Development and validation of a model that uses enhanced administrative data to predict mortality in patients with sepsis

Abstract

Objective: We aimed to determine whether a sepsis risk-adjustment model that uses only administrative data could be used when other intensive care unit risk-adjustment methods are unavailable.

Design: Cohort study with development and validation cohorts.

Patients: The development cohort included 166,931 patients at 309 hospitals that cared for at least 100 patients with sepsis between 2004 and 2006. The validation cohort included 357 adult sepsis patients who were enrolled in Project IMPACT, 2002-2009.

Measurements and main results: We developed a multilevel mixed-effects logistic regression model to predict mortality at the patient level. Predictors included patient demographics (age, sex, race, insurance type), site and source of sepsis, presence of 25 individual comorbidities, treatment (within the first 2 days of hospitalization) with mechanical ventilation and/or vasopressors, and/or admission to the intensive care unit (within 2 days of hospitalization). We validated this model in 357 sepsis patients who were admitted to the intensive care unit at a single academic medical center and who had a valid Acute Physiology and Chronic Health Evaluation II score, a valid Simplified Acute Physiology Score II, and a valid Mortality Probability Model III score. Overall, 33,192 patients (19.9%) died in the hospital. In the development cohort, the predicted mortality ranged from 0.002 to 0.938 with a mean of 0.199. The model's area under the receiver operating characteristic curve was 0.78. In the validation cohort, all models had modest discriminatory ability and the areas under the receiver operating characteristic curves of all models were statistically similar (Acute Physiology and Chronic Health Evaluation II, 0.71; Simplified Acute Physiology Score II, 0.74; Mortality Probability Model III, 0.69; administrative model, 0.69; p value that the areas under the receiver operating characteristic curves are different, .35). The Hosmer-Lemeshow statistic was significant (p < .01) for Acute Physiology and Chronic Health Evaluation II, Simplified Acute Physiology Score II, and Mortality Probability Model III but was nonsignificant (p = .11) for the administrative model.

Conclusions: A sepsis mortality model using detailed administrative data has discrimination similar to and calibration superior to those of existing severity scores that require chart review. This model may be a useful alternative method of severity adjustment for benchmarking purposes or for conducting large, retrospective epidemiologic studies of sepsis patients.