Biomedical Engineering Reference
In-Depth Information
investigated. An important positive attribute of the technique is that no assumption is
made about the underlying data structure. If there is no correlation between the features
in the dataset then the method results in as many cluster units as feature vectors.
5.3.2.3 Pattern Association
A single biosensor is largely used to determine the presence of a single chemical com-
pound or the concentration level of one substance contained in a sample. This approach to
sensor design ensures high specificity for each detector but, unfortunately, provides low
analyte selectivity. The problem of low selectivity is especially prevalent in biosensors that
utilize whole cells, organelles, or tissues. To compensate for these limitations it is possible
to use
N
-sensors to detect
M
-substances. The information can originate from different sen-
sory devices during a single period of time, or from a single bioreceptor over an extended
period of time. The synergistic use of data acquired from several very selective and highly
sensitive biological and chemical detectors is the basic design principle behind many elec-
tronic noses (31,32) and tongues (53).
A variety of mathematical and statistical techniques can be used to “optimally” combine
different sources of sensory information into a single representational format (54,55). The
process, also known as data fusion, can take place either at the signal, object property, or
symbolic level. Signal-level fusion can be used in real-time applications and can be treated
as another step in the overall processing of the signal. Property and symbol-level fusion
can provide additional features to enhance recognition and interpretation capabilities. The
different levels are distinguished by the type of information provided to the system, how
the information is modeled, the degree of the sensor registration required for fusion, the
methods used for fusion, and the means by which the fusion process can improve the
“quality” of the information provided to the observer or the automated analyzer.
Two important advantages of using information from multiple biosensors include
data
redundancy
and
complementary information
. Combining redundant data from several simi-
lar bioreceptors increases system accuracy and reliability by reducing the uncertainty
associated with the operating fidelity of any single sensor. In addition, different types of
sensors monitoring the same physical phenomenon can provide complementary informa-
tion that is difficult or impossible to acquire with a single type of sensor. The complemen-
tary information can either represent redundant data for enhanced system reliability or
new information used to increase the analyte discrimination capabilities. One reoccurring
problem in integrating redundant or complementary information from multiple biosen-
sors is that of associating features, or signatures, acquired by one biological element with
those simultaneously captured by another sensor or the same sensor at another instance
of time (Table 5.1).
A constraint imposed on feature association is that corresponding feature elements from
the various sensors represent the same type of information. The match matrix is often gen-
erated using an optimization technique such as genetic algorithms, simulated annealing,
competitive neural networks, or the SOFM (56). The key issue in these approaches has
been to maximize data compression while retaining relevant information on the novel
input. In essence, the task involves finding a more abstract, compressed, feature set by
going from one representation level to another. Evidence of hidden redundancy and meas-
ures of similarity are used to guide the feature recognition algorithms. The Kohonen
SOFM is one method that has been used for dimensionality reduction and multisensor
correlation. Figure 5.10 is an illustration of a multiple biosensor system that uses a 3D
SOFM (56) to correlate the features from three separate detectors. The SOFM has been
widely used for classification purposes and several researchers have presented ways in
which the correlated data is presented on the SOFM to reflect the classification and the
