Machine Learning Models for Predicting in-Hospital Cardiac Arrest: A Comparative Analysis with Logistic Regression

Abstract

Purpose: To develop and compare multiple machine learning (ML) algorithms with traditional logistic regression for predicting in-hospital cardiac arrest (IHCA) using comprehensive electronic health record data, with the goal of improving early risk stratification beyond conventional early-warning scores and providing potential integration into hospital early warning systems for timely clinical intervention. Patients and Methods: We performed a retrospective case-control study at a large tertiary medical center, including 800 IHCA cases and 3,464 controls. Candidate predictors comprised demographics, comorbidities, vital signs, and laboratory measurements. Five models-logistic regression, decision tree, random forest, XGBoost, and multivariate adaptive regression splines (MARS)-were trained and validated. Performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. Results: XGBoost yielded strong discrimination and the highest accuracy (AUC 0.909; accuracy 0.883), while random forest showed comparable discrimination (AUC 0.910) with slightly lower accuracy (0.876). Logistic regression performed robustly but lower than ML models (AUC 0.895; accuracy 0.876). ML models consistently identified clinically meaningful predictors-including blood urea nitrogen, heart rate, and pre-existing heart failure-offering insights beyond traditional regression. Conclusion: Integrating ML approaches with conventional regression enhances IHCA risk prediction by capturing non-linear relationships and interactions while retaining the interpretability of regression. These approaches could strengthen hospital early-warning systems, enabling earlier detection and intervention, and ultimately improving patient outcomes.