Diagnostic tests are the lynchpin of medical care. From a health point of view, bad testing can lead to missed or incorrect diagnoses. From a cost point of view, bad testing is a major source of unneeded expenditures. Doctors must consider all the information about a patient when deciding which tests, if any, to deploy. Ziad Obermeyer, MD, MPhil, a physician in Emergency Medicine, is tackling this problem by leveraging the power of artificial intelligence. Obermeyer presented his work on developing algorithms to aid physicians in testing decisions at the 2018 World Medical Innovation Forum, and was awarded the Peter K. Ranney Innovation Award at the conference. He recently spoke with CRN about his research, goals and how he first got involved with artificial intelligence and emergency medicine.
Why did you decide to go into emergency medicine?
ZO: No matter what rotation I was on in medical school, I always enjoyed the parts in the emergency department (ED) the most. When it came to choosing a specialty, emergency medicine let me do the parts of all the other specialties I liked best, and offered an overview of medicine and physiology you can’t get from any other single specialty. I love the clinical part of my job — most shifts I get the sense that I left the world in a slightly better place than I found it. Emergency medicine also fit well with my research, which is all about decision-making under uncertainty.
In the ED, clinicians are often faced with many diagnostic testing decisions every day. What are the current issues with these testing decisions?
ZO: We all know there’s a lot of overuse of diagnostic testing: tests that come back negative, hospitalizations that change little about a patient’s care, etc. Underuse is harder to see: a test that would have been positive, a hospitalization that would have prevented death. I’ve worked a lot on this by looking for examples of patients who die in the days after being sent home from the ED. It turns out that among discharged patients, more than 10,000 annually die within seven days of discharge.
Your work is focused on using artificial intelligence to aid physicians in testing decisions. How can artificial intelligence help improve diagnostic testing?
ZO: In my lab, we consider one specific but very important decision: testing ED patients for subtle acute coronary syndromes with stress testing or catheterization soon after their visits. We train a machine learning algorithm to identify patients who would receive a potentially life-saving intervention (a stent or open-heart surgery) after one of those tests – the basic idea is, if we’re doing a test, we’d rather do it on someone who is going to have a positive test and then receive a revascularization intervention than someone who isn’t.
We can then look at doctors’ testing decisions through the lens of what the algorithm would have done. We find two things: First, doctors perform a lot of tests on people for whom the algorithm would have advised, “Don’t do this test!” – using only information available the moment the patient walks into the front door of the ED. It’s not surprising that there’s a lot of overuse – up to 30-40 percent of all tests – but the cool thing here is that rather than telling doctors retrospectively, “You shouldn’t have done that test because it led to nothing,” we can give the doctor guidance prospectively. So, this is one solution for overuse that doesn’t depend on asking doctors to “do less” in general, or make it harder for them to do tests; it gives them individualized guidance on specific decisions.
The more surprising thing is that the algorithm finds a lot of people who look very high risk, but weren’t ever tested by the doctor in the ED. Those people can experience dangerous arrhythmias, heart attacks, a need for urgent revascularization procedures, and even sudden death at very high rates, suggesting that they might have benefitted from testing. We can also tease out some of the cognitive errors that doctors are missing: over-relying on risk information contained in demographics, under-weighting symptoms in people with competing diseases like COPD, etc. This can help us with the problem of underuse.
What is the path for getting your algorithm to physicians in the ED to help them make testing decisions?
ZO: Like any new medical technology, there’s a clear pathway for getting this ready for the clinic: a randomized trial. There’s a lot of talk of doctors adopting and not adopting algorithms, but in medicine we have a playbook for this. We just need to run the playbook.
You presented on your algorithm at the 2018 World Medical Innovation Forum and were awarded the Peter K. Ranney Innovation Award for your achievements. What was the Forum like?
ZO: The Forum was an amazing gathering of some of the most exciting, innovative people in medicine. I made a ton of connections that will be incredibly useful in pushing this work forward.
This year’s forum focused on the advances and opportunities at the intersection of artificial intelligence and healthcare. How do you think artificial intelligence will disrupt healthcare in the future?
ZO: By tackling one problem at a time. There’s a lot of optimism about artificial intelligence in medicine, but there’s no substitute for painstaking, careful work on solving a single problem. That’s the work that will show the path forward.