Analysing electronic health records: The benefits of target trial emulation

• Published in: Health policy and technology 2021-09, Vol.10 (3), p.100545, Article 100545
• Main Authors: Bakker, Lytske J., Goossens, Lucas M.A., O'Kane, Maurice J., Uyl-de Groot, Carin A., Redekop, William K.
• Format: Article
• Language: English
• Subjects: Electronic health records; Information technology; Observational research; Type 2 diabetes
• ISSN: 2211-8837; 2211-8845
• DOI: 10.1016/j.hlpt.2021.100545

•Obtaining valid inferences from observational research using electronic health records is challenging due to selection bias, confounding and missing data.•Use of target trial emulation may increase the value of results from observational research by reducing the risk of selection bias, immortal time bias and confounding.•Target trial emulation can be a means to reduce the risk of bias by comparing the observational study performed to the perfect target trial.•When using EHR data to emulate a target trial, samples containing sufficient information on outcome measures and variables to adjust for confounding and selection bias are essential given the risk of missing data. Electronic health records (EHRs) are increasingly used in effectiveness and safety research. However, these studies are often at risk of bias. This study demonstrates the relevance, and discusses challenges, of using target trial emulation to avoid bias, such as selection bias, immortal time bias and confounding when performing observational research with EHRs. Target trial emulation can be used to identify and address some of the drawbacks of observational research in a systematic way. Potential sources of bias are identified by describing key components of an ideal randomized controlled trial and comparing this to the observational study actually performed. The methods were applied to assess treatment response to antidiabetic treatment using EHRs from patients with diabetes treated in secondary care. Using target trial emulation ensured prevalent users were excluded and patients were not included based on information generally not available when initiating a clinical trial. Furthermore, applying these methods demonstrated how the number of records eligible for use can rapidly decrease. Hereafter, adjustments were performed to address potential sources of bias and it was shown that missing variables essential for adjustment can be an important issue. Using target trial emulation, sources of selection bias and confounding were identified and adjusted for accordingly when analysing treatment response in patients with type 2 diabetes. However, when using EHR data to emulate a target trial, samples containing sufficient information on outcome measures and variables to adjust for confounding and selection bias are essential given the risk of missing data.