AI Helped Diagnose a Rare Disease. Why the Warnings?
FND Nope
Moderator
In a recent BBC article, the parents of 19-month-old Lily described how ChatGPT helped them identify a rare genetic disorder affecting only around 70 known people worldwide. Lily had undergone heart surgery at four months old, and her pupils remained persistently dilated. Her mother, Rosie, entered these symptoms into ChatGPT, which suggested Multisystemic Smooth Muscle Dysfunction Syndrome (MSMDS), a condition associated with the ACTA2 gene.
When Rosie raised the possibility with doctors, some had never heard of the condition, while others considered it too unlikely because of its rarity. Subsequent testing confirmed the diagnosis. It is a remarkable example of what can happen when someone investigates a possibility that has not received sufficient clinical attention.
Yet the BBC article quickly turns towards warnings about using AI for medical advice. Nick Meade, chief executive of Genetic Alliance UK, said he would "urge caution", explaining that rare diseases are particularly vulnerable to problems with AI models. The article also cites research from the University of Oxford and quotes GP Dr Rebecca Payne, who warned that AI is not ready to replace physicians and can produce incorrect diagnoses or fail to recognise medical emergencies.
These are legitimate concerns, but they also raise a question that the article does not adequately explore: what exactly are we warning people against?
Investigating a diagnosis is not practising medicine
There is an important distinction between asking AI to suggest possible explanations for unexplained symptoms and relying on it to make clinical decisions. Lily's parents did not perform genetic testing themselves, prescribe treatment or attempt to replace her doctors. They identified a possibility, brought it to medical professionals and obtained confirmation through conventional testing.
Unless someone intends to perform CRISPR gene editing on themselves, prescribe their own medication or undertake some other potentially dangerous intervention, investigating a possible diagnosis should not be confused with practising medicine. Naturally, patients still need clinicians to assess the information they find. AI can misunderstand symptoms, invent references and generate convincing explanations that are entirely incorrect. Not everyone has the medical knowledge needed to recognise these mistakes.
But these limitations are arguments for responsible use and clinical verification, not for discouraging patients from exploring possibilities. We should be helping people understand how to distinguish a hypothesis from an established diagnosis, rather than treating the act of asking questions as inherently dangerous.
Doctors cannot know everything
One of the most revealing details in Lily's story is that some doctors had never heard of her condition. Others considered it too rare to be likely. Neither response is particularly surprising. There are thousands of rare diseases, and no individual clinician can reasonably be expected to know all of them, recognise every unusual presentation or remain familiar with every development in medical literature.
Doctors are human. They work within their own areas of expertise, under time constraints, and with incomplete information. They can overlook findings, pursue an incorrect hypothesis or simply encounter something they have never seen before. We also cannot assume that every consultation involves exhaustive consideration of all reasonable differential diagnoses.
That is precisely where AI has potential. It can help identify unfamiliar conditions, locate relevant literature and suggest diagnostic possibilities that might otherwise go unexplored. It does not need to be correct every time to be useful. It needs to generate possibilities that can subsequently be evaluated against the available evidence.
Meade's observation about rare diseases being poorly represented in AI training data is relevant. But rarity also presents a problem for human clinicians, as Lily's experience demonstrates. The fact that a condition is unfamiliar does not make it impossible, and the fact that AI sometimes gets rare conditions wrong does not mean it cannot occasionally identify one correctly.
What about patients who already have a diagnosis?
This question becomes particularly relevant when patients have received diagnoses such as functional neurological disorder (FND) or irritable bowel syndrome (IBS), yet continue to experience unexplained symptoms, develop new clinical findings or receive investigation results that warrant further consideration. These conditions occupy areas of medicine where important questions remain about mechanisms, diagnostic boundaries and the relationship between symptoms and underlying pathology.
An existing diagnosis should not become a reason to disregard new evidence. If a patient diagnosed with FND develops additional neurological abnormalities, or someone diagnosed with IBS has investigations suggesting a gastrointestinal motility disorder, those findings deserve assessment on their own merits. The question is not whether AI can overturn a diagnosis, but whether it can help identify possibilities that have not yet been adequately investigated.
This is especially important where patients have spent years seeking answers. AI may help them recognise unfamiliar clinical terminology, understand the limitations of previous investigations or identify conditions worth discussing with their doctors. It may also help them formulate more precise questions about the evidence supporting their existing diagnosis.
Of course, a suggested alternative is not automatically a better explanation. The same standards of evidence should apply regardless of whether a hypothesis originates from a consultant, a medical textbook or ChatGPT. What matters is whether the proposed explanation can be investigated and whether the findings support it.
Why aren't more doctors using these tools?
Perhaps the more interesting question is why patients should be expected to discover these possibilities themselves. If a parent can use a publicly available AI system to identify an unfamiliar genetic condition, surely there is a case for clinicians to explore how the same technology might assist their own diagnostic investigations.
A doctor brings something to this process that AI cannot: direct clinical observation, physical examination, experience interpreting investigations and the ability to assess a patient's presentation in context. These insights can make AI-generated suggestions considerably more useful, while also helping identify when the system has misunderstood the evidence.
There is research worth considering here. A 2024 randomised study published in JAMA Network Open investigated whether access to GPT-4 improved physicians' diagnostic reasoning. Access to the model did not significantly improve their scores compared with conventional resources, although GPT-4 alone performed well on the study's clinical vignettes. The results illustrate both the potential of these systems and the importance of understanding how to use them effectively.
AI is literally a tool. In the hands of an experienced clinician, it can become something considerably more useful than an automated symptom checker. The clinician can assess the suggestions against the examination, laboratory findings, imaging and medical history. The AI can help broaden the search beyond the conditions that immediately come to mind. Neither replaces the other.
Caution should work both ways
The BBC article cites Dr Payne's warning that AI can produce incorrect diagnoses and fail to recognise emergencies. That deserves attention. But there is another risk that receives considerably less discussion in the article: the possibility that conventional medical care can also overlook a diagnosis, particularly when the condition is unfamiliar or unusually rare.
These risks are not mutually exclusive. We can acknowledge that AI sometimes produces incorrect information without assuming that doctors always identify the correct explanation. We can recognise the importance of clinical expertise without treating that expertise as exhaustive. And we can encourage patients to verify AI-generated suggestions without discouraging them from asking questions in the first place.
Lily's case does not establish that AI is generally more accurate than doctors, nor does it tell us how often similar investigations produce misleading results. What it demonstrates is that an AI-generated hypothesis can sometimes lead to a diagnosis that had not previously been established through conventional care.
For patients living with unexplained symptoms, that possibility matters. They do not necessarily need AI to tell them what disease they have. Sometimes they simply need help identifying something worth investigating, particularly when nobody else appears to be looking for it.
Responsible use should be encouraged, and clinical verification remains essential. But warnings about the limitations of AI should not obscure the limitations of human knowledge. If these tools can help patients and clinicians navigate those gaps, the question should be how to use them effectively, rather than whether patients should be investigating at all.
