Micheli-Tzanakou, E. “ Neural Networks in Biomedical Signal Processing.

”
The Biomedical Engineering Handbook: Second Edition.
Ed. Joseph D. Bronzino
Boca Raton: CRC Press LLC, 2000
 58
Neural Networks
in Biomedical
Signal Processing

58.1 Neural Networks in Sensory Waveform Analysis

Multineuronal Activity Analysis • Visual Evoked Potential
58.2 Neural Networks in Speech Recognition
Evangelia Micheli- 58.3 Neural Networks in Cardiology
Tzanakou 58.4 Neural Networks in Neurology
Rutgers University 58.5 Discussion

Computing with neural networks (NNs) is one of the faster growing fields in the history of artificial
intelligence (AI), largely because NNs can be trained to identify nonlinear patterns between input and
output values and can solve complex problems much faster than digital computers. Owing to their wide
range of applicability and their ability to learn complex and nonlinear relationships—including noisy or
less precise information—NNs are very well suited to solving problems in biomedical engineering and,
in particular, in analyzing biomedical signals.
NNs have made strong advances in continuous speech recognition and synthesis, pattern recognition,
classification of noisy data, nonlinear feature detection, and other fields. By their nature, NNs are capable
of high-speed parallel signal processing in real time. They have an advantage over conventional technol-
ogies because they can solve problems that are too complex—problems that do not have an algorithmic
solution or for which an algorithmic solution is too complex to be found. NNs are trained by example
instead of rules and are automated. When used in medical diagnosis, they are not affected by factors
such as human fatigue, emotional states, and habituation. They are capable of rapid identification, analysis
of conditions, and diagnosis in real time.
The most widely used architecture of an NN is that of a multilayer perceptron (MLP) trained by an
algorithm called backpropagation (BP). Backpropagation is a gradient-descent algorithm that tries to minimize
the average squared error of the network. In real applications, the network is not a simple one-dimensional
system, and the error curve is not a smooth, bowl-shaped curve. Instead, it is a highly complex, multidimen-
sional curve with hills and valleys (for a mathematical description of the algorithm, see Chapter 182).
BP was first developed by P. Werbos in 1974 [1], rediscovered by Parker in 1982 [2], and popularized
later by Rummelhart et al. in 1986 [3]. There exist many variations of this algorithm, especially trying to
improve its speed and performance in avoiding getting stuck into local minima—one of its main drawbacks.
In my work, I use the ALOPEX algorithm developed by my colleagues and myself (see Chapter 182)
[4–10], and my colleagues and I have applied it in a variety of world problems of considerable complexity.
This chapter will examine several applications of NNs in biomedical signal processing. One- and two-
dimensional signals are examined.

© 2000 by CRC Press LLC

58.1 Neural Networks in Sensory Waveform Analysis
Mathematical analysis of the equations describing the processes in NNs can establish any dependencies
between quantitative network characteristics, the information capacity of the network, and the proba-
bilities of recognition and retention of information. It has been proposed that electromyographic (EMG)
patterns can be analyzed and classified by NNs [11] where the standard BP algorithm is used for
decomposing surface EMG signals into their constituent action potentials (APs) and their firing patterns
[12]. A system such as this may help a physician in diagnosing time-behavior changes in the EMG.
The need for a knowledge-based system using NNs for evoked potential recognition was described in
a paper by Bruha and Madhavan [13]. In this paper, the authors used syntax pattern-recognition algo-
rithms as a first step, while a second step included a two-layer perceptron to process the list of numerical
features produced by the first step.
Myoelectric signals (MES) also have been analyzed by NNs [14]. A discrete Hopfield network was used
to calculate the time-series parameters for a moving-average MES. It was demonstrated that this network
was capable of producing the same time-series parameters as those produced by a conventional sequential
least-squares algorithm. In the same paper, a second implementation of a two-layered perceptron was
used for comparison. The features used were a time-series parameter and the signal power in order to
train the perceptron on four separate arm functions, and again, the network performed well.
Moving averages have been simulated for nonlinear processes by the use of NNs [15]. The results
obtained were comparable with those of linear adaptive techniques.
Moody and colleagues [16] used an adaptive approach in analyzing visual evoked potentials. This
method is based on spectral analysis that results in spectral peaks of uniform width in the frequency
domain. Tunable data windows were used. Specifically, the modified Bessel functions Io – sin h, the gaus-
sian, and the cosine-taper windows are compared. The modified Bessel function window proved to be
superior in classifying normal and abnormal populations.
Pulse-transmission NNs—networks that consist of neurons that communicate with other neurons via
pulses rather than numbers—also have been modeled [7,17]. This kind of network is much more realistic,
since, in biological systems, action potentials are the means of neuronal communication. Dayhoff [18]
has developed a pulse-transmission network that can temporally integrate arriving signals and also display
some resistance to temporal noise.
Another method is optimal filtering, which is a variation of the traditional matched filter in noise
[19]. This has the advantage of separating even overlapping waveforms. It also carries the disadvantage
that the needed knowledge of the noise power spectral density and the Fourier transform of the spikes
might not be always available.
Principal-components analysis also has been used. Here the incoming spike waveforms are represented by
templates given by eigenvectors from their average autocorrelation functions [20]. The authors found that
two or three eigenvectors are enough to account for 90% of the total energy of the signal. This way each
spike can be represented by the coordinates of its projection onto the eigenvector space. These coordinates
are the only information needed for spike classification, which is further done by clustering techniques.

