Search

Introduction In a large biobank of over half a million people, we have several pairs of participants who appear to share their genome. As more individuals are sequenced, more pairs are likely to be found. If these are twins then this is great news, but it isn’t quite that simple. Objectives and Approach Where 2 people share a genome we need to be able to confirm that these pairs are twins. However, there are a number of issues which could cause 2 people to appear to share a genome; for example being recruited twice, donating blood on another’s behalf, etc. We already identify and exclude participant data based on these conditions. We developed our methodology by looking at the first identified pair in great detail, looking for evidence which specifically ruled out possible alternate explanations, and then applying and refining the method on later pairs. Results We were able to demonstrate the pair were almost certainly twins using their biochemistry and family questionnaire data as principal sources. We also identified a number of variables which were useful in indicating the likelihood of a twin, and now form part of a methodology which we are still developing. Even more usefully, we identified a number of variables that seemed like useful measures but proved extremely misleading. To date we have 26 pairs of possible twins, with 9 confirmed as twins and the remainder looking likely to be twins but falling short of a threshold for confidence. We also have 75 pairs which confirm duplicate participants we have already excluded. Conclusion/Implications We formed two lessons: even very simply linkages come with pitfalls, and you should gather more administrative data than you think. We’re proposing the collection of additional familial relationship data in our third resurvey. We are also looking into machine learning and statistical techniques to better identify twins and duplicates.

Introduction We have a large cohort study of half a million people, which continually incorporates new data through health insurance, centre for disease control records, death certificates, resurveys, and ongoing quality assurance and participant information updates. To support our researchers we need data which is correct, up-to-date, and unchanging. Objectives and Approach We must provide the new data, fixes and corrections to researchers, without missing anything or introducing issues. We make sequential iterations of our data available to researchers on a biannual basis; allowing a static version that can be referenced with regards to earlier work and providing the newest version of the data for new work. Due to the very large size of the data/code base and the small size of the team managing it, delivering this without error is a struggle. To mitigate this we developed testing scripts which catch issues and flag for resolution prior to release to researchers. Results We currently have 32 tests which catch all known issues which occur during a rebuild. On any occasion where a new type of issue is encountered, tests which would catch that issue and related issues are developed. As a result our last few releases have gone far more smoothly, with few if any issues reported after a release and certainly no previously encountered issues! Examples of current tests include: detection of a failed health insurance import; that we have the same number of participants; failure to increment version number between releases; checking that disease numbers have not changed dramatically over the shared timeframe. Conclusion/Implications Producing multiple static releases is a good way to balance the needs of a researcher for both static and current data, but it does introduce opportunities for both human and computer errors. Mitigating this risk with automated testing is convenient and effective.

ABSTRACT Objectives In our Chinese biobank of half a million people, we use data gathered from health insurance agencies to supplement our follow-up. We have 217,000 participants with insurance records including a breakdown of what the insurance paid for, totalling 1.6 million insurance records and 60 million chargeable items. The objective was to find ways of using this information to enhance our Electronic Health Records (EHRs) by adding usable and reliable treatment data, as a basis for future research. Approach Machine translations of every charge description were produced so that early investigation could be done by analysts who were not Chinese speakers. Key phrases were produced by specialist clinicians in an iterative process. We began by focussing on haemodialysis treated ESRD, heart failure, and coronary revascularisation. With our refined techniques, key phrase searches were developed which could be tied into ongoing validation procedures elsewhere in the study (e.g. cancer) or which could be validated using existing data from other sources (e.g. death reporting). Results Machine translation provided both problems and unexpected solutions. While it could be inaccurate (‘Divine Comedy’, ‘semen’, ‘corpse cuisine’), more often than not it provided unexpected advantages, converting regional, archaic, or otherwise uncommon Chinese terms into the most common English equivalent. The majority of chargeable elements in our insurance records are not treatment data per se, but instead hospital fees, generic care, and records of tests without result data. This makes identification of relevant treatment data challenging. Targeted key phrase searches proved successful, demonstrating that it was possible to use this data to answer research questions, even teasing out details which would otherwise not be available to us (e.g. ESRD, location and type of revascularisation). Validation of these findings is ongoing. For example, we found that 395 of our participants have been charged for ‘corpse cuisine’ (more accurately ‘corpse preparation’). Comparing these figures to our death records (an independently gathered source) we confirmed that 326 are known to be dead, and we added the remaining 69 to our list for active follow-up. Similarly, will we be seeking hospital records for the 528 patients who are receiving cancer treatment with no record of cancer. Conclusion Our methods for dealing with treatment data are still being refined, but early results are looking promising. We are investigating standardisation to ICD-10-PCS codes, developing more treatment-based diagnoses, and feeding our findings back into our ongoing validation program.