SEARCH · Search PubMed

Results for “Neural network”

Search indexed PubMed citations on genomics, clinical trials, systematic reviews and public health. Explore titles, authors and supplied subject terms, then open the PubMed record.

SourceResearch connectionFrom yearThrough year

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 865 records · Page 48Linked to original sources

PubMedDate not supplied

Using neural networks to predict the onset of diabetes mellitus.

Classification is an important decision making tool, especially in the medical sciences. Unfortunately, while several classification procedures exist, many of the current methods fail to provide adequate results. In recent years, artificial neural networks have been suggested as an alternative tool for classification. Here, we use neural networks to predict the onset of diabetes mellitus in Pima Indian women. The modeling capabilities of neural networks are compared to traditional methods like logistic regression and to a specific method called ADAP, which has been used to predict diabetes. The results indicate that neural networks are indeed a viable approach to classification. Furthermore, we attempt to provide a basis upon which neural networks can be used for variable selection in statistical modeling.

Classification↗

PubMed1998

Experiments to determine whether recursive partitioning (CART) or an artificial neural network overcomes theoretical limitations of Cox proportional hazards regression.

New computationally intensive tools for medical survival analyses include recursive patitioning (also called CART) and artificial neural networks. A challenge that remains is to better understand the behavior of these techniques in effort to know when they will be effective tools. Theoretically they may overcome limitations of the traditional multivariable survival technique, the Cox proportional hazards regression model. Experiments were designed to test whether the new tools would, in practice, overcome these limitations. Two datasets in which theory suggests CART and the neural network should outperform the Cox model were selected. The first was a published leukemia dataset manipulated to have a strong interaction that CART should detect. The second was a published cirrhosis dataset with pronounced nonlinear effects that a neural network should fit. Repeated sampling of 50 training and testing subsets was applied to each technique. The concordance index C was calculated as a measure of predictive accuracy by each technique on the testing dataset. In the interaction dataset, CART outperformed Cox (P < 0.05) with a C improvement of 0.1 (95% CI, 0.08 to 0.12). In the nonlinear dataset, the neural network outperformed the Cox model (P < 0.05), but by a very slight amount (0.015). As predicted by theory, CART and the neural network were able to overcome limitations of the Cox model. Experiments like these are important to increase our understanding of when one of these new techniques will outperform the standard Cox model. Further research is necessary to predict which technique will do best a priori and to assess the magnitude of superiority.

Databases, Factual↗

PubMed2005

An alternative approach for neural network evolution with a genetic algorithm: crossover by combinatorial optimization.

In this work we present a new approach to crossover operator in the genetic evolution of neural networks. The most widely used evolutionary computation paradigm for neural network evolution is evolutionary programming. This paradigm is usually preferred due to the problems caused by the application of crossover to neural network evolution. However, crossover is the most innovative operator within the field of evolutionary computation. One of the most notorious problems with the application of crossover to neural networks is known as the permutation problem. This problem occurs due to the fact that the same network can be represented in a genetic coding by many different codifications. Our approach modifies the standard crossover operator taking into account the special features of the individuals to be mated. We present a new model for mating individuals that considers the structure of the hidden layer and redefines the crossover operator. As each hidden node represents a non-linear projection of the input variables, we approach the crossover as a problem on combinatorial optimization. We can formulate the problem as the extraction of a subset of near-optimal projections to create the hidden layer of the new network. This new approach is compared to a classical crossover in 25 real-world problems with an excellent performance. Moreover, the networks obtained are much smaller than those obtained with classical crossover operator.

Algorithms↗

PubMed1995

Analysis of training set parallelism for backpropagation neural networks.

