UC San Diego Health is using AI in its emergency departments to analyse EMR data to gauge a patient’s risk of developing sepsis, managing to achieve a 20 per cent reduction in mortality due to sepsis over a two-year period.
The healthcare organisation has also used generative AI in its Epic EMR in a project with Microsoft to automatically draft responses to patient queries, and is studying how ambient AI documentation can reduce “pajama time” or the after-hours documentation burden that is responsible for much of the clinician burnout crisis.
UC San Diego chief medical officer and chief clinical and innovation officer (CCIO) Chris Longhurst, who also heads up the health service’s Jacobs Center for Health Innovation, told the Australasian Institute of Digital Health’s HIC conference in Brisbane this month that it is one of the first medical centres in the US to implement AI into its electronic health record and into clinical practice.
“It’s been 25 years in the United States since we published To Err is Human and this was really a landmark publication, because, building on some work that had been done before, it suggested that 44,000 to 88,000 Americans die every year from medical errors, or a 747 worth of people going down on a daily basis.
“It was followed shortly by Crossing the Quality Chasm in 2001 which suggested that digitisation, electronic health records and computerised prescriber order entry would be part of the solution for reducing these medical errors. So we’ve spent two decades putting in EHRs, and we’ve solved all these problems, right? No more patient safety issues, right?
“One in four patients discharged from a Boston area hospital over a period of study experienced an adverse event, and at least one in four of those were preventable. So here we’ve had 25 years of efforts around safety, and we have got to acknowledge they are not working.”
One example is an early use of AI for chest x-rays that was put into production just as the Covid pandemic arrived in the US. A UC San Diego radiologist called Albert Hsiao, who studies machine learning, got his hands on a chest Xray set from Wuhan, and trained his machine learning algorithms to identify covid pneumonia.
Dr Longhurst said that on Friday, March 13, 2020, when the White House was declaring the pandemic, he was sitting in a room with Dr Hsiao and the hospital’s cloud team and clinical research IT team, talking about how to move this capability into production. Rather than the normal 10 or 12 months, it went live in two weeks.
A survey showed that doctors using the AI thought it was helping them make an earlier diagnosis than they would have otherwise, with one in five clinicians saying it helped them make a different clinical decision.
While this was an early study, it still showed up as one of only four in a review of clinical AI models published in 2022 that showed or even looked at clinical impact.
“To me, this really shows a big gap, which is we are studying a lot of AI models, but we’re not studying enough clinical impact and what matters to the patients and the families that we serve.”
This has led UC San Diego to develop a new role at the health service: a chief health AI officer in Karandeep Singh, who started in January this year. UC San Diego joins UC San Francisco Health and UC Davis Health, which implementing a chief health AI officer role in their health systems in 2023.
Dr Singh this year tweeted about what he calls the AI paradox, which is that “there’s a huge gap where the models we research are rarely implemented and the models that are implemented are rarely researched”. Dr Longhurst said Dr Singh’s role as chief health AI officer is to help close that gap.
AI for patient flow
UC San Diego subsequently began using AI to help with patient flow in the hospital, in particular how to deal with the problem of emergency department boarding, and how to make different operational decisions.
It used a quite simple open source algorithm that is able to forecast demand, and is now outperforming the old Excel spreadsheet formulas that had the job for so long.
“It’s gotten so helpful that we can now look 24, 48 and even 72 hours out and make operational decisions that are different,” Dr Longhurst said.
“We’ve now integrated these type of machine learning models into over two dozen of our reports and dashboards to help us forecast, simulate and make better operational decisions.”
More importantly, it is being used for an area of real impact: sepsis. Sepsis kills about 350,000 Americans every year and about 3000 in San Diego county alone. Australian Commission on Safety and Quality in Healthcare figures show that about 55,000 Australians are diagnosed with sepsis and 8700 die on an annual basis.
Diagnosing sepsis is a particularly difficult thing to do, however. “85 per cent of sepsis develops at home, and so the emergency department is where some of the most critical diagnoses are made, and that’s where we started our journey on the AI front,” he said.
EMRs do have tools that help health services create their own sepsis prediction methods, but it doesn’t seem to be improving things much. However, Dr Singh and his colleagues looked at half dozen hospitals that implemented one of these tools and found that often, these models were built off data from four to six hours prior to the diagnosis of sepsis.
“But there was clinical suspicion. And so if your clinician is ordering a lactate because they’re concerned about the possibility of sepsis, and then a lactate order goes into the model for predicting sepsis, you’re now firing an alert using AI that suggests that you might want to think about sepsis in your patient for whom you’ve just ordered lactate, which isn’t terribly helpful.”
Dr Longhurst’s team then recruited a data scientist well known in the field of sepsis AI to work with Dr Singh’s team and its medical director of sepsis outcomes Gabe Wardi to work out how AI could be used to make transformative improvements in sepsis care.
The looked at ways of centralising potential sepsis warning signs and alerts and did some workflow redesign. Dr Longhurst said he believed 50 per cent of outcomes are due to the algorithm, but the other 50 per cent are due to integration into clinical workflow processes.