Multineuronal Activity Analysis

When dealing with single- or multineuron activities, the practice is to determine how many neurons (or
units) are involved in the “spike train” evoked by some sensory stimulation. Each spike in a spike train
represents an action potential elicited by a neuron in close proximity to the recording electrode. These
action potentials have different amplitudes, latencies, and shape configurations and, when superimposed
one each other, create a complex waveform—a composite spike. The dilemma that many scientists face
is how to decompose these composite potentials into their constituents and how to assess the question
of how many neurons their electrode is recording from. One of the most widely used methods is window
discrimination, in which different thresholds are set, above which the activity of any given neuron is
assigned, according to amplitude. Peak detection techniques also have been used [21]. These methods
perform well if the number of neurons is very small and the spikes are well separated. Statistical methods

© 2000 by CRC Press LLC

of different complexity also have been used [22-24] involving the time intervals between spikes. Each
spike is assigned a unique instant of time so that a spike train can be described by a process of time
points corresponding to the times where the action potential had occurred. Processes such as these are
called point processes, since they are characterized by only one number. Given this, a spike train can be
treated as a stochastic point process that may or may not be stationary. In the former case, its statistics
do not vary in the time of observation [25]. In the second case, when nonstationarity is assumed, any
kind of statistical analysis becomes formidable.
Correlations between spike trains of neurons can be found because of many factors, but mostly because
of excitatory or inhibitory interactions between them or due to a common input to both. Simulations
on each possibility have been conducted in the past [26]. In our research, when recording from the optic
tectum of the frog, the problem is the reversed situation of the one given above. That is, we have the
recorded signal with noise superimposed, and we have to decompose it to its constituents so that we can
make inferences on the neuronal circuitry involved. What one might do would be to set the minimal
requirements on a neural network, which could behave the same way as the vast number of neurons that
could have resulted in a neural spike train similar to the one recorded.
This is a very difficult problem to attack with no unique solution. A method that has attracted attention
is the one developed by Gerstein et al. [22,27]. This technique detects various functional groups in the
recorded data by the use of the so-called gravitational clustering method. Although promising, the analysis
becomes cumbersome due to the many possible subgroups of neurons firing in synchrony. Temporal
patterns of neuronal activity also have been studied with great interest. Some computational methods
have been developed for the detection of favored temporal patterns [28-31]. My group also has been
involved in the analysis of complex waveforms by the development of a novel method, the ST-scan method
[32]. This method is based on well-known tomographic techniques, statistical analysis, and template
matching. The method proved to be very sensitive to even small variations of the waveforms due to the
fact that many orientations of them are considered, as it is done in tomographic imaging. Each histogram
represents the number of times a stroke vector at a specific orientation is cutting the composite waveform
positioned at the center of the window. These histograms were then fed to a NN for categorization. The
histograms are statistical representations of individual action potentials. The NN therefore must be able
to learn to recognize histograms by categorizing them with the action potential waveform that they
represent. The NN also must be able to recognize any histogram as belonging to one of the “learned”
patterns or not belonging to any of them [33]. In analyzing the ST-histograms, the NN must act as an
“adaptive demultiplexer.” That is, given a set of inputs, the network must determine the correspondingly
correct output. This is a categorization procedure performed by a perceptron, originally described by
Rosenblatt [34]. In analyzing the ST-histograms, the preprocessing is done by a perceptron, and the error
is found either by an LMS algorithm [35] or by an ALOPEX algorithm [36,39,40].

Visual Evoked Potentials

Visual evoked potentials (VEPs) have been used in the clinical environment as a diagnostic tool for many
decades. Stochastic analysis of experimental recordings of VEPs may yield useful information that is not
well understood in its original form. Such information may provide a good diagnostic criterion in
differentiating normal subjects from subjects with neurological diseases as well as provide an index of
the progress of diseases.
These potentials are embedded in noise. Averaging is then used in order to improve the signal-to-
noise (S/N) ratio. When analyzing these potentials, several methods have been used, such as spectral
analysis of their properties [42], adaptive filtering techniques [43,44], and some signal enhancers, again
based on adaptive processes [45]. In this latter method, no a priori knowledge of the signal is needed.
The adaptive signal enhancer consists of a bank of adaptive filters, the output of which is shown to be a
minimum mean-square error estimate of the signal.
If we assume that the VEP represents a composite of many action potentials and that each one of these
action potentials propagates to the point of the VEP recording, the only differences between the various

© 2000 by CRC Press LLC

action potentials are their amplitudes and time delays. The conformational changes observed in the VEP
waveforms of normal individuals and defected subjects can then be attributed to an asynchrony in the
arrival of these action potentials at a focal point (integration site) in the visual cortex [36]. One can
simulate this process by simulating action potentials and trying to fit them to normal and abnormal
VEPs with NNs.
Action potentials were simulated using methods similar to those of Moore and Ramon [37] and Bell
and Cook [38]. Briefly, preprocessing of the VEP waveforms is done first by smoothing a five-point filter
that performs a weighted averaging over its neighboring points:

( ) [ ( ) ( ) ( ) ( ) ( )] 9
S n = F n − 2 + 2F n − 1 + 3F n + 2F n + 1 + F n + 2 (58.1)

The individual signals νj are modulated so that at the VEP recording site each νj has been changed in
amplitude am(j) and in phase ph(j). The amplitude change represents the propagation decay, and the
phases represent the propagation delays of signals according to the equation

() () [
v j i = am j ⋅ AP i − ph j ( )] (58.2)

For a specific choice of am(j) and ph(j), j = 1, 2, … , N, the simulated VEP can be found by

VEP = b + k ∑v
j =1
α
j (58.3)

where k is a scaling factor, b is a dc component, and α is a constant [39,40].

