AI system at Canadian hospital slashes unexpected deaths by 26%, delivering real-time alerts for high-risk patients

Hospitals are constantly searching for ways to quickly identify patients who might be at risk of sudden deterioration. This time gap could be the difference between saving a life or losing one. The good news is that AI is proving to be a boon in this sector. A new Canadian study has a solution - an artificial intelligence tool designed to give healthcare workers an early warning. This system, referred to as CHARTwatch, was introduced at St. Michael’s Hospital in Toronto to help doctors and nurses catch warning signs of patient decline and respond sooner.

The study found that the hospital's General Internal Medicine (GIM) unit saw a decrease in certain types of patient deaths when using this system. Specifically, the focus was on “non-palliative” deaths, meaning deaths that occur without the patient being in a state of palliative care. Palliative care is specialized treatment provided to patients with serious, often life-threatening illnesses, aimed at improving the quality of life rather than curing the condition. The goal of the AI tool was to prevent unexpected or unplanned deaths, rather than those occurring while patients are receiving palliative care. 

The study, spanning from 2016 to 2022, involved over 13,000 patient admissions in the GIM unit. The results revealed a 26% relative reduction in non-palliative deaths during this intervention period, dropping from 2.1% to 1.6%. While this percentage reduction may seem miniscule, it translates to a meaningful impact on patient outcomes. In a hospital setting, even small reductions in mortality rates could mean dozens of lives being saved over time. 

CHARTwatch sends real-time alerts to clinicians, implying that they can take prompt action when a patient shows signs of decline, or rather, sudden decline. Among high-risk patients identified by the system, non-palliative deaths decreased from 10.3% to 7.1%. Furthermore, more proactive care was provided to patients following CHARTwatch's implementation. For instance, their antibiotics and corticosteroids doses were increased accordingly, and their vitals were monitored more often than usual.

This is obviously a positive development, but researchers have included words of caution in the research paper. The study was not randomized, and other unknown factors could also have influenced the results. Additionally, the study focused on a single hospital unit, and the results could be very different in a bigger setting. Nonetheless, this study is a key form of evidence that machine learning tools will play a vital role in medicare, contributing to better patient care and hopefully saving multiple lives in the process.