AI at the Mayo Clinic: The leading edge of curating data and building apps

ORLANDO, FLA. – We know that the usefulness of AI-driven solutions is tied to both the quality and quantity of the data used. It’s not just a matter of garbage in, garbage out. You’ve got to have a massive amount of high-quality data to train the algorithms – that’s how to get results that will work with a broad range of patients. For this reason, the Mayo Clinic is building one of the largest repositories of clinical data in the world.

“Everything starts with data,” asserted Dr. John Halamka, president of the Mayo Clinic Platform, which is focused on transforming healthcare through the use of AI, connected devices and a network of partners.

He explained in a presentation at the recent HIMSS conference that you need accurate data and a lot of it. A colleague at another organization told him that his facility had 5,000 patient records with which they will build AI algorithms. “That’s not enough breadth,” warned Halamka.

For its part, the Mayo Clinic has 11.2 million patients with electronic records. And it’s not stopping there. The hospital chain is building a global, federated network of partner hospitals and patient records that can be drawn upon for building apps.

Already, it has 242 algorithms under development. The goal is to improve the art and science of medicine around the world.

“You need a global network to deliver on a global basis,” said Halamka. So, the Mayo Clinic has been creating alliances with other large hospitals and health organizations to share data. They include Toronto’s University Health Network, along with the Apollo chain of 73 hospitals in India and the Albert Einstein hospital in Brazil.

The data are de-identified and they never actually leave the host site – instead, metrics about the anonymous patients are shared.

That protects patient privacy. And the sharing of data over a wide range of geographies and ethnicities helps avoid bias in the data, as much as possible, when building AI models.

Nevertheless, said Halamka, “every algorithm will have a bias. We create and test the algorithms, recognize the bias, and then adjust.”

For example, the Mayo Clinic has a cardiology algorithm that can predict patient mortality. “It’s very accurate for people with low BMI, but not for those with a high BMI. Is it ethical to use it?,” he asked.

“Yes, for people with low BMI. That’s why you have to test and account for biases.”

And in the early days of generative AI – which is right now – it must also be recognized that the quality of genAI depends a great deal on the prompts or questions that are asked. Halamka talked about one instance of generative AI where “it could be accurate, or it could kill the patient.”

He cited the case of a 59-year-old patient with chest pain, shortness of breath and left leg radiation. ChatGPT was asked for a diagnosis and responded that the patient had likely suffered a mild cardiac infarction and that anti-coagulants should be introduced immediately.

ChatGPT was then asked, what possible diagnosis may have been missed? It responded with “aortic aneurysm”.

The problem, however, is that if only the first diagnosis had been followed, anti-coagulants could have killed a patient with an aortic dissection – it would have aggravated the bleeding.

Such hair-raising possibilities and serious risks beg the question, where should organizations start when creating AI-powered apps?

Best to start with low-risk apps, said Halamka, suggesting things like clinical documentation as the low-hanging fruit.

“With ambient listening, we’re reducing the pajama time for clinicians who otherwise spend hours documenting at home.”

He added, “It works well. You read the notes, edit and sign off. It’s low risk and high-benefit, even if it misses something.”

Email management is another useful app – AI can scan one’s emails and draft responses, saving the user a great amount of time. A clinician can then check the AI’s work, edit as needed, and send the emails off.

AI can even improve the writing of people – especially those who studied the sciences in school rather than liberal arts. Halamka said in this respect, AI has done wonders for his own writing.

On the other hand, the Mayo Clinic is avoiding high-risk apps, such as deploying clinical AI to patients. “The results could have serious repercussions for patients,” he said, especially if there are errors in the results. It’s early days for the technology, and it’s imperfect.

Still, AI is a moving target and improvements are emerging quickly. Halamka said that if AI is able, in two years, to read chest X-rays, and if an algorithm can spot polyps five times better than a GI specialist, why not make use of these tools? At least as intelligent assistants.

Halamka stated that a monumental use of AI would be to analyze the genome for various diseases, correlate with the experiences and outcomes of patients, and to then prescribe the best tests, medications and therapies.

He asked the audience how many have had their genome sequenced – and a few raised their hands. “So, you’ve had your three billion base pairs sequenced,” said Halamka. “That costs $500. But to interpret it would cost $1 million.” That seems like a good project for Mayo, he added.

The possibilities of such a project are mind-boggling. Halamka recounted the experience of his father-in-law, who was diagnosed with Stage Four pancreatic cancer – the point at which it can no longer be cured. The problem, he said, is that pancreatic cancer usually isn’t detected until Stage Four.

But using AI, “what if you can detect it at Stage 0?”

“If you have a risk for pancreatic cancer, wouldn’t you want a test for it?”

That comment opened a lively discussion with the audience. One member asked, “Just because you can do something, such as a test, should you do it?”

Halamka noted that it was an excellent question, and that some people would answer yes and some no.

For example, he mentioned the experience of his own mother, who has frontal lobe dementia and is in a memory clinic. She is now in her 80s and there is no known cure for her disease.

“Would you want to know in your 60s that you’ll have this in your 80s?,” asked Halamka. “Some people would, and some people wouldn’t.”

For its part, the Mayo Clinic is currently running a trial on a cohort that has an increased risk of pancreatic cancer. According to the hospital, the model is performing with 97 percent accuracy.

“Most people with an increased risk want to know,” said Dr. Matt Callstrom, chair of the Mayo Clinic’s Department of Radiology in Rochester, Minn.

Halamka observed that if you can catch pancreatic cancer in its early stages, it would save lives and reduce costs for the healthcare system. For these reasons, he added, it would make sense for insurers to fund trials of this kind.