The ALOPEX process was used again in order to adjust the parameters (amplitude and phase) so that
the cost function reaches a minimum and therefore the calculated waveform coincides with the experi-
mental one.
The modified ALOPEX equation is given by

() ( ) () () ()
pi n = pi n − 1 + γ∆pi n ∆R n + µri n (58.4)

where pi(n) are the parameters at iteration n, and γ and µ are scaling factors of the deterministic and
random components, respectively, which are adjusted so that at the beginning γ is small and µ is large.
As the number of iterations increases, γ increases while µ decreases. The cost function R is monitored
until convergence has been achieved at least 80% or until a preset number of iterations has been covered.
The results obtained show a good separation between normal and abnormal VEPs. This separation is
based on an index λ, which is defined as the ratio of two summations, namely, the summation of
amplitudes whose ph(i) is less than 256 ms and the summation of amplitudes whose ph(i) is greater than
256 ms. A large value of λ indicates an abnormal VEP, while a small λ indicates a normal VEP.
The convergences of the process for a normal and an abnormal VEP are shown in Figs. 58.1 and 58.2,
respectively, at different iteration numbers. The main assumption here is that in normal individuals, the
action potentials all arrive at a focal point in the cortex in resonance, while in abnormal subjects there
exists as asynchrony of the arrival times. Maier and colleagues [41] noted the importance of source
localization of VEPs in humans in studying the perceptual behavior in humans. The optimization proce-
dure used in this section can help that task, since individual neuronal responses are optimally isolated.
One of the interesting points is that signals (action potentials) with different delay times result in composite
signals of different forms. Thus the reverse solution of this problem, i.e., extracting individual signals from
a composite measurement, can help resolve the problem of signal source localization. The multiunit

© 2000 by CRC Press LLC

 (a)

(b)

(c)

FIGURE 58.1 Normal VEP. (a) The

fitting at the beginning of the process,
(b) after 500 iterations, and (c) after
1000 iterations. Only one action poten-
tial is repeated 1000 times. The x axis is
×10 ms; the y axis is in millivolts.

© 2000 by CRC Press LLC

 (a)

(b)

(c)

FIGURE 58.2 Abnormal VEP. (a) Fit-

ting at t = 3 iterations, (b) after 500
iterations, and (c) after 2000 iterations.
One action potential was used 1000
times.

© 2000 by CRC Press LLC

recordings presented in the early sections of this paper fall in the same category with that of decomposing
VEPs. This analysis might provide insight as to how neurons communicate.

58.2 Neural Networks in Speech Recognition

Another place where NNs find wide application is in speech recognition. Tebelski et al. [46] have studied
the performance of linked predictive NNs (LPNNs) for large-vocabulary, continuous-speech recognition.
The authors used a six-state phoneme topology, and without any other optimization, the LPNN achieved
an average of 90%, 58%, and 39% accuracy on tasks with perplexity of 5, 11, and 402, respectively, which
is better than the performance of several other simpler NNs tested. These results show that the main
advantages of predictive networks are mainly that they produce nonbinary distortion measures in a simple
way and that they can model dynamic properties such as those found in speech. Their weakness is that
they show poor discrimination, which may be corrected by corrective training and function work
modeling.
Allen and Kanam [47] designed in NN architecture that locates word boundaries and identifies words
from phoneme sequences. They tested the model in three different regimes with a highly redundant
corpus and a restricted vocabulary, and the NN was trained with a limited number of phonemic variations
for the words in the corpus. These tests yielded a very low error rate. In a second experiment, the network
was trained to identify words from expert transcriptions of speech. The error rate for correct simultaneous
identification of words and word boundaries was 18%. Finally, they tested the use of the output of a
phoneme classifier as the input to the word and word boundary identification network. The error rate
increased almost exponentially to 49%.
The best discrimination came from hybrid systems such as the use of an MLP and a hidden Markov
model (HMM). These systems incorporate multiple sources of evidence such as features and temporal
context without any restrictive assumptions of distributions or statistical independence. Bourland et al.
[48] used MLPs as linear predictions for autoregressive HMMs. This approach, although more compatible
with the HMM formalism, still suffers from several limitations. Although these authors generalized their
approach to take into account time correlations between successive observations without any restrictive
assumptions about the driving noise, the reputed results show that many of the tricks used to improve
standard HMMs are also valid for this hybrid approach.
In another study, Intrator [49] used an unsupervised NN for dimensionality reduction that seeks
directions emphasizing multimodality and derived a new statistical insight to the synaptic modification
equations learning as in the Bienenstock, Cooper, and Munro (BCM) neurons [50]. The speech data
consisted of 20 consecutive time windows of 32 ms with 30-ms overlap aligned at the beginning of the
time. For each window, a set of 22 energy levels was computed corresponding to the Zwicker critical-
band filters [51]. The classification results were compared with those of a backpropagation network.
These results showed that the backpropagation network does well in finding structures useful for classi-
fication of the trained data, but these structures are more sensitive to voicing. Classification results using
a BCM network, on the other hand, suggest that for the specified task, structures that are more sensitive
to voicing can be extracted, even though voicing imposes several effects on the speech signal. These
features are more speaker-invariant.
Phan et al. [52] have attempted to solve the “cocktail party effect,” which describes phenomena in
which humans can selectively focus attention on one sound source among competing sound sources.
This is an ability that is hampered for the hearing-impaired. A system was developed that successfully
identifies a speaker in the presence of competing speakers for short utterances. Features used for identi-
fication are monaural, whose feature space represent a 90% data reduction from the original data. This
system is presently used off-line and also has been applied successfully to intraspeaker speech recognition.
The features used in the preprocessing were obtained by wavelet analysis. This multiresolution analysis
decomposes a signal into a hierarchical system of subspaces that are one-dimensional and are square
integrable. Each subspace is spanned by basis functions that have scaling characteristics of either dilation

