This article was originally published in Undark Magazine.

Ten years ago, 12-year-old Rory Staunton dove for a ball in gym class and scraped his arm. He woke up the next day with a 104-degree Fahrenheit fever, so his parents took him to the pediatrician and eventually the emergency room. It was just the stomach flu, they were told. Three days later, Rory died of sepsis after bacteria from the scrape infiltrated his blood and triggered organ failure.

“How does that happen in a modern society?” his father, Ciaran Staunton, asked me.

Each year in the United States, sepsis kills more than a quarter million people—more than stroke, diabetes, or lung cancer. One reason for all this carnage is that if sepsis is not detected in time, it’s essentially a death sentence. Consequently, much research has focused on catching sepsis early, but the condition’s complexity has plagued existing clinical support systems—electronic tools that use pop-up alerts to improve patient care—with low accuracy and high rates of false alarm.

That may soon change. Back in July, Johns Hopkins researchers published a trio of studies in Nature Medicine and npj Digital Medicine showcasing an early-warning system that uses artificial intelligence. The system caught 82 percent of sepsis cases and significantly reduced mortality. While AI—in this case, machine learning—has long promised to improve health care, most studies demonstrating its benefits have been conducted using historical data sets. Sources told me that, to the best of their knowledge, when used on patients in real time, no AI algorithm has shown success at scale. Suchi Saria, the director of the Machine Learning and Healthcare Lab at Johns Hopkins University and the senior author of the studies, said in an interview that the novelty of this research is how “AI is implemented at the bedside, used by thousands of providers, and where we’re seeing lives saved.”

The Targeted Real-Time Early Warning System scans through hospitals’ electronic health records—digital versions of patients’ medical histories—to identify clinical signs that predict sepsis, alert providers about at-risk patients, and facilitate early treatment. Leveraging vast amounts of data, TREWS provides real-time patient insights and a unique level of transparency in its reasoning, according to the Johns Hopkins internal-medicine physician Albert Wu, a co-author of the study.

Wu says that this system also offers a glimpse into a new age of medical electronization. Since their introduction in the 1960s, electronic health records have reshaped how physicians document clinical information; nowadays, however, these systems primarily serve as “an electronic notepad,” he added. With a series of machine-learning projects on the horizon, both from Johns Hopkins and other groups, Saria says that using electronic records in new ways could transform health-care delivery, providing physicians with an extra set of eyes and ears—and helping them make better decisions.

It’s an enticing vision, but one in which Saria, the CEO of the company developing TREWS, has a financial stake. This vision also discounts the difficulties of implementing any new medical technology: Providers might be reluctant to trust machine-learning tools, and these systems might not work as well outside controlled research settings. Electronic health records also come with many existing problems, from burying providers under administrative work to risking patient safety because of software glitches.

Saria is nevertheless optimistic. “The technology exists; the data is there,” she says. “We really need high-quality care-augmentation tools that will allow providers to do more with less.”

Currently, there’s no single test for sepsis, so health-care providers have to piece together their diagnoses by reviewing a patient’s medical history, conducting a physical exam, running tests, and relying on their own clinical impressions. Given such complexity, over the past decade, doctors have increasingly leaned on electronic health records to help diagnose sepsis, mostly by employing a rules-based criteria—if this, then that

These alerts also appear on providers’ computer screens as a pop-up, which forces them to stop whatever they’re doing to respond. So, despite these rules-based systems occasionally reducing mortality, there’s a risk of alert fatigue, where health-care workers start ignoring the flood of irritating reminders. According to M. Michael Shabot, a surgeon and the former chief clinical officer of Memorial Hermann Health System, “It’s like a fire alarm going off all the time. You tend to be desensitized. You don’t pay attention to it.”

Read: The burnout crisis in American medicine

Already, electronic records aren’t particularly popular among doctors. In a 2018 survey, 71 percent of physicians said that the records greatly contribute to burnout, and 69 percent said that they take valuable time away from patients. Another 2016 study found that, for every hour spent on patient care, physicians have to devote two extra hours to electronic health records and desk work. James Adams, the chair of the Department of Emergency Medicine at Northwestern University, calls electronic health records a “congested morass of information.”

But Adams also says that the health-care industry is at an inflection point to transform the files. An electronic record doesn’t have to simply involve a doctor or nurse putting data in, he says; instead, it “needs to transform to be a clinical-care-delivery tool.” With their universal deployment and real-time patient data, electronic records could warn providers about sepsis and various other conditions—but that will require more than a rules-based approach.

