Enhanced Risk Adjustment from Adding Laboratory Test Measures to Administrative Data in the VA Amresh Hanchate, PhD Boston University & VA Bedford / CHQOER Collaborators Ann Borzecki, VA Bedford / Boston Univ. Michael Shwartz, Boston Univ. / VA Boston Arlene Ash, Boston University Amy Rosen, VA Bedford / Boston Univ, Funded by VA HSR&D Borzecki (PI) Background Accurate risk adjustment is critical for quality assessment and provider comparisons Although ideal, chart-based clinical data are costly and time-consuming to collect Risk adjustment continues to be based on administrative data, despite known limitations – CMS Hospital Compare Adding Automated Clinical Data In recent years, common for hospitals to have automated clinical information – Laboratory test results – Vital sign results AHRQ initiative across different states to pool standardized laboratory test data from participating community hospitals Early studies of enhanced risk adjustment indicate improved prediction of inpatient or 30-day mortality – Pine et al (JAMA 2007) - 188 PA hospitals – Tabak et al (Med Care 2008) – 266 participating hospitals – Escobar et al (Med Care 2008) – Northern California Kaiser Permanente VA Automated Clinical Data Largest integrated health care system in the US with considerable automated clinical data covering all hospitals Better standardization of data collection and aggregation Render et al (Critical Care 2005; 2008) – ICU mortality Objectives To examine how adding laboratory test data to a standard administrative data-based risk adjustment model improves prediction of inpatient mortality changes relative rankings at hospital-level Table 1. Study Cohorts Fiscal Years 2004-2007 # admissions # hospitals AMI 30,119 141 CHF 87,755 145 COPD 62,339 149 GI Hemorrhage 38,448 144 Cirrhosis & Alc. Hepatitis 14,328 143 Hip Fracture 8,745 133 Pneumonia 74,148 151 Renal Failure 25,859 147 Acute Stroke 25,582 145 Admission Cohort Methods Outcome: In-hospital Mortality Administrative data-based risk adjustment (Admin) model: – Age, sex – Admission secondary diagnosis ICD-9 codes grouped into AHRQ Comorbidity categories Enhanced risk adjustment (ERA) model: – Adds following 6 laboratory test results to the administrative data-based model – sodium, BUN, WBC, creatinine, bilirubin and hematocrit – measured within 24 hours of admission time – dichotomous indicators (abnormal test range) – dichotomous indicator for missing test measure Methods Outcome: In-hospital Mortality Administrative data-based risk adjustment (Admin) model: – Age, sex – Admission secondary diagnosis ICD-9 codes grouped into AHRQ Comorbidity categories Enhanced risk adjustment (ERA) model: – Adds following 6 laboratory test results to the administrative data-based model – sodium, BUN, WBC, creatinine, bilirubin and hematocrit – measured within 24 hours of admission time – dichotomous indicators (abnormal test range) – dichotomous indicator for missing test measure Methods - 2 Each cohort split evenly into estimation and validation samples Estimated hierarchical logistic regression of administrative data-based and ERA models using the estimation sample Estimates applied to validation sample to measure prediction accuracy