AI-Powered Tool Reduces Hospital Death Risk by 26%, Study Reveals

AI-powered tech could reduce the risk of death in hospitalized patients by 26%, a new study by Unity Health Toronto, ICES, and the University of Toronto (TO) has revealed.

The study, published in the Canadian Medical Association Journal, evaluated the performance of CHARTWatch, an AI-powered early warning system developed by the general internal medicine (GIM) ward at St. Michael’s Hospital (Toronto, Canada). The results suggested that the tool could prevent unexpected deaths by identifying hospitalised patients who are likely to experience further health complications.

Methodology and Results

The study included 13,649 GIM patients aged 55-80, with more than 9,000 in the pre-intervention period and over 4,000 using CHARTWatch. Additionally, 8,470 patients in subspecialty units that did not utilize the AI-based system were also included in the research. The objective of the study was to evaluate CHARTWatch’s effectiveness by comparing GIM’s patient death rate to that of other subspecialty units not making use of the AI tool.

CHARTWatch provided physicians with twice-daily real-time alerts and sent daily patient updates to the palliative care team. In addition, the system introduced a specialized care strategy for high-risk patients, which included extensive patient checkups, improved nurse-physician communication channels, and alerts for doctors to reassess patients.

Throughout the 19-month intervention period, nearly 500 GIM patients were identified as high-risk, contrasting with 1,656 high-risk patients in the 43-month pre-intervention period. Notably, GIM experienced a lower rate of non-palliative deaths compared to the pre-intervention group (1.6% vs. 2.1%), highlighting the tool’s potential impact on patient outcomes in complex care settings.

How does CHARTWatch work?

CHARTWatch is a machine learning model that can predict patient deterioration using real-time data from electronic medical records, communicating predictions to physicians and nurses and providing a clinical pathway for high-risk patients regularly monitored by a multidisciplinary implementation team.

The model is time-aware, which means it can add predictions in early risk scores, changes in risk scores since the last assessment, and summaries of risk score changes over time. Patients were divided into three risk groups to help clinicians better understand the danger.

Predictions were communicated to physicians and charge nurses through text messaging and email three times a day. Updates subsequently increased to hourly in January 2021.

Image: Clinical outcomes in the general internal medicine (GIM) and subspecialty cohorts

Source: “Clinical evaluation of a machine learning–based early warning system for patient deterioration”; Canadian Medical Association Journal

Could AI Be a Patient Lifesaver?

CHARTWatch’s rigorous strategic protocol appeared to be key in ensuring that high-risk patients participating in the research were provided with strengthened healthcare services. Following these promising results, the authors of the study are optimistic regarding CHARTWatch’s capability to improve patient health and prevent premature deaths.

Lead author Dr Amol Verma, a clinician-scientist at St. Michael’s Hospital and Temerty professor of AI research at the University of Toronto, stated: “Our findings suggest that AI-based early warning systems are promising for reducing unexpected deaths in hospitals.”

“At the same time period in the other units in our hospital that were not using CHARTWatch, we did not see a change in these unexpected deaths,” he added. “That was a promising sign.”

Prospects and Challenges

In recent years, the early implementation of AI in healthcare has shown promise in preventing and treating different types of health conditions such as cancer, gum diseases and depression.

Prospects are similarly bright when it comes to exploring AI’s capability to reduce the risk of death in hospitalised patients. In this regard, Dr Verma highlighted CHARTWatch’s capability to successfully complement human clinical judgement as well as its remarkable potential to save real lives.

However, Verma also acknowledged the current limitations of the research. “Our study was not a randomized controlled trial across multiple hospitals. It was within one organization, within one unit”, he said. “So, before we say that this tool can be used widely everywhere, I think we do need to do research on its use in multiple contexts.”