AI and rare disease
How AI can help filter rare disease
A disease is defined as rare in Europe when it affects fewer than 1 in 2000. Rare diseases present a difficult healthcare conundrum. On the one hand, rare diseases need to be identified but on the other, we need to ensure our healthcare systems aren’t overloaded searching for rare diseases when a diagnosis could be something common.
This is where Artificial intelligence (AI) is making a contribution. AI is a big step forward in modern technology that’s capable of thinking for itself by making decisions without human input through a series of neural networks (which develop and grow like a human brain over time). This means they are able to process requests and create sub requests on their own, allowing huge sets of complicated data to be processed in a matter of seconds with no human intervention.
Humans are a remarkably adaptable and successful species and our current healthcare systems have undoubtedly contributed to our increasing longevity. However rare diseases present a particular challenge by presenting with similarities to common conditions.
How can AI help?
Amyotrophic Lateral Sclerosis (ALS) is a very rare condition that requires specific tests to determine if it is present or if it is a different disease. But how does a healthcare system get to the point of being able to identify the signs and symptoms and order the right test(s)? In a condition such as ALS, early intervention is the key to prolonged quantity and quality of life and AI could help by providing doctors with an alert as to what test(s) could/should be performed through analysis of carefully encrypted data via differential privacy techniques.
The machine learning portion of the AI would also be capable of learning from the data it gathers over time. This would mean that in theory it would become more and more accurate over time at spotting certain diseases and if we make sure that we focus on feeding it with as rich a set of data as possible, the results could be far-reaching.
Where are we now?
Many companies are already exploring the potential of AI and machine learning.
Face2gene was a genetic search and reference application for doctors in which AI scanned through image data of patient faces to spot signs of genetic disorders such as Down’s Syndrome. Another example of AI in action is a software package developed by a company called Diploid, it is called Moon and it enables diagnosis of rare disease through it’s software, reducing the time of potential diagnosis from weeks to minutes.
The Mayo Clinic partnered with an AI company called nference to create Qrative whose mission statement is to uncover the patients who have an unmet need for diagnosis and care and encourage pharmaceutical companies to “purpose” their treatments on these patients so they get treatment on time, rather than repurpose them later on in their lifecycle. For example, Thalidomide is used as a treatment for Multiple Myeloma now when it was designed to treat morning sickness in the 1960’s.
The final, and most obvious example of AI in action is IBM’s Watson computer system. Watson is a well-established and renowned AI computing system and unsurprisingly, it’s being used to help uncover rare disease in Germany via the University Hospital of Marburg’s Centre for Undiagnosed and Rare Diseases. As with Qrative’s system, Watson allows analysis and partial diagnosis of rare conditions in a matter of minutes rather than a matter of weeks, months or sometimes even years. This, obviously, is dependent on the quality of the data input and as you can imagine, it can often only offer partial diagnosis as human intervention is always required to be confident.
What does this mean for our doctors?
Precious healthcare resources’ time is freed up from scouring through clinical papers and medical textbooks (or even Google!). These AI systems do however still require human intervention. As a computer system can’t see a patient in reality, they can’t make an assumption based on incomplete or anecdotal evidence. They are dependent on the accuracy of the data they are fed and as such, they provide a list of potential conditions to review. These are ranked in ascending order as potential diagnoses.
By gaining this list of potential routes to investigate in a matter of minutes as opposed to weeks or longer, appointment time can be focused on explaining the options to the patient and arranging the care pathways and treatment options for the future. This means that a doctor can focus on their role guiding and informing the patient rather than on being a data cruncher.