An Artificial Neural Network to Localize Atrioventricular Accessory Pathways in Patients Suffering from the Wolff‐Parkinson‐White Syndrome

Abstract

The electrocardiographic localization of atrioventricular accessory pathways has been extensively described in the literature by a number of well-known electrophysiologists and surgeons. These descriptions, often represented as decision trees, are useful, but do not apply in all cases. To formalize the process of determining the proper localization, this expert human knowledge could be represented in an expert system. But since reasoning is partly based on the use of heuristic knowledge, and are often not represented in the written description of the human expert, the results will be suboptimal. On the other hand, by using a self-learning neural network approach, the causal relations between input (polarity of the delta waves) and output (the correct localization) do not have to be defined by the expert. It is derived by the neural network, by analyzing a learning set of cases consisting of the ECG plus the corresponding correct localization. In our set of 60 cases, 2 hours of training were required to learn how to localize all cases correctly. From a control set of 25 cases, 23 were interpreted by the system satisfactorily. the neural network approach can be useful in situations where causal relations between the electrocardiogram and underlying mechanism are partly undefined.