(B-009) Validity of Hospital Cancer Registry and Electronic Health Record Cancer Diagnoses Among a Large Academic-Community Health System in the United States

Background: Electronic health records (EHR) and cancer registries are increasingly being leveraged to assess the treatment and outcomes of cancer survivors. As cancer survivors in the US regularly receive care across health systems, missing or misleading documentation of clinical data may occur.

Objectives: To evaluate the validity of registry and EHR cancer diagnoses in a large academic-community health system.

Methods: Using a large academic-community health system that spans 12 acute care hospitals, we linked the hospital cancer registries between 2011-2023 with all available EHR data within the research data warehouse using patient identifiers (medical record number, name, date of birth). We limited the population to patients with only one cancer diagnosis in the cancer registry. Cancer diagnoses in the cancer registry (ICD-O-3) and EHR (ICD-10) were standardized by cancer sites. We calculated the match rate between the cancer registry and EHR. We identified the number of cancer diagnoses in the EHR that were not in the cancer registry. We then conducted a manual EHR review of a random sample of 50 patients to describe reasons for matching and missingness. As the health system initiated a new EHR across the 12 hospitals beginning in 2022, we stratified cancer diagnosis based on whether the registry cancer diagnosis was pre- or post-EPIC initiation.

Results: Our cohort included 17,257 patients (mean age 67 years, 54% female, 74% white, top 5 cancers breast [21%], prostate [13%], skin [9.3%], colorectal [6.8%], lung [6.1%], 10,930 patients pre-EPIC initiation, 6,327 patients post-EPIC initiation). Prior to EPIC initiation, the overall match rate was 73% (95%CI: 72-74), and no cancer type had a match rate of more than 90%. The overall match rate following EPIC initiation was 92% (95%CI: 91-92); the match rate among the three most prevalent cancer sites was 97% (95%CI: 96-98) [breast], 99% (95%CI: 98-99) [prostate], and 95% (95%CI: 94-97) [skin]. A large proportion of the population had cancer site ICD-10 diagnoses that were not present in the cancer registry: 61% and 25% of patients pre-and-post EPIC initiation, respectively. Among the 50-patient random sample, all the cancer registry diagnoses were correct, while the EHR site diagnoses not in the registry were due to original diagnoses at another hospital, metastasis of original cancer, diagnostic rule-out, and diagnostic error.

Conclusions: Our study demonstrates that the linkage between EHR and cancer registries has promise in supporting the development of large cohorts to assess characteristics, cancer treatments, and health outcomes of cancer survivors. Nevertheless, investigators must also consider system factors (e.g., new EHR), missingness in the hospital registry, and reasons for EHR diagnoses not in the registry.