Veterans Affairs intensive care unit risk adjustment model: validation, updating, recalibration

A valid metric is critical to measure and report intensive care unit (ICU) outcomes and drive innovation in a national system.

To update and validate the Veterans Affairs (VA) ICU severity measure (VA ICU).

A validated logistic regression model was applied to two VA hospital data sets: 36,240 consecutive ICU admissions to a stratified random sample of moderate and large hospitals in 1999-2000 (cohort 1) and 81,964 cases from 42 VA Medical Centers in fiscal years 2002-2004 (cohort 2). The model was updated by adding diagnostic groups and expanding the source of admission variables.

C statistic, Hosmer-Lemeshow goodness-of-fit statistic, and Brier's score measured predictive validity. Coefficients from the 1997 model were applied to predictors (fixed) in a logistic regression model. A 10 x 10 table compared cases with both VA ICU and National Surgical Quality Improvement Performance metrics. The standardized mortality ratios divided observed deaths by the sum of predicted mortality.

The fixed model in both cohorts had predictive validity (cohort 1: C statistic = 0.874, Hosmer-Lemeshow goodness-of-fit C statistic chi-square = 72.5; cohort 2: 0.876, 307), as did the updated model (cohort 2: C statistic = 0.887, Hosmer-Lemeshow goodness-of-fit C statistic chi-square = 39). In 7,411 cases with predictions in both systems, the standardized mortality ratio was similar (1.04 for VA ICU, 1.15 for National Surgical Quality Improvement Performance), and 92% of cases matched (+/-1 decile) when ordered by deciles of mortality. The VA ICU standardized mortality ratio correlates with the National Surgical Quality Improvement Performance standardized mortality ratio (r2 = .74). Variation in discharge and laboratory practices may affect performance measurement.

The VA ICU severity model has face, construct, and predictive validity.