Phenotyping of Risk Factors for Adverse Events: New Methods for Explaining Prognostic Estimates in Cardiology

The implementation of machine learning in clinical practice is limited by decision‑making opacity and the difficulty of interpreting prognostic models. Multivariate logistic regression yields transparent prognostic models that make it possible to describe how the predictors influence the occurrence of adverse events. Risk factors (RFs) and their combinations (phenotypes), which increase the likelihood of adverse events, can be considered the most comprehensible predictors for clinicians. The objective of this study is to develop novel methods for RF phenotype identification that ensure high interpretability and accuracy, and to validate them on predicting in‑hospital mortality (IHM) and atrial fibrillation (AF) in STEMI patients after PCI. Four methods are proposed: (1) minimizing logloss, (2) maximizing AUC of RF combinations, (3) building phenotypes as conjunctions of RFs, and (4) representing phenotypes as linear combinations of RFs. Comparison with existing approaches showed that methods 2 and 4 outperformed the continuous‑predictor model: method 2 achieved AUC 0.814 vs. 0.78 for AF, and method 4 achieved AUC 0.915 vs. 0.902 for IHM. Thus, the proposed methods enable accurate prognostic models that combine statistical performance with clinical transparency, facilitating the interpretation of adverse outcome risks.