Medical Artificial Intelligence · us
AI Risk Scoring Paired With Expedited Referral Shortens Time to Autoimmune Disease Diagnosis in Trial
A randomized trial linked risk scoring based on medical records to expedited rheumatology assessment; among those ultimately diagnosed, the median wait fell from 138 days to 43 days. The findings support the benefit of the overall care pathway, but cannot isolate the model’s contribution.
A positive antinuclear antibody test result is often just the start of a diagnostic journey. Some people need prompt rheumatology assessment, while others do not have an autoimmune disease; deciding who should receive priority appointments is a challenge after testing. A randomized trial at Vanderbilt’s health system in the United States shows that combining an interpretable risk model with expedited referral can shorten the time it takes patients to receive a diagnosis.
The study, called AURA, was published in NEJM AI on November 6. Antinuclear antibody (ANA) testing can help assess systemic autoimmune diseases such as lupus, but healthy people may also test positive. Test results alone therefore cannot determine who has a disease or who needs priority evaluation. The research team sought to use other clues in electronic health records to help identify higher-risk patients.
According to Vanderbilt’s November 11 account, the model used logistic regression and incorporated age, sex, ANA titer, platelet count, other autoantibodies, and diagnosis codes related to relevant symptoms. This type of model can show the factors influencing the score. When an intervention-group patient’s predicted 10-year risk of systemic autoimmune disease reached 20%, the team notified the ordering physician and offered an opportunity for expedited rheumatology assessment; physicians and patients could decline.
The trial was conducted within a single health system and randomized 2,385 people. Some participants were excluded after randomization, leaving 954 in the intervention group and 1,006 in the control group for the final analysis. Both groups received usual care, and physicians could arrange referrals at any time. Risk scores were also calculated for the control group, but no action was taken based on those scores. This allowed the study to test whether linking scoring to actual referral arrangements could change the diagnostic process.
The results showed that the intervention group received autoimmune disease diagnoses sooner, with a hazard ratio for diagnosis of 1.9, a 95% confidence interval of 1.3 to 2.7, and a P value below 0.001. Among those ultimately diagnosed, the median time from a positive ANA result to diagnosis was 43 days, compared with 138 days in the control group. These two medians describe only those who received a diagnosis and cannot be treated as waiting times for all participants.
Another comparison retained uncertainty: among people who underwent rheumatology assessment within six months of a positive ANA result, 23% in the intervention group received an autoimmune disease diagnosis, compared with 16% in the control group. The difference was not statistically significant (P = 0.07). The finding of faster diagnosis therefore cannot be directly extended to mean that the model has been shown to increase the proportion of assessed patients who receive a diagnosis.
This study provides randomized trial evidence for embedding a predictive tool in a clinical workflow, but it tested the combined intervention of a model plus expedited assessment and cannot separately estimate their individual effects. The study also came from only one health system, and some participants were not included in the analysis after randomization; the reasons for these exclusions and their potential effects must be considered when interpreting the findings. Whether earlier diagnosis can improve treatment outcomes and reduce organ damage, and whether other hospitals can provide the same capacity for specialist appointments, still require further validation.