What doctors need, according to Shabot, is an algorithm that can integrate various streams of clinical information to offer a clearer, more accurate picture when something’s wrong.

Machine-learning algorithms work by looking for patterns in data to predict a particular outcome, like a patient’s risk of sepsis. Researchers train the algorithms on existing data sets, which helps the algorithms create a model for how that world works and then make predictions on new data sets. The algorithms can also actively adapt and improve over time, without the interference of humans.

TREWS follows this general mold. The researchers first trained the algorithm on historical electronic-records data so that it could recognize early signs of sepsis. After this testing showed that TREWS could have identified patients with sepsis hours before they actually got treatment, the algorithm was deployed inside hospitals to influence patient care in real time.

In these studies, TREWS monitored patients in the emergency department and inpatient wards, scanning through their data—vital signs, lab results, medications, clinical histories, and provider notes—for early signals of sepsis. (Providers could do this themselves, Saria says, but it might take them about 20 to 40 minutes.) If the system suspected organ dysfunction based on its analysis of millions of other data points, it flagged the patient and prompted providers to confirm sepsis, dismiss the alert, or temporarily pause the alert.

“This is a colleague telling you, based upon data and having reviewed all this person’s chart, why they believe there’s reason for concern,” Saria says. “We very much want our frontline providers to disagree, because they have ultimately their eyes on the patient.” And TREWS continuously learns from these providers’ feedback. Such real-time improvements, as well as the diversity of data TREWS considers, are what distinguish it from other electronic-records tools for sepsis.

Saria says that TREWS’s high adoption rate shows that providers will trust AI tools. But Fei Wang, an associate professor of health informatics at Weill Cornell Medicine, is more skeptical about how these findings will hold up if TREWS is deployed more broadly. Although he calls these studies first-of-a-kind and thinks their results are encouraging, he notes that providers can be conservative and resistant to change: “It’s just not easy to convince physicians to use another tool they are not familiar with,” Wang says. Any new system is a burden until proven otherwise. Trust takes time.

TREWS is further limited because it only knows what’s been inputted into the electronic health record—the system is not actually at the patient’s bedside. As one emergency-department physician put it, in an interview for the third study, the system “can’t help you with what it can’t see.” And even what it can see is filled with missing, faulty, and out-of-date data, according to Wang.

The most impressive aspect of TREWS, according to Zachary Lipton, an assistant professor of machine learning and operations research at Carnegie Mellon University, is not the model’s novelty, but the effort it must have taken to deploy it on 590,736 patients across five hospitals over the course of the study. “In this area, there is a tremendous amount of offline research,” Lipton says, but relatively few studies “actually make it to the level of being deployed widely in a major health system.” It’s so difficult to perform research like this “in the wild,” he adds, because it requires collaborations across various disciplines, from product designers to systems engineers to administrators.

As such, by demonstrating how well the algorithm worked in a large clinical study, TREWS has joined an exclusive club. But this uniqueness may be fleeting. Duke University’s Sepsis Watch algorithm, for one, is currently being tested across three hospitals following a successful pilot phase, with more data forthcoming. In contrast with TREWS, Sepsis Watch uses a type of machine learning called deep learning. Although this can provide more powerful insights, how the deep-learning algorithm comes to its conclusions is unexplainable—a situation that computer scientists call the black-box problem. The inputs and outputs are visible, but the process in between is impenetrable.

Wang suggests that that’s a dangerous conclusion. “How can you confidently say your algorithm is accurate?” he asks. After all, it’s difficult to know anything for sure when a model’s mechanics are a black box. That’s why TREWS, a simpler algorithm that can explain itself, might be a more promising approach. “If you have this set of rules,” Wang says, “people can easily validate that everywhere.”

Indeed, providers trusted TREWS largely because they could see descriptions of the system’s process. Of the clinicians interviewed, none fully understood machine learning, but that level of comprehension wasn’t necessary.

Shariat also worries that the sheer volume of alerts, with about two out of three being false positives, might contribute to alert fatigue—and potentially overtreatment with fluids and antibiotics, which can lead to serious medical complications such as pulmonary edema and antibiotic resistance. Saria acknowledges that TREWS’s false-positive rate, although lower than that of existing electronic-health-record systems, could certainly improve, but says it will always be crucial for clinicians to continue to use their own judgment.