Session number: 1086
Abstract title: Impact of Comorbidity on Predicted Hospital Mortality in Intensive Care Unit Patients
Author(s):
JA Johnston - Veterans Affairs Medical Center, Cincinnati, OH; Division of General Internal Medicine, University of Cincinnati Medical Center, Cincinnati, OH.
J Tsevat - Veterans Affairs Medical Center, Cincinnati, OH; Division of General Internal Medicine, University of Cincinnati Medical Center, Cincinnati, OH.
DP Wagner - Department of Health Evaluation Sciences, University of Virginia School of Medicine, Charlottesville, VA.
D Welsh - Center for Practice Management and Outcomes Research, Veterans Affairs Ann Arbor HSR&D, Ann Arbor, MI.
S Timmons - Veterans Affairs Medical Center, Cincinnati, OH.
ML Render - Veterans Affairs Medical Center, Cincinnati, OH; Division of Pulmonary and Critical Care, University of Cincinnati Medical Center, Cincinnati, OH.
Objectives: (1) To determine which of 30 comorbid conditions are independently associated with in-hospital mortality in a cohort of ICU patients. (2) To compare 3 measures of comorbidity and their impact on in-hospital mortality predictions using an automated risk adjustment system. (3) To determine the contribution of comorbidity to predicted in-hospital mortality relative to other clinical predictors.
Methods: We conducted a retrospective cohort study of 17,893 veterans from 17 VAMCs and 44 ICUs admitted in 1996-97. Comorbidity measures, defined using ICD-9-CM codes according to the method of Elixhauser and colleagues, included an APACHE-weighted comorbidity score, a count of Elixhauser comorbidities, and the 30 individual comorbidity variables. We created multivariable logistic regression models using comorbidity, age, laboratory values, diagnosis, and admission source to determine the independent effect of each comorbid condition on mortality. We compared models with different measures of comorbidity with and without additional clinical predictors by the c-statistic and the method of Hanley and McNeil, and determined the relative contributions of each model variable to uniquely attributable model chi-square. Results were validated using a bootstrapping technique; calibration was assessed across deciles of mortality risk.
Results: Subjects were predominantly men (97.8%) with a mean age of 64 years, an average of 2.6 comorbid conditions, and an overall in-hospital mortality of 11.2%. After including other model predictors, 14 out of 30 comorbidity measures were significantly associated with in-hospital mortality. Models using independently weighted comorbidities performed better (p<0.001 using method of Hanley and McNeil) than models using an APACHE-weighted score or a count of Elixhauser comorbidities. In the final multivariable model, comorbidity accounted for less (8.4%) of the model’s uniquely attributable chi-square statistic than laboratory values (67.7%) and diagnosis (17.7%), but more than age (4.0%) and admission source (2.1%). This final model had excellent discrimination (c-statistic 0.878) and was well calibrated across all deciles of mortality risk.
Conclusions: Independently weighted comorbid conditions identified through computerized discharge abstracts can contribute significantly to ICU risk adjustment models.
Impact statement: Using independently weighted comorbidity variables in conjunction with an automated risk adjustment tool is feasible, easy to implement, and can improve mortality predictions for ICU patients within the VA system.