Validity of Using Prescription Medications to Classify Disease – A Retrospective Observational Study Using Routinely Collected Electronic Health Records from the UK

Christian Schnier,1 John Busby,1 Aziz Sheikh,2 Jennifer K Quint,3 David B Price,4– 6 Liam G Heaney1,7 1School of Medicine, Dentistry and Biomedical Sciences, Queen’s University, Belfast, UK; 2Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, UK; 3School of Public Health, Imperial College London, London, UK; 4Optimum Patient Care Global Ltd, Cambridge, UK; 5Observational and Pragmatic Research Institute, Singapore, Republic of Singapore; 6Centre of Academic Primary Care, Division of Applied Health Sciences, University of Aberdeen, Aberdeen, UK; 7Belfast Health & Social Care NHS Trust, Belfast, UKCorrespondence: Christian Schnier, School of Medicine, Dentistry and Biomedical Sciences Wellcome Wolfson Institute for Experimental Medicine, Queen’s University Belfast, 97 Lisburn Road, Belfast, United Kingdom, Email [email protected]: Epidemiological studies rely on valid classifications of patients’ disease status. However, in the absence of perfect information on every patients’ health status, researchers use proxy information with variable and often unknown validity.Methods: To investigate the validity of using prescription records for disease classification, we conducted a retrospective observational study on a UK-wide database of medical prescriptions and clinical records (Optimum Patient Care Research Database). We used electronic health records of 25,000 randomly selected patients for each year between 2004 and 2020 (total N=425,000) and compared disease classification of 18 different chronic conditions based on clinical records for a period of three years (gold standard) with disease classification based on prescription records for the same period. We then used logistic regression to analyse if positive and negative predicted values (PPV and NPV) were associated with known predictors of disease.Results: Results showed large variations in PPV ranging from 8% (heart failure) to 94% (all type diabetes) and smaller variations in NPV ranging from 96% (anxiety) to 100% (Type 1 diabetes). Age, sex, ethnicity, and year but not socio-economic status were associated with variations in validity, especially in classifying dementia, diabetes, and depression.Discussion: Varying validity can partly be explained by different (stratum-specific) prevalence of disease. Additionally, conditions like heart failure can be treated with medication that can also be prescribed for other conditions or can be treated without medication. However, varying validity can also be attributed to imperfect clinical records, which we used as gold standard. As a consequence of low validity, the apparent prevalence based on using prescription records was between 1.3 times lower (all-type diabetes) and up-to 11 times higher (heart failure) than the true prevalence based on the clinical records.Conclusion: Studies using prescription data to classify disease status run a substantial risk of misclassification bias.Plain Language Summary: In medical research, we often group people into those who have a disease and those who are healthy. However, in large studies we usually cannot medically examine every person, so we use readily available information, medical health records, to estimate the disease status. The problem with estimating the disease status from these records, though, is that the information is incomplete. To study how well we can estimate disease status from prescription records, we compared disease status of 425,000 patients based on clinical records with disease status of the same patients based on prescription records. For the 18 health conditions we included, we found wide variability. For most health conditions we found that if we did not find a prescription record, we could be reasonably confident that the patient indeed did not have the health condition. However, we also found that if we did find a prescription for a health condition, this did not necessarily mean that this patient indeed had a clinical diagnosis. We found that some of these failures to group patients based on prescription records can be explained by the frequency of disease and by different treatment options. We found that using prescription records to predict disease could lead to an estimate of up-to 11 times higher burden of disease than actually true. Finally, we demonstrated that comparisons of disease burden between different patient groups (eg, between males and females) can go wrong if prescription records are used to estimate the disease status.Keywords: electronic health records, misclassification bias, validation study, diagnostic code lists, Optimum Patient Care Research Database