Predicting Cardiac Arrest in Children with Heart Disease: A Novel Machine Learning Algorithm.

Area of Science:

• Pediatric Cardiology
• Biomedical Informatics
• Machine Learning in Healthcare

Background:

• Children with heart disease face a higher risk of cardiac arrest.
• Despite advancements in resuscitation, survival rates remain low for this population.
• Predictive algorithms using machine learning offer potential for cardiac arrest prevention.

Purpose of the Study:

• To evaluate a novel machine learning algorithm for predicting cardiac arrest in pediatric patients with heart disease.
• To compare the performance of machine learning against traditional logistic regression for cardiac arrest prediction.

Main Methods:

• Retrospective analysis of electronic health records (EHR) from a pediatric cardiac intensive care unit (CICU) (2010-2019).
• Collected demographic, laboratory, and vital sign data for patients who experienced cardiac arrest and a control group.
• Compared a functional gradient boosting machine learning algorithm with time series data against logistic regression.

Main Results:

• The machine learning algorithm achieved a peak performance (AUROC 0.85) one hour prior to cardiac arrest.
• The algorithm utilized 11 variables from EHR data (vital signs and lab values).
• Performance was comparable to a previously developed multivariable logistic regression model.

Conclusions:

• A novel machine learning algorithm can predict cardiac arrest in pediatric CICU patients with performance similar to traditional models.
• The algorithm, developed on retrospective EHR data, shows promise for real-time clinical decision support.
• Further prospective, multicenter validation is necessary before clinical implementation.