Neural Networks, Heart Attack and Bayesian Decisions: An Application of the Boltzmann Perceptron Network

A decision aid is proposed for the diagnosis of the most commonly occurring cause of emergency admission to hospital in the developed world-acute myocardial infarction, or heart attack. The motivation for the proposal lies in the Bayesian ( minimum risk)decision theory which is briefly reviewed. The fact that many feedforward artificial neural networks are known to estimate the conditional class probabilities required for Bayesian decision theory is explored and one candidate-the Boltzmann Perceptron Network-is selected as possessing the most desirable properties. A brief account of the theory (based upon the so-called Boltzmann machine) underlying this little known network is presented. The Boltzmann Perceptron Network is trained to diagnose the presence or absence of myocardial infarction on data gathered from a large UK teaching hospital and is found to perform as well as senior registras with specific cardiological training (diagnostic accuracy in excess of 80%). In addition, the Boltzmann Perceptron Network is found to provide greater user confidence than the multi-layer Perceptron