Rare-Disease Diagnostic AI Enters a Randomized Physician Trial, as RaDaR Shifts the Question From Answering to Clinical Collaboration

The difficulty of rare diseases often lies not in how unfamiliar any single disease name is, but in how clues are scattered across years of illness, different specialties, and fragmented medical records. When physicians must organize a differential diagnosis from among thousands of possible diseases within limited time, an AI tool that can read clinical narratives and propose a shortlist, if it truly improves judgment, is more than an upgraded search engine.

On June 23, the research team published a preprint on arXiv introducing RaDaR (Rare Disease navigatoR), a rare-disease diagnostic reasoning large language model. It is a 32B-parameter, open-source-oriented model that the researchers say was trained on 49,170 publicly available free-text cases and 104,666 synthetic cases, with an added reasoning-enhancement process. The goal is for the model to directly process clinical text rather than rely only on already organized phenotype codes.

In public benchmarks and data from four external validation centers, RaDaR was configured to generate a candidate list of up to five possible diagnoses, rather than give a single conclusion. The paper reports that it ranked near the top in comparisons with multiple open-source models; on external real-world electronic medical record data, the diagnostic accuracy listed by the researchers varied substantially by center, ranging from about 48% to 84%. That variation itself also shows that medical-record writing styles, specialty mix, and disease distribution can profoundly affect the usability of medical AI.

Closer to the clinical setting were two workflow evaluations. The first was a retrospective disease-course analysis: among 113 patients with confirmed rare diseases who had undergone multiple undiagnosed medical visits, RaDaR included the final diagnosis among its candidates earlier than the first clinical suspicion of that rare disease documented in the medical record in 61.06% of cases; the researchers estimated that this could have been an average of 1.87 months earlier. This is not prospective evidence of clinical outcomes, but it raises a concrete question: whether the model can put the right direction in front of physicians a little earlier along a long diagnostic journey.

The second was a randomized physician-assistance trial. The paper says 84 physicians from 28 hospitals were randomly assigned to two groups, of whom 76 completed the trial; the control group could use general web search, while the RaDaR group could use both the model and the web. Among those who completed the trial, the diagnostic accuracy of the RaDaR-assisted group was 49.44%, compared with 28.00% in the control group, a difference of 21.44 percentage points, and the average answer time did not increase significantly.

However, improved accuracy does not mean the handoff can be made with confidence. The research team also reported a safety signal: when the RaDaR-assisted group answered incorrectly, it had a higher proportion of high-confidence ratings, suggesting that AI may make some erroneous judgments feel more certain. For clinical decision-support systems, this kind of “calibration” issue is critical; if physicians treat model outputs as authoritative answers rather than differential diagnoses to be verified, the risk shifts from missed diagnosis toward overconfidence.

This study is still an arXiv preprint and cannot yet be equated with peer-reviewed clinical evidence. It also has not shown that RaDaR can improve patient outcomes, shorten real-world diagnostic journeys, or pass regulatory review. Its value lies in pushing rare-disease AI evaluation one step forward: from static case-answering toward external medical records, disease-course timing, and physician interaction. If the next step is hospital adoption, the questions will become more practical and more stringent: data privacy, local deployment, error tracking, accountability, and how the model can maintain reliable performance across different healthcare systems.