of the two models Hospital-level risk adjusted rates calculated using CMS Hospital Compare method Table 2. Comparison of Risk Adjustment Models for Inpatient Mortality: Discrimination and Calibration Admission Cohort AMI CHF COPD GI Hemorrhage Cirrhosis & Hep. Hip Fracture Pneumonia Renal Failure Acute Stroke Obs. Mort. Obs. Mort. Rate (%) in Rate (%) in Top Decile Bottom Decile Observed C-statistic In-hospital Mortality Admin ERA Admin ERA Admin ERA Rate (%) Model Model Model Model Model Model 6.3 3.6 2.5 2.8 8.8 6.2 6.4 6.2 6.3 0.75 0.72 0.71 0.77 0.66 0.72 0.74 0.74 0.71 0.80 0.78 0.75 0.82 0.78 0.73 0.77 0.79 0.78 1.0 0.8 0.5 0.3 3.0 1.4 0.9 1.3 2.0 0.8 0.4 0.5 0.1 0.8 0.7 0.8 0.6 0.8 19.8 11.2 8.0 10.4 18.7 15.4 19.6 17.9 18.2 25.0 14.3 9.3 13.0 34.4 17.7 21.7 23.2 22.6 Table 2. Comparison of Risk Adjustment Models for Inpatient Mortality: Discrimination and Calibration Admission Cohort AMI CHF COPD GI Hemorrhage Cirrhosis & Hep. Hip Fracture Pneumonia Renal Failure Acute Stroke Obs. Mort. Obs. Mort. Rate (%) in Rate (%) in Top Decile Bottom Decile Observed C-statistic In-hospital Mortality Admin ERA Admin ERA Admin ERA Rate (%) Model Model Model Model Model Model 6.3 3.6 2.5 2.8 8.8 6.2 6.4 6.2 6.3 0.75 0.72 0.71 0.77 0.66 0.72 0.74 0.74 0.71 0.80 0.78 0.75 0.82 0.78 0.73 0.77 0.79 0.78 1.0 0.8 0.5 0.3 3.0 1.4 0.9 1.3 2.0 0.8 0.4 0.5 0.1 0.8 0.7 0.8 0.6 0.8 19.8 11.2 8.0 10.4 18.7 15.4 19.6 17.9 18.2 25.0 14.3 9.3 13.0 34.4 17.7 21.7 23.2 22.6 Table 2. Comparison of Risk Adjustment Models for Inpatient Mortality: Discrimination and Calibration Admission Cohort AMI CHF COPD GI Hemorrhage Cirrhosis & Hep. Hip Fracture Pneumonia Renal Failure Acute Stroke Obs. Mort. Obs. Mort. Rate (%) in Rate (%) in Top Decile Bottom Decile Observed C-statistic In-hospital Mortality Admin ERA Admin ERA Admin ERA Rate (%) Model Model Model Model Model Model 6.3 3.6 2.5 2.8 8.8 6.2 6.4 6.2 6.3 0.75 0.72 0.71 0.77 0.66 0.72 0.74 0.74 0.71 0.80 0.78 0.75 0.82 0.78 0.73 0.77 0.79 0.78 1.0 0.8 0.5 0.3 3.0 1.4 0.9 1.3 2.0 0.8 0.4 0.5 0.1 0.8 0.7 0.8 0.6 0.8 19.8 11.2 8.0 10.4 18.7 15.4 19.6 17.9 18.2 25.0 14.3 9.3 13.0 34.4 17.7 21.7 23.2 22.6 Table 2. Comparison of Risk Adjustment Models for Inpatient Mortality: Discrimination and Calibration Admission Cohort AMI CHF COPD GI Hemorrhage Cirrhosis & Hep. Hip Fracture Pneumonia Renal Failure Acute Stroke Obs. Mort. Obs. Mort. Rate (%) in Rate (%) in Top Decile Bottom Decile Observed C-statistic In-hospital Mortality Admin ERA Admin ERA Admin ERA Rate (%) Model Model Model Model Model Model 6.3 3.6 2.5 2.8 8.8 6.2 6.4 6.2 6.3 0.75 0.72 0.71 0.77 0.66 0.72 0.74 0.74 0.71 0.80 0.78 0.75 0.82 0.78 0.73 0.77 0.79 0.78 1.0 0.8 0.5 0.3 3.0 1.4 0.9 1.3 2.0 0.8 0.4 0.5 0.1 0.8 0.7 0.8 0.6 0.8 19.8 11.2 8.0 10.4 18.7 15.4 19.6 17.9 18.2 25.0 14.3 9.3 13.0 34.4 17.7 21.7 23.2 22.6 Table 2. Comparison of Risk Adjustment Models for Inpatient Mortality: Discrimination and Calibration Admission Cohort AMI CHF COPD GI Hemorrhage Cirrhosis & Hep. Hip Fracture Pneumonia Renal Failure Acute Stroke Obs. Mort. Obs. Mort. Rate (%) in Rate (%) in Top Decile Bottom Decile Observed C-statistic In-hospital Mortality Admin ERA Admin ERA Admin ERA Rate (%) Model