Where t represents a given node, i represents any classification of labels, and p(i|t) represents the proportion of label classification i in node t . Since the effect of information entropy and Gini coefficient is the same in practical application, information entropy is selected to calculate impurities in the decision tree constructed in this study.
Random forest (RF) is a very representative bagged ensemble algorithm. All its basic evaluators are decision trees, and the forest composed of classification trees is the random forest classifier. Each decision tree provides a different solution to the problem. The solutions of all the decision trees are eventually combined (usually by voting or averaging) into a single final model output [21], which is usually a more stable and accurate prediction. In this study, the model predicted whether a patient would die in hospital by providing the probability of death for each patient. The probability is determined by the ratio of the decision tree with positive results to the total number of decision trees.
For the selection of feature variables, we adopted the feature importance attribute of RF and LASSO regression. The feature importance of RF indicates the contribution of each predictor to the model, so the most relevant predictor can be selected to represent most of the performance of the model. Least Absolute Shrinkage and Selection Operator (LASSO) imposes a constraint on the sum of the absolute values of the model parameters, applying a regularization process in which it penalizes the coefficients of the regression variables and sets some of them precisely to zero [22]. Variables with non-zero results were selected as important variables.