomedical Digital Signal Processing
In the case when the two input signals are the same, the crosscorrelation function
becomes the autocorrelation function of that signal. Thus, the autocorrelation func-
tion is defined as
N 7 =
DY. (x(n) - x} (xn + &)- x)
Tyxlk) (11.12)
W
& (x) -x)}?
Figure 11.6 presents the crosscorrelation of respiratory signals recorded simulta~
neously from a human subject using impedance pneumography. In Figure 11.6(a),
the signals were acquired at different points along the midaxillary line of the sub-
ject. The subject was breathing regularly at the beginning of the recording, moving
‘without breathing in the middle of the recording, and breathing regularly again at
the end. In Figure 11.6(b), each combination of two recorded channels were cross-
correlated in order to try to differentiate between movement and regular breathing.
11.2.2 Correlation in the frequency domain
‘The original definition for the crosscorrelation was for continuous signals. Thus if
h(e) and g(¢) are two continuous signals, then their crosscorrelation function is
defined as
con = J ght ddt (11.13)
‘The Fourier transform of the crosscorrelation function satisfies
Corr (@) = Glw)* Hw) (1.14)
where G(@)* is the complex conjugate of G(@).
Thus if we consider h and g to be digitized, we may approximate the crosscorre-
lation function as
N-1
cant) = 5, 3) Mn) (11.15)
i
‘This equation is also known as the biased estimator of the crosscorrelation func-
tion, Between the DFTs of the two input discrete signals and the DFT of the biased
estimator, we have the relationship
Corr (k) = G(k)* H(k) (11.16) Other Time- and Frequency-Domain Techniques 225
eta yA
(a) (b)
Figure 11.6 Respiratory signals recorded using impedance pneumography. (a) From top to bot-
tom, signals 1, 2, 3, and 4 were simultaneously recorded along the midaxillary line of a human
subject using & sampling rate of 5 sps. (b) From top to bottom, the traces represent the
crosscorrelation of the channels 1-2, 1-3, 1-4, 2-3, 2-4, 3-4, and the averaged correlation
coefficient. The results show, in part, that channels 2 and 3 are highly correlated during normal
breathing without motion, but not well correlated during motion without breathing.
Thus, the crosscorrelation of the two discrete signals can also be computed as the
inverse DFT of the product of their DFTs. This can be implemented using the FFT
and inverse FFT (IFFT) algorithms for increasing the computational speed in the
way given by
cgh(n) = IFFT ( FFT*(g) * FFT(h)) (11.17)
11.2.3 Correlation function
‘The C-language program in Figure 11.7 computes the crosscorrelation function of
two 512-point input sequences x[{512] and y[512] and stores the output data into
rxy{512]. The idea of this program was used to implement the C-language function 226 Biomedical Digital Signal Processing
to compute the crosscorrelation between an ECG signal and a template. In such a
case, the array yl ] must have the same dimension as the template.
7* The crosscorrelation function of x1] and y(] is */
/* output into rxy[ ] */
void corr (float *x,float *y)
(
int i,m,nz
float s,sl,s2,xm,ym,t;
float rxy(512)7
12;
s=sl=s2=xm=ym=0.0;
4=0;i Sminutes Sminutes <|
Cycles per minute phe N oul
(a) (b)
Figure 11.11 Electrogastrogram (EGG) of a diabetic patient. (a) Compressed spectral array
(CSA). (b) Gray-scale plot.
11.4.5 Power spectrum function
Figure 11.12 is a C-language function that computes the power spectrum for an N-
point data sequence (N= 512). The input data are taken from an input buffer as
integers, converted to floating-point format, and then used to compute the power
spectrum. The output data are scaled and saved in an output file as integers after
conversion from the floating-point format. The program presented is based on
Welch’s idea of periodogram averaging, Thus, this function divides the input se-
quence of data into two 256-point segments, windows each segment, and performs
the power spectrum estimation for each segment. Finally, the results obtained for
each segment are averaged and scaled according to the window effects.
11.5 LAB: FREQUENCY-DOMAIN ANALYSIS OF THE ECG
This lab provides experience in studying the frequency characteristics of signals.
Another resource for practice with these techniques is Alkin (1991).
11.5.1 Power spectral analysis of periodic signals
Use (G) enwave to create sine wave and square wave signals at several different
fundamental frequencies using several different sampling rates. Use the
(2) wr Spect command to compute and observe the frequency spectrum of each 234 Biomedical Digital Signal Processing
signal and comment on how sampling rate affects the results. How could you use a
power spectrum routine to obtain the actual Fourier coefficients?
11.5.2 Power spectral analysis of an ECG
1, Use the (P)wx Spect command to find the frequency corresponding to the
main peak in the frequency spectrum of an ECG. Is the result what you expected?
2. Select ad(v) ops from the main menu, and choose a QRS complex using
(1) emplate. Use the (P)wr Spect command to find the spectrum of the zero-
padded QRS complex. Is the frequency corresponding to the main peak in the fre-
quency spectrum of the QRS complex what you expected?
3. After selecting a QRS complex template, use the (w) indow command to win-
dow the selected template, and then run the (P) wx Spect command. Document the
effects of the various windows.
4, Select a P-QRS-T segment as a template. Using the (p)wr spect command,
find the power spectrum estimate for an ECG. Find the frequency corresponding to
the main peak in the frequency spectrum of an ECG. How does it differ from the
results of part 1?
¥define WINDOW (J, a,b) (1.0-fabs((((3)-1)- (al) *(b}))/*paraen */
define SQR(a) (sqrarg=(a), sqrarg*sararg) /*modulus of a, squared*/
for (j=1; j<=mms 34+) (
INDOW (3, facm, facp) ;
w1(2*j-2] *sw; /* the real part of then data segment
is windowed */
0; /*the imaginary part is
set to zero */
wi(2*j-1] =
!
fft(wi-1,mm,1); /*the fft of the windowed signal is performed*/
/* the power spectrum estimate for the windowed data segment
is computed */
p[0] SQR(w1 [0] +SOR(w1(1]))¢
for (j=l; j