Training set parallelism and network based parallelism are two popular paradigms for parallelizing a feedforward (artificial) neural network. Training set parallelism is particularly suited to feedforward neural networks with backpropagation learning where the size of the training set is large in relation to the size of the network. This paper analyzes training set parallelism for feedforward neural networks when implemented on a transputer array configured in a pipelined ring topology. Theoretical expressions for the time per epoch (iteration) and optimal size of a processor network are derived when the training set is equally distributed among the processing nodes. These show that the speed up is a function of the number of patterns per processor, communication overhead per epoch and the total number of processors in the topology. Further analysis of how to optimally distribute the training set on a given processor network when the number of patterns in the training set is not an integer multiple of the number of processors, is also carried out. It is shown that optimal allocation of patterns in such cases is a mixed integer programming problem. Using this analysis it is found that equal distribution of training patterns among the processors is not the optimal way to allocate the patterns even when the training set is an integer multiple of the number of processors. Extension of the analysis to processor networks comprising processors of different speeds is also carried out. Experimental results from a T805 transputer array are presented to verify all the theoretical results.

Algorithms↗

PubMed1994

A neural network system for detection of atrial fibrillation in ambulatory electrocardiograms.

INTRODUCTION: A neural network classifier has been designed, which is able to distinguish atrial fibrillation (AF) from other supraventricular arrhythmias in ambulatory (Holter) ECGs. METHOD AND RESULTS: The classification algorithm uses a rhythm analysis that considers the ECG to be a time series of RR interval durations. This is combined with an analysis of baseline morphology that considers the morphological characteristics of the non-QRS portions of the waveform. A backpropagation-based neural network has been used as part of the classifier implementation. When applied to a library consisting exclusively of 42,970 examples of AF and other supraventricular rhythm disturbances validated by an experienced cardiologist, the algorithm demonstrated a sensitivity of 82.4% for 10-beat runs of paroxysmal atrial fibrillation (PAF) and a specificity of 96.6%. Since this system has been implemented as a postprocessor to a conventional automated Holter system, operating only on segments of ECG that are known to contain supraventricular arrhythmias rather than ventricular arrhythmias or sinus rhythm, it can be added to most existing Holter processing systems without significantly increasing the average time to process a tape. CONCLUSION: A neural network system has been designed, which can potentially provide, for the first time, an accurate, quantitative technique to determine the natural history of PAF and to evaluate potential treatments for PAF.

Atrial Fibrillation↗

PubMed2005

Application of artificial neural networks to describe the combined effect of pH and NaCl on the heat resistance of Bacillus stearothermophilus.

A model for prediction of bacterial spore inactivation was developed. The influence of temperature, pH and NaCl on the heat resistance of Bacillus stearothermophilus spores was described using low-complexity, black box models based on artificial neural networks. Literature data were used to build and train the neural network, and new experimental data were used to evaluate it. The neural network models gave better predictions than the classical quadratic response surface model in all the experiments tried. When the neural networks were evaluated using new experimental data, also good predictions were obtained, providing fail-safe predictions of D values in all cases. The weights and biases values of neurons of the neural network that gave the best results are presented, so the reader can use the model for their own purposes. The use of this non-linear modelling technique makes it possible to describe more accurately interacting effects of environmental factors when compared with classical predictive microbial models.

Food Microbiology↗

PubMed1996

CNN: a speaker recognition system using a cascaded neural network.

The main emphasis of this paper is to present an approach for combining supervised and unsupervised neural network models to the issue of speaker recognition. To enhance the overall operation and performance of recognition, the proposed strategy integrates the two techniques, forming one global model called the cascaded model. We first present a simple conventional technique based on the distance measured between a test vector and a reference vector for different speakers in the population. This particular distance metric has the property of weighting down the components in those directions along which the intraspeaker variance is large. The reason for presenting this method is to clarify the discrepancy in performance between the conventional and neural network approach. We then introduce the idea of using unsupervised learning technique, presented by the winner-take-all model, as a means of recognition. Due to several tests that have been conducted and in order to enhance the performance of this model, dealing with noisy patterns, we have preceded it with a supervised learning model--the pattern association model--which acts as a filtration stage. This work includes both the design and implementation of both conventional and neural network approaches to recognize the speakers templates--which are introduced to the system via a voice master card and preprocessed before extracting the features used in the recognition. The conclusion indicates that the system performance in case of neural network is better than that of the conventional one, achieving a smooth degradation in respect of noisy patterns, and higher performance in respect of noise-free patterns.

