Biomedical Engineering Reference
In-Depth Information
are often used to extract suitable features from the transducer output. These techniques
use various criteria for identifying which features are best for the classification. An opti-
mal feature selection technique must satisfy some criterion that minimizes a measure of
representation error. In addition, the feature vector used to describe the pattern produced
by the signal(s) is often a reduced dimensional representation. The primary objective of
dimensionality reduction is to satisfy practical constraints imposed by the computing soft-
ware and hardware. Furthermore, the classification result is often more accurate when the
pattern embedded in the raw signal is simplified through reduction with only the most
important or critical features being represented. The choice and number of required fea-
tures for a classification task are problem specific and governed by the level of permissi-
ble information loss.
Classification algorithms are used to assign the extracted feature vectors to prede-
fined classes (11). A class is a set of feature vectors that have been grouped based on a
similarity measure. The metric used to define “similarity” depends on the objective of
the classification task. Often the distribution of classes within a feature space is deter-
mined using a clustering algorithm. Most pattern classification algorithms are statisti-
cal, syntactic, or connectionist. An example of a statistical algorithm is the minimum
distance classifier, which assigns a feature vector to a class based on its Euclidean dis-
tance from the class prototypes located in the feature space. Other approaches involve
linear discriminate operators, entropy criteria methods, and maximum likelihood esti-
mation classifiers.
Syntactic techniques, in contrast, classify feature vectors based on formal grammars and
symbolic descriptors. The low-level pattern structures, called primitives, correspond to
simple features in the sensor signal. Each primitive may require multiple attributes
defined by statistical metrics to describe it. The sensor signal is decomposed into primi-
tives by means of formal grammars. Each class of patterns is represented by a distinct
grammar that can be expressed in symbolic form as sentences. Two basic approaches to
performing syntactic pattern recognition are to define distinct grammars that reflect class
structure, and use graphs to match relational descriptions.
Connectionist algorithms are largely ANN approaches that establish classes based on
training a densely interconnected network of simple computing units called neurons (12).
These approaches are preferred in pattern analysis applications where there only exists a
minimal amount of a priori class information. The network parameters are adjusted by
presenting a collection of input-output examples and iteratively adjusting the internal
parameters to provide the desired output. Connectionist approaches can efficiently deter-
mine nonlinear class boundaries in multidimensional feature spaces.
5.2.2
Artificial Intelligence and Neural Networks
Artificial intelligence has emerged as a major tool for solving complex problems arising
from ill-defined or massive amounts of data. Common AI tools include Boolean logic,
deductive and inductive reasoning, soft or fuzzy logic, expert systems, knowledge-base
systems, genetic algorithms, support vector machines, and ANNs. Neural network struc-
tures (12-15) have gained popularity in recent decades as a viable mechanism for embed-
ding computational intelligence into sensors and enabling the analysis of large-volume
high-dimensional datasets. A neural network is defined as a computing system made up
of numerous simple, highly interconnected processing elements which process informa-
tion by responding to external inputs. The various classes of ANNs are identified by the
type of neuron (static, dynamic, Gaussian), structure of the network (feed forward, feed-
back, clustering map), and the learning algorithm (supervised, unsupervised, competi-
tive) employed. A large number of neural networks have been described in the literature
