Elsevier Health, partners unveil conversational AI decision support tool

2024-03-01

Preview

来源: FierceHealthcare

Based on early feedback from its test users, Elsevier expects ClinicalKey AI to meaningfully improve users' ability to quickly find important medical information at the point of care.

Elsevier Health has launched an generative AI-driven version of its flagship clinical decision support tool, ClinicalKey, in partnership with AI company OpenEvidence.

The latest add-on module offers core and new content in an advanced conversational search interface to help doctors more quickly find the information they need. ClinicalKey is used by thousands of health systems and medical schools globally.

While the original tool will remain available, ClinicalKey AI is now available for purchase as an add-on. More than 30,000 doctors across the U.S. have tested the underlying AI technology to date.

ClinicalKey AI pulls from validated sources beyond Elsevier's own point of care content available through ClinicalKey, including journal abstracts, medical reference texts and government publications. The tool generates a summarized response cited with references. It also considers patient context, such as comorbidities and current medications, when prompted. It provides personalized responses based on the clinician’s profile or specialty, and offers the ability to access previous queries and responses. The tool is HIPAA-compliant.

“One thing that we did realize over time is it’s a big search problem that we have to solve,” Jan Herzhoff, Ph.D., president at Elsevier Health, told Fierce Healthcare.

Last November, Elsevier first announced its partnership with OpenEvidence to work on building out the tool. They also partnered with Cone Health and the University of New Mexico to help test the accuracy and usefulness of the tool as well as offer subject matter experts the chance to review user feedback and help implement improvements.

Elsevier worked to understand user behavior on non-AI clinical decision support tools, including its own. Based on its own data and market research, it determined that user queries generally fall into one of three categories: basic, medium or hard complexity.

Questions of a basic complexity level, such as those related to a drug or dosage, are the most common and take a user about four minutes to find an answer. These types of questions go unanswered about 10% of the time, per Herzhoff.

Questions of medium complexity, like those related to interactions between drugs, make up about a third of all queries and can take users about 10 minutes to find an answer. These don’t get answered about 20% of the time.

Finally, questions of hard complexity, like those related to comorbidities, can take a user hours or even days to find an answer. More than 35% don’t get answered. Based on early feedback from its test users, Elsevier expects ClinicalKey AI to make a meaningful difference on these stats, particularly in this category.

“It finds the needle in the haystack in all of our content,” Herzhoff said.

Because all of the content on which the model was trained is licensed, it adds to the credibility of the tool, he added. And the fact that ClinicalKey AI cites references greatly reduces the frequency of hallucinations.

"Generative AI isn't about replacing a doctor's training, intuition or expertise, it's about amplifying it,” Dusadee Sarangarm, M.D., chief medical information officer at the University of New Mexico Health, said in an announcement. “It is having the world's medical library at your fingertips, instantly summarizing massive amounts of research and literature at the bedside."

From the get-go, Elsevier wanted to build an architecture into the tool that would allow authors of books that the tool cites to receive royalty payments any time their work is referenced, Herzhoff said.

“As soon as things are generated or it’s a black box and you don’t really know where it’s coming from, you can’t really do that anymore,” he noted. Several authors are suing ChatGPT over allegations of copyright infringement. However, “seeing the answer coming through from the highest quality journals and sources … is very reassuring," he said.

When it comes to the anticipated response from providers, Herzhoff says the company is optimistic. Last year, in a survey (PDF) of clinicians, Elsevier found that about 11% of clinical decisions are currently assisted by generative AI. Nearly half of physicians globally find using such tools desirable in the future.

ClinicalKey AI is not yet integrated with EHRs, but Herzhoff hopes it will be in the future.