AURA, an autoimmune AI referral-triage tool, was associated with a median time to diagnosis of 43 days versus 138 days in a randomised Vanderbilt trial, according to a Johns Hopkins release published on 6 November. The system alerted physicians when patients with a positive antinuclear antibody test had a predicted ten-year systemic autoimmune disease risk above 20%.
The result concerns getting selected patients to specialist evaluation, not software diagnosing a disease. A positive antinuclear antibody, or ANA, test can occur in healthy people and in people with other conditions, the release explains. With rheumatology referrals as its target, AURA addresses a clinical queue in which the initial screening result alone does not establish who has an autoimmune condition.
The study published in NEJM AI, led by April Barnado with colleagues including Ryan Moore, Henry Domenico and Daniel Byrne, evaluated the model within Vanderbilt’s electronic health record workflow. The trial figures below are those reported in the university release.
Autoimmune AI alerts offered expedited rheumatology evaluation
The prospective trial ran from 20 June 2023 to 9 March 2025 in Vanderbilt’s Epic system. Researchers randomly assigned 954 patients to an intervention group and 1,006 to a control group. Control patients also had risk scores calculated, but those scores were not proactively shared with their clinical teams.
AURA used existing record information including age, sex, laboratory values such as platelet counts, diagnostic billing codes and ANA titre. A positive ANA result triggered the risk assessment. For high-risk patients in the intervention group, an automated notification offered the ordering physician an expedited rheumatology evaluation.
Both groups continued to receive standard care, with physicians using their own judgement about referrals. Chart reviews every six months checked whether patients had seen rheumatology and received an autoimmune diagnosis. This tests an alert added to clinical practice rather than replacing the clinician’s assessment with a model output.
That distinction also separates the trial from RPI’s Quest2Dx Parkinson’s screening work, which TSN covered as research-dataset testing rather than a clinical deployment. Both concern identifying people who may need further evaluation, but their evidence comes from different settings.
The diagnosis percentages refer to patients seen within six months
Johns Hopkins reports median diagnosis times of 43 days for the intervention group and 138 days for standard-care controls. Separately, among patients evaluated by rheumatology within six months of a positive ANA test, 23% of the intervention group and 16% of the control group received an autoimmune diagnosis. Those percentages describe the evaluated subgroup, not all 954 and 1,006 assigned patients.
The release gives another denominator for access to care: 203 of the 270 high-risk intervention patients, or 75%, were seen in rheumatology clinic. That is neither the fraction of the whole trial diagnosed nor a measure of the model’s diagnostic accuracy. These figures answer different questions about timing, specialist attendance and diagnoses among people evaluated.
The screening-versus-diagnosis boundary is also central to Google’s smartwatch hypertension study. A risk signal can prompt clinical follow-up without being a standalone diagnostic result.
Referral speed is an endpoint, not a diagnosis by software
The trial’s practical contribution is to evaluate what happens after a score enters a real clinical workflow. A prediction that remains hidden cannot prompt the same response as an alert offering an expedited appointment. Comparing those pathways therefore examines the combination of risk assessment, notification and access to specialist review, rather than the model in isolation. It would be too broad to turn that finding into a promise that any hospital installing the model will reproduce the reported time difference.
The denominators sharpen that limitation. Faster median diagnosis and a higher diagnosis proportion among people seen within six months are useful but distinct observations. Neither says that everyone referred had disease, that everyone flagged obtained an appointment or that the tool can replace a rheumatologist. The 203-of-270 attendance figure keeps the gap between identifying risk and completing a specialist visit visible.
Capacity is also part of this interpretation. Unlike the NHS use of AI MRI software to shorten scans, AURA changes which patients are prioritised for evaluation rather than the speed of the underlying examination. The study supports attention to that referral pathway, while a single health system leaves performance in other workflows an open question.
The researchers made the model open source to enable geographic validation and pragmatic trials elsewhere, according to Johns Hopkins. Results from those other settings—not an assumption of universal transfer—are the next evidence to watch.
Sources
- Johns Hopkins: AURA trial release, 6 November 2026
- Barnado et al., NEJM AI: Randomized Evaluation of an Autoimmune Disease Prediction Model — AURA, DOI 10.1056/AIoa2600112
Related stories
RPI’s Quest2Dx AI Screens for Parkinson’s Using Health Questionnaires follows the referral question into another disease area, with an important contrast: dataset evaluation rather than this prospective clinical trial.
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
