A Machine Learning Approach for Predicting In-Hospital Cardiac Arrest Using Single-Day Vital Signs, Laboratory Test...

Area of Science:

• Medical Informatics
• Clinical Prediction Models
• Cardiology

Background:

• Predicting in-hospital cardiac arrest (IHCA) is vital for reducing mortality.
• Current models using only vital signs may not fully capture patient risk.
• Improving IHCA prediction requires integrating diverse patient data.

Purpose of the Study:

• To enhance in-hospital cardiac arrest (IHCA) prediction models.
• To evaluate the impact of incorporating laboratory test results and ICD-10 diagnosis codes.
• To improve clinical decision-making for better patient outcomes.

Main Methods:

• Retrospective cohort study of 62,061 adult internal medicine patients (Jan 2010-Aug 2022).
• Utilized eXtreme Gradient Boosting (XGBoost) model.
• Trained model with vital signs, 14 laboratory tests, and ICD-10 diagnosis block (ICD10BD).

Main Results:

• The combined model (vitals, labs, ICD10BD) achieved AUCs of 0.934 (GW) and 0.896 (ICU).
• Models with vitals and labs showed AUCs of 0.925 (GW) and 0.878 (ICU).
• Vitals-only models had lower AUCs: 0.839 (GW) and 0.828 (ICU).

Conclusions:

• Integrating laboratory results and diagnosis codes with vital signs significantly improves IHCA prediction.
• The enhanced XGBoost model offers a more comprehensive risk assessment.
• This approach can enhance clinical decision-making and improve patient outcomes in hospitals.