Background: In a cohort study in users of antimuscarinic drugs for overactive bladder (OAB), we investigated the validity of cancer (CA) diagnoses in general practitioner (GP) medical records and the value added by other databases linkable to the Clinical Practice Research Datalink.
Objectives: To evaluate a process for identifying and verifying cases of common CAs diagnosed in 2004-2012.
Methods: Of 50,843 new users of OAB drugs without a history of CA, provisional CA cases were identified by electronic search for diagnostic codes in GP medical records. Patient profiles were reviewed and discussed by physician-epidemiologists. Confirmed (CONF) cases had evidence of CA treatment, repeated use of a CA diagnostic code, or a subsequent "cancer care review" code. Non-cases in practices linkable to CA registry (CR) and hospital episode statistics (HES) data could become confirmed cases by using the linked data.
Results: The electronic search identified 1,486 provisional cases in GP records, 825 from linked [LP] and 661 from non-linked [NLP] practices. Manual review of GP records confirmed 792 (96%) cases in LP and 616 (93%) in NLP. In LP, 1,077 cases were confirmed; of which, 305 (28%) were not identifiable through GP records and were obtained by linkage to CR and HES data (54 from CR, 122 from HES, and 129 from both), while 184 CONF cases (17%) were identifiable only through GP records. Among 732 CONF cases identified in LP during years with CR data (2004-2010), 492 (67%) were in GP records, 619 (85%) in CR, and 593 (81%) in HES; 22 (3%) were in GP records only, 54 (7%) in CR only, and 57 (8%) in HES only; 226 (31%) were in 2 sources, and 373 (51%) in all 3. Of these 732 cases, 33% were not in GP records, with lower values for CA treated by GPs (breast 12%; prostate 21%) and higher values for others (lung 54%; pancreatic 55%; renal 66%).
Conclusions: The majority of CAs in GP records (similar for LP and NLP) are confirmed by profile review or other data sources. A substantial proportion of CAs in LP would be missed without CR and HES data (and are likely missed in NLP). The effect of incomplete identification of CA cases on relative risk estimates in safety studies will need to be assessed further.
