Each year in the US, there are thousands of children who are born with a severe congenital deformation, where one of the two ventricles in the heart is severely underdeveloped. If they survive infancy, these children can go on to live full and normal lives. The mortality rate for this condition is 15%, and 63% of these deaths are due to cardio-respiratory arrests. This alarming rate of arrest events persists despite vigilant ICU care with the best available intensive monitoring equipment. Our overall goal is to improve current patient monitoring systems by developing machine learning algorithms that can predict the onset of an arrest event, hours before it occurs. This early warning indication can be provided to nurses and doctors who can intervene to prevent these life-threatening events from occurring, improving outcomes for these critically ill children. Preliminary studies at Texas Children's Hospital have resulted in a computer algorithm that can estimate the odds of arrest in single ventricle children, 1-2 hours prior to overt symptoms. The algorithm is based on a logistic regression risk model, and was developed using over 55,000 hours of vital sign observations.
The specific aims of the proposed research are: (1) To test the hypothesis that this computer algorithm can provide an early warning of arrest, with sufficient accuracy for clinical use; (2) To improve the performance of this algorithm by incorporating additional sensor technologies; (3) To understand the relationship between the risk score provided by this algorithm and other post-surgical complications that commonly occur in these children during the their hospitalization.
Aim 1 is a validation study of this algorithm on a large, prospective, and independent cohort, in order to measure its true predictive performance. Performance metrics to be measured will be the ROC area and positive and negative likelihood ratios. This will help us determine the optimal threshold for the detection of an arrest event.
Aims 2 and 3 focus on improving the detection accuracy of the algorithm using Near Infrared Reflectance Spectroscopy (NIRS) monitoring, and relating the risk of arrest to outcomes such as mechanical circulatory support, re-operation, arrhythmia, and necrotizing enterocolitis. Successful completion of these aims will result in the first clinically validated, real- time early warning system for anticipating acute arrest events in children with single ventricle physiology. The techniques and technologies developed in this work are immediately translatable to other diseases and conditions for both adults and children.
Thousands of children each year are born in the US with only a single functional ventricle, putting them at substantial risk of acute cardiac arrest and ultimately death. This project will validate a new monitoring technology which can predict the onset of arrest 1-2 hours before it occurs in this population. This technology will provide nurses and doctors the valuable time that they need to mitigate problems leading up to arrest before these problems become life threatening, improving the chance of survival for these critically ill children.
| Hendryx, Emily P; Rivière, Béatrice M; Sorensen, Danny C et al. (2018) Finding representative electrocardiogram beat morphologies with CUR. J Biomed Inform 77:97-110 |
| Vu, Eric L; Rusin, Craig G; Penny, Dan J et al. (2017) A Novel Electrocardiogram Algorithm Utilizing ST-Segment Instability for Detection of Cardiopulmonary Arrest in Single Ventricle Physiology: A Retrospective Study. Pediatr Crit Care Med 18:44-53 |
