Research led by Stanford University shows an artificial intelligence (AI) driven model is able to accurately predict hospitalized patients who are likely to deteriorate, allowing them to get extra care more quickly to try and prevent this outcome.
Writing in JAMA Internal Medicine, Robert Gallo, a research fellow at Stanford, and colleagues found the AI intervention was able to reduce the risk of emergency outcomes for patients by 10.4 percentage points compared with the average risk for deterioration in vulnerable patients not assessed by the tool.
When hospitalized patients suddenly deteriorate, they are at high risk for negative outcomes such as admission to intensive care units, cardiac arrest or similar. Being able to quickly predict who is likely to deteriorate and when could lead to much better outcomes for patients, as well as saved healthcare costs.
“Clinicians often care for many hospitalized patients concurrently and may not recognize early signs preceding a patient’s clinical deterioration… Automated early warning scores help alert clinicians to impending patient clinical deterioration so that preventive or rescue actions can be taken to avoid adverse outcomes,” explained Gallo and colleagues.
“Rich real-time patient electronic health record data (eg, vital signs, diagnoses, laboratory results, nursing flowsheets) can be used to create predictive models… Given the ease of training and integrating these models, these early warning scores have been quickly and widely implemented,” they added. However, studies to test the accuracy and efficacy of these models are limited.
Gallo and team assessed whether an AI-driven tool could help physicians and other healthcare professionals to predict patient deterioration quickly enough to try and prevent it. The cohort study included 9938 patients who were hospitalized at a single hospital between 2021 and 2022, of which 963 patients were in the primary analysis group.
The Epic Deterioration Index (EDI) early warning score uses 31 clinical measures as well as machine learning to predict whether a patient will experience a combined outcome of: rapid response team activation, transfer to intensive care, cardiopulmonary arrest, or death in hospital.
The team found that use of the model was helpful at preventing patient deterioration and reduced the risk of the EDI outcome by 10.4 percentage points.
“Amid the limited evidence base for early warning scores despite widespread adoption, this study provides evidence for their effectiveness and supports further testing of these interventions in other care settings,” write the authors.
