The study addresses the problem of explaining the prediction results of machine learning models, developing methods for phenotyping risk factors, and predicting in-hospital mortality (IHM) in patients with ST-segment elevation myocardial infarction (STEMI) after percutaneous coronary intervention (PCI). For IHM prediction, risk factors were identified and their phenotypes were formed, providing not only transparency of decision-making but also improving prediction quality. Two methods were proposed for extracting risk factors and forming their phenotypes: entropy minimization and search for a separating curve based on AUC maximization. These methods were applied to a dataset of 4673 electronic health records of patients with STEMI and IHM prediction after emergency PCI. Risk factor phenotypes were identified, and IHM predictive models were developed based on them. The results demonstrated that the multifactor logistic regression (MLR) prediction model, with phenotypes identified by the entropy minimization method as predictors, was inferior in prediction accuracy to the MLR model with continuous predictors (AUC - 0.885 vs. 0.902, p-value = 0.036). The MLR model based on phenotypes formed by the separating lines method provided higher quality IHM prediction (AUC - 0.915 vs. 0.902 and 0.885, p-value = 0.029 and p-value <0.000001, respectively).