Introduction: Nearly 19,000 patients admitted to the pediatric intensive care unit annually will experience cardiac arrest and less than forty percent of those will survive to hospital discharge without neurological injury. Cardiac arrest prediction models utilizing machine l

earning algorithms have shown superior predictive power in comparison to regression methods (AUROC 0.975 vs. 0.865 (Kennedy, Pediatric Critical Care Medicine, 2015)), but lack the precision needed for real-time utilization in the int

ensive care unit. Generating a more robust feature set, we hypothesize, will enhance the predictive power and potentially

allow for a live, clinically useful prediction tool. One such avenue for such data is the bedside telemetry system and continuous electrocardiogram. These systems are ubiquitous in pediatric intensive care units and are non-invasive in nat

ure, which makes them an ideal source for data with which to make clinical predictions. However, ECG telemetry data is prone to motion artifact and interference and requires substantial preprocessing prior to use in machine learning applications. Methods: ECG signals were obtained from patients that experienced cardiac arrest (defined as those that received at least two minutes of CPR) and age matched controls that did not receive CPR. Electrocardiogram signals were pre-processed using median filtering and R waves were identified using a peak finding algorithm (Figure 1). Each ECG complex was aligned by the R wave and a mean representative waveform was generated for each minute of ECG data present (Figure 2). The representative waveform was then used to perform interval calculations, amplitude measurements, and encode these morphological elements as features to be used in a machine learning model for pediatric cardiac arrest (Figure 3). Results: While this study is still in progress, a sample of an age matched control in comparison to a cardiac arrest patient is provided as a three-dimensional mesh for visual comparison (Video 1 & 2). Conclusion: Subtle variations and changes in the ECG waveform morphology have not previously been used as a feature in predictive models of cardiac arrest. Anecdotal evidence suggests that this variability may distinguish arrest cases from non-arrest cases.



Author: Thomas Fogarty III

Coauthor(s): Jeffrey Kim MD, Ronald Bronicki MD, Jorge Coss-Bu MD, and Curtis Kennedy MD PhD

Status: Work In Progress