Algorithms for artificial neural networks are usually developed assuming in the most of the models that information propagates fordward and backward across the neural network. Fordwards propagation is modeled easily in ANN and biologically plausible in biological neurons, however for backwards propagation the plausibility of the algorithms developed for ANN seems remote in biological neurons. Based on the presynaptic changes induced by the arachidonic acid released by postsynaptic neurons during long-term potentiation (LTP) in the dentate gyrus, we show a computer simulation model where a backward feedback is performed across local synapses by arachidonic acid. Our simulation model shows how arachidonic acid could be playing the role of retrograde messenger during LTP.

In Chap. 5, Judith E. Dayhoff describes her joint work with her PhD advisor, George Gerstein, on designing and applying novel data analysis methods for detecting repeating spike patterns and cooperating groups of neurons (gravitational clustering) in single- and multiple-neuron spike train recordings.

The evolution of protein interactions cannot be deciphered without a detailed analysis of interaction interfaces and binding modes. We performed a large-scale study of protein homooligomers in terms of their symmetry, interface sizes, and conservation of binding modes. We also focused specifically on the evolution of protein binding modes from nine families of homooligomers and mapped 60 different binding modes and oligomerization states onto the phylogenetic trees of these families. We observed a significant tendency for the same binding modes to be clustered together and conserved within clades on phylogenetic trees; this trend is especially pronounced for close homologs with 70% sequence identity or higher. Some binding modes are conserved among very distant homologs, pointing to their ancient evolutionary origin, while others are very specific for a certain phylogenetic group. Moreover, we found that the most ancient binding modes have a tendency to involve symmetrical (isologous) homodimer binding arrangements with larger interfaces, while recently evolved binding modes more often exhibit asymmetrical arrangements and smaller interfaces.

Background. Available blood assays for venous thromboembolism (VTE) suffer from diminished specificity. Compared with single marker tests, such as D-dimer, a multi-marker strategy may improve diagnostic ability. We used direct mass spectrometry (MS) analysis of serum from patients with VTE to determine whether protein expression profiles would predict diagnosis. Methods and Results. We developed a direct MS and computational approach to the proteomic analysis of serum. Using this new method, we analyzed serum from inpatients undergoing radiographic evaluation for VTE. In a balanced cohort of 76 patients, a neural network-based prediction model was built using a training subset of the cohort to first identify proteomic patterns of VTE. The proteomic patterns were then validated in a separate group of patients within the cohort. The model yielded a sensitivity of 68% and specificity of 89%, which exceeded the specificity of D-dimer assay tested by latex agglutination, ELISA, and immunoturbimetric methods (sensitivity/specificity of 63.2%/60.5%, 97.4%/21.1%, 97.4%/15.8%, respectively). We validated differences in protein expression between patients with and without VTE using more traditional gel-based analysis of the same serum samples. Conclusion. Protein expression analysis of serum using direct MS demonstrates potential diagnostic utility for VTE. This pilot study is the first such direct MS study to be applied to a cardiovascular disease. Differences in protein expression were identified and subsequently validated in a separate group of patients. The findings in this initial cohort can be evaluated in other independent cohorts, including patients with inflammatory conditions and chronic (but not acute) VTE, for the diagnosis of VTE.

We demonstrate a model in which synchronously firing ensembles of neurons are networked to produce computational results. Each ensemble is a group of biological integrate-and-fire spiking neurons, with probabilistic interconnections between groups. An analogy is drawn in which each individual processing unit of an artificial neural network corresponds to a neuronal group in a biological model. The activation value of a unit in the artificial neural network corresponds to the fraction of active neurons, synchronously firing, in a biological neuronal group. Weights of the artificial neural network correspond to the product of the interconnection density between groups, the group size of the presynaptic group, and the postsynaptic potential heights in the synchronous group model. All three of these parameters can modulate connection strengths between neuronal groups in the synchronous group models. We give an example of nonlinear classification (XOR) and a function approximation example in which the capability of the artificial neural network can be captured by a neural network model with biological integrate-and-fire neurons configured as a network of synchronously firing ensembles of such neurons. We point out that the general function approximation capability proven for feedforward artificial neural networks appears to be approximated by networks of neuronal groups that fire in synchrony, where the groups comprise integrate-and-fire neurons. We discuss the advantages of this type of model for biological systems, its possible learning mechanisms, and the associated timing relationships.