Algorithms↗

PubMed1999

Embedding recurrent neural networks into predator-prey models.

We study changes of coordinates that allow the embedding of ordinary differential equations describing continuous-time recurrent neural networks into differential equations describing predator-prey models-also called Lotka-Volterra systems. We transform the equations for the neural network first into quasi-monomial form (Brenig, L. (1988). Complete factorization and analytic solutions of generalized Lotka-Volterra equations. Physics Letters A, 133(7-8), 378-382), where we express the vector field of the dynamical system as a linear combination of products of powers of the variables. In practice, this transformation is possible only if the activation function is the hyperbolic tangent or the logistic sigmoid. From this quasi-monomial form, we can directly transform the system further into Lotka-Volterra equations. The resulting Lotka-Volterra system is of higher dimension than the original system, but the behavior of its first variables is equivalent to the behavior of the original neural network. We expect that this transformation will permit the application of existing techniques for the analysis of Lotka-Volterra systems to recurrent neural networks. Furthermore, our results show that Lotka-Volterra systems are universal approximators of dynamical systems, just as are continuous-time neural networks.

Journal Article↗

PubMed2001

Neural networks for improved target differentiation and localization with sonar.

This study investigates the processing of sonar signals using neural networks for robust differentiation of commonly encountered features in indoor robot environments. Differentiation of such features is of interest for intelligent systems in a variety of applications. Different representations of amplitude and time-of-flight measurement patterns acquired from a real sonar system are processed. In most cases, best results are obtained with the low-frequency component of the discrete wavelet transform of these patterns. Modular and non-modular neural network structures trained with the back-propagation and generating-shrinking algorithms are used to incorporate learning in the identification of parameter relations for target primitives. Networks trained with the generating-shrinking algorithm demonstrate better generalization and interpolation capability and faster convergence rate. Neural networks can differentiate more targets employing only a single sensor node, with a higher correct differentiation percentage (99%) than achieved with previously reported methods (61-90%) employing multiple sensor nodes. A sensor node is a pair of transducers with fixed separation, that can rotate and scan the target to collect data. Had the number of sensing nodes been reduced in the other methods, their performance would have been even worse. The success of the neural network approach shows that the sonar signals do contain sufficient information to differentiate all target types, but the previously reported methods are unable to resolve this identifying information. This work can find application in areas where recognition of patterns hidden in sonar signals is required. Some examples are system control based on acoustic signal detection and identification, map building, navigation, obstacle avoidance, and target-tracking applications for mobile robots and other intelligent systems.

Algorithms↗

PubMed1995

Prospective assessment of an artificial neural network for the detection of peripheral vascular disease from lower limb pulse waveforms.

The diagnostic performance of an artificial neural network pulse classification system for the detection of peripheral vascular disease was investigated prospectively. Lower limb photoelectric plethysmographic pulses, and Doppler ankle/brachial pressure index (ABPI) measurements (pre- and post-exercise) were obtained from 200 patients referred to a vascular investigation laboratory. A single toe pulse was processed and used as input data to a neural network which had been trained previously with a set of pulses from 100 legs. The neural network outputs represented the diagnostic arterial disease classifications defined by the ABPI. From the 200 patients entered prospectively, 266 legs were available for neural network assessment. A network sensitivity of 92% and specificity of 63% were achieved with a diagnostic accuracy of 80%. By using a higher confidence for the classification decision a small, but insignificant overall improvement was obtained. When a borderline classification was introduced 100% sensitivity and 100% negative predictive value were obtained, though 31% of legs were unclassifiable. Nevertheless, the very high sensitivity and negative predictive value could make this quick and simple technique the one of choice for the first stage in screening large numbers of subjects.

Adult↗

PubMed2003

Speaker verification using committee neural networks.

