AI ‘early warning’ system shows promise in preventing hospital deaths, study says

An AI early-warning system designed to predict which hospital patients are at risk of deterioration has been linked to a reduction in unexpected deaths, according to a new study.

The research, published Monday in the Canadian Medical Association Journal, reported a 26% decrease in non-palliative deaths among patients in the general internal medicine unit at St. Michael’s Hospital when the AI tool was utilized.

“We’ve seen a lot of excitement around artificial intelligence in medicine, but there’s been less real-world deployment of these tools,” said Dr. Amol Verma, lead author of the study, who is a general internal medicine specialist and scientist at the Toronto hospital. “This is one of the first examples of an AI tool being rigorously tested and showing potential to improve patient care,” added Verma, who is also a professor of AI research and education in medicine at the University of Toronto.

The AI technology, known as CHARTwatch, continuously monitored more than 100 different data points for each patient. When the system detected that a patient’s condition was worsening, it alerted doctors and nurses to intervene quickly.

“The tool uses information already routinely collected in a patient’s electronic medical record,” Verma explained, referring to factors like age, medical history, vital signs, blood pressure, heart rate, and lab results. The AI system updated its predictions hourly, based on changing patient data.

If the clinical team agreed with the AI's assessment, they took appropriate action, such as transferring the patient to intensive care, administering antibiotics for infections like sepsis, or increasing the frequency of monitoring. For patients nearing the end of life, the AI could also help provide earlier palliative care to ease their suffering, Verma said.

“Importantly, the AI doesn’t dictate specific treatments. That’s still left to the judgment of the doctors and nurses,” he emphasized. “It simply signals, ‘Pay attention to this patient.’”

In a busy hospital environment, where healthcare workers are responsible for numerous patients undergoing various tests and treatments, these alerts are valuable. “It’s impossible for humans to constantly monitor 20 or 30 patients at once,” Verma said.

Co-senior author Muhammad Mamdani noted that AI can process large amounts of data, and combining that with human clinical judgment can lead to better outcomes. He advised clinicians to trust their instincts if they believe a patient is in danger, even if the AI indicates otherwise. However, if the AI predicts a serious outcome that the clinician does not anticipate, he urged them to trust the AI.

The study compared non-palliative death rates in the general internal medicine unit from Nov. 1, 2020, to June 1, 2022, when the AI tool was in use, to a previous period from Nov. 1, 2016, to June 1, 2020, before the AI was implemented. The non-palliative death rate decreased from 2.1% to 1.6% during the AI period.

To rule out other factors that could influence the results, the researchers examined cardiology, respirology, and nephrology units, which did not use the AI tool. These units showed no change in non-palliative deaths between the two time periods. The study also controlled for variables such as patient age and excluded COVID-19 data to account for potential pandemic-related impacts.

Overall, the study analyzed 13,649 admissions in the general internal medicine unit and 8,470 admissions in the comparison units.

While the findings are promising, Verma cautioned that they should be viewed with care and that a randomized controlled trial is needed to validate the AI tool’s effectiveness, similar to how medications are tested.

Ross Mitchell, a professor at the University of Alberta who was not involved in the study, described the results as “very encouraging” but emphasized the need for further research. “This specific technology, CHARTwatch, should be studied in more hospitals across Canada to understand its broader impact,” said Mitchell, who holds the Alberta Health Services Chair in AI in health.