Automated synthesis of prediction models for neural network based myocardial infarction classifiers

Parameter and architectural selection for multiple layered perceptron (MLP) classifiers involve a number of heuristic design procedures. The aim in the design process of such classifiers is to achieve maximum generalization and avoid over-fitting of the training data. It has been the objective of this study to develop a symbolic prediction model to calculate the point at which training should cease for a given neural network (NN) based 12-lead ECG classifier to ensure maximum generalization. This prediction model has been obtained by means of genetic programming (GP), where a GP individual has been evolved to generate a symbolic model that predicts the optimal number of training epochs for three different ECG myocardial infarction classifiers: Anterior myocardial infarction (AMI), inferior myocardial infarction (IMI), and combined myocardial infarction (CMI). The GP model demonstrated to be a very accurate method showing no significant differences between the optimal number of epoch values and the predicted values for both train and test data sets for the three aforementioned pathologies.