Towards the prediction of transient ST changes

This paper studies the ECG signal prior to a transient ST change. Two hypotheses are proposed. The first is that various types of ST changes can be differentiated using the signal just prior to the ST event. The second is that ischemic ST changes can be differentiated from non-events, again using the signal prior to the ST event. A machine learning approach, based on Gaussian mixture models and maximum likelihood Bayesian classification, is used to analyze the ECG signal. Two sets of feature extraction techniques, reconstructed phase space and Karhunen Loeve transform, are applied, both of which capture morphological characteristics of the ECG signal. The results in addressing the first hypothesis show that information indicative of the type of ST change is present in the signal prior to the onset of the ST event; however the classification accuracy is low. The second hypothesis cannot be affirmed with the results presented here