Abstract T MP88: Algorithms for Identification of Acute Stroke Hospitalizations in Medicare Data

Background: Medicare claims (MC) are a source of nationwide data on various disease conditions including stroke. We developed and tested algorithms to accurately identify acute stroke hospitalizations in MC data using a population-based acute stroke hospitalization database. Methods: Data were obtained from year 2000 Minnesota Stroke Survey (MSS), a surveillance study of stroke incidence and prevalence in the Minneapolis-St. Paul metro area. Stroke cases were ascertained with acute stroke discharge codes (ICD-9 431, 432, 434, 436, 437) and validated with clinical criteria after medical record abstraction. Stroke was defined as a neurological deficit of vascular origin lasting at least 24 hours or until death. MSS subjects aged 65-74 years (n=464) were linked to MC hospitalization claims. A random sample of non-stroke MC hospitalization claims was included in the dataset (n=1,384). We randomly selected 2/3 of the data for training and 1/3 for testing. Classification and Regression Tree (CART) models were used for algorithm development. We report algorithm sensitivity (SN), specificity (SP), likelihood ratio positive and negative (LR+/LR-), discriminative performance (AUC) and total misclassification error (ME). Results: The CART algorithm showed high discriminative performance on training and test data (Table). Test data performance was: SN=91%, SP=95%, LR+=19.97, LR-=0.09, AUC 0.93, ME=6%. The algorithm automatically selected these MC variables as informative about stroke hospitalizations: diagnostic group code, hospital discharge codes, hospital discharge destination, number of private room days, physical and occupational therapy charges. Conclusion: Our CART algorithm shows good performance and is available for validation with other data sets. Potential utility of developed algorithms has broader implications for stroke epidemiology and health services research.