A Comparison of Administrative and Physiologic Predictive Models in Determining Risk Adjusted Mortality Rates in Critically Ill Patients

BACKGROUND Hospitals are increasingly compared based on clinical outcomes adjusted for severity of illness. Multiple methods exist to adjust for differences between patients. The challenge for consumers of this information, both the public and healthcare providers, is interpreting differences in risk adjustment models particularly when models differ in their use of administrative and physiologic data. We set to examine how administrative and physiologic models compare to each when applied to critically ill patients. METHODS We prospectively abstracted variables for a physiologic and administrative model of mortality from two intensive care units in the United States. Predicted mortality was compared through the Pearsons Product coefficient and Bland-Altman analysis. A subgroup of patients admitted directly from the emergency department was analyzed to remove potential confounding changes in condition prior to ICU admission. RESULTS We included 556 patients from two academic medical centers in this analysis. The administrative model and physiologic models predicted mortalities for the combined cohort were 15.3% (95% CI 13.7%, 16.8%) and 24.6% (95% CI 22.7%, 26.5%) (t-test p-value<0.001). The r(2) for these models was 0.297. The Bland-Atlman plot suggests that at low predicted mortality there was good agreement; however, as mortality increased the models diverged. Similar results were found when analyzing a subgroup of patients admitted directly from the emergency department. When comparing the two hospitals, there was a statistical difference when using the administrative model but not the physiologic model. Unexplained mortality, defined as those patients who died who had a predicted mortality less than 10%, was a rare event by either model. CONCLUSIONS In conclusion, while it has been shown that administrative models provide estimates of mortality that are similar to physiologic models in non-critically ill patients with pneumonia, our results suggest this finding can not be applied globally to patients admitted to intensive care units. As patients and providers increasingly use publicly reported information in making health care decisions and referrals, it is critical that the provided information be understood. Our results suggest that severity of illness may influence the mortality index in administrative models. We suggest that when interpreting "report cards" or metrics, health care providers determine how the risk adjustment was made and compares to other risk adjustment models.