Although an individual neural network has proven capabilities that are powerful for pattern detection and function approximation, real-life applications of neural networks often require an entire system for the training and usage of such neural networks. We describe systems for using neural networks in decision making roles such as medical diagnosis and pattern recognition. In our medical applications, the neural network output is treated as a composite variable subject to statistical validation such as an ROC plot analysis, use of re-sampled training to measure performance variance, and avoidance of overtraining. Another system for use of neural networks lies in our approach for training on boundaries rather than individual data points in pattern classification and image analysis problems. We discuss optimizing the neural network and training using these systems

Artificial neural networks now are used in many fields. They have become well established as viable, multipurpose, robust computational methodologies with solid theoretic support and with strong potential to be effective in any discipline, especially medicine. For example, neural networks can extract new medical information from raw data, build computer models that are useful for medical decision making, and aid in the distribution of medical expertise. Because many important neural network applications currently are emerging, the authors have prepared this article to bring a clearer understanding of these biologically inspired computing paradigms to anyone interested in exploring their use in medicine. They discuss the historical development of neural networks and provide the basic operational mathematics for the popular multilayered perceptron. The authors also describe good training, validation, and testing techniques, and discuss measurements of performance and reliability, including the use of bootstrap methods to obtain confidence intervals. Because it is possible to predict outcomes for individual patients with a neural network, the authors discuss the paradigm shift that is taking place from previous "bin-model" approaches, in which patient outcome and management is assumed from the statistical groups in which the patient fits. The authors explain that with neural networks it is possible to mediate predictions for individual patients with prevalence and misclassification cost considerations using receiver operating characteristic methodology. The authors illustrate their findings with examples that include prostate carcinoma detection, coronary heart disease risk prediction, and medication dosing. The authors identify and discuss obstacles to success, including the need for expanded databases and the need to establish multidisciplinary teams. The authors believe that these obstacles can be overcome and that neural networks have a very important role in future medical decision support and the patient management systems employed in routine medical practice. Cancer 2001;91:1615-35. (C) 2001 American Cancer Society.

We view the output of a classification neural network as a composite variable that can be subjected to the same kind of statistical analysis as any other clinical variable used in classification decisions. We show that receiver operating characteristic (ROC) methodology, long used in medicine, can be used in neural network performance evaluation and in sharpening final decisions by adjusting outputs for prevalence and misclassification costs. We explore the use of ensembles of neural networks to estimate classification confidence intervals. Since it is possible to predict outcomes for individual patients with neural networks, we suggest a paradigm shift from previous "bin-model" approaches, in which patient outcome and management decisions are assumed from wide statistical groups into which the patient fits, to decisions customized to the individual patient.

After the most prominent signal in an infrared image of the sky is extracted, the question is whether the signal corresponds to an aircraft. We present a new approach that avoids metric similarity measures and the use of thresholds, and instead attempts to learn similarity measures like those used by humans. In the absence of sufficient real data, the approach allows one to specifically generate an arbitrarily large number of training exemplars projecting near the classification boundary. Once trained on such a training set, the performance of our neural network-based system is comparable to that of a human expert and far better than a network trained only on the available real data. Furthermore, the results obtained are considerably better than those obtained using an Euclidean discriminator.

