9

Signal Averaging
Pradeep Tagare

Linear digital filters like those discussed in previous chapters perform very well
when the spectra of the signal and noise do not significantly overlap. For example,
a low-pass filter with a cutoff frequency of 100 Hz generally works well for attenuating noise frequencies greater than 100 Hz in ECG signals. However, if highlevel noise frequencies were to span the frequency range from 50100 Hz, attempting to remove them using a 50Hz low-pass filter would attenuate some of the
components of the ECG signal as well as the noise. High-amplitude noise corruption within the frequency band of the signal may completely obscure the signal.
Thus, conventional filtering schemes fail when the signal and noise frequency
spectra significantly overlap. Signal averaging is a digital technique for separating
a repetitive signal from noise without introducing signal distortion (Tompkins and
Webster, 1981). This chapter describes the technique of signal averaging for
increasing the signal-to-noise ratio and discusses several applications.
9.1 BASICS OF SIGNAL AVERAGING
Figure 9.1(a) shows the spectrum of a signal that is corrupted by noise. In this case,
the noise bandwidth is completely separated from the signal bandwidth, so the
noise can easily be discarded by applying a linear low-pass filter. On the other
hand, the noise bandwidth in Figure 9.1(b) overlaps the signal bandwidth, and the
noise amplitude is larger than the signal. For this situation, a low-pass filter would
need to discard some of the signal energy in order to remove the noise, thereby distorting the signal.
One predominant application area of signal averaging is in electroencephalography. The EEG recorded from scalp electrodes is difficult to interpret in part because it consists of a summation of the activity of the billions of brain cells. It is
impossible to deduce much about the activity of the visual or auditory parts of the
brain from the EEG. However, if we stimulate a part of the brain with a flash of
light or an acoustical click, an evoked response occurs in the region of the brain
that processes information for the sensory system being stimulated. By summing
184

Signal Averaging

185

the signals that are evoked immediately following many stimuli and dividing by
the total number of stimuli, we obtain an averaged evoked response. This signal
can reveal a great deal about the performance of a sensory system.
Signal averaging sums a set of time epochs of the signal together with the superimposed random noise. If the time epochs are properly aligned, the signal waveforms directly sum together. On the other hand, the uncorrelated noise averages out
in time. Thus, the signal-to-noise ratio (SNR) is improved.
Signal averaging is based on the following characteristics of the signal and the
noise:

Signal

Noise

Frequency

(a)

Amplitude

Amplitude

1. The signal waveform must be repetitive (although it does not have to be periodic). This means that the signal must occur more than once but not necessarily
at regular intervals.
2. The noise must be random and uncorrelated with the signal. In this application,
random means that the noise is not periodic and that it can only be described
statistically (e.g., by its mean and variance).
3. The temporal position of each signal waveform must be accurately known.

Noise
Signal

Frequency

(b)

Figure 9.1 Signal and noise spectra. (a) The signal and noise bands do not overlap, so a conventional low-pass filter can be used to retain the signal and discard the noise. (b) Since the signal
and noise spectra overlap, conventional filters cannot be used to discard the noise frequencies
without discarding some signal energy. Signal averaging may be useful in this case.

It is the random nature of noise that makes signal averaging useful. Each time
epoch (or sweep) is intentionally aligned with the previous epochs so that the digitized samples from the new epoch are added to the corresponding samples from the
previous epochs. Thus the time-aligned repetitive signals S in each epoch are added
directly together so that after four epochs, the signal amplitude is four times larger
than for one epoch (4S). If the noise is random and has a mean of zero and an average rms value N, the rms value after four epochs is the square root of the sum of
squares (i.e., (4N2)1/2 or 2N). In general after m repetitions the signal amplitude is
mS and the noise amplitude is (m) 1/2N. Thus, the SNR improves as the ratio of m to

186

Biomedical Digital Signal Processing

m1/2 (i.e., m1/2). For example, averaging 100 repetitions of a signal improves the
SNR by a factor of 10. This can be proven mathematically as follows.
The input waveform f(t) has a signal portion S(t) and a noise portion N(t). Then
f(t) = S(t) + N(t)

(9.1)

