(52) Improving The Accuracy of Automated Gout Flare Ascertainment Using Natural Language Processing of Medical Records and Linked Medicare Claims Data

Publication Number: 1097

Background: Recurrent gout flares are a crucial outcome in studies of gout treatment. Our previous gout flare claims algorithm did not perform well because flares are episodic with limited data to distinguish encounters for flares vs. routine follow-up for gout.

Objectives: To determine whether integrating medical concepts extracted from the narrative notes from the electronic health record (EHR) data using natural language processing (NLP) can improve the identification of gout flares.

Methods: Using Medicare data (2007-16) linked with EHR data from two tertiary care centers, we selected individuals with age≥65 and ≥1 gout ICD code who initiated the urate-lowering therapy (ULT). Patients were required to have a 12-month baseline period with continuous insurance enrollment before initiation. The baseline period and on-treatment follow-up were segmented into 1-month units. We then retrieved EHR notes for months with gout ICD codes in claims. We selected a random sample of patients and reviewed each of their notes for the presence of ongoing gout flare based on the physician documentation of gout flare as the gold standard. Months containing 1+ notes with ongoing gout flares were considered months with a gout flare event. We then used 60% of patients to train predictive models with LASSO using claims variables only, EHR medical concepts only, and both combined. We evaluated the added value of medical concepts by comparing the area under the curve (AUC) of ROCs in the 40% validation data. The top three variables by the magnitude of coefficients were listed.

Results: We extracted and labeled 3,482 notes from 500 patients for the presence of acute gout flares, representing 839 months of follow-up time, and 280 months with gout flares. We used 93 claims variables (44 time-varying) and 84 EHR medical concepts in LASSO. The claims-only model retained 20 variables (top 3: steroids, colchicine, diabetes) with a validation AUC of 0.693. The medical concept-only model retained 17 concepts (“gout”, “pain”, “arthralgia”) with a validation AUC of 0.688. The combined model retained 32 claims variables (steroids, NSAIDs, colchicine) and 13 medical concepts (“gout”, “pain”, “arthralgia”) with a validation AUC of 0.731.

Conclusions: Although claims variables alone and EHR medical concepts alone achieved similar discrimination, their combined use resulted in a 5.5% improvement. The proposed phenotyping algorithm for gout flares was designed to attain month-level precision facilitating use with current methods for comparative effectiveness studies. Limitations include the requirement for linked claims-EHR and the need for transportability assessment in other health care systems.