Received December 18, 2000; accepted January 3, 2001. BACKGROUND. Transrectal prostate biopsy decisions often have been based on absolute cutoff values for total and free prostate-specific antigen (PSA). The authors decided that it would be more appropriate to develop risk profiles for the individual patient to allow him to decide whether to undergo a prostate biopsy. METHODS. To develop risk profiles, the authors first used multivariate logistic regression analysis to analyze 2054 males who were part of the Tyrol (Austria) PSA Screening Project. Second, artificial neural network (ANN) analyses were performed using data from 3474 males who also were part of the Tyrol PSA Screening Project and who had undergone prostate biopsy. These analyses were compared with standard cutoff levels of specificity for the detection of prostate carcinoma. RESULTS. To the authors’ knowledge, this was the first time that multivariate logistic regression analysis was used to decide whether to perform prostate biopsies based on risk profiles rather than on single cutoff levels. For the detection of prostate carcinoma, at sensitivity levels of 90 –95%, the ANN was 150 –200% more specific than the standard cutoff points. For screened volunteers with total PSA levels below 4 ng/mL, ANN showed a lower cancer predictive ability in comparison with volunteers with total PSA levels above 4 ng/mL. However, the ANN was approximately 150 –200% more specific than the standard cutoff levels in both groups. CONCLUSIONS. At high sensitivity levels, ANN increased the specificity for prostate carcinoma detection in a PSA-based screened population. The improvement in specificity between standard cutoff levels and ANN ranged between 150 –200% and was not affected by the presence of benign prostatic hyperplasia or prostatitis. Cancer 2001;91:1667–72. © 2001 American Cancer Society.

Artificial neural networks (ANNs) are a type of artificial intelligence software inspired by biological neuronal systems that can be used for nonlinear statistical modeling. In recent years, these applications have played an increasing role in predictive and classification modeling in medical research. We review the basic concepts behind ANNs and examine the role of this technology in selected applications in prostate cancer research.

In target recognition in uncontrolled environments the test target may not belong to the prestored targets or target classes. Hence, in such environments the use of a typical classifier which finds the closest class still leaves open the question of whether the test target truly belongs to that class. To decide whether a test target matches a stored target, common approaches calculate a degree of similarity between the two targets using a similarity measure such as Euclidean distance, and make a decision based on whether the distance exceeds a (prespecified) threshold. Based on psychophysical studies, this is very different from, and far inferior to, human capabilities. In this paper we show a new approach where a neural network learns a decision boundary between the confirmation vs. rejection of a match with the help of a human critic. The decision boundary is a multidimensional surface, and models the human similarity measure for the recognition task at hand, thus avoiding metric similarity measures and thresholds. A case study in automatic aircraft recognition is shown. In the absence of sufficient real data, the approach allows us to specifically generate an arbitrarily large number of training exemplars projecting near the classification boundary. The performance of the trained network was comparable to that of a human expert, and far better than a network trained only on the available real data. Furthermore, the result were considerably better than those obtained using a Euclidean discriminator.

This chapter addresses topics on the dynamic behavior of neural networks as they oscillate and produce specific timing patterns in their activity. A network of simple processing units is capable of producing prolonged self-sustained oscillations and even chaotic behavior. Modulation of a controlled parameter causes the temporal dynamics to increase in complexity until chaos is reached. An external stimulus—a pattern—can be applied to a chaotic network, resulting in a simpler, limit cycle attractor, which can be recognized in a pattern-to-oscillation map. Since random networks tend to have only one observed dynamic attractor, we have designed a weight perturbation schedule to develop multiple dynamic attractors from different initial states of the network. The result is to create different basins of attraction for different patterns or pattern groups. We can observe a tremendous flexibility not only in evoked attractors (usually oscillations) but in their basins of attraction—the collections of states that lead to the same attractor. Attractor training has been done in networks with time-delay mechanisms, where an a priori chosen dynamic attractor can be trained into the network. A comparison to temporal processing in biological systems is discussed.

The purpose of the study was to build and evaluate a decision support system for financial models, incorporating a neural network approach. Since neural network models may employ a variety of different training techniques, the authors have developed an approach to choosing and optimizing the structures and procedures used during training, to optimize the fit between the trained neural network and the financial decision-support goals. They evaluate alternative methods for training a network to forecast gold market prices. Essential to this evaluation is the identification of an appropriate trading model to evaluate system performance, without constraining the details of the financial decisions that can be made with the resulting trained neural network. The methods explored were neural network models using both standard and non-standard training techniques, and varying parameters used during weight adjustment and in the training schedule. They illustrate techniques for choosing network structure and inputs, data segmentation, error measures for training, error measures for validation, and optimizing validation set lengths. These techniques are applied with respect to the aim of forecasting the price of gold