Identifying cancer relapse using SEER-Medicare data.

INTRODUCTION: Tumor registries capture valid information at the time of cancer diagnosis, but often do not conduct longitudinal follow-up evaluations. However, investigators may be interested in questions relating to subsequent relapsed disease. Linking administrative data to registry data, as in the creation of the SEER (Surveillance, Epidemiology, and End Results) and Medicare data set, can provide the ability to infer the occurrence of relapse in selected situations.
METHODS: The authors created different algorithms to detect relapse of acute myelogenous leukemia (AML). A retrospective cohort of patients with AML was identified, and both their billing data and medical records were obtained. The algorithms were then applied to the billing data, the results were compared with medical record review.
RESULTS: Eighty-nine patients were identified, of whom 22 were treated for relapsed AML. The sensitivity of the best algorithm for detecting relapse was 86%, and the specificity 99%, with a positive predictive value of 95% and a negative predictive value of 96%.
CONCLUSIONS: Identification of relapse from SEER-Medicare data using clinical algorithms is feasible for cancers where a majority of patients receive treatment for relapse, without a "watch and wait" strategy, and where that treatment is with a modality that can be detected in billing data (ie, intravenous chemotherapy, radiation, surgery, or all three). Optimal analytic situations are ones in which the investigator is mostly interested in positive predictive value, less interested in sensitivity, and wants to evaluate outcomes among those patients who receive treatment for their relapsed disease. However, the accuracy of such an approach for cancers other than AML has not yet been established.