Probabilistic Prediction of Myocardial Infarction

In this research, predictive risk models of myocardial infarction are constructed for a patient based on their clinical history and EKG information. The data used for the study consists of 697 patients who were presented to the emergency rooms of hospitals in Sheffield, England and Edinburgh, Scotland. We find that both logistic regression and single layer neural networks are reasonable models of prediction. These results are verified with analysis of areas under the ROC curves of the models, which have approximately 0.94 for logistic regression and 0.96 for neural networks. It is demonstrated that while a logistic regression with a larger number of variables is slightly more predictive, a smaller logistic regression model is a feasible tool for predictive purposes. These models are an important contribution to decision support systems where a physician has the ability and time to research the full clinical history of a patient, but also in an emergency room setting, where time is of the essence and less comprehensive medical profile is available.