Beyond the Noise: Clinical Alarm Fatigue Reduction
Improving patient care by reducing nuisance alarms with AI.
Project Overview
Updated September 19, 2024
The Problem
Excessive false clinical alarms are one of the top safety issues for hospital patients. In a typical hospital setting, a nurse may deal with up to 350 clinical alarms a day for each patient for vitals monitoring1. A significant portion of these alarms are unnecessary false alarms, also known as “nuisance alarms”. Studies estimate the number of nuisance alarms to be 80% to 95% of all clinical alarms2.
The impact of nuisance alarms is well-recognized as a cause of patient anxiety, sleep deprivation and delirium, which can be causes of prolonged hospitalization and patient morbidity. They also desensitize clinicians, waste valuable clinical time and cause clinical staff to get irritated and burn out.
Across all standard vital signs, the worst excessive alarm offender is pulse oximetry blood oxygen saturation, or SpO2. In one pediatric ICU study, 44% of all clinical alarms were from SpO2 alarms, and only 7% of these were clinically useful3. Finding a way to reduce the number of SpO2 nuisance alarms would reduce clinician alarm fatigue and, ultimately, patient harm.
How We Are Solving It
Led by Medtronic Canada in collaboration with Excelar, Providence Health Care Ventures, and the University of Toronto; this project will advance a machine learning (ML) algorithm to reduce nuisance SpO2 alarms by at least 40%.
Medtronic’s ML algorithm can reliably predict what alarms are actionable versus unnecessary and can be safely silenced. While other commercial solutions rely on delaying or managing alarms by moving the alarm away from the patient and onto mobile devices that clinicians carry around, they do not actually eliminate nuisance alarms. An example of an actionable alarm is when a patient is in respiratory distress (cannot breathe properly) causing the SpO2 to drop under a threshold. An example of a clinically irrelevant nuisance alarm is a SpO2 drop below the pre-set device threshold (e.g. 90%) for a very short period of time. This could be caused by a rapid arm or hand movement, when the SpO2 sensor has been temporarily removed and repositioned, or when the patient is hovering around an SpO2 value just above the threshold.
This project will validate and refine the algorithm in a real-world care environment using live sensors and clinical systems data. Providence Health Care Ventures and Excelar’s CareFlow will extend their current systems into a standardized platform that will make it faster and more cost effective to develop, test, contextualize, validate, implement, and observe/evaluate the performance of AI models and algorithms under real world clinical testing conditions for health tech developers. The Medtronic AI algorithm will be the first such module validated using this platform. The University of Toronto will also work to refine and improve the precision of the alarm algorithm by applying other ML and alarm personalization, where alarms are built for specific types of patients, to provide a more personalized healthcare approach.
The software’s architecture also has the potential to be easily expanded to other SpO2 vendors and alarms for other vitals, like respiratory rate and blood pressure.
This resulting solution will aim to greatly reduce problems associated with nuisance alarms in hospitals and provide a blueprint to leverage the extraordinary potential of AI to improve care.
Project Lead
Project Partners
-
“We are delighted to be the health partner in this project and the potential impact it will have for consortium members in the near future, and for the creation, adoption, and refinement of AI in actual health care operations. This is something critical to the decade ahead as we grapple with demographic changes, resource challenges, and increasingly complex health needs.”
President, Providence Health Care Ventures