Hybrid phenotype-guided modeling across algorithm–feature regimes with application to ICU mortality prediction for Acinetobacter baumannii

Acinetobacter baumannii causes severe intensive care unit (ICU) infections with high mortality, yet most prediction tools rely on risk scores or supervised machine learning (ML) and overlook hidden patient subgroups. This study applied a hybrid machine learning framework combining unsupervised clustering and supervised prediction to refine ICU mortality estimation and evaluate whether incorporating phenotype information enhances performance. Patient phenotypes were identified using clustering, and cluster membership was incorporated as an additional predictive feature. ML models were trained with and without cluster membership under two feature settings: full-feature models including all variables and reduced-feature models limited to significant predictors identified within the clusters. The resulting phenotypes were clinically distinct and strongly associated with mortality, demonstrating that data-driven patient grouping can provide complementary prognostic information. Incorporating phenotype membership improved predictive accuracy in a context-dependent manner, varying by feature regime and the learning algorithm. This study introduces a novel framework for phenotype-guided critical care modeling that bridges unsupervised and supervised learning, advancing personalized critical care and supporting global efforts to reduce preventable ICU mortality.