Randomized Evaluation of an Autoimmune Disease Prediction Model — AURA (AUtoimmune Risk Assessment)

BACKGROUND: No evidence-based guidance exists for triaging individuals with positive antinuclear antibody (ANA) tests for rheumatology referral. We developed and deployed an interpretable clinical prediction model to predict which individuals with incident positive ANA tests will develop systemic autoimmune disease and evaluated its effect on time to autoimmune disease diagnosis in a 2-year randomized pragmatic study. METHODS: The model was integrated into the electronic health record (EHR) at a single academic health system. All ANA-positive individuals were randomly assigned 1:1 independently of risk scores. In the intervention group, individuals with high-risk scores were invited for expedited rheumatology evaluation. In the control group, risk scores were calculated but not acted on. All participants received standard of care and could be referred to rheumatology. Chart reviews every 6 months assessed whether patients had undergone rheumatology evaluation and received a diagnosis of autoimmune disease. The primary outcome, time from positive ANA test to diagnosis, was analyzed using a Cox proportional-hazards model. Among individuals evaluated by rheumatology within 6 months of the positive ANA, the proportion diagnosed with an autoimmune disease was compared between study groups. RESULTS: Of 2385 randomly assigned individuals, 954 were analyzed in the intervention group and 1006 in the control group after postrandomization exclusions. Individuals in the intervention group had a significantly shorter time to autoimmune disease diagnosis from the date of positive ANA (hazard ratio, 1.9; 95% confidence interval, 1.3 to 2.7; P<0.001). Among individuals diagnosed with an autoimmune disease, the median time to diagnosis was 43 days in the intervention group and 138 days in the control group. Among those evaluated by rheumatology within 6 months of a positive ANA test, 23% in the intervention group and 16% in the control group were diagnosed with an autoimmune disease (P=0.07). CONCLUSIONS: In this randomized pragmatic study, the intervention incorporating the clinical prediction model was associated with a significantly shorter time to autoimmune disease diagnosis. (Funded by the National Institutes of Health and others.).