Security is a major problem in web based access or remote access to data bases. In the present study, the technique of committee neural networks was developed for speech based speaker verification. Speech data from the designated speaker and several imposters were obtained. Several parameters were extracted in the time and frequency domains, and fed to neural networks. Several neural networks were trained and the five best performing networks were recruited into the committee. The committee decision was based on majority voting of the member networks. The committee opinion was evaluated with further testing data. The committee correctly identified the designated speaker in (50 out of 50) 100% of the cases and rejected imposters in (150 out of 150) 100% of the cases. The committee decision was not unanimous in majority of the cases tested.

Algorithms↗

PubMed2001

Classification of signal-time curves from dynamic MR mammography by neural networks.

The aim of this study was to test the performance of artificial neural networks for the classification of signal-time curves obtained from breast masses by dynamic MRI. Signal-time courses from 105 parenchyma, 162 malignant, and 102 benign tissue regions were examined. The latter two groups were histopathologically verified. Four neural networks corresponding to different temporal resolutions of the signal-time curves were tested. The resolution ranges from 28 measurements with a temporal spacing of 23s to just 3 measurements taken 1.8, 3, and 10 minutes after contrast medium administration. Discrimination between malignant and benign lesions is best if 28 measurement points are used (sensitivity: 84%, specificity: 81%). The use of three measurement points results in 78% sensitivity and 76% specificity. These results correspond to values obtained by human experts who visually evaluated signal-time curves without considering additional morphologic information. All examined networks yielded poor results for the subclassification of the benign lesions into fibroadenomas and benign proliferative changes. Neural networks can computationally fast distinguish between malignant and benign lesions even when only a few post-contrast measurements are made. More precise specification of the type of the benign lesion will require incorporation of additional morphological or pharmacokinetic information.

Adult↗

PubMed2001

Convergence in discrete-time neural networks with specific performance.

We analyze convergence in discrete-time neural networks with specific performance such as decay rate and trajectory bounds in terms of componentwise absolute (exponential) stability. Simple necessary and sufficient stability and positive invariance conditions are presented, which allow us to design a convergent network with prescribed performance. Our approach is based on a decomposition of competitive-cooperative connectivity or inhibitory-excitatory interaction that abounds in neural networks, without assuming symmetry of the connection matrices. The key idea is that through the decomposition, we can always relate a competitive-cooperative network with a cooperative dynamical system. The latter possesses significant order-preserving properties that are basic to our analysis. The explicit division of connection weights into inhibitory and excitatory types offers a higher potential for relating formal neural network models to neurophysiology.

Algorithms↗

PubMed1994

Artificial neural networks for the diagnosis of atrial fibrillation.

Different forms of artificial intelligence have been applied to pattern recognition in medicine. Recently, however, a relatively new technique involving software-based neural networks has become more readily available. Deterministic logic is currently applied to rhythm analysis in computer-assisted ECG interpretation methods developed in the University of Glasgow. The aim of the present study is to compare an artificial neural network with deterministic logic for separating sinus rhythm (SR) with supraventricular extrasystoles (SVEs) and/or ventricular extra-systoles (VEs) from atrial fibrillation (AF) at a particular point in the diagnostic logic of the Glasgow Program. A total of 2363 ECGs with 1495 AF and 868 SR + (SVEs and/or VEs) are used for training and testing a variety of neural networks, and the optimum design is selected. Methods for combining the results of the neural-network classification and the deterministic interpretation are also developed. A further 717 ECGs are used to test the selected network. The results show that the use of an artificial neural network can improve the sensitivity of reporting AF from 88.5% using the deterministic approach to 92%, without sacrificing specificity (92.3%).

Atrial Fibrillation↗

PubMed2004

Reproducing chaos by variable structure recurrent neural networks.

In this paper, we present a new approach for chaos reproduction using variable structure recurrent neural networks (VSRNN). A neural network identifier is designed, with a variable structure that will change according to its output performance as compared to the given orbits of an unknown chaotic systems. A tradeoff between identification errors and computational complexity is discussed.

Algorithms↗

PubMedDate not supplied

Performance of a neural network trained to make third-molar treatment-planning decisions.

