Published codelists improve reproducibility, but choosing the right one remains a challenge.
A new study, published in the International Journal of Population Data Science (IJPDS), highlights an unexpected challenge for researchers using routinely collected healthcare data: although publishing clinical codelists has improved transparency and reproducibility, choosing the most appropriate codelist is becoming an increasingly complex task.
Healthcare conditions recorded in electronic health records are identified using structured clinical codes. Rather than relying on a single code, researchers typically create "codelists" – collections of codes that define conditions such as hypertension, asthma or lung cancer. These codelists play a crucial role in determining which patients are included in research studies.
Historically, codelists were rarely published, making it difficult to compare studies or reproduce research findings. Today, researchers increasingly share their codelists alongside published papers or deposit them in repositories such as the Health Data Research UK (HDR UK) Phenotype Library, improving transparency and supporting more reproducible research.
The study compared codelists held in the HDR UK Phenotype Library for eight physical and mental health conditions: asthma, autism, coronary heart disease, chronic obstructive pulmonary disease, depression, hypertension, obesity and type 2 diabetes. Focusing on Read Codes, the coding system previously used in UK primary care, the researchers examined how similar codelists created by different research groups actually were.
They found considerable variation. For example, eight published codelists for hypertension contained anywhere between 19 and 130 Read Codes. Overall, agreement between codelists ranged from only fair to, at best, moderate. One reason for these differences appeared to be whether researchers included codes that were not direct diagnoses, such as administrative codes recording attendance at a diabetes or asthma clinic.
Professor Bruce Guthrie, one of the study's authors, said: "Publication of codelists is an important step towards making research more reproducible and comparable. However, researchers are now faced with the different problem of choosing which of the multiple published codelists they should use, with little information available about which codelist is most accurate in identifying people with a particular condition."
The authors note that their study did not examine how differences between codelists affect which patients are ultimately identified as having a condition, an important question for future research. In the meantime, they recommend that researchers carefully consider which codelist best fits their study and, wherever possible, choose codelists that have been formally validated.
Click here to read the full article
Dr. Brian Kelly, Medical Statistician, University of Edinburgh, Scotland, UK