© 2000 by CRC Press LLC

or compression depending on the resolution. The implementation of these basis functions is incorporated
in a recursive pyramidal algorithm in which the discrete approximation of a current resolution is
convolved with quadrature mirror filters in the subsequent resolution [53]. After preprocessing, speech
waveform is analyzed by the wavelet transform. Analysis is limited to four octaves. For pattern-recognition
input configuration, the wavelet coefficients are mapped to a vector and used as inputs to a neural network
trained by the ALOPEX algorithm. White noise was superimposed on the features as well in order to
establish a threshold of robustness of the algorithm.

58.3 Neural Networks in Cardiology

Two-dimensional echocardiography is an important noninvasive clinical tool for cardiac imaging. The
endocardial and epicardial boundaries of the left ventricle (LV) are very useful quantitative measures of
various functions of the heart. Cardiac structures detection in ECG images are also very important in
recognizing the image parts.
A lot of research in the last couple of years has taken place around these issues and the application of
neural networks in solving them. Hunter et al. [54] have used NNs in detecting echocardiographic LV
boundaries and the center of the LV cavity. The points detected are then linked by a “snake” energy-
minimizing function to give the epicardial and endocardial boundaries. A snake is an energy-minimizing
deformable curve [55] fined by an internal energy that is the controller of the differential properties in
terms of its curvature and its metrics. The most robust results were obtained by a 9 × 9 square input
vector with a resolution reduction of 32:1. Energy minimization is carried out by simulated annealing.
The minimum energy solution was obtained after 1000 iterations over the entire snake. The use of NNs
as edge detectors allows the classification of points to be done by their probability of being edges rather
than by their edge strength, a very important factor for echocardiographic images due to their wide
variations in edge strength.
A complex area in electrocardiography is the differentiation of wide QRS tachycardias in ventricular
(VT) and supraventricular tachycardia (SVT). A set of criteria for this differentiation has been proposed
recently by Brugada et al. [56]. One important aspect of applying NNs in interpreting ECGs is to use
parameters that not only make sense but are also meaningful for diagnosis. Dassen et al. [57] developed
an induction algorithm that further improved the criteria set by Brugada et al., also using NNs.
Nadal and deBossan [58] used principal-components analysis (PCA) and the relative R-wave to R-wave
intervals of P-QRS complexes to evaluate arrhythmias. Arrhythmias are one of the risks of sudden cardiac
arrest in coronary care units. In this study, the authors used the first 10 PCA coefficients (PCCs) to reduce
the data of each P-QRS complex and a feedforward NN classifier that splits the vector space generated
by the PCCs into regions that separate the different classes of beats in an efficient way. They obtained
better classification than using other methods, with correct classification of 98.5% for normal beats,
98.5% for ventricular premature beats, and 93.5% for fusion beats, using only the first two PCCs. When
4 PCCs were used, these classifications improved to 99.2%, 99.1%, and 94.1%, respectively. These results
are better than those obtained by logistic regression when the input space is composed of more than two
classes of beats. The difficulties encountered include the elimination of redundant data in order not to
overestimate the importance of normal beat detection by the classifier.
In another work, Silipo et al. [59] used an NN as an autoassociator. In previous work they had proved
that NNs had better performances than the traditional clustering and statistical methods. They considered
beat features derived from both morphologic and prematurity information. Their classification is ade-
quate for ventricular ectopic beats, but the criteria used were not reliable enough to characterize the
supraventricular ectopic beat.
A lot of studies also have used NNs for characterization of myocardial infarction (MI). Myocardial
infarction is one of the leading causes of death in the United States. The currently available techniques
for diagnosis are accurate enough, but they suffer from certain drawbacks, such as accurate quantitative

© 2000 by CRC Press LLC

 FIGURE 58.3 Selection of region of interest. The intensity pattern in the box is used as input to the NN.

