Using AI to Improve Rare Disease Diagnosis: Key Insights from #WeekInRare

By Paul Howard, PhD Senior Director, Public Policy, Amicus Therapeutics

Each year, Global Genes convenes one of the largest gatherings of people affected by or working in rare diseases. People living with a rare disease, their caregivers, advocates, healthcare professionals, researchers, partners, and allies convene at the RARE Advocacy Summit to learn from each other. The last week of September I was grateful to attend the Global Genes #WeekinRare in Kansas City to hear about the latest developments from leading innovators in the field.

One of the major through-lines of this year’s conference focused on harnessing data sharing and AI to shorten the diagnostic odyssey and accelerate delivery of effective treatment options for the more than 90% of rare diseases that lack them today.

Achieving this outcome is critical: There are more than 10,000 rare diseases, many of which also have symptoms that can mimic more common diseases. Even when there is a common genetic origin for a given rare disease, its symptoms and progression can vary widely across individuals. As a result, getting to the right diagnosis can be a daunting and painful process – physically, financially, and psychologically.  

Key data challenges in diagnosing and treating rare disease

A doctor’s first encounter with a patient who has a rare disease can either be the beginning of a prolonged diagnostic odyssey or – with the right AI tools – put that patient on the fast track to an early diagnosis, potential treatment, or enrolling in a clinical trial testing a new medicine.

Melissa Haendel, Director of Precision Health and Translational Informatics and Sarah Graham Kenan, Distinguished Professor at the UNC School of Medicine, delivered a fascinating talk on the promises and pitfalls of informatics for rare disease diagnosis, and how data integration (or what she calls “making all the data count”) can help us shorten the diagnostic odyssey facing patients and caregivers. First, the majority of rare diseases affect children, but our primary screening tool for identifying children at high risk of a rare disease diagnosis – newborn screening (NBS) – will inevitably miss many diseases that aren’t included in state NBS programs. Conversely, late- or adult-onset rare diseases present their own diagnostic challenges because treating physicians may have never seen them before, forcing patients to navigate between doctors and specialists for years (and sometimes decades) before reaching a definitive diagnosis.

Simply put, primary care physicians are not trained to diagnose rare diseases, and the health IT systems they use to record routine clinical encounters and facilitate billing aren’t optimized to organize information that helps them to “connect the dots” in a way that can easily lead to a successful rare disease diagnosis.

There are many centers of excellence (the National Organization for Rare Disorders maintains one directory of such centers) doing amazing work in diagnosing and treating people living with a rare disease, but these centers are often located in large urban areas – making them difficult to reach  for people living in rural and underserved communities, or for anyone who struggles to take time away from work or family to travel.

In short, physicians who do not work at major medical centers need help not just organizing complex data sets but also powerful analytic tools that can help them make sense of it. Dr. Haendel thought doctors could benefit from a “zebra button” in medical charts to help make sense of a patient’s complex symptoms (this is because physicians are taught to think of “horses not zebras” and focus on more common diagnoses first).  

Medical chatbots can also help physicians record interviews with patients, and then summarize and organize the information in tandem with previous office visits, tests, and scans. Large Language Models (like ChatGPT) can help scan medical literature to identify other potential case reports, clinical trials, or journal articles that might provide greater insight into whether a patient’s rare genetic variation might be pathogenic, or benign. 

A lot of these AI tools are under development, and regulators at the FDA, EMA (Europe), and MHRA (UK) are thinking very hard about how to develop guardrails for medical AI that focus on preventing bias in the models (from biased training data), strengthening explainability (so that the physician and the patient can understand why the model might suggest a certain diagnosis or treatment), and monitoring real world patient outcomes (AI models are subject to model “drift” when the data they are trained from becomes outdated, i.e., the patient population changes or physicians use the tools in ways the developers didn’t anticipate).     

While AI isn’t going to replace doctors anytime soon (and probably never will) these tools can help us do things that humans struggle with, curating patient data from tests, clinical observations and case studies, patient reported data from registries, and the latest studies from the medical literature.  If you’re looking for a match for a one in a million rare disease, human beings are never going to have the time to scan a million medical records or tens of thousands of journal articles looking for a digital needle in petabytes of data. The AI tools can find the needle and even help you understand why it’s there.  

Integrating AI into clinical decision support systems is a key part of the solution

Today, the systems we employ for electronic health records (EHR) are just not designed for organizing and presenting information at the scale required to differentiate rare diseases. Some of this data is well coded and labeled, while other data is unstructured – in notes, imaging, or other non-standardized formats.

We use AI as a catch-all term but there are algorithms that can be optimized for very different tasks based on their training data – natural language processing, facial recognition, imaging analysis, etc. And the models are getting better at explaining their own work (called explainable AI, or XAI) and helping to check the work of other algorithms, through confidence scores.   

There are also other types of data – pictures, audio, and video recordings – available outside of traditional medical encounters that (if analyzed effectively) can help us return a faster diagnosis or track how diseases evolve over time (including in response to different types of treatments).

Companies like FDNA, who spoke at the Global Genes conference, utilize machine learning analysis of facial features common to certain rare diseases (face2gene) to help physicians narrow down possible diagnoses.

A supporting tool rather than a replacement for human understanding

AI technologies can offer a great deal of value in the delivery of healthcare to people struggling with complex chronic and rare diseases. However, patient experience and the knowledge, empathy, and intuition of physicians will always be critical to delivering care that is compassionate, ethical, and personalized.

Dr. Haendel pointed out that parents are often the first to notice that something is not right with their children; they bring a wealth of insight and intuition that is not captured in health records or is simply too subtle for current tests to detect. Parents and patient advocates have been transformational for many rare disease research efforts, and it is critical people living with rare diseases remain co-directors of their own care and that we design AI systems that reflect human values (while minimizing our human biases).

That brings us to where AI is today. There are very real concerns about AI’s potential to exacerbate existing social and medical biases, propensity to hallucinate, and risk of violating patient privacy and physician autonomy.

Still, these problems aren’t unique to AI.  In our health care system today, access to high quality care can feel like a zip code lottery, implicit and explicit biases impact patient outcomes, and delayed or incorrect diagnoses happen all too frequently.

Integrated with the right safeguards, and human beings in the loop every step of the way, AI can help ‘rehumanize’ a healthcare system that already feels deeply impersonal.  

Powering a brighter future for families affected by rare diseases

So how do we get there? At Amicus Therapeutics, we are not an “AI-first” company, but we recognize the value that these systems can have for improving patient outcomes and want to help develop responsible and transparent standards that can advance the field. We are one of the founding members of Rare Disease Diagnostic Artificial Intelligence Consortium (RADIC). We’ve published research in Orphanet on how machine learning algorithms can help find patients with an elevated risk of having Fabry disease.    

We’re doing this work to support our commitment to health equity, because we know that cutting edge technologies have to be developed with a focus on ensuring that every person who has a rare disease has access to the care and treatments they need, with patients as our partners.