• Brian J Moore, Susan White, Raynard Washington, Natalia Coenen, and Anne Elixhauser.
• *IBM Watson Health, Ann Arbor, MI †The Ohio State University, Columbus, OH ‡Department of Public Health, PA §IBM Watson Health, Santa Barbara, CA ∥Agency for Healthcare Research and Quality, Center for Quality Improvement and Patient Safety, Rockville, MD.
• Med Care. 2017 Jul 1; 55 (7): 698-705.

ObjectiveWe extend the literature on comorbidity measurement by developing 2 indices, based on the Elixhauser Comorbidity measures, designed to predict 2 frequently reported health outcomes: in-hospital mortality and 30-day readmission in administrative data. The Elixhauser measures are commonly used in research as an adjustment factor to control for severity of illness.Data SourcesWe used a large analysis file built from all-payer hospital administrative data in the Healthcare Cost and Utilization Project State Inpatient Databases from 18 states in 2011 and 2012.MethodsThe final models were derived with bootstrapped replications of backward stepwise logistic regressions on each outcome. Odds ratios and index weights were generated for each Elixhauser comorbidity to create a single index score per record for mortality and readmissions. Model validation was conducted with c-statistics.ResultsOur index scores performed as well as using all 29 Elixhauser comorbidity variables separately. The c-statistic for our index scores without inclusion of other covariates was 0.777 (95% confidence interval, 0.776-0.778) for the mortality index and 0.634 (95% confidence interval, 0.633-0.634) for the readmissions index. The indices were stable across multiple subsamples defined by demographic characteristics or clinical condition. The addition of other commonly used covariates (age, sex, expected payer) improved discrimination modestly.ConclusionsThese indices are effective methods to incorporate the influence of comorbid conditions in models designed to assess the risk of in-hospital mortality and readmission using administrative data with limited clinical information, especially when small samples sizes are an issue.

19 keywords...

hide…