跳到正文
原文
MedPage Today·· 14 天前精选AI 评分76

AI工具缩短自身免疫疾病确诊时间

Positive Antinuclear Antibody Test? This Tool Helps Show What That Means.

AI 导读

一项随机对照试验表明,使用AURA工具评估自身免疫风险,可将确诊时间缩短约70%,从阳性抗核抗体测试到确诊平均只需43天,而非常规护理需138天。

推荐理由

研究显示,AI工具能显著缩短自身免疫疾病的确诊时间,有助于优化临床诊疗流程。

正文 · 原文

— Dramatically cuts time to definitive diagnosis of autoimmune disease in randomized trial

by John Gever, Contributing Writer, MedPage Today

November 6, 2026 • 4 min read

  • Antinuclear antibodies (ANAs) are a hallmark of certain autoimmune conditions, but they may also be found in healthy people and those with other disorders, such that their presence often leads to a "diagnostic odyssey."
  • To speed up the journey to a definitive diagnosis, researchers developed a predictive tool based on a short list of demographic and clinical characteristics to indicate which patients with ANAs likely have an autoimmune disease.
  • This randomized trial found that, for patients who did have an autoimmune disease, the tool cut the median time to diagnosis by about 70% compared with usual care.

For patients ultimately determined to have an autoimmune disease following a positive antinuclear antibody (ANA) test, a digital aid made what is often a diagnostic odyssey more like a trip to the corner store, results of a randomized trial indicated.

Called AURA, use of the (relatively) simple tool cut the median time to autoimmune disease diagnosis by nearly 70% -- 43 days from the positive ANA test, versus 138 days with usual care, according to April Barnado, MD, MSCI, of Vanderbilt University in Nashville, Tennessee, and colleagues.

When applied to some 2,000 ANA-positive patients, most of whom were not ultimately diagnosed with an autoimmune disease, the tool had an area under the receiver operating characteristic curve (AUC) of 0.79 (95% CI 0.73-0.86) for identifying those who did, the researchers reported in NEJM AI.

"By leveraging routinely collected, real-time EHR [electronic health record] data within a robust randomized pragmatic design, this study demonstrates a scalable framework for moving predictive analytics from theoretical validation into operational clinical implementation," Barnado and colleagues wrote. "Although future multicenter trials are required to establish external generalizability across diverse community practices, these findings illustrate that tightly integrated, interpretable structured-data models can meaningfully accelerate clinical care timelines for high-risk patients."

AURA stands for AUtoimmune Risk Assessment and combines a variety of demographic and clinical factors to produce a "risk score" for the likelihood that an ANA-positive patient has an autoimmune condition such as lupus or systemic sclerosis. ANAs are a hallmark of these diseases but very far from specific to it. One review from 2018 indicated that 2.5% of the general population has "significantly elevated" ANAs, with the authors writing that "[m]ost people with a positive ANA count are not diagnosed with autoimmune disease, and the probability of future disease is low."

Still, the presence of ANAs certainly raises the likelihood that autoimmunity is the reason. The clinical challenge is to identify those patients for whom it's worthwhile to pursue the additional rounds of testing and physical exams involved in a rheumatologic workup. Diagnostic odysseys are notoriously common for ANA-related diseases, with their highly variable clinical presentations -- Barnado's group noted that 7 years is the average time from symptom onset to diagnosis for systemic lupus.

In an earlier report, Barnado and colleagues described how they distilled data from about 2,000 ANA-positive individuals into a model that would pinpoint those most likely to have an autoimmune disease, via a numeric risk score. These data were of the kind routinely included in patients' EHRs: absolute ANA titer, numbers of billing codes for symptoms consistent with autoimmune disease, platelet count, and presence of a second autoantibody, along with age and sex. This initial training-and-validation model yielded an AUC of 0.83 for triaging patients into a rheumatologic evaluation.

For the current, prospective study, the researchers applied the model to a different set of ANA-positive patients seen at Vanderbilt, randomizing about 2,400 to either have their data analyzed with the model, with the resulting risk score conveyed to the treating clinicians via notification in Vanderbilt's EHR system or to usual care in which clinicians used only their own judgment.

Roughly 200 patients in each group were excluded after randomization, either because they were already scheduled for rheumatologic exams or because they had already seen a rheumatologist outside the Vanderbilt system. That left 954 assessed with AURA and 1,006 in usual care. Of the former, AURA identified 270 as high risk for autoimmune disease, making them eligible for an expedited rheumatology referral.

As expected, most patients in the final cohort were not ultimately diagnosed with an autoimmune disease: Within 800 days of the first positive ANA test, only about 8% of those evaluated with AURA ended up with such a diagnosis, versus roughly 5% of the usual-care group. This difference made sense, Barnado's group explained, because of the expedited rheumatologic exams available to those identified as high risk with AURA, who would thus be diagnosed sooner. One could also imagine that in the usual-care group, absence of an EHR flag to suggest a rheumatologic referral would mean that some high-risk patients would never get one.

Confining the analysis to those in the highest quartile of risk scores indicated that 20% were eventually diagnosed with an autoimmune disease, and nearly all of them got that diagnosis within 3 months of their first ANA-positive test. Notably, patients in the usual-care group were also scored with AURA but this was not relayed to clinicians; only about 12% of those with high scores received an autoimmune diagnosis, and it took considerably longer for those who did.

The researchers boasted that their tool's simplicity, both in the small number of variables employed and its regression-based statistical machinery, should make it more useful in the clinic than those powered by big AI models. "Unlike black-box deep-learning algorithms that can obscure decision-making," they wrote, "our model allows providers to easily verify the clinical factors driving the risk score."

An important limitation was the single-center data source, both for the model's development and its testing in the current study. Also, the authors acknowledged that it was impossible to distinguish "the independent predictive utility of the clinical prediction model from the effect of the early rheumatology referral invitation sent to providers" -- merely sending all ANA-positive patients for rheumatologic evaluation would likely speed up diagnosis, though costly and inefficient.

来源:MedPage Today · medpagetoday.com