Study Finds AI-Based Early Warning System Cut Hospital Deaths by 26%

AI machine learning helped predict when patients’ conditions would deteriorate ahead of time, warning medical staff about the potential risks.

Increased use of early warning alert systems powered by artificial intelligence (AI) at hospitals and other medical facilities may help save more patients’ lives, according to the findings of a new study.

Despite concerns over the ethical use of AI applications, a study published on September 16 in the Canadian Medical Association Journal (CMAJ) indicates early warning health systems that use the predictive technology reduced the number of hospital deaths by more than a quarter.

A lot of negative media attention has been centered on emerging AI technology in recent months, in the context of inaccurate responses, stolen intellectual property, and the potential for lost jobs across many industries. However, researchers from St. Michael’s Hospital and Unity Health Toronto in Canada, led by Dr. Amol A. Verma, indicate that the use of AI to provide early warnings to medical providers may have a profound impact on the overall quality of care provided to hospitalized patients and improve outcomes.

Verma’s team looked at data on nearly 14,000 patients treated in the general internal medicine (GIM) unit at St. Michael’s from November 2020 to June 2022, and compared them to more than 8,000 patients treated before the AI intervention period, from November 2016 to June 2020.

The AI system, known as Chartwatch, calculated health predictions for patients with heart, lung and kidney conditions, by measuring 100 inputs from patient medical records. The inputs are routinely checked during the daily process of care and include measures like vital signs, heart arteries, blood pressure and other lab tests.

However, instead of waiting for nurses to take measurements and notice patterns that might lead to worsened conditions, Chartwatch monitors changes in the medical record and makes predictions every hour.

The AI program attempts to predict whether the patient is likely to get worse in the near future. It then sends alerts to nurses monitoring the patients remotely and on-the-ground intervention rapid response teams.

In the AI predictive group, 1.6% of patients died compared to 2.1% of patients in the pre-intervention group.

The Chartwatch early warning system reduced unexpected deaths by 26% and reduced deaths overall by 16%, the researchers concluded.

The data also showed Chartwatch helped to reduce deaths among high-risk patients. About 7% of patients in the high-risk AI group died compared to 10% in the non-AI group.

Healthcare providers’ failure to recognize when patients begin to deteriorate is the leading cause of unplanned transfers to the intensive care unit (ICU), the researchers noted. It’s also linked to longer hospital stays and higher death rates compared to transfers from the emergency room.

Researchers hope that using the AI alert system can help predict when patients begin to deteriorate and reduce patient deaths.

“Implementing a machine learning–based early warning system in the GIM unit was associated with lower risk of non-palliative death than in the pre-intervention period,” Verma said. “Machine learning–based early warning systems are promising technologies for improving clinical outcomes.”