By Dorothy Brooks
New data show that an artificial intelligence (AI)-driven tool, deployed in UC San Diego Health emergency departments (EDs), can not only pick up on the subtle signs of sepsis at an early stage, but the data also strongly suggest that use of the tool can significantly reduce mortality — a big win in the battle against a disease that affects 1.7 million American adults each year, leading to at least 350,000 deaths, according to the data from the Centers for Disease Control and Prevention.1
The tool, dubbed Composer, was developed by researchers at UC San Diego (UCSD) School of Medicine, with the ED being the primary target for initial implementation, explains Gabriel Wardi, MD, MPH, FACEP, chief of the Division of Critical Care in the Department of Emergency Medicine at UCSD School of Medicine and one of the authors of the research. “Because the majority of sepsis patients come through the emergency department, that’s where we thought the biggest bang for our buck would be,” he observes.
To measure the impact of the tool, the researchers analyzed more than 6,000 patients admitted for sepsis at two of the health system’s EDs between January 2021 and April 2023. Comparing the outcomes of patients during a 705-day pre-intervention period with the outcomes of patients in a 145-day post-intervention period, the researchers found that use of the AI algorithm was associated with 17% reduction in in-hospital sepsis mortality and a 10% increase in sepsis bundle compliance — an indication that the tool helped clinicians accelerate evidence-based care on the recommended timetable.2
Notably, this is the first study to show a boost in patient outcomes through the use of an AI deep-learning model, according to the researchers. “Most [researchers] have done something with retrospective data, but very few have actually tested [such models] in a prospective fashion to show either an improvement or a lack of improvement in patient-centered outcomes,” states Wardi.
The AI algorithm is integrated into the electronic medical record (EMR) so that it can begin to automatically monitor patient data as soon as individuals present to the ED. Researchers note that the tool is looking at more than 150 different patient variables that could be linked to sepsis from a range of sources, including lab results, vital sign trends, medical history, demographic data, and current medications.
If the tool’s algorithm determines from these variables that a patient is at high risk for a sepsis infection, it notifies the ED nurse through the EMR with a practice advisory. “We chose [to notify] our nurses in the ED rather than our physicians because the nurses are oftentimes the ones that have the most contact with patients,” states Wardi. Further, while emergency physicians may be taking care of more than a dozen patients, the ED nurses at UCSD are capped at four patients, he says.
When such an advisory is received by an ED nurse, he or she then can discern whether the information needs to be immediately relayed to the physician in charge of the patient’s care, explains Wardi. For instance, there may be some cases where a workup for sepsis is already in process, or there may be other factors that point in a different direction. However, in cases where sepsis is a concern, the tool provides the nurses with something objective that they can send to the emergency physician in the form of a direct message that can be viewed in the EMR.
Along with the best practice advisory that goes to the nurse, the physician does receive a much more subtle indication that a patient has crossed the threshold in the algorithm indicating that he or she is at high risk for sepsis. “There is a little blue cross that pops up [in the EMR],” says Wardi, noting that when developing their approach for implementation, investigators felt it would be more effective to target the nursing staff with the practice advisory, in part, because physicians often receive so many of these pop-up notifications.
Wardi emphasizes that when developing the tool, the goal was not to key in on the slam-dunk cases of sepsis risk that are picked up easily by clinicians, but rather on the cases where there is diagnostic uncertainty about what could be going on. As an example, he describes the case of a patient with symptoms that could be suggestive of heart failure, pancreatitis, dehydration, or maybe an infection. Then lab results start coming in, and vital signs begin to exhibit concerning trends. “If the [AI model], that’s continually ingesting data, crosses the threshold [indicating a risk for sepsis], then it will immediately send something to the nurse in the form of a best practice advisory,” he explains.
Alternatively, Wardi notes that if a patient comes in who is febrile, tachycardic, and hypotensive at triage, you do not need a machine learning model or anything particularly fancy to pick up on sepsis risk. “We chose not to focus on those cases [with the tool] because there’s not much added benefit once the treatment team has already determined they have a pretty clear case of sepsis,” he says.
Wardi notes that the model went through rigorous testing before implementation. “What we did is something called a silent trial where we basically had the model running in real time, but the providers and nurses were not seeing the results,” he says. “Then, over a nine-month period, we looked at a bunch of cases to find when the model did well and when the model misfired so that we could further tweak it [prior to going live].”
Investigators also integrated input from bedside nurses in the development of the practice advisory that pops up when the algorithm deems a patient is at risk for sepsis. “Their feedback was really instrumental in making sure it was something that was useful,” says Wardi.
Nurses, in fact, feel empowered by the tool, and physicians have been receptive to the process as well, states Wardi, although he acknowledges there were a handful of individuals who initially expressed concerns about relying on a model to determine how they make care decisions. “I explained to them that this is not meant to replace any of their medical decision-making; it’s meant to be almost a second set of eyes that’s looking over your shoulder saying this person is at risk,” he says. “It doesn’t mean you start treatment, but you at least consider it … because oftentimes in a busy emergency department, you start treating someone and going down a certain pathway. Pulling back and kind of redirecting the ship, so to speak, is something that doesn’t always happen enough.”
With the Composer tool now in use in all of the health system’s EDs, the next step involves expanding it to the inpatient units. “In most places, and we’re no different, outcomes tend to be worse for patients who develop sepsis in the hospital, so hopefully by using this tool, we can catch these patients as they start to decompensate a little bit sooner,” says Wardi.
Investigators also hope to eventually test the tool on a larger scale. “Expanding it to see if other sites are able to achieve similar results is a big goal of ours,” he observes.
- Centers for Disease Control and Prevention. Sepsis. What is Sepsis? https://www.cdc.gov/sepsis/wha...
- Boussina A, Shashikumar SP, Malhotra A, et al. Impact of a deep learning sepsis prediction model on quality of care and survival. NPJ Digit Med 2024;7:14.
