OP0010 Use of claims and electronic medical record data to predict ra disease activity

Background Prior studies have demonstrated challenges in developing and validating claims-based algorithms that accurately predict RA disease activity.1 2 The ability to adjust for and predict RA disease activity would be a powerful epidemiological tool for studies that lack direct disease activity measures such as the DAS28. Objectives We used machine-learning methods to incorporate claims and electronic medical record (EMR) data to develop models to predict DAS28 (CRP) as a continuous measure, and to distinguish moderate-to-high disease activity from low activity/remission. Methods We identified 300 adults (≥18 years of age) with RA enrolled in a single academic centre cohort with ≥1 year of linked Medicare insurance claims preceding a DAS28 (CRP) measurement between 2006 and 2010. Of these, 95 had Medicare Part D pharmacy data. From claims we included demographics, co-morbidities, joint replacement surgery, physical therapy visits, numbers of RA-related codes, laboratory values and imaging studies, and healthcare utilisation. For those with Part D pharmacy data we included medications (steroids, analgesics, DMARDs) and switches between drugs. From the EMRs we obtained smoking status, BMI, blood pressure, medication use, laboratory values for seropositivity (RF or anti-cyclic citrullinated peptide antibodies), haematocrit, ESR and CRP. We constructed models with claims only, claims with medications and claims with EMR data. We examined these models with DAS28 (CRP) as a continuous measure and as a binary outcome (moderate/high activity vs low activity/remission). We used adaptive least absolute shrinkage and selection operator (LASSO), which avoids model overfitting by penalising large coefficients and selects a subset of variables by shrinking some coefficients to zero. We used adjusted R2 to compare continuous model fit and C-statistics to compare binary models. Results In models that included DAS28 as a continuous measure, using claims alone explained 11% of the DAS28 variability. Adding medications and EMR data to claims improved the adjusted R2 by 6% (table 1). In models that included DAS28 as a binary outcome (moderate/high activity vs low activity/remission), our claims-only model yielded a C-statistic of 0.68, which increased to 0.79 after inclusion of medications and EMR data. Conclusions Incorporating medications, EMR data and laboratory values into a claims-based index did not significantly improve the ability to predict DAS28 scores as a continuous measure. However, models that include claims, medications and EMR data may be used to reasonably distinguish moderate-to-high disease activity from low disease activity/remission. References [1] Sauer BC, et al. Arthritis Res Ther2017;19:86. [2] Desai RJ, et al. Arthritis Res Ther2015;17:83. Disclosure of Interest C. Feldman Grant/research support from: Bristol-Myers Squibb, Pfizer, K. Yoshida Grant/research support from: Tuition support from Harvard T.H. Chan School of Public Health (partially supported by training grants from Pfizer, Takeda, Bayer and PhRMA)., B. Pan: None declared, M. Frits: None declared, N. Shadick Grant/research support from: BRASS registry, Amgen, Bristol-Myers Squibb, and Mallinckrodt, Consultant for: Bristol-Myers Squibb, M. Weinblatt Grant/research support from: Bristol-Myers Squibb, Amgen, Crescendo Bioscience, Sanofi, Consultant for: Bristol-Myers Squibb, Amgen, Crescendo Bioscience, AbbVie, Eli Lilly, Pfizer, Roche, Merck, Samsung, Novartis, S. Connolly Shareholder of: Bristol-Myers Squibb, Employee of: Bristol-Myers Squibb, E. Alemao Shareholder of: Bristol-Myers Squibb, Employee of: Bristol-Myers Squibb, D. Solomon Grant/research support from: Bristol-Myers Squibb, Pfizer, Amgen, Genentech