Premium
Addition of time‐dependent covariates to a survival model significantly improved predictions for daily risk of hospital death
Author(s) -
Wong Jenna,
Taljaard Monica,
Forster Alan J.,
Escobar Gabriel J.,
van Walraven Carl
Publication year - 2013
Publication title -
journal of evaluation in clinical practice
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.737
H-Index - 73
eISSN - 1365-2753
pISSN - 1356-1294
DOI - 10.1111/j.1365-2753.2012.01832.x
Subject(s) - covariate , decile , proportional hazards model , medicine , concordance , survival analysis , demography , statistics , emergency medicine , mathematics , sociology
Abstract Rational, aims and objectives The study aims to determine the extent to which the addition of post‐admission information via time‐dependent covariates improved the ability of a survival model to predict the daily risk of hospital death. Method Using administrative and laboratory data from adult inpatient hospitalizations at our institution between 1 April 2004 and 31 March 2009, we fit both a time‐dependent and a time‐fixed Cox model for hospital mortality on a randomly chosen 66% of hospitalizations. We compared the predictive performance of these models on the remaining hospitalizations. Results All comparative measures clearly indicated that the addition of time‐dependent covariates improved model discrimination and prominently improved model calibration. The time‐dependent model had a significantly higher concordance probability (0.879 versus 0.811) and predicted significantly closer to the number of observed deaths within all risk deciles. Over the first 32 admission days, the integrated discrimination improvement (IDI) and net reclassification improvement (NRI) were consistently above zero (average IDI of +0.0200 and average NRI of 62.7% over the first 32 days). Conclusions The addition of time‐dependent covariates significantly improved the ability of a survival model to predict a patient's daily risk of hospital death. Researchers should consider adding time‐dependent covariates when seeking to improve the performance of survival models.