Prediction of acute myocardial infarction with artificial neural networks in patients with nondiagnostic electrocardiogram

Abstract

Background: Myocardial infarction remains one the leading causes of mortality and morbidity and involves a high cost of care. Early prediction can be helpful in preventing the development of myocardial infarction with appropriate diagnosis and treatment. Artificial neural networks have opened new horizons in learning about the natural history of diseases and predicting cardiac disease. Methods: A total of 935 cardiac patients with chest pain and nondiagnostic electrocardiogram (ECG) were enrolled and followed for 2 weeks in two groups based on the appearance of myocardial infarction. Two types of data were used for all patients: nominal (clinical data) and quantitative (ECG findings). Two different artificial neural networks – radial basis function (RBF) and multi-layer perceptron (MLP) – were used to classify the two groups. Results: The RBF neural network had an accuracy of 83% with ECG findings and an accuracy of 78% with clinical features. When and clinical data were used in an MLP neural network trained with a genetic algorithm, ECG results led to a classification accuracy of 96% and clinical data yielded an accuracy of 84.5%. Conclusion: Both neural network structures predicted MI within about 2 weeks of hospital referral with an acceptable degree of accuracy in patients with nondiagnostic ECG. The MLP neural network significantly outperformed the RBF network because of the use of the genetic algorithm, which provided a global strategy to accurately determine MLP weights (clinical trials registry: NCT01870258). Key words: Artificial neural networks, Electrocardiography, Myocardial infarction, Multi-layer perceptron (MLP), Radial basis function (RBF).