Can AI Improve Diagnosis of Rare Diseases?

Rare diseases can often be difficult to diagnose because their symptoms can overlap with a variety of conditions. One such disease is acute hepatic porphyria (AHP), a rare genetic disease with gastrointestinal symptoms that resemble endometriosis, appendicitis, and other disorders. It can take up to 15 years to properly diagnose, leading to worsening symptoms and, in some cases, irreversible organ damage.

To improve diagnosis of rare diseases like AHP, researchers from UC San Francisco and UCLA, invented a predictive algorithm to analyze health care records and identify suspected porphyria patients. The algorithm searches through electronic health records to identify disease patterns and flag patients who may be at risk for AHP. Although AHP only affects about 1 in 100,000 people, using AI to help identify patients with this condition may one day pave the way to including AI in the overall diagnostic process.

What is acute hepatic porphyria (AHP) and why is it so difficult to diagnose?

AHP is part of a family of metabolic disorders called porphyrias. They are typically genetic disorders and are somewhat rare. One of the classic symptoms of AHP are acute attacks of abdominal pain. There are other symptoms like nausea, vomiting and limb weakness, but the acute attacks of abdominal pain are the most notable.

Why is it so hard to diagnose?

Part of the reason AHP is so challenging to diagnose is because there are many disorders that present with abdominal pain. The condition is also more common in young women, who are frequently misdiagnosed and the pain is attributed to other kinds of issues like gynecological or gastrointestinal complaints.

Some patients can go years without a definitive diagnosis. There are FDA-approved treatments to prevent these attacks, but the challenge is how to identify the people who can benefit.

How did your work with AI and AHP begin?

We were contacted by a company that developed a preventative treatment and wanted to test a machine learning-based approach to identify undiagnosed patients.

How does using AI help with the diagnosis of the condition?

We wanted to train an algorithm to help make the diagnosis. AI can sift through the full medical record – lab tests and results, medication orders, clinical notes, demographics, procedures and other diagnoses – looking for the signs and symptoms indicating a reasonable possibility of AHP. Using these signals, the algorithm can calculate the probability of AHP.

How did you train the algorithm to look for AHP?

We had access to 10 years of anonymized patient data from the UCSF and UCLA medical record systems. From this combined dataset, we focused on just those patients who had come to UCSF or UCLA reporting acute abdominal pain. From this group, we identified patients who were subsequently diagnosed with AHP, our cases. Those who were ruled out for AHP, were our controls. These cases and controls were used to train machine learning models to distinguish cases from controls using signals buried within their electronic medical records.

How accurate was the algorithm?

We trained two different prediction models. The first predicted which patients would eventually be referred to an AHP specialty clinic for further evaluation. The second predicted which of these patients would eventually be diagnosed with AHP. These two models were assembled together to make predictions on the overall population of patients suffering from acute abdominal pain. Taken together, the model accuracies were between 89% to 93%, which is pretty good considering how uncommon this disease is.

Why use AI to identify patients?

We want to lessen unnecessary suffering in people who have this condition and speed up diagnosis. We were able to determine that if the algorithm had been running on these patients, we might have saved about a year’s worth of time using AI diagnosis compared to the usual clinical pipeline.

What’s next for this work?

The next phase of this research is to extend into other disease spaces like rare blood and vascular disorders, and eventually to more common disorders like irritable bowel disease and hypertension.

What’s the future of AI in diagnosing patients?

The real gold standard for testing machine learning algorithms is to see how well they make predictions on brand new patients. When this kind of early warning system is developed and then rigorously tested, it could eventually be a key part of the resources built into a medical record system and help health care providers consider a diagnosis and not have patients undergo unnecessary testing or treatments.