Multivariable models using administrative data and biomarkers to adjust for case mix in the ICU.

BACKGROUND: Routinely collected laboratory biomarkers could improve control of confounding from disease severity in non-interventional studies of general intensive care unit (ICU) patients. Their ability to predict both short- and long-term mortality was evaluated.

METHODS: The performance of age, sex, Charlson co-morbidity index, and baseline values of ten predefined blood biomarkers as predictors of 30-day and 1-year mortality was evaluated in 5505 general ICU stays.

RESULTS: Regression models based on age, sex, Charlson index, and biomarkers were somewhat less accurate in predicting 30-day mortality (c-index 0.83, Brier score 0.27) compared to the SAPS II score (c-index = 0.88, Brier score = 0.09) and in predicting 1-year mortality (c-index = 0.82, Brier score = 0.31) compared to the SAPS II score (c-index = 0.85, Brier score = 0.13). Cystatin C improved predictive ability slightly compared to creatinine, but age and Charlson comorbidity index were more important predictors. Using multiple imputation to replace missing biomarker values notably improved predictive ability of the models.

CONCLUSIONS: Automatically collected baseline variables are almost as predictive of both short- and long-term mortality in general ICU patients, as the SAPS II score. This can facilitate internal control of confounding in non-interventional studies of mortality using administrative data.