by Paul Sisson
Sepsis is a deadly reaction to infection, one that can cause runaway inflammation and a cascade of organ damage. It is estimated to kill at least 350,000 Americans per year.
Early detection of this condition is key to saving lives. Medical research shows that the more quickly medical teams can detect the signs of sepsis forming, the more quickly they can start life-saving antibiotic treatments and administer intravenous fluids, supporting the body’s efforts to regain equilibrium.
A team of researchers and clinicians at UC San Diego Health has been working to see if artificial intelligence can provide an edge in early sepsis diagnosis, and, in a paper published Tuesday, the group finds that an internally developed system called “COMPOSER,” a machine-learning model trained with more than 100,000 digital records of sepsis patients, can reduce mortality rates.
Every hour, COMPOSER takes a look at the electronic health care records of UCSD emergency patients, examining myriad factors such as prescribed medications and the most recently collected vital statistics to predict who, based on its prior massive collation of health records, might be in the early stages of sepsis.
This learning algorithm’s true value, explained Dr. Gabriel Wardi, an emergency medicine and sepsis specialist and one of the paper’s co-authors, is in the gray area between sepsis and many other conditions that can mimic the condition.
“The majority of sepsis that comes through the emergency department, you could probably have a medical student or an undergraduate pick it up,” Wardi said. “They have a high fever, their heart’s beating really quickly, their blood pressure is low; you don’t need an AI algorithm to say, ‘hey, this person is septic.’”
But some cases are not so clear-cut. Patients often arrive with imprecise symptoms, such as general weakness, which could indicate early sepsis or many other medical problems.
“Oftentimes, what will happen is they’ll get an initial workup, some blood work will be done, maybe some imaging studies, but (emergency departments) are busy places, it’s difficult for providers that have done that initial assessment to say, ‘hey, let me come back and kind of put everything together,” Wardi said. “That’s the benefit of the algorithm.
“It helps our providers in these situations where there is some diagnostic uncertainty. It’s almost like a spidey sense, (saying) ‘hey, this person appears to be at a high risk of developing sepsis in the next few hours, why don’t you go take a look at them again and see if they’d benefit from fluids, from antibiotics, from cultures to see what’s going on?’”
Published in the journal npj Digital Medicine, the COMPOSER study involved a total of 6,217 emergency patients at UCSD’s two emergency departments in Hillcrest and La Jolla. Researchers compared the results of more than 5,000 emergency patients seen from Jan. 1, 2021 through Dec. 6, 2022 to those of 1,152 ER patients treated while COMPOSER was active during a nearly five-month period from Dec. 7, 2022, through April 30, 2023.
Researchers calculated a sepsis mortality rate of 9.5 percent for those overseen by COMPOSER, 1.9 percentage points lower than the predicted mortality rate of 11.39 percent. The group seen without COMPOSER active had an adjusted mortality rate of 10.3 percent.
Results, the paper acknowledges, should be taken as correlations rather than proven cause and effect because the trial was not randomized.
Dr. Karin Molander, a director of Sepsis Alliance, a nonprofit working to reduce the condition’s prevalence nationwide, reviewed the paper and said that while there is clearly deeper work still to be done, the findings are encouraging.
While she said that physicians will still want to be able to verify the facts that AI advice is based upon, Molander said the idea of having a “super smart” assistant with the time to comb through voluminous medical records and tease out trends, then compare those trends to the experiences of thousands of other patients, is exciting.
“Knowing that there is an AI in the background that does not require sleep or a bathroom break or a meal helping you kind of monitor the system, that sounds pretty good,” Molander said. “But the challenge is making sure that it does not hallucinate or, God forbid, confabulate.
“It cannot come to errant conclusions.”
Researchers calculated that the 1.9 percent drop in mortality, while small, translated to 22 patients surviving sepsis when they otherwise would have died. The impact seems to have been more visible at the UC San Diego Medical Center in Hillcrest, which was said to have “had a significant decrease in mortality” during the COMPOSER trial period while Jacobs Medical Center in La Jolla “did not observe a significant change.”
Getting observable results required significant work to tune the COMPOSER algorithm, building in a way for it to signal uncertainty and avoid sending too many false alarms to already busy caregivers. The algorithm can’t do anything on its own. All actual care ordered for patients must come from licensed medical professionals.
Co-author Shamim Nemati, a UCSD associate professor of bioinformatics with a doctorate in electrical engineering and computer science from Massachusetts Institute of Technology, said the key was finding a way to parse conditions that mimic sepsis but are actually something else.
“We had to teach the algorithm to distinguish between sepsis and the look-alikes, you know, cirrhosis of the liver, liver failure, (gastrointestinal) bleeds,” Nemati said.
Already, UCSD is expanding COMPOSER’s real-time analysis beyond emergency departments, having the system begin looking for signs of sepsis in admitted patients. And it is set to be installed in UCSD’s newest acquisition, the former Alvarado Hospital Medical Center now known as the health system’s “East Campus” on I-8 near San Diego State University.
Further revisions being made under the hood, Nemati said, include allowing the algorithm to seek more data when there is not enough for it to make a prediction.
“Because COMPOSER has a notion of its own uncertainty, it has the capability, when uncertainty is high, to ask for additional diagnostic tests,” Nemati said. “So, a nurse receives a pop up that says, ‘hey, this patient is at risk for sepsis, can you do a fresh set of vitals?”
UCSD is also experimenting with advanced wearable patient sensors that could improve the quality of the real-time data flowing into electronic patient records, providing more precise measurements to make predictions. Significant gains, the technologist said, have also been made by using large-language models such as ChatGPT, the model that has given term paper anxiety to every teacher in the world, to understand the doctor notes recorded in patient charts.
“We’re currently using those notes written at the point of ED triage to figure out what are the diagnostic hypotheses that physicians have in mind,” Nemati said. “That’s going to help us to reduce false alarms by helping to understand, you know, what are the other potential explanations for the abnormalities that we are observing in the patient data.”
UC San Diego Health has recently pushed forward on multiple fronts in the realm of artificial intelligence, hiring its first chief AI officer, partnering with Microsoft Inc. to use ChatGPT to help doctors respond to routine patient questions and pursuing a new hub to help centralize data to allow deeper integration of AI into front-line patient care.