Let f(t) be sampled every T seconds. The value of any sample point in the time
epoch (i = 1, 2,, n) is the sum of the noise component and the signal component.
f(iT) = S(iT) + N(iT)

(9.2)

Each sample point is stored in memory. The value stored in memory location i
after m repetitions is
m

k=1

f(iT) =

S(iT) +

k=1

N(iT)

for i = 1, 2, , n

(9.3)

k=1

The signal component for sample point i is the same at each repetition if the signal
is stable and the sweeps are aligned together perfectly. Then
m

S(iT)

= mS(iT)

(9.4)

k=1

The assumptions for this development are that the signal and noise are uncorrelated
and that the noise is random with a mean of zero. After many repetitions, N(iT) has
an rms value of n.
m

N(iT)

m n2 = m

(9.5)

k=1

Taking the ratio of Eqs. (9.4) and (9.5) gives the SNR after m repetitions as
SNRm =

mS(iT)
= m SNR
m n

(9.6)

Thus, signal averaging improves the SNR by a factor of m . Figure 9.2 is a graph
illustrating the results of Eq. (9.6).

Signal Averaging

187

Increase in SNR

100

10

1
0

100

200

300

400

500

Number of sweeps

Figure 9.2 Increase in SNR as a function of the number of sweeps averaged.

Figure 9.3 illustrates the problem of signal averaging. The top trace is the ECG
of the middle trace after being corrupted by random noise. Since the noise is
broadband, there is no way to completely remove it with a traditional linear filter
without also removing some of the ECG frequency components, thereby distorting
the ECG. Signal averaging of this noisy signal requires a way to time align each of
the QRS complexes with the others. By analyzing a heavily filtered version of the
waveform, it is possible to locate the peaks of the QRS complexes and use them
for time alignment. The lower trace shows these timing references (fiducial points)
that are required for signal processing.
Figure 9.4 shows how the QRS complexes, centered on the fiducial points, are
assembled and summed to produce the averaged signal. The time-aligned QRS
complexes sum directly while the noise averages out to zero. The fiducial marks
may also be located before or after the signal to be averaged, as long as they have
accurate temporal relationships to the signals.
One research area in electrocardiography is the study of late potentials that require an ECG amplifier with a bandwidth of 500 Hz. These small, high-frequency
signals of possible clinical significance occur after the QRS complex in body surface ECGs of abnormals. These signals are so small compared to the other waveforms in the ECG that they are hidden in the noise and are not observable without
signal averaging. In this application, the fiducial points are derived from the QRS
complexes, and the averaging region is the time following each QRS complex.

188

Biomedical Digital Signal Processing

Figure 9.3 The top trace is the ECG of the center trace corrupted with random noise. The
bottom trace provides fiducial marks that show the locations of the QRS peaks in the signal.

(a)

(b)

(c)

(d)
Figure 9.4 Summing the time-aligned signal epochs corrupted with random noise such as those
in (a), (b), and (c), which were extracted from Figure 9.3, improves the signal-to-noise ratio. The
result of averaging 100 of these ECG time epochs to improve the SNR by 10 is in (d).

9.2 SIGNAL AVERAGING AS A DIGITAL FILTER

Signal Averaging

189

Amplitude

Signal averaging is a kind of digital filtering process. The Fourier transform of the
transfer function of an averager is composed of a series of discrete frequency components. Figure 9.5 shows how each of these components has the same spectral
characteristics and amplitudes. Because of the appearance of its amplitude response, this type of filter is called a comb filter.
The width of each tooth decreases as the number of sweep repetitions increases.
The desired signal has a frequency spectrum composed of discrete frequency components, a fundamental and harmonics. Noise, on the other hand, has a continuous
distribution. As the bandwidth of each of the teeth of the comb decreases, this filter
more selectively passes the fundamental and harmonics of the signal while rejecting the random noise frequencies that fall between the comb teeth. The signal averager, therefore, passes the signal while rejecting the noise.

Frequency

Figure 9.5 Fourier transform of a signal averager. As the number of sweeps increase, the width
of each tooth of the comb decreases.

