W.R.M. Dassen, M. Egmont-Petersen, R.G.A. Mulleneers. "Artificial neural
networks in cardiology; a review," chapter 29, In P.E. Vardas (Ed.), Cardiac
arrhythmias, pacing and electrophysiology, Kluwer Academic Publishers,
Dordrecht, 1998, pp. 205-211.
In daily practice cardiologists often have to make a diagnosis based on measured or estimated data. Sometimes this is relatively simple, for example when a patient presents with: (a) typical chest pain, (b) characteristic ECG abnormalities and finally (c) a typical rise in cardiac enzymes, the diagnosis of acute myocardial infarction can be made with a high positive predictive accuracy, regardless of whether other parameters have been measured or not. Also in the interpretation of the electrocardiogram a diagnosis can sometimes be made easily, with almost 100% certainty. In a wide-QRTS tachycardia with evident AV dissociation a ventricular origin is almost certain.
However, occasionally the information is not so clear and a differential diagnosis needs to be established. This requires a combination of reasoning and pattern recognition. Depending on the clinical situation the cardiologist should recognise a certain pattern. When the characteristic signs mentioned above are not so dominantly present, the relationship between these parameters and the weight of each factor used to determine a conclusion can rapidly become very complex. If a linear relationship between the parameters and the conclusion is no longer present, models based on classical statistical methods constitute only rough approximations.
A mathematical technique developed to overcome this limitation is called an artificial neural network. A major advantage of using an artificial neural network to model the relationship between possible signs and symptoms and the diagnosis is the fact that this relationship does not have to be a linear one. Furthermore, it does not even have to be known. Based on a number of correctly classified samples, referred to as the training set, the neural network itself determines the underlying relationships.
In this chapter, after a brief definition of classification and a summary of the history and the way neural networks function, a compilation will be presented of those areas in cardiology in which the application of neural networks have been evaluated. The chapter will end with an overview of those aspects which still impede general application of this method for clinical decision support.
Reprints, please contact me: firstname.lastname@example.org