Biomedical Engineering Reference
In-Depth Information
only selecting a “smart material,” but requires an integrated system that is greater than the
sum of the individual parts. Sophisticated biosensor designs now incorporate hardware and
software technologies that enable adaptation, learning from past experience, and stimulus
identification. These systems draw from the disciplines of computer science and electrical
engineering. The computational algorithms may be embedded in physical sensors using pro-
grammable microcontrollers or implemented off-line on computer workstations where the
captured information is graphically displayed for human analysts.
In this chapter, the role of machine learning and how it can be applied to biosensor cal-
ibration and pattern classification is discussed from an engineering and informatics per-
spective. In this context, artificial neural networks (ANNs) are introduced as a viable AI
tool for analyzing large volumes of multidimensional data from single sensors or sensor
arrays. Finally, the computational mechanism of two commonly used ANNs, the radial
basis function (RBF) network and the self-organizing feature map (SOFM), are summa-
rized and their role in enabling smart functions is illustrated.
5.2
The Role of Machine Learning in Developing Smart Biosensor Functions
The discipline of
machine learning
is concerned with extracting meaningful information
from sensor data using statistical, analytical, and connectionist algorithms. The various
algorithms used to perform machine learning tasks can be organized into a taxonomy that
is based on the desired system output. These learning algorithms include: supervised,
unsupervised, semisupervised, reinforcement, transduction, inductive transfer, and
approximate inference techniques. Pattern recognition is a subarea of research within
machine learning that seeks to classify data (patterns) based on either a priori information
or on statistical information extracted directly from the dataset.
The functional capabilities of an individual biosensor or a sensor array can be signifi-
cantly enhanced by taking advantage of pattern recognition algorithms to help categorize
and analyze the large number of features extracted from the acquired signals. Pattern
analysis involves sensing biomolecules from one or more sites, preprocessing the acquired
transducer data, extracting relevant features that enable classification, clustering features
into groups with similar characteristics called classes, and interpreting the underlying
meaning of the final class assignment (8).
5.2.1
Pattern Recognition
Pattern recognition is the process of classifying the analyte, detecting chemical composi-
tion, or estimating the pattern's descriptive parameters (9). The goal of
classification
is to
assign the sensor signal pattern, or feature vector, to a predefined class.
Detection
strives to
determine the temporal or spatial occurrence of a specific instance of the desired pattern.
In contrast,
parameter estimation
involves determining specific parameters that assist the
analyst in defining a given pattern. In large datasets, parameter estimation is used to either
remove noise and redundant information in the original data or measure the similarity
between patterns. Data compression and data fusion algorithms are often used to reduce
the amount of data vectors for interpretation or analysis, while pattern discrimination
techniques help establish measures of similarity between two or more patterns.
An essential step in pattern analysis is the transformation of measured signals from the
transducer to a limited selection of features that best represent meaningful patterns in the
signal (10). Feature vectors (also called patterns) contain all salient information necessary
to describe the input analyte or chemical substance. Statistical and syntactic approaches
