Topic Models for Mortality Modeling in Intensive Care Units

Mortality prediction is an important problem in the intensive care unit (ICU) because it is helpful for understanding patients’ evolving severity, quality of care, and comparing treatments. Most ICU mortality models primarily consider structured data and physiological waveforms (Le Gall et al., 1993). An important limitation of these structured data approaches is that they miss a lot of vital information captured in providers’ free text notes and reports. In this paper, we propose an approach to mortality prediction that incorporates the information from free text notes using topic modeling.