PCN295 A CLAIM-BASED ALGORITHM PREDICTING BREAST CANCER STAGE AT DIAGNOSIS

Abstract

To develop and validate a predictive algorithm to identify breast cancer stages using treatment information obtained from claims data. The SEER-Medicare database contains linked “gold standard” cancer stages from SEER cancer registries and claims data from Medicare. We built a classification tree model based on variables identified through diagnosis, procedure and medication codes from inpatient, outpatient, physician and pharmacy claims. Female fee-for-service patients older than 66 and diagnosed with stage0-IV breast cancer between 2008 and 2013 were randomly assigned to training and validation sets. A classification tree model was based on variables identified through diagnosis, procedure and medication codes from inpatient, outpatient, physician and pharmacy claims. The performance of the classification model was measured by sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC). The algorithm identified stage 0 breast cancer with 85% sensitivity, 97% specificity, 84% PPV, 97% NPV, and 91% AUC; stage I with 83% sensitivity, 75% specificity, 74% PPV, 85% NPV, and 79% AUC; stages I/II with 93% sensitivity, 73% specificity, 90% PPV, 79% NPV, and 83% AUC; stages II/III with 63% sensitivity, 91% specificity, 78% PPV, 83% NPV, and 77% AUC; stage IV with 79% sensitivity, 99% specificity, 67% PPV, 99% NPV, and 89% AUC. Our algorithm had excellent predictive power for stage 0 and IV breast cancer, and good performance for stage I cases. Stage II and III identification were less successful due to the similarities in treatment recommendations. The overall accuracy significantly improved with combined estimations of stages I and II, as well as stages II and III.