Mortality prediction in a cardiac intensive care unit using machine learning and penalized regression: comparison with APACHE II and GRACE

Abstract Introduction : To evaluate and compare the prognostic performance of traditional scores and machine learning–based and penalized regression models for in-hospital mortality prediction in patients admitted to a heterogeneous cardiac intensive care unit. Methods : This retrospective observational study included adult patients admitted to a tertiary university cardiac intensive care unit in south of Brazil. Mortality prediction models based on Extreme Gradient Boosting (XGBoost) and Firth’s penalized logistic regression were developed using routinely collected clinical and laboratory data. Model performance was compared with APACHE II in the overall cohort and with APACHE II and GRACE in the acute coronary syndrome subgroup using area under the receiver operating characteristic curve (AUC). Results : A total of 1,855 patients were included, with an in-hospital mortality rate of 8.2%. In the overall population, XGBoost demonstrated superior discriminative performance (AUC 0.900) compared with APACHE II (AUC 0.840), while penalized logistic regression also showed high discrimination (AUC 0.889). In the acute coronary syndrome subgroup, both XGBoost and penalized regression outperformed GRACE. The advantage of XGBoost was most evident in regions of low false-positive rates. Renal dysfunction and hemodynamic instability were the main prognostic determinants. Conclusion : Machine learning–based and penalized regression models, particularly XGBoost, demonstrated robust and clinically relevant performance for mortality prediction in a heterogeneous cardiac intensive care unit. These approaches may complement traditional prognostic scores by improving early risk stratification and supporting clinical decision-making in high-complexity settings.