Monitoring and Prediction of Cardiac Arrest in Pediatric ICU Patients with Machine Learning

• Program: Biomedical Engineering
• Course: EN.580.480 Precision Care Medicine

Project Poster

Open full size poster in new tab (PDF)

Project Post Summary:

Cardiac arrest stands as a predominant cause of mortality among children admitted to the pediatric intensive care unit (PICU) before discharge. To enable early detection of children at risk, we employed machine learning (ML) techniques to predict in-hospital cardiac arrests (IHCA) up to five hours in advance. Our data encompasses 240-Hz electrocardiogram (ECG), 60-Hz photoplethysmography (PPG), 0.5-Hz physiological time series, medication, demographics, and precursor events (respiratory failure) from 73 patients (n_IHCA = 14) admitted to the PICU at the Johns Hopkins Hospital. Features were sampled as non-overlapping five-minute windows. We derived 23 heart rate variability metrics from ECG waveforms and 21 summary statistics from 3 PPG morphological features. We also computed summary statistics for 10 vital signs, including respiratory rate, blood oxygen saturation, and ST segments. We represented 42 therapeutic drug classes and the occurrence of respiratory failure as binary features. Subsequently, four ML models—logistic regression, support vector machine, random forest, and XGBoost—were trained and evaluated based on nested cross-validation. Training various models with all features revealed that random forest has the best performance results across different feature combinations. For all-feature models, random forest achieved the following on the test sets: auPRC (area under PRC) = 0.937, auROC (area under ROC) = 0.940, accuracy = 0.848, specificity = 0.970, sensitivity/recall = 0.658, positive-predictive-value (PPV/precision) = 0.873, and negative-predictive-value (NPV) = 0.840. We demonstrated that our ML models can predict IHCA with notably high performance up to five hours before onset. The models developed here achieve and demonstrate actionable early warning of impending IHCA in pediatric patients using multimodal signals and electronic health record data that are collected routinely in the PICU.