Model Model Model Model Model 6.3 3.6 2.5 2.8 8.8 6.2 6.4 6.2 6.3 0.75 0.72 0.71 0.77 0.66 0.72 0.74 0.74 0.71 0.80 0.78 0.75 0.82 0.78 0.73 0.77 0.79 0.78 1.0 0.8 0.5 0.3 3.0 1.4 0.9 1.3 2.0 0.8 0.4 0.5 0.1 0.8 0.7 0.8 0.6 0.8 19.8 11.2 8.0 10.4 18.7 15.4 19.6 17.9 18.2 25.0 14.3 9.3 13.0 34.4 17.7 21.7 23.2 22.6 Table 3. Comparing Facility Performance using Risk Adj. Mortality (RAM) from Admin. and ERA Models Admission Cohort Concordance: # common Discordance: Average rank facilities out of 10 difference between Admin. lowest/highest RAM and Lab. models for facilities facilities from each model without concordance Lowest RAMHighest RAMLowest RAM Highest RAM AMI CHF COPD GI Hem. Cirrhosis Hip Fracture Pneumonia Renal Fail Acute Stroke 5 4 5 5 2 4 1 2 5 7 3 4 7 5 5 3 3 5 13 21 39 31 15 16 32 20 42 15 25 47 10 11 10 21 30 28 Note: Only facilities with a minimum of 50 admissions for each admission cohort included. Table 3. Comparing Facility Performance using Risk Adj. Mortality (RAM) from Admin. and ERA Models Admission Cohort Concordance: # common Discordance: Average rank facilities out of 10 difference between Admin. lowest/highest RAM and Lab. models for facilities facilities from each model without concordance Lowest RAMHighest RAMLowest RAM Highest RAM AMI CHF COPD GI Hem. Cirrhosis Hip Fracture Pneumonia Renal Fail Acute Stroke 5 4 5 5 2 4 1 2 5 7 3 4 7 5 5 3 3 5 13 21 39 31 15 16 32 20 42 15 25 47 10 11 10 21 30 28 Note: Only facilities with a minimum of 50 admissions for each admission cohort included. Table 3. Comparing Facility Performance using Risk Adj. Mortality (RAM) from Admin. and ERA Models Admission Cohort Concordance: # common Discordance: Average rank facilities out of 10 difference between Admin. lowest/highest RAM and Lab. models for facilities facilities from each model without concordance Lowest RAMHighest RAMLowest RAM Highest RAM AMI CHF COPD GI Hem. Cirrhosis Hip Fracture Pneumonia Renal Fail Acute Stroke 5 4 5 5 2 4 1 2 5 7 3 4 7 5 5 3 3 5 13 21 39 31 15 16 32 20 42 15 25 47 10 11 10 21 30 28 Note: Only facilities with a minimum of 50 admissions for each admission cohort included. Table 3. Comparing Facility Performance using Risk Adj. Mortality (RAM) from Admin. and ERA Models Admission Cohort Concordance: # common Discordance: Average rank facilities out of 10 difference between Admin. lowest/highest RAM and Lab. models for facilities facilities from each model without concordance Lowest RAMHighest RAMLowest RAM Highest RAM AMI CHF COPD GI Hem. Cirrhosis Hip Fracture Pneumonia Renal Fail Acute Stroke 5 4 5 5 2 4 1 2 5 7 3 4 7 5 5 3 3 5 13 21 39 31 15 16 32 20 42 15 25 47 10 11 10 21 30 28 Note: Only facilities with a minimum of 50 admissions for each admission cohort included. Summary Adding laboratory test measures to administrative data-based risk-adjustment models improved the ability to identify patients at higher risk for inpatient death, and significantly changed relative hospital rankings Limitations Missing rates of laboratory test results sizable (up to 28%) and varied significantly by hospitals Abnormal ranges of laboratory test results may vary for admission cohort Robustness of hospital performance changes to be measured