Bayesian approaches to identify persons with osteoarthritis in electronic health records and administrative health data in the absence of a perfect reference standard

Boston University Medical Campus, Boston MA

Investigators

Linked publications & trials

Abstract

PROJECT SUMMARY/ABSTRACT Osteoarthritis (OA) is a leading contributor to the Global Burden of Disease Study’s Years Lived with Disability (YLDs) measure, and due to aging and increasing rates of obesity, its ranking is steadily rising. With no treatments that delay progression of OA available, the cost of joint replacements is straining Medicare budgets and contributing to enormous economic impact and public health burdens. We are in critical need of a better understanding of factors that prevent or delay worsening OA and of its population impact. Electronic health records (EHR) and administrative health data are a major resource for data-driven approaches in real-world evidence studies and are increasingly used to study disease risk factors and treatments and the genetics of disease such as those from the UK Biobank, where these databases are used to identify cases of disease that tied to genetic susceptibilities. With data on millions of patients, these databases also allow inquiries into health care utilization and costs, inequities of care, growing prevalence of OA and its burden, treatments, adherence of care to guidelines and attendant comorbidities. The validity of research using administrative data; however, relies on accurate characterization and identification of disease cases. The long-term goal of this research is to improve OA case ascertainment in EHR and administrative health data. The central hypothesis of this proposal is that by using multiple data elements including imperfect diagnosis and procedures codes and understanding the conditional dependence among the data elements, the accuracy and predictive values of OA case finding algorithms can be substantially improved. Using insurance claims data from one of the largest administrative health databases in the US, MarketScan, and EHR data from Boston Medical Center, this proposal aims (1) to develop an algorithm in a large administrative database to estimate the probability of OA in an individual accounting for the conditional dependence of its multiple diagnosis and procedure codes; (2) in an EHR database, to compare our approach with conventional diagnosis/procedure code-based algorithms validated against chart review. The contribution of this work is significant because it is the first OA algorithm to exploit the conditional dependencies between its data elements to improve accuracy. Further, the proposed methodology is significant in that it can be broadly applied to other conditions and diseases that can substantially improve the quality of real-world evidence observational studies using administrative health data.