The objective of the current study is to compare the relative performance of decision
tree, neural network, and logistic regression for predicting 30-day and 1-year mortality
in a real-word, unfiltered dataset (n = 47, 391) of patients hospitalized with acute
myocardial infarction. Area under the ROC curve (AUC) was used for evaluating performance
of a learning algorithm. For 30-day mortality, we achieved an average of 0.788 for
decision tree models, 0.837 for neural net models and 0.836 for regression models
on training set (on validation sets: 0.774, 0.835 and 0.834, respectively). For 1-year
mortality, the averages were 0.754 for decision tree models, 0.8194 for neural net
models and 0.8191 for regression models (on validation sets: 0.743, 0.8179 and 0.8176,
respectively). Differences were non-significant between neural network and regression,
but both significantly outperformed decision trees. The machine learning methods investigated
in the present study could not outperform traditional regression modelling for mortality
prediction in myocardial infarction. (C) 2019 Elsevier B.V. All rights reserved.
