While artificial intelligence technology is increasingly being used — formally and informally — to support medical diagnoses, its utility in emergency medical settings remains an open question. Can AI support doctors in situations where split-second decision making can mean the difference between life and death? Researchers at Drexel University broached the question with clinicians at Children's National Medical Center in Washington, D.C., to better understand how and when the technology could help them save lives.
Led by Angela Mastrianni, PhD, a Drexel graduate who is a postdoctoral fellow at NYU Langone Health and Aleksandra Sarcevic, PhD, a professor in Drexel's College of Computing & Informatics and director of the Interactive Systems for Healthcare Research Lab , the team looked at two types of scenarios in which AI technology is used to support emergency medical doctors in making treatment decisions.
In the first scenario, key information used in decision making — including patient age, how the injury occurred, and vital signs — was synthesized and presented to the team in real-time. In the second scenario, treatment recommendations were provided, in addition to the synthesized information.
In an experiment that involved 35 emergency care providers from six health systems, the researchers found that participants were more likely to make correct decisions when both AI information and recommendations were provided, compared to receiving no AI support.
However, they also found that participants were divided on their perception of receiving recommendations from an AI assistant during medical emergencies. Although most preferred to receive AI recommendations and synthesis, some physicians had concerns that the recommendations could infringe on their agency and bias decision making.
The authors recently presented its findings at the American Computing Machinery's Conference on Computer-Supported Cooperative Work & Social Computing (CSCW).
"Although our study involved a small sample of health care providers, this is the sort of inquiry that will be important as the emergency medical community considers how AI technology can support its lifesaving work," Sarcevic said. "There is no question that the technology can augment the work of humans in medical settings, but understanding when and where it is appropriate and accepted will be key to navigating its adoption."
To arrive at their findings, the team first designed a prototype of an AI-enabled decision-support display — dubbed "DecAide" — for use in a pediatric trauma resuscitation setting. By surveying and interviewing a variety of emergency medicine care providers, the team gained an understanding of the types of information that providers use to support decision making during resuscitations and how best to present it.
With this guidance, the display took shape as a concise listing of key patient information, highlighting abnormalities and color coding any changes in vital signs. One version presented only this information, while a second also offered a recommendation — such as a blood transfusion or neurosurgical procedure — along with its probability of success based on a risk calculation model drawing on resuscitation data from the primary research site, Children's National Hospital.
The team evaluated the participants' interaction with the system by creating 12 scripted vignette scenarios during which information was gradually presented about trauma patients. In a timed virtual exercise, 35 providers were each presented these scenarios under three conditions: in one vignette, the decision-support display offered no information or guidance from AI; during another it offered AI-synthesized information and in the third, both AI-synthesized information and a recommendation were offered. Participants were asked to make real-time assessments in each scenario and decide whether or not the patient needed a life-saving intervention, such as a blood transfusion, brain surgery, a chest tube or needle decompression, intubation or chest surgery.
The team recorded each decision in the participants' treatment and diagnosis process — more than 800 instances in total — comparing them to the ground truth data from which the vignettes were created, to determine diagnostic accuracy. Each participant also completed a survey about how they used the information display. To test the effect of information trust and bias in the decision-making process, in one out of every eight decisions presented to the participants, the researchers programmed the display to provide an incorrect recommendation.
Participants made the correct decisions in 64.4% of the scenarios when both AI information synthesis and a recommendation were provided. The rate fell to 56.3% when only information synthesis was provided without a recommendation and 55.8% when no support was provided.
The technology did not appear to slow decision making, as the time taken for participants to make decisions remained relatively consistent through all three of the display conditions in the experiment. And in many instances, participants made their decision before AI-enabled recommendation was provided on the display.
The use and perception of the decision-making support varied widely, however. Eighteen participants noted that they considered the recommendations, but only after they had already made their decision. Twelve participants ignored the AI recommendations altogether, either because they lacked nuance or the participants did not trust the system because the data driving its recommendation was not provided. Overall, the participants expressed fewer concerns about the presentation of AI-synthesized information.
"We are seeing a gradual adoption of decision support systems in medical specialties such as radiology, but there is still quite a bit of hesitancy in using this new technology in dynamic and time-critical medical settings, like emergency medicine," Mastrianni said. "While there is evidence that AI models can diagnose illness at high levels of accuracy, we know that more research is needed to understand how best to integrate it in clinical settings so that providers begin to trust and use this new technology."
The team suggests that continued research in this area should include larger participant pools with representatives from a wider range of medical specialties and types of hospitals. They note that before any such tools are adopted, additional information and support is needed for medical leaders deciding whether and how to implement them and how to create clear policies around their use.