“We studied this for two years and over that two-year period, we found an almost 20 per cent reduction in mortality due to sepsis in emergency departments. That translates to 50 lives saved every year just at our two hospitals in San Diego.”
It also showed that the AI was able to help reduce health equity problems by being more useful in early warnings of sepsis in UC San Diego’s Hillcrest Medical Center facility, which tends to a lower socio-economic group than its fancier Jacobs Medical Center in La Jolla.
“Our clinicians are pretty good at diagnosing sepsis in cancer patients, but on average, our patients in Hillcrest were younger, presented with different symptoms, often HIV positive or suffering from other immunocompromised diseases, and so the impact of the lives saved was actually greater in that hospital that had more health equity gaps.
“I think this is really important, because locally developed AI can help to solve problems in your local health system, where you have health equity opportunities, where you have these gaps and want to make a difference.”
It is with this idea of impact at scale that UC San Diego founded the Center for Health Innovation about three or four years ago, with a large donation helping to build a mission control centre to help it better manage patient flow.
Dr Longhurst said this mission control centre will use AI algorithms to look not just at inpatients but patients being monitored at home as part of its chronic disease management program.
Arrival of ChatGPT
Dr Longhurst remembers the arrival of ChatGPT in November 2022 and marvelling at the potential impact it could have on healthcare. Others did too, including a sub Reddit that was using a chat bot to answer patient queries.
As described in a paper in JAMA Internal Medicine, on average, the chat bot appeared to produce higher quality and more empathetic responses than human doctors.
“The click bait headline said chat bots smarter and better, but that wasn’t my takeaway,” Dr Longhurst said.
“My takeaway was that this was all about time. Even though AI was putatively blinded in ranking all of these questions and responses, it was really obvious which one was which, because the doctor’s answers were always three or four sentences, and the chat bot answers were three or four paragraphs.”
This led to the idea of using GPT in the hospital’s Epic EMR. Partnering with Microsoft, UC San Diego then became the first in the country to Implement a large language model in production to help solve a real problem. It involves pyjama time.
“This is the problem: particularly since the pandemic, messages from patients to our clinicians have skyrocketed, and it’s unreimbursed care,” Dr Longhurst said. “Of course, our clinicians want to help answer those questions between visits, but we don’t have a way to bill for them effectively, and so for the doctors, that’s extra work after hours.”
A pilot was set up to test whether AI could automate these replies by generating draft messages, offering clinicians the option of either starting with the AI generated draft, or start their own blank reply. If the Start with Draft button was chosen, a statement was appended to every message saying part of it had been generated automatically and had reviewed and edited by your doctor.
The pilot ran over a summer and was hugely oversubscribed, as all of the doctors wanted to use the technology, but while Dr Longhurst had hypothesised that it would save time for doctors, it ultimately did not. What it did do was improve patient satisfaction.
“It turns out that before the AI generated drafts, it would take about 30 seconds to read a short patient message, figure out your macro type of response and hit sent. After the AI generated drafts, it took about 30 seconds to read the short message, read the three paragraph response, make a few edits and click send.
“So it was not a time saver in our study. We’re doing follow up work now that is confirming what we suspected, which is they’re perceived by patients as being higher quality. In fact, we’ve got anecdata suggesting it’s helping avoid visits in some cases, and our doctors told us it lowered cognitive burden.”
Stanford University ran a parallel study measuring burnout, and saw a decrease in physician burnout.
“Why? Well, we all know it’s a lot easier to edit somebody else’s first draft than it is to start with that blank screen. And so lowering cognitive burden and decreasing that contribution to burnout is an important outcome. Based off of that, we’ve decided to expand the pilot.”
UC San Diego is also looking deeply at ambient listening and the use of AI scribes, not just to improve documentation but to actually improve consultations through what Dr Longhurst called a “re-humanisation” of the exam room.
Kaiser Permanente is now rolling out ambient AI documentation technology across its network, which is finding that there is indeed a reduction in pyjama time, or the after hours work that interrupts family time for clinicians who take their onerous documentation work home with them.
But it also found that there was a reduction in time taking notes during the actual appointment too. “This is our opportunity to re-humanise the exam room visit,” he said.
Promise of AI
While Dr Longhurst said he believes the promise of AI in the next two to three years is “probably overhyped”, the promise in the next seven to 10 years is actually under-appreciated.
“I absolutely believe this is the most important tool to come to medicine since antibiotics,” he said. “It is going to have a huge impact on healthcare delivery in those systems that have the digital infrastructure ready to leverage it.
“In fact, I am on record as saying that AI documentation, AI scribes, will be the standard of care in the next several years.
“It will be based off small language models, because I don’t think everything that we record and examine needs to go to the cloud, nor does it need GPT4 to help transcribe those notes.
“But the main prediction here is really important, is that I think that this will be one of the main tools for diagnostic decision support.”
What a great write-up! This was one of my favourite keynotes from the #HIC24 conference. Thank you for making it accessible for everyone who wasn’t able to attend and hear this one in person
Completely agree Simon. Make sure you tune in to Pulse💗podcast this week for a very cool chat with Chris.
I took notes of Chris’s talk at the conference but this summary by Kate far exceeds mine. Really helpful Kate! It was a brilliant talk by Chris.