BACKGROUND: In an ongoing postauthorization drug safety study, we evaluated the validity of algorithms to identify cancer outcomes among individuals with type 2 diabetes mellitus (T2DM) taking a glucose-lowering drug (GLD) in the United States (US) Medicare administrative claims database.
OBJECTIVES: To estimate the positive predictive values (PPV) of claims-based algorithms for identifying (1) female invasive breast cancer (Br CA) and (2) in situ or invasive bladder cancer (Blad CA) among individuals with T2DM in US Medicare.
METHODS: Patients were fee-for-service US Medicare enrollees with ≥ 6 months of continuous enrollment, aged ≥ 65 years, and initiating a GLD in 2014-2019. Amongst other criteria, patients were excluded if there was evidence of any previous invasive cancer (both cancer cohorts) or in situ bladder cancer (Blad CA cohort), or any of the following in the 180 days prior to study entry: breast biopsy (Br CA cohort) or hematuria, cystoscopy/bladder biopsy, or urine cytology (Blad CA cohort). Case-finding algorithms included 2 recorded diagnoses of Br CA (ICD-9-CM 174.* or ICD-10-CM C50.*; males excluded) or Blad CA (ICD-9-CM 233.7 or 188.* or ICD-10-CM D09.0 or C67.*) in any setting on separate days and within 60 days of each other. Medical records were requested for a sample of algorithm-identified cases (Br CA, n = 275; Blad CA, n = 241), targeting a final sample of up to 125 cases of each cancer outcome for adjudication review. Using abstracted medical record information and prespecified case definitions for each cancer outcome, adjudicators classified each algorithm-identified case as a confirmed case, a confirmed non-case, or a provisional case (insufficient information to assign case status). The PPVs and 95% confidence intervals (CI) were estimated as the proportion of confirmed cases (1) among all cases in the adjudication review (provisional cases treated as false positives) and (2) among only confirmed cases and confirmed non-cases (provisional cases omitted).
RESULTS: We obtained 36% (Br CA, n = 99) and 46% (Blad CA, n = 110) of requested patient records. Adjudicators reviewed records for 97 Br CA and 98 Blad CA algorithm-identified cases, with 77 and 91 classified as confirmed cases, respectively. The PPVs (95% CI) for the algorithms (1) among all cases included in the adjudication review and (2) among only confirmed cases and confirmed non-cases (excluding provisional cases) were: Br CA (1) 79.4% (70.0%-86.9%), (2) 84.6% (75.5%-91.3%); Blad CA (1) 92.9% (85.8%-97.1%), (2) 97.8% (92.4%-99.7%).
CONCLUSIONS: The algorithms applied in this study have high validity for identifying female invasive breast cancer and invasive or in situ bladder cancer among patients with T2DM taking a GLD in US Medicare.
