Modified decision tree as a tool for interpreting predictive machine learning models in clinical medicine

A modified decision tree method has been developed, based on the multi-level categorization of predictors and the identification of risk factors for in-hospital mortality. When building the decision tree, the search for splitting conditions is optimized using only these risk factors. The method was tested for predicting in-hospital mortality in patients with ST-segment elevation myocardial infarction. It was found that the modified decision tree has better prediction quality metrics and a simpler structure than the classical decision tree. A model for the minimal extraction of production rules was developed to explain the inferences generated by the model. The application of the identified production rules for handling continuous predictors significantly improves prediction quality. The model based on the modified decision tree algorithm is an effective prognostic tool. It allows for a highly accurate estimation of the probability of in-hospital mortality in patients with ST-segment elevation myocardial infarction and provides a clinical interpretation of the prediction results.