Integrating Clinical Data Warehouses: How Can Multi- System Care for Older Veterans Be Measured Consistently? AcademyHealth Annual Research Meeting Tuesday, June 10, 2008 Presenter: James F. Burgess, Jr., Ph.D. VA Center for Organization, Leadership and Management Research and Boston University School of Public Health
Co-Authors/Collaborators Matt Maciejewski (VA Center for Health Services Research in Primary Care and U. of North Carolina School of Pharmacy) Mark Perkins (VA Center for Outcomes Research in Older Adults) Nancy Sharp (VA Center for Outcomes Research in Older Adults and U. of Washington) Chuan-Fen Liu (VA Center for Outcomes Research in Older Adults and U. of Washington) –Supported by VA HSR&D IIR
Outline Problem of merging data from Clinical Data Warehouses with data from different health care systems Possible approaches to matching VA and Medicare services by type of care Introduce the study motivating this issue Methodology of our chosen approach Context of identifying “primary care” in this particular VA case to research generally
Merging Data from Different Health Care Systems Data generating processes vary, especially nature of encounter data –By location, by provider, by diagnosis or grouped diagnoses, by procedure, others Two kinds of payment incentives (data collected to be paid, pay for reporting or other payments or incentives for particular data to be collected) Origins of data (primarily paper/electronic) Auditing or other scrutiny helps accuracy
Why Do We Want to Match VA and Medicare Services by Care Type? To identify continuity of primary care, we need to: –Identify primary care in Medicare in the absence of a variable that specifically identifies primary care –Classify VA and Medicare encounters as either primary care or something else Processes to generate measures that are an essential part of the actual patient care workflow are most accurate
Dual Use, Continuity of Care and Duplication of Care Study Purpose –Examine how continuity of primary care is impacted by use of VA and Medicare services –Evaluate duplication of preventive and high cost services Sample –Veterans obtaining primary care at CBOCs and/or VAMC primary care clinics in 2000 Follow Up Years:
Matching VA and Medicare Data Two basic approaches: matching on cost or workload counts (we do counts) Aligning incentives and organizational structures in the two systems VA a provider focused on treatment, Medicare a payor focused on billing Most physicians in VA employed by VA, most Medicare billing MDs are not employed by the billing hospitals
Philosophies of Matching Try to make VA look like Medicare –Use CPTs and match as though VA data is billing data (severely undercounts VA work) Try to make Medicare look like VA –Classify Medicare work into VA-type “Clinic Stop” categories (these are often used for VA research) Create a hybrid and transform both –Pick and choose from advantages and disadvantages of data in each sector and select a comparison point that directly reflects neither system
General Approach Classify VA and Medicare encounter into “Care Type” based on hierarchical algorithm Roll up encounters: –by subject –by care type –by fiscal year For each subject, join VA care type counts and Medicare care type counts Use combination of provider specialty and Procedure (CPT-4) codes to classify
Validation of Algorithm VA definition of “Primary care” vs. encounters that algorithm would call “primary care” VA definition of primary care (VA’s DSS system) –Encounter at clinic stop 323, 301, 318, 350, or 319 Algorithm’s definition of primary care (PC) –Primary care provider (Family Practice(FP)/PC Physician, FP/PC Nurse Practitioner, or FP Physician Assistant) –E&M CPT4 code associated with PC office visit –Other CPT4 code not Medicine or not E&M code associated with specialty care visit
Positive/Negative Predictive Probability for PC E/M Code Given PC Stop Code Encounter Primary Care E/M Code PC Stop Code Encounter YESNOType of Prob. Percent YES190,9868,452 Positive Pred. Prob. 95.7% NO147,961376,554 Negative Pred. Prob. 71.8%
Positive/Negative Predictive Probability for PC E/M Code Given PC Stop Code Encounter By Year MeasureYEAR VA PC Stop Code Encounter vs. Primary Care E/M code PV+94.6%95.7%96.0%96.3%96.6% PV-69.5%71.0%71.7%73.4%74.2%
Positive/Negative Predictive Probability for PC Provider Type Given PC Stop Code Encounter Primary Care Provider Type PC Stop Code Encounter YESNOType of Prob. Percent YES147,74051,698 Positive Pred. Prob. 74.1% NO178,584345,931 Negative Pred. Prob. 66.0%
Positive/Negative Predictive Probability for PC Provider Type Given PC Stop Code Encounter By Year MeasureYEAR VA PC Stop Code Encounter vs. Primary Care Provider Type PV+74.6%76.8%75.2%71.3%71.0% PV-70.0%64.6%64.2%65.0%65.4%
Positive/Negative Predictive Probability for PC Care Type Given PC Stop Code Encounter Primary Care Care Type PC Stop Code Encounter YESNOType of Prob. Percent YES103,87095,568 Positive Pred. Prob. 52.0% NO19,634504,881 Negative Pred. Prob. 96.3%
Positive/Negative Predictive Probability for PC Care Type Given PC Stop Code Encounter By Year MeasureYEAR VA PC Stop Code Encounter vs. Primary Care Care Type PV+51.8%53.3%52.5%50.7%51.8% PV-96.7%96.0% 96.2%96.3%
Primary Care Type Classification between Medicare and VA Classification AlgorithmMedicare (N = 739 K) VA (N = 724 K) Number of records % % VA Specific Stop codeN/A 199, Primary care E/M codes249, , Primary care provider type 197, , Primary care type (E/M and provider type) 103, ,
Conclusions and Implications Extreme caution in interpretation of terms like “primary care” that we think we understand is important when comparing across systems Generalizing from studies using VA clinical data warehouse systems to identify types of patient care services to non-VA services is difficult Comprehensive care for Medicare eligible veterans using VA and Medicare systems would benefit from a joint clinical data warehouse