Can the Use of Bayesian Analysis Methods Correct for Incompleteness in Electronic Health Records Diagnosis Data? Development of a Novel Method Using Simulated and Real-Life Clinical Data

Background: Patient health information is collected routinely in electronic health records (EHRs) and used for research purposes, however, many health conditions are known to be under-diagnosed or under-recorded in EHRs. In research, missing diagnoses result in under-ascertainment of true cases, which attenuates estimated associations between variables and results in a bias toward the null. Bayesian approaches allow the specification of prior information to the model, such as the likely rates of missingness in the data. This paper describes a Bayesian analysis approach which aimed to reduce attenuation of associations in EHR studies focussed on conditions characterized by under-diagnosis. Methods: Study 1: We created synthetic data, produced to mimic structured EHR data where diagnoses were under-recorded. We fitted logistic regression (LR) models with and without Bayesian priors representing rates of misclassification in the data. We examined the LR parameters estimated by models with and without priors. Study 2: We used EHR data from UK primary care in a case-control design with dementia as the outcome. We fitted LR models examining risk factors for dementia, with and without generic prior information on misclassification rates. We examined LR parameters estimated by models with and without the priors, and estimated classification accuracy using Area Under the Receiver Operating Characteristic. Results: Study 1: In synthetic data, estimates of LR parameters were much closer to the true parameter values when Bayesian priors were added to the model; with no priors, parameters were substantially attenuated by under-diagnosis. Study 2: The Bayesian approach ran well on real life clinic data from UK primary care, with the addition of prior information increasing LR parameter values in all cases. In multivariate regression models, Bayesian methods showed no improvement in classification accuracy over traditional LR. Conclusions: The Bayesian approach showed promise but had implementation challenges in real clinical data: prior information on rates of misclassification was difficult to find. Our simple model made a number of assumptions, such as diagnoses being missing at random. Further development is needed to integrate the method into studies using real-life EHR data. Our findings nevertheless highlight the importance of developing methods to address missing diagnoses in EHR data.