measure of severity, extent, and precise location of the infarction. Since the acoustic properties in the
involved region are mostly changing, one can study them by the use of ultrasound.
Baxt [60] used an NN to identify MI in patients presented to an emergency department with anterior
chest pain. An NN was trained on clinical pattern sets retrospectively derived from patients hospitalized
with a high likelihood of having MI, and the ability of the NN was compared with that of physicians
caring for the same patients. The network performed with a sensitivity of 92% and a specificity of 96%.
These figures were better than the physicians, with a sensitivity of 88% and a specificity of 71%, or any
other computer technology (88% sensitivity, 74% specificity) [61].
Diagnosis of inferior MI with NNs was studied by Hedén et al. [62] with sensitivity of 84% and a
specificity of 97%, findings that are similar to those of Pahlm et al. [63].
Yi et al. [64] used intensity changes in an echocardiogram to detect MI. Once an echocardiogram is
obtained, it is digitized and saved in a file as a gray-scale image (512 × 400) (Fig. 58.3). A window of the
region of interest is then selected between the systole and diastole of the cardiac cycle and saved in a
different file. The window can be either of a constant size or it can be adaptive (varying) in size. This
new file can be enlarged (zoom) for examination of finer details in the image. All image artifacts are
filtered out, and contrast enhancement is performed. This new image is then saved and serves as input
to the NN. A traditional three-layer NN with 300 input nodes (one node for each pixel intensity in the
input file), a varying number of hidden nodes, and two output nodes were used. The output node
indicates the status of the patient under testing. A “one” indicates normal and a “zero” an abnormal case.
The weights of the connections were calculated using the optimization algorithms of ALOPEX. One
sub-ALOPEX was used for the weights between hidden nodes and input nodes, and a second sub-ALOPEX
was used for those from the output to the hidden layer.
The network was trained with a population of 256 patients, some with scars and some normal. These
patients were used to obtain “templates” for each category. These templates were then used for comparison
with the test images. None of these testing images was included in the training set. The cost function
used for the process was the least-squares rule, which, although slow, produces reliable results.
A similar process is used for the output layer. The noise was made adaptable. The intensities of the
images are normalized before being submitted to the NN. A cutoff of 0.2 was used. Therefore, anything
above 0.8 is normal, below 0.2 is scar, and all the in-between values are classified as unknown. Due to
the fact that the scar training set was very small compared with the normals, a better classification of
normals than scars was observed. A study was also made as to how the number of hidden nodes influences
the results for the same standard deviation of noise.

© 2000 by CRC Press LLC

 In another study, Kostis et al. [65] used NNs in estimating the prognosis of acute MI. Patients who
survive the acute phase of an MI have an increased risk of mortality persisting up to 10 years or more.
Estimation of the probability of being alive at a given time in the future is important to the patients and
their physicians and is usually ascertained by the use of statistical methods. The purpose of the investi-
gation was to use an NN to estimate future mortality of patients with acute MI. The existence of a large
database (Myocardial Infarction Data Acquisition Systems, or MIDAS) that includes MI occurring in the
state of New Jersey and has long-term follow-up allows the development and testing of such a computer
algorithm. Since the information included in the database does not allow the exact prediction of vital
status (dead or alive) in all patients with 100% accuracy, the NN should be able to categorize patients
according to the probability of dying within a given period of time.
Since information included in the database is not sufficient to allow the exact prediction of vital status
(dead or alive) in all patients with 100% accuracy, we developed an NN able to categorize patients
according to the probability of dying within a given period of time rather than predicting categorically
whether a patient will be dead or alive at a given time in the future. It was observed that there were many
instances where two or more patients had identical input characteristics while some were dead and some
alive at the end of the study period. For this reason, it is difficult to train a standard NN. Since there is
no unique output value for all input cases, the network had difficulty converging to a unique set of
solutions. To alleviate this problem, a conflict-resolution algorithm was developed. The algorithm takes
templates with identical input vectors and averages each of their input characteristics to produce a single
case. Their output values of vital status are averaged, producing, in effect, a percentage probability of
mortality for the particular set of input characteristics. As each new subject template is read into the
program, its input characteristics are compared with those of all previous templates. If no match is found,
its input values and corresponding output value are accepted. If a match is found, the output value is
brought together with the stored output value (percentage probability of mortality and the number of
templates on which it is based), and a new output value is calculated, representing the percentage average
mortality of the entire characteristic group. Since each member of the group is an identical input case,
no input characteristic averaging is necessary, thus preserving the statistical significance of the average
mortality with respect to that exact case.
This new algorithm, using the two-hidden-layer perceptron optimized by ALOPEX, has converged to
greater than 98% using several thousand input cases. In addition, 10 output nodes were used in the final
layer, each corresponding to a range of percent chance or mortality (e.g., node 1: 0% to 10%; node 2:
10% to 20%; etc.). The outputs of the network are designed to maximize one of the 10 potential output
“binds,” each corresponding to a decile of mortality between 0% and 100%. The output node containing
the correct probability value was set to a value of 1.0; the others to 0.0. In this manner, the network
should be able to provide percentage probability of mortality and also to resolve input-case conflicts. An
SAS program was written to present the predicted probability of mortality separately in patients who are
dead or alive at the end of the follow-up period. The NNs constructed as described above were able to
be trained and evaluated according to several definitions of network response: matching the output value
at every output node, identifying the correct location of which node was to contain the peak output
value, and matching the output value at the peak output location.
The network was tested on known and then on unknown cases. A correspondence of the observed to
the predicted probability of being alive at a given time was observed. The categorical classifications (dead
or alive) yielded an overall accuracy of 74%. A reciprocal relationship between sensitivity and specificity
of the rules for determination of vital status was observed.

