Medical-AI trial: a large language model to help diagnose rare diseases — can it shorten the diagnostic odyssey? (cluster-randomized) (NCT07650799)

Rare diseases each affect few patients, but there are thousands of them, and together they are far from rare. Many present with non-specific symptoms, and reaching the right diagnosis can mean moving from doctor to doctor over years. This is called the diagnostic odyssey, a heavy burden on patients and families.

A large language model (LLM) that can handle vast medical knowledge raises the hope of listing candidate diseases from symptoms and shortening that odyssey.

This study tests such a rare-disease diagnostic LLM in real practice. Per the registry summary, it evaluates, as a multicenter, prospective, cluster-randomized, parallel-controlled real-world effectiveness study, whether the LLM can improve diagnostic quality, efficiency, and health-economic outcomes for physicians managing patients with suspected rare or diagnostically unresolved disease.

The primary outcomes are the diagnostic accuracy of the top-3 candidates the LLM lists and the rare genetic-disease diagnostic yield (the proportion reaching a diagnosis). The size is 1,055, randomized by cluster (groups of sites or physicians).

The key here is confirming, in a randomized comparison, not only accuracy on paper but real-world outcomes: whether it aids physicians in actual practice, shortens the odyssey, and helps even accounting for cost. Because rare-disease information is scattered and specialists are limited, an LLM that pulls knowledge together to present candidates may be an especially good fit.

On the other hand, there is a risk of plausible-sounding errors (hallucination), so its real power as a tool supporting physician judgment must be measured carefully. That said, this study evaluates diagnostic quality, efficiency, and economics; it does not establish replacement of diagnosis by the LLM or its effectiveness.