Methods for selection of adequate neural network structures with application to early assessment of chest pain patients by biochemical monitoring.

A methodology for selecting, training and estimating the performance of adequate artificial neural network (ANN) structures and incorporating them with algorithms that are optimized for clinical decision making is presented. The methodology was applied to the problem of early ruling-in/ruling-out of patients with suspected acute myocardial infarction using frequent biochemical monitoring. The selection of adequate ANN structures from a set of candidates was based on criteria for model compatibility, parameter identifiability and diagnostic performance. The candidate ANN structures evaluated were the single-layer perceptron (SLP), the fuzzified SLP, the multiple SLP, the gated multiple SLP, the multi-layer perceptron (MLP) and the discrete-time recursive neural network. The identifiability of the ANNs was assessed in terms of the conditioning of the Hessian of the objective function, and variability of parameter estimates and decision boundaries in the trials of leave-one-out cross-validation. The commonly used MLP was shown to be non-identifiable for the present problem and available amount of data, despite artificially reducing the model complexity with use of regularization methods. The investigation is concluded by recommending a number of guidelines in order to obtain an adequate ANN model.