Abstract

Using Physiologic Biosignals to Improve Cardiac Arrest Prediction in Critically Ill Children

Author: Thomas Fogarty III

Introduction: Nearly 2% of all patients admitted to the pediatric intensive care unit will experience cardiac arrest. Outcomes from cardiac arrest are poor – less than 60% will survive to hospital discharge and of those that do survive, a majority will have severe neurological consequences. Despite the availability of high resolution data streams from telemetry systems, no prediction algorithms have been developed to enhance provider awareness of impending cardiac arrest. With rapid recognition and response, the outcomes of cardiac arrest are significantly improved. Preliminary machine learning approaches to predict cardiac arrest have yielded models with ROC statistics as high as 0.975 using basic telemetry data and clinical variables obtained from paper charts (Kennedy, 2015). Our hypothesis is that predictive models for cardiac arrest that include ECG waveform and heart rate variability features in time series will outperform models that lack these features. Methods: A total of 105 cardiac arrest cases were identified in our PICU over the 5 years since our EMR was introduced served as our case and age/sex matched controls were included on a 1:1 basis. Electrocardiogram waveform morphology and heart rate variability data were encoded as time series features in a semi-automated fashion using the MATLAB software suite (R2016a). Clinical data was obtained through query of the electronic data warehouse in our institution. Various machine learning techniques will be used to build a predictive model for cardiac arrest. Results: Study in progress. Conclusion: Integration of time series EMR and telemetry data will likely provide a more robust framework for clinical event prediction than either on their own. Automated methods to extract ECG features from telemetry systems are needed to develop event prediction models into clinically useful alarms. This study will provide a framework for such automation and demonstrate its potential utility in cardiac arrest prediction.

Co Author/Co-Investigator Names/Professional Title: Thomas Fogarty MD, Jeffrey Kim MD, Ronald Bronicki MD, Jorge Coss-Bu MD, Curtis Kennedy MD PhD