Prospective, population-based studies that recruit participants in mid-life are valuable resources for dementia research. Follow-up in these studies is often through linkage to routinely-collected healthcare datasets. We investigated the accuracy of these datasets for dementia case ascertainment in a validation study using data from UK Biobank—an open access, population-based study of > 500,000 adults aged 40–69 years at recruitment in 2006–2010. From 17,198 UK Biobank participants recruited in Edinburgh, we identified those with ≥ 1 dementia code in their linked primary care, hospital admissions or mortality data and compared their coded diagnoses to clinical expert adjudication of their full-text medical record. We calculated the positive predictive value (PPV, the proportion of cases identified that were true positives) for all-cause dementia, Alzheimer’s disease and vascular dementia for each dataset alone and in combination, and explored algorithmic code combinations to improve PPV. Among 120 participants, PPVs for all-cause dementia were 86.8%, 87.3% and 80.0% for primary care, hospital admissions and mortality data respectively and 82.5% across all datasets. We identified three algorithms that balanced a high PPV with reasonable case ascertainment. For Alzheimer’s disease, PPVs were 74.1% for primary care, 68.2% for hospital admissions, 50.0% for mortality data and 71.4% in combination. PPV for vascular dementia was 43.8% across all sources. UK routinely-collected healthcare data can be used to identify all-cause dementia in prospective studies. PPVs for Alzheimer’s disease and vascular dementia are lower. Further research is required to explore the geographic generalisability of these findings. Electronic supplementary material The online version of this article (10.1007/s10654-019-00499-1) contains supplementary material, which is available to authorized users.

IntroductionProspective, population-based studies can be rich resources for dementia research. Follow-up in many such studies is through linkage to routinely collected, coded health-care data sets. We evaluated the accuracy of these data sets for dementia case identification.MethodsWe systematically reviewed the literature for studies comparing dementia coding in routinely collected data sets to any expert-led reference standard. We recorded study characteristics and two accuracy measures—positive predictive value (PPV) and sensitivity.ResultsWe identified 27 eligible studies with 25 estimating PPV and eight estimating sensitivity. Study settings and methods varied widely. For all-cause dementia, PPVs ranged from 33%–100%, but 16/27 were >75%. Sensitivities ranged from 21% to 86%. PPVs for Alzheimer's disease (range 57%–100%) were generally higher than those for vascular dementia (range 19%–91%).DiscussionLinkage to routine health-care data can achieve a high PPV and reasonable sensitivity in certain settings. Given the heterogeneity in accuracy estimates, cohorts should ideally conduct their own setting-specific validation.

ObjectiveIn UK Biobank (UKB), a large population-based prospective study, cases of many diseases are ascertained through linkage to routinely collected, coded national health datasets. We assessed the accuracy of these for identifying incident strokes.MethodsIn a regional UKB subpopulation (n = 17,249), we identified all participants with ≥1 code signifying a first stroke after recruitment (incident stroke-coded cases) in linked hospital admission, primary care, or death record data. Stroke physicians reviewed their full electronic patient records (EPRs) and generated reference standard diagnoses. We evaluated the number and proportion of cases that were true-positives (i.e., positive predictive value [PPV]) for all codes combined and by code source and type.ResultsOf 232 incident stroke-coded cases, 97% had EPR information available. Data sources were 30% hospital admission only, 39% primary care only, 28% hospital and primary care, and 3% death records only. While 42% of cases were coded as unspecified stroke type, review of EPRs enabled a pathologic type to be assigned in >99%. PPVs (95% confidence intervals) were 79% (73%–84%) for any stroke (89% for hospital admission codes, 80% for primary care codes) and 83% (74%–90%) for ischemic stroke. PPVs for small numbers of death record and hemorrhagic stroke codes were low but imprecise.ConclusionsStroke and ischemic stroke cases in UKB can be ascertained through linked health datasets with sufficient accuracy for many research studies. Further work is needed to understand the accuracy of death record and hemorrhagic stroke codes and to develop scalable approaches for better identifying stroke types.