A Functional Approach to Modeling Time-Dependent Covariates in the Intensive Care Unit

• Activity Number: 353
• Type: Contributed
• Date/Time: Tuesday, July 31, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Jonathan Gellar*+ and Parichoy Pal Choudhury and Yenny Webb-Vargas and Ciprian Crainiceanu
• Companies: and Johns Hopkins Bloomberg School of Public Health and Johns Hopkins Bloomberg School of Public Health and The Johns Hopkins University
• Keywords: Functional data analysis ; Intensive care unit ; Functional principal component analysis ; Time-varying covariates ; Penalized regression ; Acute lung injury

Abstract

Studies that take place in hospital settings often include exposures that are measured longitudinally, throughout a subject's hospital stay. Traditional approaches to modeling this type of exposure usually involve collapsing each subject's data into a single summary measure, such as the median, maximum, or cumulative exposure. We propose a new method that incorporates each subject's entire covariate history, by treating the longitudinally measured exposure as a function of time. After performing a functional principal component decomposition of the covariate function, penalized functional regression is used to assess the exposure-outcome relationship. As an application, we investigated how patterns in organ function relate to in-hospital mortality in the intensive care unit (ICU). We address two challenges in applying a functional model: (1) the number of days in the ICU varies between subjects, and (2) subjects may leave and later return to the ICU, resulting in gaps in their data. These methods are not only able to identify previously unobserved features of the exposure-outcome relationship, but they also are shown to be more predictive of the outcome than traditional approaches.