Unlocking the complexities of cancer surveillance: How reliable are the UK’s healthcare databases in accurately recording 19 different cancer types? But here’s where it gets controversial—differences in data sources and recording practices can significantly impact our understanding of cancer epidemiology. This detailed comparison explores how primary care records, hospital data, cancer registries, and death records come together—and where they might fall short—to influence research, policy, and patient care. If you think all cancer data are created equal, think again. The question is: how complete and accurate are these sources, and what does that mean for future research? Let's delve into the nuances of these datasets, their strengths, limitations, and implications for the health sciences community.
Introduction
The Clinical Practice Research Datalink (CPRD) provides researchers with access to two extensive healthcare datasets—CPRD Aurum and CPRD GOLD—that contain anonymized electronic health records from primary care providers across the UK. Both datasets are prominent tools in epidemiological research, but variations in coverage, coding methods, and collection periods can influence the reliability of cancer diagnoses recorded within them. CPRD Aurum mainly encompasses primary care practices in England, while GOLD covers practices across the entire UK, including Scotland, Wales, and Northern Ireland. To improve the completeness of cancer data, these datasets can be linked to other national sources such as the National Cancer Registration and Analysis Service (CR), Hospital Episode Statistics (HES), and the Office for National Statistics (ONS) death records. Such linkages provide a clearer picture of how many cancer cases might be missing from primary care records alone, which is especially critical when accurate incidence and survival estimates are needed.
While prior research confirms that CPRD datasets generally excel in capturing accurate clinical information regarding diagnoses, prescriptions, and mortality, the accuracy of cancer recording has yielded inconsistent findings—varying notably by cancer type. For example, one study involving over 116,000 patients found that about 10% of cancer cases identified in linked HES or CPRD GOLD records lacked confirmation in the Cancer Registry, while up to 32% of cases recorded in the registry were missing from primary care data. These discrepancies suggest that depending solely on GP records might underrepresent certain cancers, especially when detailed information like stage or grade is required. Consequently, linking primary care data with hospital and registry sources could enhance case detection accuracy, making it vital for studies that depend on precise cancer ascertainment.
Most existing research has focused on a handful of cancers, but this study aims to set a new benchmark by comprehensively evaluating how well primary care databases and their linked sources record 19 distinct cancer types. The key goals are to assess and compare incidence estimates derived from CPRD Aurum and GOLD, both with and without linkage to HES and CR, and to estimate survival probabilities using death registry data—providing a holistic view of data strengths and gaps across various cancer types.
Methods
Study Design
This retrospective cohort analysis focuses on estimating annual incidence rates and survival outcomes for 19 different cancers in the English population. By analyzing data separately from CPRD Aurum and GOLD, and cross-referencing these with linked hospital, registry, and death data, the study examines how well each source captures cancer diagnoses across the country.
Data Sources
CPRD Aurum, introduced in 2017, covers roughly 24.75% of the UK population, primarily using data processed through the EMIS Web software from practices based in England, with records dating back to 1995. CPRD GOLD, older and covering about 4.17% of the UK, contains data from practices across the UK using the Vision software, dating back to 1987. Both datasets include anonymized demographic details and diagnosis codes. When linked with HES (which records hospital admissions from 1997 onward), CR (comprehensive tumor registry data from across England, beginning around 2013), and ONS death records, these primary care datasets offer a more complete picture of cancer incidence and outcomes—assuming no missing data.
Study Population
The analysis focused on patients diagnosed with one of the 19 cancers, including leukemia (ALL and AML), bladder, brain, breast, colorectal, esophageal, gastric, head and neck, lung, melanoma, multiple myeloma (MM), neuroendocrine, ovarian, pancreatic, prostate, renal, thyroid, and uterine cancers. Diagnostic coding through ICD-10 and specific medical codes facilitated identification. For gender-specific cancers like prostate (males only) and ovarian/uterine (females only), analyses were restricted accordingly.
Participants were included if they met quality standards, were eligible for record linkage, and had at least a single registration in the dataset during the period from January 1, 2011, to December 31, 2018. To compare diagnosis capture, incident cases were identified as the first occurrence of a cancer code within each data source or linked dataset, acknowledging that some recurrent cancers might be misclassified as new cases due to limited historical data. Patients with death records from ONS during follow-up contributed to survival analyses.
Analytical Approach
Data were divided into categories based on the information source: records solely in CPRD, hospital data only, registry data only, combined hospital and registry data, and fully linked datasets combining all sources. This stratification allowed assessment of how much each source contributed to total case counts, especially when contrasted with the
