Prediction of Cardiac Arrest in the Pediatric ICU
Team: Team Owl
- Program: Biomedical Engineering
- Course: Precision Care Medicine
Project Description:
Only one out of four children who have a cardiac arrest in the hospital will survive until discharge. To address this problem, we have used machine learning to predict cardiac arrest up to three hours in advance. This increased warning time will allow clinicians to intervene earlier, improving patient outcomes.
Project Photo:
Team Owl’s logo of an ECG wave and an owl’s beak
Project Poster
Project Post Summary:
In-hospital cardiac arrest (IHCA) occurs in approximately 15,200 pediatric patients annually. However, only one out of four patients survives until discharge. We have created machine learning models to predict IHCA up to three hours in advance. Our data consists of 240 Hz ECG waveform data, 0.5 Hz vital signs time series data, and medications data from 1,145 patients in the pediatric intensive care unit. We generated five minute windows on our data to perform feature engineering. We created 23 heart rate variability (HRV) metrics from ECG R wave peaks (NN intervals) after filtering outliers and removing ectopic beats. A total of 96 summary statistics were calculated for 12 vital signs, such as respiratory rate and blood pressure. For the medications data, drugs were classified into 46 therapeutic classes, and boolean features were created. We trained six machine learning models: logistic regression, support vector machine, random forest, XGBoost, LightGBM, and a soft voting ensemble. After evaluation on held out testing data, the XGBoost model performed the best, with 0.971 auROC, 0.798 auPRC, 99.5% sensitivity, and 69.6% specificity. Upon investigating feature importances, we determined that a combination of heart rate variability metrics, vitals signs data (e.g. a low respiratory rate), and several therapeutic classes of medications are key indicators of pediatric IHCA. We have created high-performing models to predict pediatric IHCA up to three hours in advance by combining ECG waveforms, vital signs time series data, and medications data. This increased warning time will allow clinicians to intervene earlier, improving patient outcomes.