Selecting a neural network structure for ECG diagnosis

In this paper we discuss a number of approaches we have adopted to optimise a feedforward neural network for ECG diagnostic classification. Our goal was to design a classifier for categorising the Frank lead ECG as normal or one of six disease conditions. The database used in this study contained 500, 100% accurate Frank lead ECG recordings with each recording represented by 229 features. We chose to build a classifier which modelled the Bayesian posterior probabilities by using a neural network with no hidden units and a softmax output stage. Maximum likelihood estimates of the network parameters were established by minimising the log-likelihood error function. Two techniques were used to optimise the classifier. Firstly, outputs for multiple networks trained on random subsamples of the training data were combined and secondly, we looked at reducing the number of features processed by the classifier by using principle component analysis. We found that combining multiple networks increased accuracy but reducing the dimension of the feature data did not increase overall accuracy. Multiple runs of ten fold cross validation were used to determine the accuracy of the classifier and the best configuration of our classifier achieved an accuracy of 70.9±0.6%