Biomedical Engineering Reference
In-Depth Information
including adaptive linear neuron, multilayered perceptron, backpropagation network,
RBF network, SOFM, Hopfield net, bidirectional associative memory, Boltzmann machine,
adaptive resonance theory, and neocognitron to name a few.
The functional capabilities of most neural networks used in practice can be viewed from
three different perspectives (14). The first is that the fundamental feedforward network
structure can be trained to implement any Boolean logic function. The second perspective
is that the same network architecture can approximate any nonlinear function and, there-
fore, be used to establish a nonlinear mapping between sensor inputs and outputs. The
third perspective is that the neural network can learn to partition multidimensional input
(or feature) spaces for pattern classification problems. In terms of biosensor data, functional
approximation and pattern classification are the two main applications of neural networks.
In general, the methodology does not require explicit knowledge of the network struc-
ture and permits the internal system parameters to be learned through a process of repet-
itive exposure to experimentally acquired examples of input-output vector pairs. Most
neural network algorithms separate the training phase from the final operation phase. It
is during the training phase that the system develops an acceptable mapping function or
data partitions. The major drawbacks are the model structure hidden in a “black box” and
the computational time necessary to optimize the parameters through training.
Neural networks can be applied to adaptive signal processing and control (16), or used
to perform complex chemical analysis and analyte identification (17). The most direct
application of these nonlinear maps to analytical chemistry involves creating nonlinear,
multivariate calibration models where the spectral intensities at different wavelengths are
used as the input variables and the corresponding concentrations of the analyte species in
the same sample are the output attributes.
Simple feedforward neural networks have been used successfully to establish a calibra-
tion standard for a number of sensors including an optical fiber pH sensor (18-20). The
network implemented by Taib and Narayanaswamy (18) was tested with noisy data and
produced an average error of 0.2 pH units for additive noise levels of up to 13% of the
input signal. Raimundo and Narayanaswamy (21) used a similar neural network to per-
form multivariate calibration on an optical fiber chemical sensor for determining the
ammonia and relative humidity in air. Ferreira et al. (22) developed an alcohol fermenta-
tion control system based on biosensor measurements interpreted by feedforward neural
networks. The neural network established a correlation between the experimentally
acquired glucose and sucrose measurements. The nonlinear mapping determined by a
similar training procedure was also used to successfully model the forward and inverse
dynamics of a nonisothermic, continuously stirred tank reactor (23).
The pattern classification capability of neural networks has also been extensively
studied for analyzing large volumes of sampled data or high-dimensional data vectors. A
multipoint optical evanescent wave U-bend sensor that exploits a network pattern recogni-
tion system was described by Lyons et al. (24). The signals generated from the optical time
domain reflectometer sensing fiber was complex due to the cross-coupling effects arising
from the various interfering parameters. However, the authors determined that the neural
network used to perform the signal analysis resulted in 100% correct classification for all test
data analyzed. Blank and Brown (25) also investigated how neural networks can be applied
to problems in analytical chemistry by comparing the performance of a feedforward
network with more well-established chemometric techniques that exploit matrix regression
methods. Bachmann et al. (26) also used a neural network approach to perform chemomet-
ric data analysis on the outputs from an amperometric multielectrode biosensor.
The effectiveness of feedforward neural networks for identifying single molecules, accord-
ing to their fluorescence lifetime, was reported by Bowen et al. (27). For a nonideal single mol-
ecule fluorescence data sample, the neural network approach was performed better than the
