Supplementing Claims Data with Electronic Medical Records to Improve Estimation and Classification of Rheumatoid Arthritis Disease Activity: A Machine Learning Approach.

objectivePrevious attempts to estimate rheumatoid arthritis (RA) disease activity using claims data only did not yield high performance. We aimed to assess whether supplementing claims data with readily available electronic medical record (EMR) data might result in improvement.

methodsWe used a subset of the Brigham and Women's Hospital Rheumatoid Arthritis Sequential Study (BRASS) that had linked Medicare claims. The disease activity score in 28 joints with C-reactive protein (DAS28-CRP) was considered the gold standard of measure. Variables in the linked Medicare claims, as well as EMR recorded in the preceding one-year period were used as potential explanatory variables. We constructed three models: "Claims-Only," "Claims + Medications," and "Claims + Medications + Labs (laboratory data from EMR). We selected variables via adaptive LASSO. Model performance was measured with adjusted R2 for continuous DAS28-CRP and C-statistics for binary category classification (high/moderate vs low disease activity/remission).

resultsWe identified 300 patients with laboratory data and linked Medicare claims. The mean age was 68 years and 80% were female. The mean (SD) DAS28-CRP was 3.6 (1.6) and 51% had high or moderate DAS28-CRP. For the continuous estimation, the adjusted R2 was 0.02 for Claims-Only, 0.09 for Claims + Medications, and 0.18 for Claims + Medications + Labs. The C-statistics for discriminating the binary categories were 0.61 for Claims-Only, 0.68 for Claims + Medications, and 0.76 for Claims + Medications + Labs.

conclusionAdding EMR-derived variables to claims-derived variables resulted in modest improvement. Even with EMR variables, we were unable to estimate continuous DAS28-CRP satisfactorily. However, in claims-EMR models, we were able to discriminate between binary categories of disease activity with reasonable accuracy.