Tentin tekstisisältö

= TAMPEREEN TEKNILLINEN YLIOPISTO

LTT - 4106 - Processing of Physiological Signals
Exam 14.05.2007
Answer the guestions in English or in Finnish. For each guestion a maximum of 6 points can

be earned (thus: 5 * 6 = 30 points in total). Possible points from the exercises will be added to
these points.

 

1 Basics.

a) Define the term and give an example of a biomedical signal, which may be considered to
be (at least almost):

1) Deterministic (1p)
2) Stochastic (1p)

b) While developing biomedical signal processing methods one usually needs to collect a
representative set of experimental data to be used in optimization of the methods. In
which terms the experimental data needs to be representative and why? (2p)

c) EEG is sampled at 200Hz. You use DFT to transform the signal into the freguency
domain. How long a data window you need (in seconds) to have a freguency resolution of
0.5HZz? (1p)

d) The recorded EEG signal is found to have low entropy. Define what this intuitively
means. (1p)

 

2 Home health monitoring -

A home health monitoring system consists of a weight scale, a noninvasive blood pressure
meter and a beat-to-beat heart rate meter. The subjects are instructed to measure their weight
every morning, blood pressure every morning and evening, and heart rate continuously during
the awakening time. The data are recorded for one year. The data are stored and transferred
automatically to a database. Your challenge is to analyze the data and calculate the correlation
between the signals and to identify possible regular rhythms such as a week rhythm in the
signals.

a) List four typical special challenges for signal analysis in the home health monitoring
setting (such as described above). Provide an example how each of these challenges
may occur (e.g. by using the setting described above). (4p)

b) Describe a possible strategy for estimating the correlation between heart rate and
blood pressure signals. (1p)

c) Describe two possible strategies for estimating the power spectral density of the blood
pressure signal. (1p)

 

3 Artefacts and noise
For the application of recording EEG signals, name and describe one potential source or cause
of each of the following types of an artefact:
a) high-freguency noise,
b) periodic artefact, and
c) ashort-lasting, abrupt, artefact.
In each case, explain (in detail):
1) what the cause of the artefact could be, and
2) how you could remove or prevent the occurrence of the artefact. (6p)

 

1/2
j- TAMPEREEN TEKNILLINEN YLIOPISTO

4 Artificial neural networks and wavelets

a) In solving biomedical signal processing problems, use of artificial neural networks
(ANNS) is one often-used method. For what kind of problems ANNS are typically
useful for? (think, for example, of the type of problem to be solved, and properties of
available data) (2 p)

b) Another popular technigue concerns the use of wavelets to describe biomedical
signals. What are the differences in freguency resolution of wavelet and Fourier
representations? (2 p)

c) Give 2 examples in which the description of a biomedical signal using wavelets may
be more appropriate than using a Fourier representation, and explain why wavelets are
more appropriate. (2 p)

 

5 Decision support systems and performance evaluation.

When building a system to automatically detect a condition of 'cardiac failure' in an intensive

care unit we have knowledge available about how a clinician may decide whether a patient

state of 'cardiac failure' is present or whether the patient is 'normal'. In this example, he makes

a decision on the basis of recorded signals 'Cardiac Index (C1)', Filling Pressure (FP)', Body

Temperature (BT)? and 'Urine Output (UO)'. His knowledge is summarised as follows

The patient has cardiac failure, if:

e The patient has low Cardiac Index (CI < 2.0) and high filling pressure (FP > 10)
e (optional) The patient has a low body temperature (T < 32C)
e (optional) The patient has low urine output (less than 0.5 - 1.0 ml/kg/h, over 2h)

a) With this knowledge available, if you had to implement a system recognizing the cardiac
failure, what kind of pattern recognition / classification system would you use in first
instance? Why? (2p)

b) Suppose you have built a system that automatically classifies recorded data into 'Cardiac
Failure' or 'Normal' using the knowledge above and you test it with a test data set. You get
the following confusion matrix:

 

true patient state | Normal Cardiac Failure

patient state according to
developed system

 

Normal
Cardiac Failure

 

 

 

 

 

What are the sensitivity, specificity, accuracy, and positive and negative prediction values
of this system? (2p)

c) From the confusion matrix in b) it appears your system is not perfect, although you have
developed it with help from a clinician and used his knowledge in it. Give two reasons
why the performance of the system may be less than optimal. (2p)

[END]

2/2