The authors developed and tested 12 neural networks of different architectures to make lower-third-molar treatment-planning decisions, using a software-based neural network (Neudesk 1.2, Neural Computer Sciences, Southampton, UK). Network training was undertaken using clinical histories from 119 patients (with 238 lower third molars) referred for treatment planning (79 females and 40 males, mean age 25 years) together with output data consisting of actual treatments planned by a senior oral surgeon. Both the input clinical data and the consultant decisions were treated on a tooth-wise basis and were coded to numerical values. Binary data (e.g., present/absent) were coded to 1 and 0, while quantitative data (e.g., age) were scaled to fall between 0 and 1. A network based on the optimal architecture was trained and then interrogated with test data derived from a further 174 patients (119 females and 55 males, mean age 26 years) with 348 lower third molars. Network decisions were dichotomized with a threshold of 0.8. With no knowledge of the network decisions, the senior oral surgeon indicated his preferred treatments. The teeth were then assigned to "gold-standard" categories of indications present or absent based on National Institutes of Health consensus criteria. Against this, the network achieved a sensitivity of 0.78, which was slightly inferior to that of the oral surgeon (0.88), although this difference was not significant, and a specificity of 0.98, compared with 0.99 for the oral surgeon (p = NS). Agreement between the oral surgeon and network decisions was very high (kappa = 0.850). This study demonstrates that it is possible to train a neural network to provide reliable decision support for lower-third-molar treatment planning.

Adult↗

PubMed2005

Wavelet neural network modeling in QSPR for prediction of solubility of 25 anthraquinone dyes at different temperatures and pressures in supercritical carbon dioxide.

A wavelet neural network (WNN) model in quantitative structure property relationship (QSPR) was developed for predicting solubility of 25 anthraquinone dyes in supercritical carbon dioxide over a wide range of pressures (70-770 bar) and temperatures (291-423 K). A large number of descriptors were calculated with Dragon software and a subset of calculated descriptors was selected from 18 classes of Dragon descriptors with a stepwise multiple linear regression (MLR) as a feature selection technique. Six calculated and two experimental descriptors, pressure and temperature, were selected as the most feasible descriptors. The selected descriptors were used as input nodes in a wavelet neural network (WNN) model. The wavelet neural network architecture and its parameters were optimized simultaneously. The data was randomly divided to the training, prediction and validation sets. The predictive ability of the model was evaluated using validation data set. The root mean squares error (RMSE) and mean absolute errors were 0.339 and 0.221, respectively, for the validation data set. The performance of the WNN model was also compared with artificial neural network (ANN) model and the results showed the superiority of the WNN over ANN model.

Anthraquinones↗

PubMed1997

Automated segmentation and classification of multispectral magnetic resonance images of brain using artificial neural networks.

We present a fully automated process for segmentation and classification of multispectral magnetic resonance (MR) images. This hybrid neural network method uses a Kohonen self-organizing neural network for segmentation and a multilayer backpropagation neural network for classification. To separate different tissue types, this process uses the standard T1-, T2-, and PD-weighted MR images acquired in clinical examinations. Volumetric measurements of brain structures, relative to intracranial volume, were calculated for an index transverse section in 14 normal subjects (median age 25 years; seven male, seven female). This index slice was at the level of the basal ganglia, included both genu and splenium of the corpus callosum, and generally, showed the putamen and lateral ventricle. An intraclass correlation of this automated segmentation and classification of tissues with the accepted standard of radiologist identification for the index slice in the 14 volunteers demonstrated coefficients (ri) of 0.91, 0.95, and 0.98 for white matter, gray matter, and ventricular cerebrospinal fluid (CSF), respectively. An analysis of variance for estimates of brain parenchyma volumes in five volunteers imaged five times each demonstrated high intrasubject reproducibility with a significance of at least p < 0.05 for white matter, gray matter, and white/gray partial volumes. The population variation, across 14 volunteers, demonstrated little deviation from the averages for gray and white matter, while partial volume classes exhibited a slightly higher degree of variability. This fully automated technique produces reliable and reproducible MR image segmentation and classification while eliminating intra- and interobserver variability.

Adult↗