Researchers hopeful AI could help diagnose extremely sick patients

When a patient is struggling with a common symptom like shortness of breath or abdominal pain in the hospital, clinicians will consider a number of different diagnoses for the problem. 

When the patient is seriously ill, every moment counts. 

New research published in the journal of the American Medical Association this month finds that often, the right diagnosis comes late, or can be completely missed. 

The investigation found that about 23% of patients being treated at academic research centers who died or were transferred to the intensive care unit experienced a diagnostic error. The errors resulted in death, temporary or permanent harm in 79% of those situations. The patient died in one in 15 cases.

Investigators assessed the records of 2,428 patients who’d been treated between January 1, 2019 and December 31, 2019. The records were chosen from the more than two dozen academic medical centers that are participating in the Hospital Medicine ReEngineering Network, a national collaborative dedicated to improving patient care.

The study’s lead author, UCSF School of Medicine’s Andrew Auerbach, believes artificial intelligence could help lower those numbers. 

This interview has been edited for clarity and length.

Interview highlights

I understand that to compile this data, two doctors reviewed thousands of medical records to see if there were errors in providing a diagnosis. What were the top level takeaways for you? 

The biggest takeaways are just how common diagnostic errors were. For purposes of framing what we mean, these are diagnoses that were delayed, in which case they were missed for a period of time then eventually discovered, and then there were a few cases where the diagnosis was not suspected, and it was a fully missed diagnosis that was found out later on through testing or an autopsy or things like that.

Were you surprised when you saw the findings?

Yes, we were hoping the rate wouldn't be nearly as high as it has as it turned out to be, which is 23% or so of our patients. The other part that was surprising was the amount of harm that these errors were causing. 18% or so of all the patients in our cohort had a harm or death related to the diagnostic error.

Why is this happening?

I think that in our cohort the patients were very, very sick and I think the time frame was very, very compressed. The time available to make a diagnosis when you're already busy and maybe have other patients you're seeing or the communication has not been good — the acuity of our cohort means that a lot of these misdiagnoses were driven by illness.

Also I think we looked very carefully and systematically — maybe more deeply than others have been able to do with different data sources. I believe very strongly this is an accurate number, but people often ask, ‘why is this so high?’ and think we just frankly looked harder and more systematically.

The study explores some solutions, such as choosing better tests and improving clinician assessment. What do you think is most promising to improve this situation? 

I hate to say it's probably not going to be any single thing. I think it's going to be a number of different solutions. It's 2024 now, you can't have a discussion about healthcare without talking about AI. Certainly AI will be a big part of the future, particularly around things like generating a differential diagnosis if the patient seems to be changing or not responding to therapy correctly. There are hundreds of other opportunities and AI that will help us take these complicated data systems — now including large volumes of text —and figure out how to present information in a way that makes clinicians and patients more able to come to a diagnosis.

So you’re thinking months rather than years for this type of stuff? 

Some of it might be months. I think some of the [Large Language Models] could be months away. Certainly things like testing out simple, but very potentially very effective things like using a diagnostic timeout model, where you you stop once a day to pause and think about the patients who aren't responding to your care as they might or who are decompensating in other ways and really going through a checklist of, ‘what haven't we ordered? What should we be reconsidering? What are the diagnoses that we weren't thinking about before or had de-prioritized before that we should be moving back up the list?’ I think those are simple things that could happen almost now.

What can patients or their families do to expedite the process or get a better diagnosis?

In the inpatient setting, which is where this study happened, I think it's really about making sure you have a good conversation with your inpatient care team every day if you have questions about what the diagnosis of that day might be. In the hospital, people present with a symptom — like shortness of breath — [and] can have multiple different diagnoses. And then as you go through your hospital stay, the diagnosis becomes clear. So the communication between your team and the doctors and the nurses about what that process is looking like is very important.

I think having conversations about what could be missing is another place that patients can be involved.