Application of generalized dynamic neural networks to biomedical data.

This paper reviews the application of continuous recurrent neural networks with time-varying weights to pattern recognition tasks in medicine. A general learning algorithm based on Pontryagin's maximum principle is recapitulated, and possibilities of improving the generalization capabilities of these networks are given. The effectiveness of the methods is demonstrated by three different real-world examples taken from the fields of anesthesiology, orthopedics, and radiology.

Classification of equinus in ambulatory children with cerebral palsy-discrimination between dynamic tightness and fixed contracture.

In this study a generalised dynamic neural network (GDNN) was designed to process gait analysis parameters to evaluate equinus deformity in ambulatory children with cerebral palsy. The aim was to differentiate dynamic calf muscle tightness from fixed muscle contracture. Patients underwent clinical examination and had instrumented gait analysis before evaluating their equinus under anaesthesia and muscle relaxation at the time of surgery to improve gait. The performance of the clinical examination, the subjective interpretation of gait analysis results, and the application of the neural network to assess ankle function were compared to the examination under anaesthesia. Evaluation of equinus by a Neural Network showed high sensitivity and specificity values with a likelihood ratio of +14.63. The results indicate that dynamic calf muscle tightness can be differentiated from fixed calf muscle contracture with considerable precision that might facilitate clinical decision-making.

Identification of neural activity based on fMRI data: a simulation study.

This study proposes a technique for determining effective connectivity among brain regions which operates at the level of neuronal dynamics. We propose an alternative time-variant dynamic causal model (TV-DCM) where neuronal dynamic activity evolves based on generalized dynamic neural networks (GDNNs). The identification of brain architecture connectivity is carried out based on a least squares criterion and on a global search technique. Computer simulations carried out in the paper show that TV-DCM may provide multiple solutions, i.e. a set of different architectures all of which approximate the data equally well. Numerical comparisons between TV-DCM and DCM are also given. In order to determine the unique causal structure of brain regions, we apply an additional criterion, i.e. an estimation of generalization error, known from the theory of neural networks. Computer simulations also confirm the validity of our techniques.

Improving generalization capabilities of dynamic neural networks.

This work addresses the problem of improving the generalization capabilities of continuous recurrent neural networks. The learning task is transformed into an optimal control framework in which the weights and the initial network state are treated as unknown controls. A new learning algorithm based on a variational formulation of Pontrayagin's maximum principle is proposed. Under reasonable assumptions, its convergence is discussed. Numerical examples are given that demonstrate an essential improvement of generalization capabilities after the learning process of a dynamic network.

Common Optimization of Adaptive Preprocessing Units and a Neural Network during the Learning Period. Application in EEG Pattern Recognition.

In this study, a proposition of simultaneous training of the neural network (multilayer perceptron) and adaptive preprocessing unit is presented. This cooperation enables the network to affect the preprocessing and as a consequence to vary the locations of pattern vectors in a feature space. Thus, during the learning process the network tries to find a good separation of classes of patterns, which results in convergence of the whole learning process. The strategy was developed in order to make efficient EEG monitoring in neonates possible. A comparison of the method presented herein with the known learning strategies for neural networks shows the need for using it as an alternative learning process. The convergence of the whole system is also discussed. Copyright 1997 Elsevier Science Ltd.