Computational discovery of physiomes in critically ill children using deep learning
Abstract
Critical illness is dynamic and complex, but physicians often diagnose and treat based on parsimonious, static sets of symptoms and signs. For example, the Berlin definition of acute respiratory distress syndrome uses static criteria, including a threshold on the partial pressure of oxygen. 1 The Pediatric Risk of Mortality (PRISM III) score considers only the extreme values of a handful of variables during a 12-24 hour period. 2 However, the increasing volume of digital health data offers an opportunity to use computational methods to learn richer descriptors of illness (physiomes 3 or phenomes 4 ) that incorporate temporal dynamics and more variables. In this work, we use deep neural networks to mine patterns from multivariate clinical time series. We apply this to a large pediatric intensive care unit (PICU) database from Children’s Hospital Los Angeles (CHLA) 3 to learn patterns that are associated with known critical illnesses.
Zhengping Che
X-Humanoid
