Biomedical Engineering Reference
In-Depth Information
or characteristic pattern of the odorant. By presenting different odorants to the sensor array
it is possible to build up a database of signatures. The goal of many of these systems is to train
or configure the pattern recognition system to produce unique clusters or classifications of
each odorant so that automated identification can occur. In this manner, ANNs have become
a significant tool in developing adaptive, intelligent electronic noses.
Among the many applications of electronic noses where neural network pattern classi-
fication has been used are in the area of food analysis (34), including determining the qual-
ity of wheat (35), monitoring the ripeness of bananas (36) and other fruits (37), evaluating
the freshness of eggs (38), and identifying spoiled beef (39).
An aroma identification system based on an array of semiconductor tin dioxide gas sen-
sors and a feedforward neural network to discriminate the five different aromatic species
is described by Brezmes et al. (40). Daqi et al. (41) use an electronic nose with modular RBF
network classifiers to recognize multiple fragrant materials. An electronic nose in combi-
nation with a neural network to detect the presence of
Mycobacterium tuberculosis
in spu-
tum samples of patients is presented by Pavlou et al. (42). The system is developed for
rapid qualitative analysis for screening patient samples and the clinical diagnosis of TB
patients. Dutta et al. (43) use an array of 32 polymer carbon black composite sensors and
several neural networks to help identify two species of
Staphylococcus aureus
bacteria. The
object-oriented data clustering approach is a combination of several techniques including
principal component analysis, fuzzy C-means, and the SOFM.
5.3
Biosensor Data Analysis Using Artificial Neural Networks
This section will explore how artificial neural networks can be used to enhance sensor per-
formance and assist the scientist with complex data analysis by first determining the non-
linear relationship between experimentally observed biosensor inputs and outputs, and
then performing pattern classification and multidimensional data visualization tasks. The
RBF network is used to determine a calibration standard for a simple bR photocell. A vari-
ant unsupervised clustering network called the SOFM is used to partition and classify high-
dimensional medical data. The concept of self-organization is extended to the task of
associating patterns from several sensors. Finally, the process of mapping high-dimensional
data vectors onto a colorized deformable spherical SOFM is described as mechanism for
visualizing complex numeric datasets and searching for hidden patterns.
5.3.1
Sensor Calibration by Functional Approximation
The signals generated by the biosensor transducer must be linked or correlated in some
manner with the biochemical identifier (target) to generate predictable and reproducible
results with only small deviations from the expected values. The target may be the pres-
ence or absence of an analyte, the concentration level of analytes, or a specific chemical
composition. The relationship between the transducer signal (output) and biochemical
identifier (input) is often described as a
calibration standard
or
function
(44). The process of
biosensor calibration may be viewed as developing a mapping function between the
inputs and outputs. Typically, the calibration model is mathematically represented as
y
F
(
x
,
w
)
(5.1)
where
y
is the sensor output vector,
x
is the vector that represents the inputs to the biosensor
(e.g., biochemical identifier), and
w
is the state vector representing the internal parameters of
