In-Hospital Cardiac Arrest Detection Performance Analysis and Comparison on Effective Feature Selection

Area of Science:

• Clinical Informatics
• Machine Learning in Healthcare
• Cardiovascular Medicine

Background:

• In-hospital cardiac arrest (IHCA) presents significant clinical challenges.
• Effective patient screening and timely treatment are crucial for improving outcomes.
• Predictive modeling can aid in early identification and intervention for IHCA.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) model for predicting IHCA risk upon hospital admission.
• To assess the impact of various feature selection methods on ML model performance for IHCA prediction.

Main Methods:

• Utilized a dataset of 25,149 patients, with 320 experiencing IHCA.
• Compared three feature selection techniques (statistical tests, regression, correlation) with four ML models (AdaBoost, XGBoost, Random Forest, Logistic Regression).
• Evaluated 16 distinct models using metrics including AUROC, AUPRC, accuracy, recall, precision, and specificity.

Main Results:

• The XGBoost model achieved the highest performance, with an AUROC of 0.987 and an accuracy of 0.992.
• Key predictors identified include age, albumin levels, sinus arrhythmia, activated partial thromboplastin time, and protein levels.
• The choice of feature selection method significantly influenced the performance of different ML models.

Conclusions:

• The XGBoost algorithm, combined with appropriate feature selection, provides a highly effective tool for predicting IHCA.
• This predictive model can assist clinicians in identifying high-risk patients for proactive management.
• Optimizing feature selection is critical for maximizing the accuracy of ML-based clinical prediction models.