Leveraging AI to assist doctors in diagnosing rare genetic disorders in children

Researchers have turned to OpenAI's reasoning model as a diagnostic assistant for rare diseases, and the results are turning heads in the medical community. In a study of previously unsolved cases, the AI helped identify 18 new diagnoses that had eluded clinicians for years. Here's what happened, why it matters, and what it means for the future of rare disease care.

The Rare Disease Diagnostic Puzzle

Rare diseases are surprisingly common as a category. While each individual condition may affect only a small number of people, collectively they touch hundreds of millions worldwide. The biggest problem? Getting a diagnosis in the first place.

Patients with rare conditions often spend years—sometimes a decade or more—seeing specialist after specialist, undergoing test after test, without ever receiving a clear answer. Doctors call this the "diagnostic odyssey," and it's exhausting, expensive, and sometimes dangerous, since untreated conditions can worsen while patients wait for answers.

Part of the difficulty is simple math. A single physician, no matter how skilled, can't hold detailed knowledge of thousands of rare conditions in their head. Symptoms that seem unrelated may actually point to a single obscure disorder. That's where large language models, with their ability to draw on vast amounts of published medical literature, offer a new kind of help.

How the Study Worked

In this research, scientists used an OpenAI reasoning model—a type of AI designed to work through complex problems step by step rather than just pattern-matching—to analyze medical cases that had gone unsolved despite traditional workups.

The model was given detailed patient information and asked to reason through possible explanations, weighing evidence and considering conditions that might not be obvious. Unlike earlier AI systems that simply suggested diagnoses based on symptom matching, reasoning models can lay out a logical chain: this combination of findings suggests this mechanism, which points toward these candidate conditions, which could be confirmed with these tests.

The headline result: the approach produced 18 new diagnoses in cases that had previously resisted explanation. These weren't trivial calls—they were conditions rare enough and subtle enough that standard diagnostic pathways had missed them.

What This Doesn't Mean

It's important to keep the enthusiasm grounded. This was a research study, not a rollout of AI doctors in hospitals. A few caveats worth noting:

• The AI didn't diagnose anyone on its own. Its suggestions serve as leads for human clinicians, who must verify findings with appropriate testing.
• Reasoning models can still be wrong. Like all large language models, they can produce confident-sounding but incorrect conclusions. In medicine, an error isn't just an inconvenience.
• One study isn't a treatment standard. Broader validation across larger patient populations, multiple hospitals, and controlled settings would be needed before this becomes routine practice.

The promising signal is that the model functioned as a "second opinion machine"—surfacing possibilities that busy clinicians, working with incomplete information, might never have considered.

Why Reasoning Models Are a Big Deal for Medicine

The shift from first-generation chatbots to reasoning models matters here. Earlier systems were good at retrieving information but weaker at connecting dots across complex, incomplete data—exactly what rare disease diagnosis requires.

A reasoning model can hold multiple hypotheses at once, notice when two seemingly unrelated symptoms fit a known syndrome, and explain why it reached a conclusion. That explainability matters enormously for doctors, who need to understand and challenge an AI's logic rather than just accept a black-box answer.

For patients stuck in a diagnostic odyssey, even a modest improvement in the speed or odds of reaching a diagnosis could change lives. Earlier identification often means earlier treatment, better outcomes, and years of uncertainty avoided.

The Road Ahead

This study points toward a future where AI tools sit alongside genetic testing, specialist consultations, and medical databases as part of the diagnostic toolkit. Realistically, that future involves AI flagging possibilities, doctors ordering confirmatory tests, and patients still making decisions with human clinicians.

Challenges remain: protecting patient privacy when feeding cases into commercial AI systems, ensuring models perform equally well across different populations, and integrating these tools into clinical workflows without adding burden to overstretched doctors. But as a proof of concept, identifying 18 previously missed diagnoses is hard to argue with.

The takeaway for most readers: AI isn't diagnosing rare diseases on its own, and you shouldn't try to self-diagnose with a chatbot. But the technology is increasingly capable of helping the professionals who do—and for rare disease patients, that could eventually mean shorter, less painful journeys to answers.

Note: This article describes research findings and is not medical advice. If you have concerns about a health condition, consult a qualified healthcare professional.