58.4 Neural Networks in Neurology

NNs found applications in neurology as well and, in particular, in characterizing memory defects, as are
apparent in diseases such as Parkinson’s and Alzheimer’s. Both diseases exhibit devastating effects and
disrupt the lives of those affected.

© 2000 by CRC Press LLC

 For several decades, in an attempt to further understand brain functions, computational neuroscientists
have addressed the issue by modeling biologic neural network structure with computer simulations. A
recent article by Stern et al. [66] reports on an important relationship of Alzheimer’s disease expression
and levels of education. A similar inverse relationship was earlier reported by Zhang et al. [67] in a
Chinese population. This inverse relationship was attributed to a protective role of the brain reserve
capacity [68]. Such a capacity becomes important, since in an educated person’s brain more synapses
exist, which might protect the individual in the expression of symptoms of the disease. It is not argued,
however, that the disease will not be acquired; rather, that it will be delayed. Zahner et al. [69] have
employed a three-layer feedforward NN trained with ALOPEX to simulate the effects of education in
dementia, age-related changes, and in general, brain damage. Our results show that the higher the level
of training of the NN, 50%, 60%, 70%, 80%, etc., the slower is the damage on the “brain.” Damage was
simulated by systematically adding noise on the weights of the network. Noise had a gaussian distribution
with varying standard deviations and mean of zero. Figure 58.4 shows the results of these simulations
as recognition rate versus standard deviation of noise added to the inputs for increasingly damaged
weights at various levels of training. Each point corresponds to an average of 100 samples. Notice how
much slower the drop in recognition is for the 70% and 80% curves as compared with the 50% and 60%
training. Also notice the distances of the starting points of the curves as we follow the progress of dementia
at different stages (Fig. 58.4, a versus b versus c). All curves demonstrate impairment with damage, but
they also show some threshold in the training level, after which the system becomes more robust.

58.5 Discussion
Neural networks provide a powerful tool for analysis of biosignals. This chapter has reviewed applications
of NNs in cardiology, neurology, speech processing, and brain waveforms. The literature is vast, and this
chapter is by no means exhaustive. In the last 10 years, NNs have been used more and more extensively
in a variety of fields, and biomedical engineering is not short of it. Besides the applications, a lot of
research is still going on in order to find optimal algorithms and optimal values for the parameters used
in these algorithms. In industry, an explosion of VLSI chip designs on NNs has been observed. The
parallel character of NNs makes them very desirable solutions to computational bottlenecks.

© 2000 by CRC Press LLC

FIGURE 58.4 Recognition rate versus standardization of noise added to the inputs. Different curves correspond
to various learning levels with damaged weights. (a) Noise added only to the inputs. (b) Noise added to the weights
to mimic “brain damage,” σ = 0.05. (c) Noise on the weights with σ = 0.1. Notice how much more robust the “brain”
is to noise with higher levels of education.

© 2000 by CRC Press LLC

References
1. Werbos P. 1974. Beyond Regression: New Tools for Prediction and Analysis in the Behavioral
Sciences. Ph.D. thesis, Harvard University, Cambridge, Mass.
2. Parker DB. 1985. Learning logic, S-81-64, file 1, Office of Technology Licensing, Stanford University,
Stanford, Calif.
3. Rumelhart DE, Hinton GE, Williams RJ. 1986. Learning internal representations by error propa-
gation. In DE Rumelhart, JL McClelland (eds), Parallel Distributed Processing, vol 2: Foundations.
Cambridge, Mass, MIT Press.
4. Harth E, Tzanakou E. 1974. A stochastic method for determining visual receptive fields. Vision Res
14:1475.
5. Tzanakou E, Michalak R, Harth E. 1984. The ALOPEX process: Visual receptive fields with response
feedback. Biol Cybernet 51:53.
6. Micheli-Tzanakou E. 1984. Nonlinear characteristics in the frog’s visual system. Biol Cybernet
51:53.
7. Deutsch S, Micheli-Tzanakou E. 1987. Neuroelectric Systems. New York, NYU Press.
8. Marsic I, Micheli-Tzanakou E. 1990. Distributed optimization with the ALOPEX algorithms. In
Proceedings of the 12th Annual International Conference of the IEEE/EMBS 12:1415.
9. Dasey TJ, Micheli-Tzanakou E. 1989. A pattern recognition application of the ALOPEX process
with hexagonal arrays. In International Joint Conference on Neural Networks 12:119.
10. Xiao L-T., Micheli-Tzanakou E, Dasey TJ. 1990. Analysis of composite neuronal waveforms into
their constituents. In Proceedings of the 12th Annual International Conference of the IEEE/EMBS
12(3):1433.
11. Hiraiwa A, Shimohara K, Tokunaga Y. 1989. EMG pattern analysis and classification by Neural
Networks. In IEEE International Conference on Systems, Man and Cybernetics, part 3, pp 1113–1115.
12. Huang Q, Graupe D, Huang Y-F, Liu RW. 1989. Identification of firing patterns of neuronal signals.
In Proceedings of the 28th IEEE Conference on Decision and Control, vol 1, pp 266–271.
13. Bruha I, Madhavan GP. 1989. Need for a knowledge-based subsystem in evoked potential neural-
net recognition system. In Proceedings of the 11th Annual International Conference of the
IEEE/EMBS, vol 11, part 6, pp 2042–2043.
14. Kelly MF, Parker PA, Scott RN. 1990. Applications of neural networks to myoelectric signal analysis:
A preliminary study. IEEE Trans Biomed Eng 37(3):221.
15. Ramamoorthy PA, Govid G, Iyer VK. 1988. Signal modeling and prediction using neural networks.
Neural Networks 1(1):461.
16. Moody EB Jr, Micheli-Tzanakou E, Chokroverty S. 1989. An adaptive approach to spectral analysis
of pattern-reversal visual evoked potentials. IEEE Trans Biomed Eng 36(4):439.
17. Dayhoff JE. 1990. Regularity properties in pulse transmission networks. Proc IJCNN 3:621.
18. Dayhoff JE. 1990. A pulse transmission (PT) neural network architecture that recognizes patterns
and temporally integrates. Proc IJCNN 2:A-979.
19. Roberts WM, Hartile DK. 1975. Separation of multi-unit nerve impulse trains by the multichannel
linear filter algorithm. Brain Res 94:141.
20. Abeles M, Goldstein MH. 1977. Multiple spike train analysis. Proc IEEE 65:762.
21. Wiemer W, Kaack D, Kezdi P. 1975. Comparative evaluation of methods for quantification of neural
activity. Med Biol Eng 358.
22. Perkel DH, Gerstein GL, Moore GP. 1967. Neuronal spike trains and stochastic point processes: I.
The single spike train. Biophys J 7:391.
23. Perkel DH, Gerstein GL, Moore GP: 1967. Neuronal spike trains and stochastic point processes:
II. Simultaneous spike trains. Biophys J 7:419.
24. Gerstein GL, Perkel DH, Subramanian KN. 1978. Identification of functionally related neuronal
assemblies. Brain Res 140:43.

