While much of the public conversation around the rise of artificial intelligence has centered on its potential harms, academic and health care researchers have been quietly but aggressively finding ways to use AI to advance the practice of medicine.
Many of those projects will be years in the pipeline, owing in part to the medical research community’s rigorous standards of testing and interpreting results. Already, though, we’re beginning to see glimpses of how future AI might be used to the good–and a recent study suggests one very promising result.
In a prospective, pre-post study, published in Nature’s npj digital medicine, researchers found that an AI algorithm they’d developed can more quickly and accurately identify patients at risk for the serious blood infection sepsis than do existing protocols. The algorithm, called COMPOSER, also saves lives: a 1.9% absolute decrease in mortality corresponding to a 17% relative decrease across the two hospital emergency departments where it was used, the researchers found.
“That’s what I think people get excited about,” says Gabriel Wardi, the chief of critical care in emergency medicine at University of California San Diego Heath and senior author of the study. “Just the burden of sepsis coming down in the United States by 1.9% is huge.”
Improvements in sepsis detection
Although there’s more work to be done, the implications are significant. Sepsis, where infection triggers a chain reaction in the body that can lead to tissue damage, organ failure, and death, develops in about 1.7 million Americans each year and is linked to 350,000 deaths annually. The World Health Organization puts the global count at nearly 50 million cases and 11 million dead each year.
More to the point, one in three people who die in a U.S. hospital had sepsis during that hospitalization, according to the Centers for Disease Control and Prevention. Until now, this was largely a condition that was recognized only after its effects were beginning to be seen and felt.
“Early sepsis detection is on everyone’s mind,” says Shamim Nemati, director of predictive health analytics at UCSD and a co-author of the study. While there are some existing rule-based alerts for the infection, he says, most of them not only are slower, but suffer from a high rate of false alarms. In busy emergency departments with multiple alerts, that’s a fair concern.
The algorithm was constructed to address that. The deep-learning AI model that continuously monitors more than 150 patient variables–things like vital signs, lab results, current medications, and medical history–once that patient has been checked in. Using multiple layers of artificial neural networks, it identifies patients at high risk for sepsis while limiting false positives. “It works silently and safely behind the scenes, continuously surveilling every patient for signs of possible sepsis,” Wardi told me.
A high-risk detection triggers a notification to the ED nursing staff through the hospital’s electronic health record, and nurses can then relay the information to physicians. No alert is sent if the detected pattern can be explained by other conditions with higher confidence.
“When the algorithm sees a complex case that is unlike anything it has seen before, it says ‘I don’t know’ instead of making an ill-informed recommendation,” says researcher Aaron Boussina, a Ph.D. student in biomedical informatics at UCSD and co-author of the paper. “That is important since false alarms erode trust in the AI (model) and contribute to alarm fatigue, in which true alerts may be ignored.”
The researchers analyzed the results of more than 6,000 patient admissions before and after COMPOSER was activated at the two EDs–what is known as a pre-post study. The 1.9% absolute reduction in mortality, Wardi says, worked out to about 60 lives saved per year at the hospitals. (The system went live in December 2022.) Nationally, a nearly 2% reduction in sepsis mortality would translate to many thousands of saved lives.
The study was limited to the two San Diego facilities, and it was a non-randomized trial. Andre Holder, a physician researcher with expertise in prediction modeling using machine learning, who was not involved with the study, told me he had no major concerns with the methodology. “The authors note the limitations of the study,” Holder says, “but the study design is most fit to identify the real-world benefit of their tool.”
One concern in studies like this is what researchers call surveillance bias–looking more for something like sepsis, can find milder cases. This can make the death rates seem lower. The researchers say that during the study period, their hospitals did not experience changes in billing coding or screening practices, nor did they observe a significant increase in overall sepsis cases or the case mix, once COMPOSER was implemented.
Much more than sepsis
The use of AI in medicine is certainly not new. The National Institutes of Health sponsored the first AI in Medicine workshop at Rutgers University in 1975, and two years ago the FDA authorized the use of 91 AI-enabled medical devices, including an algorithm that helps clinicians identify collapsed lungs on imaging.
The decades in between have been punctuated by aggressive advances in AI technology in medicine. Investigators at Cedars-Sinai developed an AI tool that may help predict who will develop the most common type of pancreatic cancer, a notoriously difficult disease to forecast. They’re also using AI in brain cell modeling and Alzheimer’s research, among many other facets of medicine.
Several experts, including some at the respected New England Journal of Medicine wrote last year that they’ve been inundated by research manuscript submissions that reference some aspect of AI, adding that there is “virtually no area in medicine and care delivery that is not already being touched by AI.” Potential uses include easing administrative workflow, reducing dosage errors when patients self-administer certain medications, helping detect and track infectious diseases and estimating the 10-year cardiovascular risk in patients from a chest x-ray–it’s a nearly endless list of possibilities.
Still, a majority of Americans remain wary of AI’s inclusion in their own health care cases, its implementation in most areas is deliberate, and not all AI tools can yet be considered “medical grade.” But some AI tools have already changed the way researchers do their work, and they may aid in such critical everyday tasks as early detection of breast cancer and self-supervised learning on retinal images, which may help detect not only eye-related but other systemic ailments, such as heart failure and Parkinson’s disease.
The Biden administration has called for careful monitoring of AI’s development in medicine, but also embraced its greater possibilities. A White House memo late last year said that by one estimate, “AI’s broader adoption could help doctors and health care workers deliver higher-quality, more empathetic care to patients in communities across the country while cutting healthcare costs by hundreds of billions of dollars annually.”
Perhaps unsurprisingly, other machine-learning algorithms to test for sepsis already exist. “But only a handful of people have actually put the algorithms into clinical practice,” Wardi says–one reason why people in the field are excited about the UCSD study’s results. The authors reported that this was the first study to show improvement in patient outcomes using an AI deep-learning sepsis prediction model.
More broadly, the notion of AI as not only helpful but essential to the future practice of medicine is beginning to take root. In a world in which artificial intelligence is often associated with darker themes, this may well prove to be part of its higher use.
“I’m cautiously optimistic,” says Nemati, “that future generations will look back at this period as the miracle years of AI in Medicine.”
Carolyn Barber, M.D., is an internationally published science and medical writer and a 25-year emergency physician. She is the author of the book Runaway Medicine: What You Don’t Know May Kill You, and the co-founder of the California-based homeless work program Wheels of Change.
More must-read commentary published by Fortune:
- Glassdoor CEO: ‘Anonymous posts will always stay anonymous’
- We analyzed 46 years of consumer sentiment data–and found that today’s ‘vibecession’ is just men starting to feel as bad about the economy as women historically have
- Housing market data suggests the most optimistic buyers during the pandemic are more likely to stop paying their mortgages
- Intel CEO: ‘Our goal is to have at least 50% of the world’s advanced semiconductors produced in the U.S. and Europe by the end of the decade’
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.
