A new Elixhauser-based comorbidity summary measure to predict in-hospital mortality

Abstract

Background: Recently, van Walraven developed a weighted summary score (VW) based on the 30 comorbidities from the Elixhauser comorbidity system. One of the 30 comorbidities, cardiac arrhythmia, is currently excluded as a comorbidity indicator in administrative datasets such as the Nationwide Inpatient Sample (NIS), prompting us to examine the validity of the VW score and its use in the NIS.

Methods: Using data from the 2009 Maryland State Inpatient Database, we derived weighted summary scores to predict in-hospital mortality based on the full (30) and reduced (29) set of comorbidities and compared model performance of these and other comorbidity summaries in 2009 NIS data.

Results: Weights of our derived scores were not sensitive to the exclusion of cardiac arrhythmia. When applied to NIS data, models containing derived summary scores performed nearly identically (c statistics for 30 and 29 variable-derived summary scores: 0.804 and 0.802, respectively) to the model using all 29 comorbidity indicators (c=0.809), and slightly better than the VW score (c=0.793). Each of these models performed substantially better than those based on a simple count of Elixhauser comorbidities (c=0.745) or a categorized count (0, 1, 2, or ≥ 3 comorbidities; c=0.737).

Conclusions: The VW score and our derived scores are valid in the NIS and are statistically superior to summaries using simple comorbidity counts. Researchers wishing to summarize the Elixhauser comorbidities with a single value should use the VW score or those derived in this study.

Figure 1

Calibration plots for logistic regression models predicting in-hospital mortality. The plots display the relationship between predicted mortality and observed mortality. Perfect calibration is represented by the 45-degree line. Predictors in baseline model were age, sex, race, length of stay, expected primary payer, and occurrence of operation. Other models include the variables in the baseline model in addition to the specified comorbidity adjustment. The solid line represents a nonparametric smooth curve (lowess algorithm). The dashed curves represent a 95% confidence band for each calibration curve. For relative comparison of distributions, histograms of predicted probabilities of in-hospital mortality are overlaid on the x-axis. Binary₂₉—all 29 Elixhauser comorbidities included as binary indicators. VW score—weighted summary score derived by van Walraven. SID₃₀—weighted summary score derived in 2009 Maryland State Inpatient Database data using all 30 original Elixhauser comorbidities. SID₂₉—weighted summary score derived in 2009 Maryland State Inpatient Database data using 29 Elixhauser comorbidities (excluded cardiac arrhythmia). Count—summary score obtained by summing the number of Elixhauser comorbidities. Count4—categorical variable defined by the presence of 0, 1, 2, or ≥3 Elixhauser comorbidities.

Figure 2

Receiver-operating characteristic (ROC) curves for logistic regression models predicting in-hospital mortality. Predictors in baseline model were age, sex, race, length of stay, expected primary payer, and occurrence of operation. Other models include the variables in the baseline model in addition to the specified comorbidity adjustment. Binary₂₉— all 29 Elixhauser comorbidities included as binary indicators. VW score—weighted summary score derived by van Walraven. SID₃₀—weighted summary score derived in 2009 Maryland State Inpatient Database data using all 30 original Elixhauser comorbidities. SID₂₉—weighted summary score derived in 2009 Maryland State Inpatient Database data using 29 Elix-hauser comorbidities (excluded cardiac arrhythmia). Count— summary score obtained by summing the number of Elix-hauser comorbidities. Count₄—categorical variable defined by the presence of 0, 1, 2, or ≥3 Elixhauser comorbidities.