Session number: 1134
Abstract title: Identifying elderly veterans for Geriatric Evaluation and Management using administrative data
Author(s):
CR Rakovski - Center for Health Quality Outcomes and Economics Research
DR Berlowitz - Center for Health Quality Outcomes and Economics Research
FF Wang - Center for Health Quality Outcomes and Economics Research
J Lucove - Brown University
AK Rosen - Center for Health Quality Outcomes and Economics Research
Objectives: Methods are needed to efficiently identify elderly patients for geriatric evaluation and management (GEM). Because GEM reduces hospitalizations and nursing home placement, patients who are low utilizers of healthcare in the base year but become high utilizers in the subsequent year may benefit from GEM. We determine whether case-mix information from administrative data can identify those likely to use high resources in the following year.
Methods: The study population was a 40% random sample of veterans who used VA healthcare services in FY'97, excluding people <65 years old, who received GEM in FY'97, were high users in FY'97 (in the top 5%), or who died in FY'98 (N=365,861). ICD-9-CM codes, age, and sex were obtained from FY'97 data. A person's utilization equaled the number of days (1-365) on which a person received inpatient or outpatient services. We developed models using diagnosis-based, case-mix measures from Adjusted Clinical Groups (ACGs) and Diagnostic Cost Groups (DCGs). We fit two logistic regression models to predict high utilization in FY'98 (>43 days, the top 5%): DCG model: age, sex and 116 HCCs (Hierarchical Condition Categories); and ACG model: age, sex and 32 ADGs (Adjusted Diagnostic Groups). We evaluated the area under the receiver operating characteristic (ROC) curve (c), model sensitivity and specificity for cutoff probabilities.
Results: The DCG model (c=0.78) had better overall specificity and sensitivity than the ACG model (c=0.76). The DCG model achieved the best balance of specificity and sensitivity at 80% specificity and 70% sensitivity. This would result in 73,172 false positives (elderly not at risk falsely identified as at risk) and 5,988 false negatives (elderly at risk but missed by the model). A cutoff with 90% specificity and 50% sensitivity would reduce false positives to 36,586, but increase those missed to 9,979.
Conclusions: Case-mix information from administrative data can identify the majority of elderly at risk for high use in the future; however, this method would also identify large numbers of elderly who are not at risk.
Impact statement: Administrative data have the potential to identify high-risk elderly. However, due to poor sensitivity and specificity, current case-mix models may not be efficient in identifying candidates for GEM.