© 2000 by CRC Press LLC

 25. Papoulis A. 1984. Probability, Random Variables and Stochastic Processes, 2d ed. New York,
McGraw-Hill.
26. Moore GP, Segundo JP, Perkel DH, Levitan H. 1970. Statistical signs of synaptic interactions in
neurons. Biophys J 10:876.
27. Gerstein GL, Perkel DH, Dayhoff JE. 1985. Cooperative firing activity in simultaneously recorded
populations of neurons: detection and measurement. J Neurosci 5:881.
28. Dayhoff JE, Gerstein GL. 1983. Favored patterns in nerve spike trains: I. Detection. J Neurophysiol
49(6):1334.
29. Dayhoff JE, Gerstein GL. 1983. Favored patterns in nerve spike trains: II. Application. J Neuro-
physiol 49(6):1349.
30. Abeles M, Gerstein GL. 1988. Detecting spatiotemporal firing patterns among simultaneously
recorded single neurons. J Neurophysiol 60(3):909.
31. Frostig RD, Gerstein GL. 1990. Recurring discharge patterns in multispike trains. Biol Cybernet
62:487.
32. Micheli-Tzanakou E, Iezzi R. 1985. Spike recognition by stroke density function calculation. In
Proceedings of the 11th Northeast Bioengineering Conference, pp 309–312.
33. Iezzi R, Micheli-Tzanakou E. 1990. Neural network analysis of neuronal spike-trains. In Annual
International Conference of the IEEE/EMBS, vol 12, no 3, pp 1435–1436.
34. Rosenblatt F. 1962. Principles of Neurodynamics. Washington, Spartan Books.
35. Davilla CE, Welch AJ, Rylander HG. 1986. Adaptive estimation of single evoked potentials. In
Proceedings of the 8th IEEE EMBS Annual Conference, pp 406–409.
36. Wang J-Z, Micheli-Tzanakou E. 1990. The use of the ALOPEX process in extracting normal and
abnormal visual evoked potentials. IEEE/EMBS Mag 9(1):44.
37. Moore J, Ramon F. 1974. On numerical integration of the Hodgkin and Huxley equations for a
membrane action potential. J Theor Biol 45:249.
38. Bell J, Cook LP. 1979. A model of the nerve action potential. Math Biosci 46:11.
39. Micheli-Tzanakou E, O’Malley KG. 1985. Harmonic context of patterns and their correlations to
VEP waveforms. In Proceedings of IEEE, 9th Annual Conference EMBS, pp 426–430.
40. Micheli-Tzanakou E. 1990. A neural network approach of decomposing brain waveforms to their
constituents. In Proceedings of the IASTED International Symposium on Computers and Advanced
Technology in Medicine, Healthcare and Bioengineering, pp 56–60.
41. Maier J, Dagnelie G, Spekreijse H, Van Duk W. 1987. Principal component analysis for source
localization of VEPs in man. Vis Res 27:165.
42. Nahamura M, Nishida S, Shibasaki H. 1989. Spectral properties of signal averaging and a novel
technique for improving the signal to noise ration. J Biomed Eng 2(1):72.
43. Orfanidis S, Aafif F, Micheli-Tzanakou E. 1987. Visual evoked potentials extraction by adaptive
filtering. In Proceedings of the IEEE/EMBS International Conference, vol 2, pp 968–969.
44. Doncarli C, Goerig I. 1988. Adaptive smoothing of evoked potentials: A new approach. In Pro-
ceedings of the Annual International Conference on the IEEE/EMBS, part 3(of 4), pp 1152–1153.
45. Davilla E, Welch AJ, Rylander HG. 1986. Adaptive estimation of single evoked potentials. In
Proceedings of the 8th IEEE/EMBS Annual International Conference, pp 406–409.
46. Tebelski J, Waibel A, Bojan P, Schmidbauer O. 1991. Continuous speech recognition by linked
predictive neural networks. In PR Lippman, JE Moody, DS Touretzky (eds), Advances in Neural
Information Processing Systems 3, pp 199–205. San Mateo, Calif, Morgan Kauffman.
47. Allen RB, Kanam CA. 1991. A recurrent neural network for word identification from continuous
phonemena strings. In PR Lippman, JE Moody, DS Touretzky (eds), Advances in Neural Informa-
tion Processing Systems 3, pp 206–212. San Mateo, Calif, Morgan Kauffman.
48. Bourland H, Morgan N, Wooters C. 1991. Connectionist approaches to the use of Markov models
speech recognition. In PR Lippman, JE Moody, DS Touretzky (eds), Advances in Neural Informa-
tion Processing systems 3, pp 213–219. San Mateo, Calif, Morgan Kauffman.

