Hanchate AD (CHQOER, Bedford, MA), Borzecki AM
(CHQOER, Bedford, MA), Loveland S
(CHQOER, Bedford, MA), Zhao S
(CHQOER, Bedford, MA), Chew P
(CHQOER, Bedford, MA), Shwartz M
(Boston University), Ash AS
(Boston University School of Medine), Rosen AK
(CHQOER, Bedford, MA)
Administrative data are commonly used for risk adjustment as they are readily available. However, they are subject to coding inaccuracies and lack of clinical detail. Chart-based data, conversely, while considered more accurate, are costly and time-consuming to collect. Recent Agency of Healthcare Quality and Research initiatives to enhance administrative data with automated laboratory test results and vital signs data are important to improving the accuracy of risk-adjustment. The Decision Support System (DSS) Laboratory Results (LAR) data are readily available, offering a unique VA opportunity for improving risk adjustment. This study examines the improvement in predicting in-hospital mortality across all VA facilities by comparing the predictive ability of administrative-based models with “enhanced” models (administrative + laboratory data) for nine admission cohorts (acute myocardial infarction, congestive heart failure, cirrhosis and alcoholic hepatitis, chronic obstructive pulmonary disease, gastrointestinal hemorrhage, hip fracture, pneumonia, acute renal failure and acute stroke).
For 184,177 admissions during FY2004-FY2007 (50% of all VA admissions), we obtained diagnosis codes from VA PTF files and laboratory test results (sodium, BUN, WBC, creatinine, bilirubin and hematocrit) from DSS-LAR files. We used logistic regression models to predict in-hospital mortality. Models were tested separately for each admission cohort, first accounting only for comorbidites using the Elixhauser comorbidity index (ICD-9-CM codes), then supplementing these with laboratory measures obtained at admission. Model performance (discrimination and calibration) was compared at the individual- and facility-level.
Model discrimination (c-statistic) increased significantly for seven cohorts when laboratory measures were included–the greatest increase was for the cirrhosis and alcoholic hepatitis cohort (from 0.68 to 0.81). The ability to distinguish patients with lowest and highest mortality (calibration) also improved significantly for seven cohorts. Facility rankings based on risk-adjusted in-hospital mortality also varied considerably. For some cohorts, inclusion of laboratory measures led to a significant narrowing of risk-adjusted mortality rates between top and bottom ranked facilities.
Using available laboratory data for risk adjustment will improve models’ ability to predict VA in-hospital mortality for many inpatient admission cohorts.
Although relatively unused despite being readily available, DSS-LAR represents a potentially valuable resource for improving quality assessment and monitoring in the VA.