AI strengthens imaging clinical decision support software

After implementing AI as part of their CDS installation, researchers at the Yale School of Medicine found that clinicians were more likely to provide the software with a structured clinical indication to order advanced imaging studies. This has led to a significant improvement in proper controls.

“Free-text entry of examination indication with AI-assisted structured indication selection improved CDS scoring and the relevance of advanced imaging order,” said presenter Dr Dorothy Sippo .

When their healthcare system implemented clinical decision support in 2019 for outpatient radiology orders, service providers initially had difficulty identifying the appropriate structured indications required by CDS software, according to Sippo. In 46% of orders, the software could not assess the relevance of the examination because the ordering clinician did not select a structured indication.

“It was also a challenge for radiologists, as they did not routinely receive a meaningful and complete clinical history,” Sippo said. “So both from the point of view of how CDS works and from the radiologists who get the information we need, we have faced challenges with our implementation of CDS.”

In September 2020, the institution upgraded its commercially available CDS software to allow prescribing physicians to enter an examination indication as free text instead of having to search for a checkbox with a structured indication. , she said. Once the AI ​​software has suggested a structured indication based on the free text, the vendor then selects the structured indication from the AI’s suggestion – either through automatic software prediction or by choosing from a list of indications provided in a pop-up window on the software. Radiologists then receive the free-text, structured indication information after the order is complete.

The group publicized the new AI functionality with a presentation to health system management, an informational email, and an instructional video. However, it was difficult to inform all vendors about the new AI functionality, she said. Some vendors were also finding ways to bypass CDS, such as closing the pop-up with the structured hints suggested by the AI. As an added complication, other updates were added to the CDS software – including content changes for oncologists – which caused further difficulties, Sippo said.

The integration of AI, however, has resulted in a significant improvement in the number of appropriate exam orders. And the AI ​​functionality was used in 38% of the orders, she said.

“We were also delighted to see that there was a decrease in the number of our unmarked orders, which is exactly what we wanted,” she said. “That’s why we implemented the AI ​​functionality.”

Impact of AI on CDS order results
Q1 2020 (before AI) Q2 2020 (before AI) Q4 2020 (after AI)
Appropriate commands 41% 37% 50%
Marginal orders 12% 9% 13%
Inappropriate orders 5% 4% 6%
Unrated orders 42% 50% 31%

The researchers learned a few lessons along the way, including that providers can enter specific, detailed clinical histories in a free text format. This allowed radiologists to receive more granular data, she said.

Additionally, AI’s ability has made it possible to collect structured information from these free-text clinical stories, Sippo said. Additionally, the authors learned that communicating clearly and widely to all vendors is essential for the new electronic health record functionality to be successfully adopted, Sippo said.

“When we saw vendors bypassing the CDS workflow, we realized they didn’t know how to use it and we needed to involve them more,” she said. “I think there is still work to be done for us in this area.”

On the downside, Sippo noted that even with the help of AI, 31% of orders were still not listed by CDS software. Researchers believe there is room for improvement, however.

“Because we know that unless there is a very close match between the free text entered and the AI’s predicted structured hint, the provider must choose from a pop-up list of predicted hints,” Sippo said.

Some of the vendors choose to bypass CDS by closing that pop-up, according to Sippo. Next month, the group plans to test a new workflow that will lower the threshold for AI to automatically select a predicted structured indication based on its analysis of the vendor’s free-text indication.

Copyright © 2021

Source link

About Terry Gongora

Check Also

Florida Consortium for Medical Marijuana Clinical Outcomes Research: A Beneficial Program

By Dr. Jeff G. Konin As Florida’s marijuana landscape changes rapidly, it’s a good time …

Leave a Reply

Your email address will not be published.