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

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Prediction Models

Background:

• Early risk stratification for in-hospital cardiac arrest (IHCA) is crucial for timely intervention.
• Conventional early-warning scores have limitations in predicting IHCA.
• Electronic health records (EHRs) offer rich data for developing advanced prediction models.

Purpose of the Study:

• To develop and compare machine learning (ML) algorithms against logistic regression for IHCA prediction.
• To enhance early risk stratification for IHCA using EHR data.
• To explore the integration of ML models into hospital early warning systems.

Main Methods:

• Retrospective case-control study at a tertiary medical center (800 IHCA cases, 3,464 controls).
• Candidate predictors included demographics, comorbidities, vital signs, and laboratory measurements.
• Five models (logistic regression, decision tree, random forest, XGBoost, MARS) were trained and validated; performance evaluated using AUC, accuracy, sensitivity, specificity, and F1 score.

Main Results:

• XGBoost achieved the highest accuracy (0.883) with strong discrimination (AUC 0.909).
• Random forest showed comparable discrimination (AUC 0.910) and slightly lower accuracy (0.876).
• ML models identified key predictors like blood urea nitrogen, heart rate, and heart failure, offering deeper insights than traditional regression.

Conclusions:

• Integrating ML with regression enhances IHCA risk prediction by capturing complex relationships.
• ML approaches can strengthen hospital early-warning systems for earlier detection and intervention.
• Improved prediction models can ultimately lead to better patient outcomes.