9.3 A TYPICAL AVERAGER

Figure 9.6 shows the block diagram of a typical averager. To average a signal such
as the cortical response to an auditory stimulus, we stimulate the system (in this
case, a human subject) with an auditory click to the stimulus input.
Simultaneously, we provide a trigger derived from the stimulus that enables the
summation of the sampled data (in this case, the EEG evoked by the stimulus) with
the previous responses (time epochs or sweeps) stored in the buffer. When the
averager receives the trigger pulse, it samples the EEG waveform at the selected
rate, digitizes the signal, and sums the samples with the contents of a memory
location corresponding to that sample interval (in the buffer). The process
continues, stepping through the memory addresses until all addresses have been
sampled. The sweep is terminated at this point. A new sweep begins with the next
trigger and the cycle repeats until the desired number of sweeps have been
averaged. The result of the averaging process is stored in the buffer which can then
be displayed on a CRT as the averaged evoked response.

190

Biomedical Digital Signal Processing

Stimulus

System

Response after
time t

Trigger
Sum of sweeps

Enable

Buffer

Figure 9.6 Block diagram of a typical signal averager.

9.4 SOFTWARE FOR SIGNAL AVERAGING

Figure 9.7 shows the flowchart of a program for averaging an ECG signal such as
the one in Figure 9.3. The program uses a QRS detection algorithm to find a fiducial point at the peak of each QRS complex. Each time a QRS is detected, 128 new
sample points are added to a buffer64 points before and 64 points after the fiducial point.
9.5 LIMITATIONS OF SIGNAL AVERAGING
An important assumption made in signal averaging theory is that the noise is
Gaussian. This assumption is not usually completely valid for biomedical signals.
Also, if the noise distribution is related to the signal, misleading results can occur.
If the fiducial point is derived from the signal itself, care must be taken to ensure
that noise is not influencing the temporal location of the fiducial point. Otherwise,
slight misalignment of each of the signal waveforms will lead to a low-pass filtering effect in the final result.

Signal Averaging

191

Clear the screen

Initialize: buffer 'sum' = 0

Read sample data from

noisy
reference
datadata
buffer
buffer

QRS?
No

Yes

Add 64 points before and

after fiducial point from
the noisy buffer to the
'sum' buffer

Display the averaged

sweep on the screen

Increment 'number' of
sweeps

No

'Number'
of
sweeps
= 256?

Yes

Exit

Figure 9.7 Flowchart of the signal averaging program.

192

Biomedical Digital Signal Processing

9.6 LAB: ECG SIGNAL AVERAGING

Load UW DigiScope, select ad(V) Ops, then A(V)erage . This module is a
demonstration of the use of signal averaging. The source file is always average.dat to which random, gaussian-distributed noise is added. A clean version of
the data is preserved to use as a trigger signal. The averaged version of the data,
displayed in the output channel, will build in size as successive waveforms are
added. A summation of the accumulated signal traces is displayed, not the average.
Thus you need to scale down the amplitude of the resultant signal in order to see
the true average at the same amplitude scale factor as the original signal. Scaling of
the output channel can be controlled as the traces are acquired. The down arrow
key divides the amplitude by two each time it is struck. For example, while adding
16 heartbeats, you would need to strike the down arrow four times to divide the
output by 16 so as to obtain the proper amplitude scale.
9.7 REFERENCES
Tompkins, W. J. and Webster, J. G. (eds.) 1981. Design of Microcomputer-based Medical
Instrumentation. Englewood Cliffs, NJ: Prentice Hall.

9.8 STUDY QUESTIONS

9.1
9.2
9.3
9.4
9.5

Under what noise conditions will signal averaging fail to improve the SNR?
In a signal averaging application, the amplitude of uncorrelated noise is initially 16 times as
large as the signal amplitude. How many sweeps must be averaged to give a resulting
signal-to-noise ratio of 4:1?
After signal averaging 4096 EEG evoked responses, the signal-to-noise ratio is 4.
Assuming that the EEG and noise sources are uncorrelated, what was the SNR before
averaging?
In a signal averaging application, the noise amplitude is initially 4 times as large as the signal amplitude. How many sweeps must be averaged to give a resulting signal-to-noise ratio
of 4:1?
In a signal averaging application, the signal caused by a stimulus and the noise are slightly
correlated. The frequency spectra of the signal and noise overlap. Averaging 100 responses
will improve the signal-to-noise ratio by what factor?