© 2000 by CRC Press LLC

 49. Intrator N. 1991. Exploratory feature extraction in speech signals. In PR Lippman, JE Moody, DS
Touretzky (eds), Advances in Neural Information Processing Systems 3, pp 241–247. San Mateo,
Calif, Morgan Kauffman.
50. Bienenstock EL, Cooper LN, Munro PW. 1992. Theory for the development of neuron selectivity:
Orientation specificity and binocular interaction in visual cortex. J Neurosci 2:32.
51. Zwicker E. 1961. Subdivision of the audible frequency range into critical bands (frequenagruppen).
J Acoust Soc Am 33:248.
52. Phan F, Zahner D, Micheli-Tzanakou E, Sideman S. 1994. Speaker identification through wavelet
multiresolution decomposition and Alopex. In Proceedings of the 1994 Long Island Conference
on Artificial Intelligence and Computer Graphics, pp 53–68.
53. Mallat SG. 1989. A theory of multiresolution signal decomposition: The wavelet representation.
IEEE Trans Pattern Anal Mach Int 11:674.
54. Hunter IA, Soraghan JJ, Christie J, Durani TS. 1993. Detection of echocardiographic left ventricle
boundaries using neural networks. Comput Cardiol 201.
55. Cohen LD. 1991. On active contour models and balloons. CVGIP Image Understanding 53(92):211.
56. Brugada P, Brugada T, Mont L, et al. 1991. A new approach to the differential diagnosis of a regular
tachycardia with a wide QRS complex. Circulation 83(5):1649.
57. Dassen WRM, Mulleneers RGA, Den Dulk K, et al: 1993. Further improvement of classical criteria
for differentiation of wide-QRS tachycardia in SUT and VT using artificial neural network tech-
niques. Comput Cardiol 337.
58. Nadal J, deBossan MC. 1993. Classification of cardiac arrhythmias based on principal components
analysis and feedforward neural networks. Comput Cardiol 341.
59. Silipo R, Gori M, Marchesi C. 1993. Autoassociator structured neural network for rhythm classi-
fication of long-term electrocardiogram. Comput Cardiol 349.
60. Baxt WB. 1991. Use of an artificial neural network for the diagnosis of myocardial infarction. Ann
Intern Med 115(II):843.
61. Goldman L, Cook SF, Brand DA, et al. 1988. A computer protocol to predict myocardial infarction
in emergency department patients with chest pain. N Engl J Med 18:797.
62. Hedén B, Edenbrandt L, Haisty WK Jr, Pahlm O. 1993. Neural networks for ECG diagnosis of
inferior myocardial infarction. Comput Cardiol 345.
63. Pahlm O, Case D, Howard G, et al. 1990. Decision rules for the ECK diagnosis of inferior myocardial
infarction. Comput Biomed Res 23:332.
64. Yi C, Micheli-Tzanakou E, Shindler D, Kostis JB. 1993. A new neural network algorithm to study
myocardial ultrasound for tissue characterization. In 19th Northeastern Bioengineering Confer-
ence, NJIT, pp 109–110.
65. Kostis WJ, Yi C, Micheli-Tzanakou E. 1993. Estimation of long-term mortality of myocardial
infarction using a neural network based on the ALOPEX algorithm. In ME Cohen and DL Hudson
(eds), Comparative Approaches in Medical Reasoning.
66. Stern Y, Gurland B, Tatemichi TK, et al. 1994. Influence of education and occupation on the
incidence of Alzheimer’s disease. JAMA 271:1004.
67. Zhang M, Katzman R, Salmon D, et al. 1990. The prevalence of dementia and Alzheimer’s disease
in Shanghai, China: Impact of age, gender and education. Ann Neurol 27:428.,
68. Satz P. 1993. Brain reserve capacity on symptom onset after brain injury: A formulation and review
of evidence for threshold theory. Neurophysiology 7(3):723.
69. Zahner DA, Micheli-Tzanakou E, Powell A, et al. 1994. Protective effects of learning on a progres-
sively impaired neural network model of memory. In Proceedings of the 16th IEEE/EMBS Annual
International Conference, Baltimore, Md. vol 2, pp 1065–1066.

© 2000 by CRC Press LLC