Physiological data are routinely recorded in intensive care, but their use for rapid assessment of illness severity has been limited. The data is high-dimensional, noisy, and changes rapidly; moreover, small changes that occur in a patient’s physiology over long periods of time are difficult to detect, yet can lead to catastrophic outcomes. A physician’s ability to recognize complex patterns across these high-dimensional measurements is limited. We propose a nonparametric Bayesian method for discovering informative representations in such continuous time series that aids both exploratory data analysis and feature construction. When applied to data from premature infants in the neonatal ICU (NICU), our model obtains novel clinical insights. Based on these insights, we devised the Physiscore, a novel risk prediction score that combines patterns from continuous physiological signals to predict infants at risk for developing major complications in the NICU. Using only 3 hours of non-invasive data from birth, Physiscore very successfully predicts morbidity in preterm infants. Physiscore performed consistently better than other neonatal scoring systems, including the Apgar, which is the current standard of care, and SNAP, a machine learning based score that requires multiple invasive tests. This work was recently published on the cover of Science Translational Medicine (Science’s new journal aimed at translational medicine work), and was covered by numerous press sources.
Speaker Biography
Suchi Saria is finishing her PhD in Computer Science at Stanford under Daphne Koller. Her main research interest lies in machine learning and data driven optimizations for health care. Her work has appeared on popular press sources such as CBS Radio, Science NOW, KCBS and The San Francisco chronicle. Her works have also been given a Best Student Paper and a Best Student Paper finalist awards. She is a recipient of the Stanford Graduate Fellowship (SGF) and two Microsoft full-tuition scholarships.