Application of neural network classifers to electrocardiographic body surface mapping

This thesis examines the capabilities of artificial neural networks for classifying electrocardiographic body surface mapping data. In particular it examines the diagnostic detection of myocardial infarctions and coronary artery disease. An overview of patten recognition, neural networks, electrocardiography, and electrocardiographic body surface mapping is presented followed by a detailed description of the experiments and analysis conducted. The experimental analysis in this thesis is divided into three sections. Firstly, a range of feed-forward artificial neural network architectures and training techniques are used to classify the body surface mapping data with the aim of identifying patients with myocardial infarctions, coronary artery disease, and normal heart function. In this initial study a number of pre-processing techniques are also explored. Secondly, a range of traditional classification techniques (linear regression, k-nearest-neighbour, and inductive learning) are applied to the same problems and compared with the neural network results. When classifying myocardial infarction it was found that artificial neural networks perform as well but no better than traditional classification techniques. This outcome provides some interesting insights into the nature of the classification problem and the information content of body surface maps. However, attempting to separate patients with coronary artery disease from patients with normal heart function neural networks were found to perform much better than traditional classification techniques. The third experimental section examines the bayesian equivalence of neural network outputs and how these probabilistic properties may be used to deal with diagnostic uncertainty. Apart from examining the theoretical connection between network outputs and a posteriori probabilities, a number of experiments are conducted to show how this information can be used to provide the physician with some important information about the classification certainty.